Changeset 125 in SHVCSoftware for trunk/source/Lib/TLibEncoder/TEncSearch.cpp
- Timestamp:
- 16 Apr 2013, 06:39:31 (12 years ago)
- File:
-
- 1 edited
-
trunk/source/Lib/TLibEncoder/TEncSearch.cpp (modified) (62 diffs)
Legend:
- Unmodified
- Added
- Removed
-
trunk/source/Lib/TLibEncoder/TEncSearch.cpp
r55 r125 4 4 * granted under this license. 5 5 * 6 * Copyright (c) 2010-201 2, ITU/ISO/IEC6 * Copyright (c) 2010-2013, ITU/ISO/IEC 7 7 * All rights reserved. 8 8 * … … 333 333 334 334 // distortion 335 m_cDistParam.bitDepth = g_bitDepthY; 335 336 uiSad = m_cDistParam.DistFunc( &m_cDistParam ); 336 337 … … 745 746 746 747 m_cDistParam.pCur = piRefPos; 748 m_cDistParam.bitDepth = g_bitDepthY; 747 749 uiDist = m_cDistParam.DistFunc( &m_cDistParam ); 748 750 uiDist += m_pcRdCost->getCost( cMvTest.getHor(), cMvTest.getVer() ); … … 794 796 if( bLuma ) 795 797 { 796 #if TRANS_SPLIT_FLAG_CTX_REDUCTION797 798 m_pcEntropyCoder->encodeTransformSubdivFlag( uiSubdiv, 5 - uiLog2TrafoSize ); 798 #else799 m_pcEntropyCoder->encodeTransformSubdivFlag( uiSubdiv, uiFullDepth );800 #endif801 799 } 802 800 } … … 1064 1062 else 1065 1063 #endif 1066 { 1067 predIntraLumaAng( pcCU->getPattern(), uiLumaPredMode, piPred, uiStride, uiWidth, uiHeight, pcCU, bAboveAvail, bLeftAvail ); 1068 } 1064 predIntraLumaAng( pcCU->getPattern(), uiLumaPredMode, piPred, uiStride, uiWidth, uiHeight, bAboveAvail, bLeftAvail ); 1069 1065 // save prediction 1070 1066 if(default0Save1Load2 == 1) … … 1118 1114 //===== transform and quantization ===== 1119 1115 //--- init rate estimation arrays for RDOQ --- 1120 if( m_pcEncCfg->getUseRDOQ() && useTransformSkip == false)1116 if( useTransformSkip? m_pcEncCfg->getUseRDOQTS():m_pcEncCfg->getUseRDOQ()) 1121 1117 { 1122 1118 m_pcEntropyCoder->estimateBit( m_pcTrQuant->m_pcEstBitsSbac, uiWidth, uiWidth, TEXT_LUMA ); … … 1169 1165 for( UInt uiX = 0; uiX < uiWidth; uiX++ ) 1170 1166 { 1171 pReco [ uiX ] = Clip ( pPred[ uiX ] + pResi[ uiX ] );1167 pReco [ uiX ] = ClipY( pPred[ uiX ] + pResi[ uiX ] ); 1172 1168 pRecQt [ uiX ] = pReco[ uiX ]; 1173 1169 pRecIPred[ uiX ] = pReco[ uiX ]; … … 1182 1178 1183 1179 //===== update distortion ===== 1184 ruiDist += m_pcRdCost->getDistPart( piReco, uiStride, piOrg, uiStride, uiWidth, uiHeight );1180 ruiDist += m_pcRdCost->getDistPart(g_bitDepthY, piReco, uiStride, piOrg, uiStride, uiWidth, uiHeight ); 1185 1181 } 1186 1182 … … 1247 1243 pcCU->getPattern()->initPattern ( pcCU, uiTrDepth, uiAbsPartIdx ); 1248 1244 1249 #if !REMOVE_LMCHROMA1250 if( uiChromaPredMode == LM_CHROMA_IDX && uiChromaId == 0 )1251 {1252 pcCU->getPattern()->initAdiPattern( pcCU, uiAbsPartIdx, uiTrDepth, m_piYuvExt, m_iYuvExtStride, m_iYuvExtHeight, bAboveAvail, bLeftAvail, true );1253 getLumaRecPixels( pcCU->getPattern(), uiWidth, uiHeight );1254 }1255 #endif1256 1257 1245 pcCU->getPattern()->initAdiPatternChroma( pcCU, uiAbsPartIdx, uiTrDepth, m_piYuvExt, m_iYuvExtStride, m_iYuvExtHeight, bAboveAvail, bLeftAvail ); 1258 1246 Int* pPatChroma = ( uiChromaId > 0 ? pcCU->getPattern()->getAdiCrBuf( uiWidth, uiHeight, m_piYuvExt ) : pcCU->getPattern()->getAdiCbBuf( uiWidth, uiHeight, m_piYuvExt ) ); … … 1266 1254 else 1267 1255 #endif 1268 #if !REMOVE_LMCHROMA 1269 if( uiChromaPredMode == LM_CHROMA_IDX ) 1270 { 1271 predLMIntraChroma( pcCU->getPattern(), pPatChroma, piPred, uiStride, uiWidth, uiHeight, uiChromaId ); 1272 } 1273 else 1274 #endif 1275 { 1276 predIntraChromaAng( pcCU->getPattern(), pPatChroma, uiChromaPredMode, piPred, uiStride, uiWidth, uiHeight, pcCU, bAboveAvail, bLeftAvail ); 1277 } 1278 1256 { 1257 predIntraChromaAng( pPatChroma, uiChromaPredMode, piPred, uiStride, uiWidth, uiHeight, bAboveAvail, bLeftAvail ); 1258 } 1279 1259 // save prediction 1280 1260 if( default0Save1Load2 == 1 ) … … 1329 1309 { 1330 1310 //--- init rate estimation arrays for RDOQ --- 1331 if( m_pcEncCfg->getUseRDOQ() && !useTransformSkipChroma)1311 if( useTransformSkipChroma? m_pcEncCfg->getUseRDOQTS():m_pcEncCfg->getUseRDOQ()) 1332 1312 { 1333 1313 m_pcEntropyCoder->estimateBit( m_pcTrQuant->m_pcEstBitsSbac, uiWidth, uiWidth, eText ); … … 1336 1316 UInt uiAbsSum = 0; 1337 1317 1338 #if CHROMA_QP_EXTENSION1339 1318 Int curChromaQpOffset; 1340 1319 if(eText == TEXT_CHROMA_U) … … 1347 1326 } 1348 1327 m_pcTrQuant->setQPforQuant ( pcCU->getQP( 0 ), TEXT_CHROMA, pcCU->getSlice()->getSPS()->getQpBDOffsetC(), curChromaQpOffset ); 1349 #else1350 if(eText == TEXT_CHROMA_U)1351 {1352 m_pcTrQuant->setQPforQuant ( pcCU->getQP( 0 ), TEXT_CHROMA, pcCU->getSlice()->getSPS()->getQpBDOffsetC(), pcCU->getSlice()->getPPS()->getChromaCbQpOffset() );1353 }1354 else1355 {1356 m_pcTrQuant->setQPforQuant ( pcCU->getQP( 0 ), TEXT_CHROMA, pcCU->getSlice()->getSPS()->getQpBDOffsetC(), pcCU->getSlice()->getPPS()->getChromaCrQpOffset() );1357 }1358 #endif1359 1328 1360 1329 #if RDOQ_CHROMA_LAMBDA … … 1398 1367 for( UInt uiX = 0; uiX < uiWidth; uiX++ ) 1399 1368 { 1400 pReco [ uiX ] = Clip ( pPred[ uiX ] + pResi[ uiX ] );1369 pReco [ uiX ] = ClipC( pPred[ uiX ] + pResi[ uiX ] ); 1401 1370 pRecQt [ uiX ] = pReco[ uiX ]; 1402 1371 pRecIPred[ uiX ] = pReco[ uiX ]; … … 1412 1381 //===== update distortion ===== 1413 1382 #if WEIGHTED_CHROMA_DISTORTION 1414 ruiDist += m_pcRdCost->getDistPart( piReco, uiStride, piOrg, uiStride, uiWidth, uiHeight, true);1383 ruiDist += m_pcRdCost->getDistPart(g_bitDepthC, piReco, uiStride, piOrg, uiStride, uiWidth, uiHeight, eText ); 1415 1384 #else 1416 ruiDist += m_pcRdCost->getDistPart( piReco, uiStride, piOrg, uiStride, uiWidth, uiHeight );1385 ruiDist += m_pcRdCost->getDistPart(g_bitDepthC, piReco, uiStride, piOrg, uiStride, uiWidth, uiHeight ); 1417 1386 #endif 1418 1387 } … … 1441 1410 1442 1411 #if HHI_RQT_INTRA_SPEEDUP 1412 #if L0232_RD_PENALTY 1413 Int maxTuSize = pcCU->getSlice()->getSPS()->getQuadtreeTULog2MaxSize(); 1414 Int isIntraSlice = (pcCU->getSlice()->getSliceType() == I_SLICE); 1415 // don't check split if TU size is less or equal to max TU size 1416 Bool noSplitIntraMaxTuSize = bCheckFull; 1417 if(m_pcEncCfg->getRDpenalty() && ! isIntraSlice) 1418 { 1419 // in addition don't check split if TU size is less or equal to 16x16 TU size for non-intra slice 1420 noSplitIntraMaxTuSize = ( uiLog2TrSize <= min(maxTuSize,4) ); 1421 1422 // if maximum RD-penalty don't check TU size 32x32 1423 if(m_pcEncCfg->getRDpenalty()==2) 1424 { 1425 bCheckFull = ( uiLog2TrSize <= min(maxTuSize,4)); 1426 } 1427 } 1428 if( bCheckFirst && noSplitIntraMaxTuSize ) 1429 #else 1443 1430 if( bCheckFirst && bCheckFull ) 1431 #endif 1444 1432 { 1445 1433 bCheckSplit = false; 1446 1434 } 1435 #else 1436 #if L0232_RD_PENALTY 1437 Int maxTuSize = pcCU->getSlice()->getSPS()->getQuadtreeTULog2MaxSize(); 1438 Int isIntraSlice = (pcCU->getSlice()->getSliceType() == I_SLICE); 1439 // if maximum RD-penalty don't check TU size 32x32 1440 if((m_pcEncCfg->getRDpenalty()==2) && !isIntraSlice) 1441 { 1442 bCheckFull = ( uiLog2TrSize <= min(maxTuSize,4)); 1443 } 1444 #endif 1447 1445 #endif 1448 1446 Double dSingleCost = MAX_DOUBLE; … … 1452 1450 UInt uiSingleCbfU = 0; 1453 1451 UInt uiSingleCbfV = 0; 1454 #if PPS_TS_FLAG1455 1452 Bool checkTransformSkip = pcCU->getSlice()->getPPS()->getUseTransformSkip(); 1456 #else1457 Bool checkTransformSkip = pcCU->getSlice()->getSPS()->getUseTransformSkip();1458 #endif1459 1453 UInt widthTransformSkip = pcCU->getWidth ( 0 ) >> uiTrDepth; 1460 1454 UInt heightTransformSkip = pcCU->getHeight( 0 ) >> uiTrDepth; … … 1464 1458 checkTransformSkip &= (!pcCU->getCUTransquantBypass(0)); 1465 1459 checkTransformSkip &= (!((pcCU->getQP( 0 ) == 0) && (pcCU->getSlice()->getSPS()->getUseLossless()))); 1466 #if INTRA_TRANSFORMSKIP_FAST1467 1460 if ( m_pcEncCfg->getUseTransformSkipFast() ) 1468 1461 { 1469 1462 checkTransformSkip &= (pcCU->getPartitionSize(uiAbsPartIdx)==SIZE_NxN); 1470 1463 } 1471 #endif1472 1464 if( bCheckFull ) 1473 1465 { … … 1625 1617 //----- determine rate and r-d cost ----- 1626 1618 UInt uiSingleBits = xGetIntraBitsQT( pcCU, uiTrDepth, uiAbsPartIdx, true, !bLumaOnly, false ); 1619 #if L0232_RD_PENALTY 1620 if(m_pcEncCfg->getRDpenalty() && (uiLog2TrSize==5) && !isIntraSlice) 1621 { 1622 uiSingleBits=uiSingleBits*4; 1623 } 1624 #endif 1627 1625 dSingleCost = m_pcRdCost->calcRdCost( uiSingleBits, uiSingleDistY + uiSingleDistC ); 1628 1626 } … … 2197 2195 if( uiTrMode == uiTrDepth ) 2198 2196 { 2199 #if PPS_TS_FLAG2200 2197 Bool checkTransformSkip = pcCU->getSlice()->getPPS()->getUseTransformSkip(); 2201 #else2202 Bool checkTransformSkip = pcCU->getSlice()->getSPS()->getUseTransformSkip();2203 #endif2204 2198 UInt uiLog2TrSize = g_aucConvertToBit[ pcCU->getSlice()->getSPS()->getMaxCUWidth() >> uiFullDepth ] + 2; 2205 2199 … … 2218 2212 2219 2213 checkTransformSkip &= (uiLog2TrSize <= 3); 2220 #if INTRA_TRANSFORMSKIP_FAST2221 2214 if ( m_pcEncCfg->getUseTransformSkipFast() ) 2222 2215 { … … 2232 2225 } 2233 2226 } 2234 #endif2235 2227 2236 2228 if(checkTransformSkip) 2237 2229 { 2238 #if INTRA_TRANSFORMSKIP_FAST==02239 Bool checkTSFast = m_pcEncCfg->getUseTransformSkipFast();2240 if(checkTSFast)2241 {2242 //fast algorithm for chroma TS2243 Bool useTransformSkip = checkTransformSkip;2244 UInt blkNumberOfTS = 0;2245 for(UInt absPartIdxSub = uiAbsPartIdx; absPartIdxSub < uiAbsPartIdx + 4; absPartIdxSub ++)2246 {2247 blkNumberOfTS += pcCU->getTransformSkip(absPartIdxSub, TEXT_LUMA);2248 }2249 if(blkNumberOfTS < 4)2250 {2251 useTransformSkip = false;2252 }2253 pcCU ->setTransformSkipSubParts( useTransformSkip, TEXT_CHROMA_U, uiAbsPartIdx, pcCU->getDepth( 0 ) + actualTrDepth);2254 pcCU ->setTransformSkipSubParts( useTransformSkip, TEXT_CHROMA_V, uiAbsPartIdx, pcCU->getDepth( 0 ) + actualTrDepth);2255 xIntraCodingChromaBlk( pcCU, uiTrDepth, uiAbsPartIdx, pcOrgYuv, pcPredYuv, pcResiYuv, ruiDist, 0 );2256 xIntraCodingChromaBlk( pcCU, uiTrDepth, uiAbsPartIdx, pcOrgYuv, pcPredYuv, pcResiYuv, ruiDist, 1 );2257 UInt singleCbfCb = pcCU->getCbf( uiAbsPartIdx, TEXT_CHROMA_U, uiTrDepth);2258 UInt singleCbfCr = pcCU->getCbf( uiAbsPartIdx, TEXT_CHROMA_V, uiTrDepth);2259 if(useTransformSkip && singleCbfCb == 0)2260 {2261 pcCU ->setTransformSkipSubParts( 0, TEXT_CHROMA_U, uiAbsPartIdx, pcCU->getDepth( 0 ) + actualTrDepth);2262 }2263 if(useTransformSkip && singleCbfCr == 0)2264 {2265 pcCU ->setTransformSkipSubParts( 0, TEXT_CHROMA_V, uiAbsPartIdx, pcCU->getDepth( 0 ) + actualTrDepth);2266 }2267 }2268 else2269 {2270 #endif2271 2230 //use RDO to decide whether Cr/Cb takes TS 2272 2231 if( m_bUseSBACRD ) … … 2356 2315 } 2357 2316 } 2358 #if INTRA_TRANSFORMSKIP_FAST==02359 }2360 #endif2361 2317 } 2362 2318 else … … 2479 2435 { 2480 2436 //--- get prediction --- 2481 predIntraChromaAng( p cCU->getPattern(), pPatChromaU, uiMode, piPredU, uiStride, uiWidth, uiHeight, pcCU, bAboveAvail, bLeftAvail );2482 predIntraChromaAng( p cCU->getPattern(), pPatChromaV, uiMode, piPredV, uiStride, uiWidth, uiHeight, pcCU, bAboveAvail, bLeftAvail );2437 predIntraChromaAng( pPatChromaU, uiMode, piPredU, uiStride, uiWidth, uiHeight, bAboveAvail, bLeftAvail ); 2438 predIntraChromaAng( pPatChromaV, uiMode, piPredV, uiStride, uiWidth, uiHeight, bAboveAvail, bLeftAvail ); 2483 2439 2484 2440 //--- get SAD --- 2485 UInt uiSAD = m_pcRdCost->calcHAD( piOrgU, uiStride, piPredU, uiStride, uiWidth, uiHeight );2486 uiSAD += m_pcRdCost->calcHAD( piOrgV, uiStride, piPredV, uiStride, uiWidth, uiHeight );2441 UInt uiSAD = m_pcRdCost->calcHAD(g_bitDepthC, piOrgU, uiStride, piPredU, uiStride, uiWidth, uiHeight ); 2442 uiSAD += m_pcRdCost->calcHAD(g_bitDepthC, piOrgV, uiStride, piPredV, uiStride, uiWidth, uiHeight ); 2487 2443 //--- check --- 2488 2444 if( uiSAD < uiMinSAD ) … … 2561 2517 UInt uiMode = modeIdx; 2562 2518 2563 predIntraLumaAng( pcCU->getPattern(), uiMode, piPred, uiStride, uiWidth, uiHeight, pcCU,bAboveAvail, bLeftAvail );2519 predIntraLumaAng( pcCU->getPattern(), uiMode, piPred, uiStride, uiWidth, uiHeight, bAboveAvail, bLeftAvail ); 2564 2520 2565 2521 // use hadamard transform here 2566 UInt uiSad = m_pcRdCost->calcHAD( piOrg, uiStride, piPred, uiStride, uiWidth, uiHeight );2522 UInt uiSad = m_pcRdCost->calcHAD(g_bitDepthY, piOrg, uiStride, piPred, uiStride, uiWidth, uiHeight ); 2567 2523 2568 2524 UInt iModeBits = xModeBitsIntra( pcCU, uiMode, uiPU, uiPartOffset, uiDepth, uiInitTrDepth ); … … 2829 2785 } 2830 2786 2831 #if INTRA_BL 2832 Void 2833 TEncSearch::estIntraBLPredQT( TComDataCU* pcCU, 2834 TComYuv* pcOrgYuv, 2835 TComYuv* pcPredYuv, 2836 TComYuv* pcResiYuv, 2837 TComYuv* pcRecoYuv ) 2838 { 2839 UInt uiDepth = pcCU->getDepth(0); 2840 UInt uiOverallDistY = 0; 2841 UInt uiOverallDistC = 0; 2842 2843 //===== set QP and clear Cbf ===== 2844 if ( pcCU->getSlice()->getPPS()->getUseDQP() == true) 2845 { 2846 pcCU->setQPSubParts( pcCU->getQP(0), 0, uiDepth ); 2847 } 2848 else 2849 { 2850 pcCU->setQPSubParts( pcCU->getSlice()->getSliceQp(), 0, uiDepth ); 2851 } 2852 2853 //===== init pattern for luma prediction ===== 2854 Bool bAboveAvail = false; 2855 Bool bLeftAvail = false; 2856 pcCU->getPattern()->initPattern ( pcCU, 0, 0 ); 2857 pcCU->getPattern()->initAdiPattern( pcCU, 0, 0, m_piYuvExt, m_iYuvExtStride, m_iYuvExtHeight, bAboveAvail, bLeftAvail ); 2858 2859 pcCU->setLumaIntraDirSubParts ( DC_IDX, 0, uiDepth ); 2860 2861 // set context models 2862 if( m_bUseSBACRD ) 2863 { 2864 m_pcRDGoOnSbacCoder->load( m_pppcRDSbacCoder[uiDepth][CI_CURR_BEST] ); 2865 } 2866 2867 // determine residual for partition 2868 Double dPUCost = 0.0; 2869 xRecurIntraCodingQT( pcCU, 0, 0, false, pcOrgYuv, pcPredYuv, pcResiYuv, uiOverallDistY, uiOverallDistC, false, dPUCost ); 2870 xSetIntraResultQT( pcCU, 0, 0, false, pcRecoYuv ); 2871 2872 //=== update PU data ==== 2873 pcCU->copyToPic( uiDepth, 0, 0 ); 2874 2875 //===== reset context models ===== 2876 if(m_bUseSBACRD) 2877 { 2878 m_pcRDGoOnSbacCoder->load(m_pppcRDSbacCoder[uiDepth][CI_CURR_BEST]); 2879 } 2880 2881 //===== set distortion (rate and r-d costs are determined later) ===== 2882 pcCU->getTotalDistortion() = uiOverallDistY + uiOverallDistC; 2883 } 2884 2885 #endif 2787 2886 2788 2887 2789 Void … … 2907 2809 for( UInt uiMode = uiMinMode; uiMode < uiMaxMode; uiMode++ ) 2908 2810 { 2909 #if !REMOVE_LMCHROMA2910 if ( !pcCU->getSlice()->getSPS()->getUseLMChroma() && uiModeList[uiMode] == LM_CHROMA_IDX )2911 {2912 continue;2913 }2914 #endif2915 2811 //----- restore context models ----- 2916 2812 if( m_bUseSBACRD ) … … 2923 2819 pcCU->setChromIntraDirSubParts ( uiModeList[uiMode], 0, uiDepth ); 2924 2820 xRecurIntraChromaCodingQT ( pcCU, 0, 0, pcOrgYuv, pcPredYuv, pcResiYuv, uiDist ); 2925 #if PPS_TS_FLAG2926 #if INTRA_TRANSFORMSKIP_FAST2927 2821 if( m_bUseSBACRD && pcCU->getSlice()->getPPS()->getUseTransformSkip() ) 2928 #else2929 if( m_bUseSBACRD && pcCU->getSlice()->getPPS()->getUseTransformSkip() && !m_pcEncCfg->getUseTransformSkipFast())2930 #endif2931 #else2932 #if INTRA_TRANSFORMSKIP_FAST2933 if( m_bUseSBACRD && pcCU->getSlice()->getSPS()->getUseTransformSkip() )2934 #else2935 if( m_bUseSBACRD && pcCU->getSlice()->getSPS()->getUseTransformSkip() && !m_pcEncCfg->getUseTransformSkipFast())2936 #endif2937 #endif2938 2822 { 2939 2823 m_pcRDGoOnSbacCoder->load( m_pppcRDSbacCoder[uiDepth][CI_CURR_BEST] ); … … 2997 2881 Pel* pReco = piReco; 2998 2882 Pel* pRecoPic; 2999 UInt uiInternalBitDepth = g_uiBitDepth + g_uiBitIncrement; 3000 UInt uiPCMBitDepth; 2883 Int shiftPcm; 3001 2884 3002 2885 if( eText == TEXT_LUMA) … … 3004 2887 uiReconStride = pcCU->getPic()->getPicYuvRec()->getStride(); 3005 2888 pRecoPic = pcCU->getPic()->getPicYuvRec()->getLumaAddr(pcCU->getAddr(), pcCU->getZorderIdxInCU()+uiAbsPartIdx); 3006 uiPCMBitDepth =pcCU->getSlice()->getSPS()->getPCMBitDepthLuma();2889 shiftPcm = g_bitDepthY - pcCU->getSlice()->getSPS()->getPCMBitDepthLuma(); 3007 2890 } 3008 2891 else … … 3018 2901 pRecoPic = pcCU->getPic()->getPicYuvRec()->getCrAddr(pcCU->getAddr(), pcCU->getZorderIdxInCU()+uiAbsPartIdx); 3019 2902 } 3020 uiPCMBitDepth =pcCU->getSlice()->getSPS()->getPCMBitDepthChroma();2903 shiftPcm = g_bitDepthC - pcCU->getSlice()->getSPS()->getPCMBitDepthChroma(); 3021 2904 } 3022 2905 … … 3038 2921 for( uiX = 0; uiX < uiWidth; uiX++ ) 3039 2922 { 3040 pPCM[uiX] = (pOrg[uiX]>>(uiInternalBitDepth - uiPCMBitDepth));2923 pPCM[uiX] = pOrg[uiX]>> shiftPcm; 3041 2924 } 3042 2925 pPCM += uiWidth; … … 3051 2934 for( uiX = 0; uiX < uiWidth; uiX++ ) 3052 2935 { 3053 pReco [uiX] = (pPCM[uiX]<<(uiInternalBitDepth - uiPCMBitDepth));2936 pReco [uiX] = pPCM[uiX]<< shiftPcm; 3054 2937 pRecoPic[uiX] = pReco[uiX]; 3055 2938 } … … 3152 3035 cDistParam.bApplyWeight = false; 3153 3036 3154 m_pcRdCost->setDistParam( cDistParam, 3037 m_pcRdCost->setDistParam( cDistParam, g_bitDepthY, 3155 3038 pcYuvOrg->getLumaAddr( uiAbsPartIdx ), pcYuvOrg->getStride(), 3156 3039 m_tmpYuvPred .getLumaAddr( uiAbsPartIdx ), m_tmpYuvPred .getStride(), … … 3182 3065 Int iHeight = 0; 3183 3066 3184 #if REF_IDX_FRAMEWORK // HM bug fix3067 #if 0 //REF_IDX_FRAMEWORK // HM bug fix 3185 3068 for( UInt uiMergeCand = 0; uiMergeCand < pcCU->getSlice()->getMaxNumMergeCand(); ++uiMergeCand ) 3186 3069 { … … 3198 3081 if ( iPUIdx == 0 ) 3199 3082 { 3200 pcCU->getInterMergeCandidates( 0, 0, uiDepth,cMvFieldNeighbours,uhInterDirNeighbours, numValidMergeCand );3083 pcCU->getInterMergeCandidates( 0, 0, cMvFieldNeighbours,uhInterDirNeighbours, numValidMergeCand ); 3201 3084 } 3202 3085 pcCU->setPartSizeSubParts( partSize, 0, uiDepth ); … … 3204 3087 else 3205 3088 { 3206 pcCU->getInterMergeCandidates( uiAbsPartIdx, iPUIdx, uiDepth,cMvFieldNeighbours, uhInterDirNeighbours, numValidMergeCand );3089 pcCU->getInterMergeCandidates( uiAbsPartIdx, iPUIdx, cMvFieldNeighbours, uhInterDirNeighbours, numValidMergeCand ); 3207 3090 } 3208 3091 xRestrictBipredMergeCand( pcCU, iPUIdx, cMvFieldNeighbours, uhInterDirNeighbours, numValidMergeCand ); … … 3242 3125 } 3243 3126 #endif 3244 3127 } 3245 3128 } 3246 3129 … … 3406 3289 #if (ENCODER_FAST_MODE) 3407 3290 TComPic* pcPic = pcCU->getSlice()->getRefPic( eRefPicList, iRefIdxTemp ); 3408 if( !testILR && pcPic-> getIsILR() && (ePartSize == SIZE_2Nx2N) )3291 if( !testILR && pcPic->isILR() && (ePartSize == SIZE_2Nx2N) ) 3409 3292 { 3410 3293 continue; … … 3509 3392 #endif 3510 3393 xCopyAMVPInfo(pcCU->getCUMvField(eRefPicList)->getAMVPInfo(), &aacAMVPInfo[iRefList][iRefIdxTemp]); // must always be done ( also when AMVP_MODE = AM_NONE ) 3511 #if !SPS_AMVP_CLEANUP3512 if ( pcCU->getAMVPMode(uiPartAddr) == AM_EXPL )3513 {3514 #endif3515 3394 xCheckBestMVP(pcCU, eRefPicList, cMvTemp[iRefList][iRefIdxTemp], cMvPred[iRefList][iRefIdxTemp], aaiMvpIdx[iRefList][iRefIdxTemp], uiBitsTemp, uiCostTemp); 3516 #if !SPS_AMVP_CLEANUP3517 }3518 #endif3519 3395 3520 3396 if(pcCU->getSlice()->getNumRefIdx(REF_PIC_LIST_C) > 0 && !pcCU->getSlice()->getNoBackPredFlag()) … … 3657 3533 Bool testIter = true; 3658 3534 TComPic* pcPic = pcCU->getSlice()->getRefPic( RefPicList(1 - iRefList), iRefIdxBi[1 - iRefList] ); 3659 if(pcPic->getIsILR() && (ePartSize == SIZE_2Nx2N)) 3535 if(pcPic->isILR() && (ePartSize == SIZE_2Nx2N)) 3536 { 3660 3537 testIter = false; //the fixed part is ILR, skip this iteration 3538 } 3661 3539 if(testIter) 3662 3540 { … … 3668 3546 { 3669 3547 #if (ENCODER_FAST_MODE) 3670 Bool testRefIdx = true; 3671 pcPic = pcCU->getSlice()->getRefPic( RefPicList(iRefList) , iRefIdxTemp ); 3672 if(pcPic->getIsILR() && (ePartSize == SIZE_2Nx2N)) 3673 testRefIdx = false; //the refined part is ILR, skip this reference pic 3674 if(testRefIdx) 3675 { 3548 Bool testRefIdx = true; 3549 pcPic = pcCU->getSlice()->getRefPic( RefPicList(iRefList) , iRefIdxTemp ); 3550 if(pcPic->isILR() && (ePartSize == SIZE_2Nx2N)) 3551 { 3552 testRefIdx = false; //the refined part is ILR, skip this reference pic 3553 } 3554 if(testRefIdx) 3555 { 3676 3556 #endif 3677 3557 uiBitsTemp = uiMbBits[2] + uiMotBits[1-iRefList]; … … 3684 3564 // call ME 3685 3565 xMotionEstimation ( pcCU, pcOrgYuv, iPartIdx, eRefPicList, &cMvPredBi[iRefList][iRefIdxTemp], iRefIdxTemp, cMvTemp[iRefList][iRefIdxTemp], uiBitsTemp, uiCostTemp, true ); 3686 #if !SPS_AMVP_CLEANUP3687 if ( pcCU->getAMVPMode(uiPartAddr) == AM_EXPL )3688 {3689 #endif3690 3566 xCopyAMVPInfo(&aacAMVPInfo[iRefList][iRefIdxTemp], pcCU->getCUMvField(eRefPicList)->getAMVPInfo()); 3691 3567 xCheckBestMVP(pcCU, eRefPicList, cMvTemp[iRefList][iRefIdxTemp], cMvPredBi[iRefList][iRefIdxTemp], aaiMvpIdxBi[iRefList][iRefIdxTemp], uiBitsTemp, uiCostTemp); 3692 #if !SPS_AMVP_CLEANUP3693 }3694 #endif3695 3568 3696 3569 if ( uiCostTemp < uiCostBi ) … … 3716 3589 } 3717 3590 #if (ENCODER_FAST_MODE) 3718 }3719 #endif3720 3591 } 3592 #endif 3593 } 3721 3594 #if (ENCODER_FAST_MODE) 3722 3595 } // for loop-iRefIdxTemp … … 3725 3598 if ( !bChanged ) 3726 3599 { 3727 #if !SPS_AMVP_CLEANUP3728 if ( uiCostBi <= uiCost[0] && uiCostBi <= uiCost[1] && pcCU->getAMVPMode(uiPartAddr) == AM_EXPL )3729 #else3730 3600 if ( uiCostBi <= uiCost[0] && uiCostBi <= uiCost[1] ) 3731 #endif3732 3601 { 3733 3602 xCopyAMVPInfo(&aacAMVPInfo[0][iRefIdxBi[0]], pcCU->getCUMvField(REF_PIC_LIST_0)->getAMVPInfo()); … … 4005 3874 return; 4006 3875 } 4007 #if ENCODER_FAST_MODE4008 Bool TEncSearch::predInterSearchILRUni( TComDataCU* pcCU, TComYuv* pcOrgYuv, TComYuv*& rpcPredYuv, TComYuv*& rpcResiYuv, TComYuv*& rpcRecoYuv )4009 {4010 rpcPredYuv->clear();4011 rpcRecoYuv->clear();4012 4013 Int iNumPredDir = pcCU->getSlice()->isInterP() ? 1 : 2;4014 4015 TComMv cMv[2];4016 TComMv cMvPred[2][33];4017 TComMv cMvTemp[2][33];4018 TComMv TempMv;4019 4020 Int iRefIdx[2]={0,0};4021 4022 Int aaiMvpIdx[2][33];4023 Int aaiMvpNum[2][33];4024 4025 UInt uiMbBits[3] = {1, 1, 0};4026 UInt uiLastMode = 0;4027 4028 UInt uiCost[2] = { MAX_UINT, MAX_UINT }; //uni, rdCost4029 UInt uiCostTemp;4030 UInt biPDistTemp = MAX_INT;4031 UInt uiBitsTemp;4032 4033 PartSize ePartSize = pcCU->getPartitionSize( 0 ); //2Nx2N4034 Int iPartIdx = 0; //one PU in CU4035 UInt uiPartAddr;4036 Int iRoiWidth, iRoiHeight;4037 4038 xGetBlkBits( ePartSize, pcCU->getSlice()->isInterP(), iPartIdx, uiLastMode, uiMbBits);4039 pcCU->getPartIndexAndSize( iPartIdx, uiPartAddr, iRoiWidth, iRoiHeight );4040 4041 for( Int iRefList = 0; iRefList < iNumPredDir; iRefList++) //list4042 {4043 RefPicList eRefPicList = ( iRefList ? REF_PIC_LIST_1 : REF_PIC_LIST_0 );4044 4045 Int iRefIdxTemp = -1;4046 Bool foundILR = false;4047 for( Int refIdx = 0; refIdx < pcCU->getSlice()->getNumRefIdx(eRefPicList); refIdx++ )4048 if( pcCU->getSlice()->getRefPic(eRefPicList, refIdx)->getIsILR() )4049 {4050 iRefIdxTemp = refIdx;4051 foundILR = true;4052 break;4053 }4054 4055 if(!foundILR) //no ILR in eRefPiclist4056 continue;4057 4058 uiBitsTemp = uiMbBits[iRefList];4059 if ( pcCU->getSlice()->getNumRefIdx(eRefPicList) > 1 )4060 {4061 uiBitsTemp += iRefIdxTemp+1;4062 if ( iRefIdxTemp == pcCU->getSlice()->getNumRefIdx(eRefPicList)-1 ) uiBitsTemp--;4063 }4064 4065 xEstimateMvPredAMVP( pcCU, pcOrgYuv, iPartIdx, eRefPicList, iRefIdxTemp, cMvPred[iRefList][iRefIdxTemp], false, &biPDistTemp);4066 aaiMvpIdx[iRefList][iRefIdxTemp] = pcCU->getMVPIdx(eRefPicList, uiPartAddr);4067 aaiMvpNum[iRefList][iRefIdxTemp] = pcCU->getMVPNum(eRefPicList, uiPartAddr);4068 4069 uiBitsTemp += m_auiMVPIdxCost[aaiMvpIdx[iRefList][iRefIdxTemp]][AMVP_MAX_NUM_CANDS];4070 4071 xMotionEstimation ( pcCU, pcOrgYuv, iPartIdx, eRefPicList, &cMvPred[iRefList][iRefIdxTemp], iRefIdxTemp, cMvTemp[iRefList][iRefIdxTemp], uiBitsTemp, uiCostTemp );4072 xCheckBestMVP(pcCU, eRefPicList, cMvTemp[iRefList][iRefIdxTemp], cMvPred[iRefList][iRefIdxTemp], aaiMvpIdx[iRefList][iRefIdxTemp], uiBitsTemp, uiCostTemp);4073 4074 if( uiCostTemp < uiCost[iRefList] )4075 {4076 uiCost[iRefList] = uiCostTemp;4077 4078 cMv[iRefList] = cMvTemp[iRefList][iRefIdxTemp];4079 iRefIdx[iRefList] = iRefIdxTemp;4080 4081 pcCU->getCUMvField(eRefPicList)->setAllMv( cMv[iRefList], ePartSize, uiPartAddr, 0, iPartIdx );4082 pcCU->getCUMvField(eRefPicList)->setAllRefIdx( iRefIdx[iRefList], ePartSize, uiPartAddr, 0, iPartIdx );4083 }4084 }4085 4086 if( uiCost[0] == MAX_UINT && uiCost[1] == MAX_UINT ) //no ILR in both list0 and list14087 return false;4088 4089 // Clear Motion Field4090 pcCU->getCUMvField(REF_PIC_LIST_0)->setAllMvField( TComMvField(), ePartSize, uiPartAddr, 0, iPartIdx );4091 pcCU->getCUMvField(REF_PIC_LIST_1)->setAllMvField( TComMvField(), ePartSize, uiPartAddr, 0, iPartIdx );4092 pcCU->getCUMvField(REF_PIC_LIST_0)->setAllMvd ( TComMv(), ePartSize, uiPartAddr, 0, iPartIdx );4093 pcCU->getCUMvField(REF_PIC_LIST_1)->setAllMvd ( TComMv(), ePartSize, uiPartAddr, 0, iPartIdx );4094 4095 pcCU->setMVPIdxSubParts( -1, REF_PIC_LIST_0, uiPartAddr, iPartIdx, pcCU->getDepth(uiPartAddr));4096 pcCU->setMVPNumSubParts( -1, REF_PIC_LIST_0, uiPartAddr, iPartIdx, pcCU->getDepth(uiPartAddr));4097 pcCU->setMVPIdxSubParts( -1, REF_PIC_LIST_1, uiPartAddr, iPartIdx, pcCU->getDepth(uiPartAddr));4098 pcCU->setMVPNumSubParts( -1, REF_PIC_LIST_1, uiPartAddr, iPartIdx, pcCU->getDepth(uiPartAddr));4099 4100 if( uiCost[0] <= uiCost[1] ) //list0 ILR4101 {4102 pcCU->getCUMvField(REF_PIC_LIST_0)->setAllMv ( cMv[0], ePartSize, uiPartAddr, 0, iPartIdx );4103 pcCU->getCUMvField(REF_PIC_LIST_0)->setAllRefIdx( iRefIdx[0], ePartSize, uiPartAddr, 0, iPartIdx );4104 4105 TempMv = cMv[0] - cMvPred[0][iRefIdx[0]];4106 pcCU->getCUMvField(REF_PIC_LIST_0)->setAllMvd( TempMv, ePartSize, uiPartAddr, 0, iPartIdx );4107 4108 pcCU->setInterDirSubParts( 1, uiPartAddr, iPartIdx, pcCU->getDepth(0) );4109 4110 pcCU->setMVPIdxSubParts( aaiMvpIdx[0][iRefIdx[0]], REF_PIC_LIST_0, uiPartAddr, iPartIdx, pcCU->getDepth(uiPartAddr));4111 pcCU->setMVPNumSubParts( aaiMvpNum[0][iRefIdx[0]], REF_PIC_LIST_0, uiPartAddr, iPartIdx, pcCU->getDepth(uiPartAddr));4112 }4113 else if( uiCost[1] < uiCost[0] ) //list1 ILR4114 {4115 pcCU->getCUMvField(REF_PIC_LIST_1)->setAllMv ( cMv[1], ePartSize, uiPartAddr, 0, iPartIdx );4116 pcCU->getCUMvField(REF_PIC_LIST_1)->setAllRefIdx( iRefIdx[1], ePartSize, uiPartAddr, 0, iPartIdx );4117 4118 TempMv = cMv[1] - cMvPred[1][iRefIdx[1]];4119 pcCU->getCUMvField(REF_PIC_LIST_1)->setAllMvd( TempMv, ePartSize, uiPartAddr, 0, iPartIdx );4120 4121 pcCU->setInterDirSubParts( 2, uiPartAddr, iPartIdx, pcCU->getDepth(0) );4122 4123 pcCU->setMVPIdxSubParts( aaiMvpIdx[1][iRefIdx[1]], REF_PIC_LIST_1, uiPartAddr, iPartIdx, pcCU->getDepth(uiPartAddr));4124 pcCU->setMVPNumSubParts( aaiMvpNum[1][iRefIdx[1]], REF_PIC_LIST_1, uiPartAddr, iPartIdx, pcCU->getDepth(uiPartAddr));4125 }4126 else4127 assert(0);4128 4129 pcCU->setMergeFlagSubParts( false, uiPartAddr, iPartIdx, pcCU->getDepth( uiPartAddr ) );4130 4131 motionCompensation ( pcCU, rpcPredYuv, REF_PIC_LIST_X, iPartIdx );4132 4133 setWpScalingDistParam( pcCU, -1, REF_PIC_LIST_X );4134 4135 return true;4136 }4137 4138 4139 #endif4140 3876 4141 3877 // AMVP … … 4168 3904 cBestMv = pcAMVPInfo->m_acMvCand[0]; 4169 3905 #if !ZERO_MVD_EST 4170 #if !SPS_AMVP_CLEANUP4171 if( pcCU->getAMVPMode(uiPartAddr) == AM_NONE || (pcAMVPInfo->iN <= 1 && pcCU->getAMVPMode(uiPartAddr) == AM_EXPL) )4172 #else4173 3906 if (pcAMVPInfo->iN <= 1) 4174 #endif4175 3907 { 4176 3908 rcMvPred = cBestMv; … … 4190 3922 } 4191 3923 #endif 4192 #if !SPS_AMVP_CLEANUP4193 if (pcCU->getAMVPMode(uiPartAddr) == AM_EXPL && bFilled)4194 #else4195 3924 if (bFilled) 4196 #endif4197 3925 { 4198 3926 assert(pcCU->getMVPIdx(eRefPicList,uiPartAddr) >= 0); … … 4201 3929 } 4202 3930 4203 #if !SPS_AMVP_CLEANUP4204 if (pcCU->getAMVPMode(uiPartAddr) == AM_EXPL)4205 {4206 #endif4207 3931 m_cYuvPredTemp.clear(); 4208 3932 #if ZERO_MVD_EST … … 4231 3955 4232 3956 m_cYuvPredTemp.clear(); 4233 #if !SPS_AMVP_CLEANUP4234 }4235 #endif4236 3957 4237 3958 // Setting Best MVP … … 4406 4127 if ( pcCU->getSlice()->getPPS()->getUseWP() && pcCU->getSlice()->getSliceType()==P_SLICE ) 4407 4128 { 4408 xWeightedPredictionUni( pcCU, pcTemplateCand, uiPartAddr, iSizeX, iSizeY, eRefPicList, pcTemplateCand, uiPartIdx,iRefIdx );4129 xWeightedPredictionUni( pcCU, pcTemplateCand, uiPartAddr, iSizeX, iSizeY, eRefPicList, pcTemplateCand, iRefIdx ); 4409 4130 } 4410 4131 … … 4413 4134 m_pcRdCost->getMotionCost( 1, 0 ); 4414 4135 DistParam cDistParam; 4415 m_pcRdCost->setDistParam( cDistParam, 4136 m_pcRdCost->setDistParam( cDistParam, g_bitDepthY, 4416 4137 pcOrgYuv->getLumaAddr(uiPartAddr), pcOrgYuv->getStride(), 4417 4138 pcTemplateCand->getLumaAddr(uiPartAddr), pcTemplateCand->getStride(), … … 4425 4146 #else 4426 4147 #if WEIGHTED_CHROMA_DISTORTION 4427 uiCost = m_pcRdCost->getDistPart( pcTemplateCand->getLumaAddr(uiPartAddr), pcTemplateCand->getStride(), pcOrgYuv->getLumaAddr(uiPartAddr), pcOrgYuv->getStride(), iSizeX, iSizeY, false, DF_SAD );4148 uiCost = m_pcRdCost->getDistPart(g_bitDepthY, pcTemplateCand->getLumaAddr(uiPartAddr), pcTemplateCand->getStride(), pcOrgYuv->getLumaAddr(uiPartAddr), pcOrgYuv->getStride(), iSizeX, iSizeY, TEXT_LUMA, DF_SAD ); 4428 4149 #else 4429 uiCost = m_pcRdCost->getDistPart( pcTemplateCand->getLumaAddr(uiPartAddr), pcTemplateCand->getStride(), pcOrgYuv->getLumaAddr(uiPartAddr), pcOrgYuv->getStride(), iSizeX, iSizeY, DF_SAD );4150 uiCost = m_pcRdCost->getDistPart(g_bitDepthY, pcTemplateCand->getLumaAddr(uiPartAddr), pcTemplateCand->getStride(), pcOrgYuv->getLumaAddr(uiPartAddr), pcOrgYuv->getStride(), iSizeX, iSizeY, DF_SAD ); 4430 4151 #endif 4431 4152 uiCost = (UInt) m_pcRdCost->calcRdCost( m_auiMVPIdxCost[iMVPIdx][iMVPNum], uiCost, false, DF_SAD ); … … 4473 4194 iRoiHeight, 4474 4195 pcYuv->getStride(), 4475 0, 0 , 0, 0);4196 0, 0 ); 4476 4197 4477 4198 Pel* piRefY = pcCU->getSlice()->getRefPic( eRefPicList, iRefIdxPred )->getPicYuvRec()->getLumaAddr( pcCU->getAddr(), pcCU->getZorderIdxInCU() + uiPartAddr ); … … 4490 4211 setWpScalingDistParam( pcCU, iRefIdxPred, eRefPicList ); 4491 4212 // Do integer search 4492 #if REF_IDX_ME_AROUND_ZEROMV || REF_IDX_ME_ZEROMV4493 #if REF_IDX_ME_AROUND_ZEROMV4494 if( pcCU->getSlice()->getRefPic(eRefPicList, iRefIdxPred )->getIsILR()) //ILR reference pic4495 {4496 xPatternSearchILR ( pcCU, pcPatternKey, piRefY, iRefStride, rcMv, ruiCost );4497 }4498 else //non ILR reference pic4499 {4500 if ( !m_iFastSearch || bBi )4501 {4502 xPatternSearch ( pcPatternKey, piRefY, iRefStride, &cMvSrchRngLT, &cMvSrchRngRB, rcMv, ruiCost );4503 }4504 else4505 {4506 rcMv = *pcMvPred;4507 xPatternSearchFast ( pcCU, pcPatternKey, piRefY, iRefStride, &cMvSrchRngLT, &cMvSrchRngRB, rcMv, ruiCost );4508 }4509 }4510 #endif4511 4213 #if REF_IDX_ME_ZEROMV 4512 if( pcCU->getSlice()->getRefPic( eRefPicList, iRefIdxPred )-> getIsILR()) //ILR reference pic4214 if( pcCU->getSlice()->getRefPic( eRefPicList, iRefIdxPred )->isILR()) //ILR reference pic 4513 4215 { 4514 4216 rcMv.setZero(); //use Mv(0, 0) for integer ME … … 4526 4228 } 4527 4229 } 4528 #endif4529 4230 #else 4530 4231 if ( !m_iFastSearch || bBi ) … … 4544 4245 { 4545 4246 #if REF_IDX_ME_ZEROMV 4546 if( pcCU->getSlice()->getRefPic( eRefPicList, iRefIdxPred )-> getIsILR()) //ILR reference pic4247 if( pcCU->getSlice()->getRefPic( eRefPicList, iRefIdxPred )->isILR()) //ILR reference pic 4547 4248 { 4548 4249 xPatternSearchFracDIFMv0( pcCU, pcPatternKey, piRefY, iRefStride, &rcMv, cMvHalf, cMvQter, ruiCost, bBi ); … … 4628 4329 setDistParamComp(0); 4629 4330 4331 m_cDistParam.bitDepth = g_bitDepthY; 4630 4332 uiSad = m_cDistParam.DistFunc( &m_cDistParam ); 4631 4333 … … 4666 4368 } 4667 4369 4668 #if REF_IDX_ME_AROUND_ZEROMV4669 //ILR integer pixel motion estimation search4670 //pcCU is CU pointer4671 //pcPatterney contains target PU infor4672 //piRefY is the PU at the same position as the source PU, but in the reference pic4673 //iRefStride is the reference stride4674 //rcMv output best integer MV4675 //ruiSAD outputs the SAD of best integer MV4676 Void TEncSearch::xPatternSearchILR( TComDataCU* pcCU, TComPattern* pcPatternKey, Pel* piRefY, Int iRefStride, TComMv& rcMv, UInt& ruiSAD )4677 {4678 Int candMvX[5] = {0, -1, 1, 0, 0};4679 Int candMvY[5] = {0, 0, 0, -1, 1};4680 Int numCand = 1;4681 Int iBestX = 0;4682 Int iBestY = 0;4683 Int uiSadBest = MAX_UINT;4684 UInt uiSad;4685 4686 Pel* piRefSrch;4687 4688 //-- jclee for using the SAD function pointer4689 m_pcRdCost->setDistParam( pcPatternKey, piRefY, iRefStride, m_cDistParam );4690 4691 // fast encoder decision: use subsampled SAD for integer ME4692 if ( m_pcEncCfg->getUseFastEnc() )4693 {4694 if ( m_cDistParam.iRows > 8 )4695 {4696 m_cDistParam.iSubShift = 1;4697 }4698 }4699 4700 for(Int candId = 0; candId < numCand; candId++)4701 {4702 piRefSrch = piRefY + (candMvY[candId] * iRefStride) + candMvX[candId];4703 m_cDistParam.pCur = piRefSrch;4704 4705 setDistParamComp(0);4706 4707 uiSad = m_cDistParam.DistFunc( &m_cDistParam ); //SAD4708 4709 //motion cost4710 uiSad += m_pcRdCost->getCost( candMvX[candId], candMvY[candId] ); //SAD rdCost4711 4712 if(uiSad < uiSadBest)4713 {4714 uiSadBest = uiSad;4715 iBestX = candMvX[candId];4716 iBestY = candMvY[candId];4717 }4718 }4719 4720 rcMv.set( iBestX, iBestY );4721 4722 ruiSAD = uiSadBest - m_pcRdCost->getCost( iBestX, iBestY );4723 return;4724 }4725 #endif4726 4370 Void TEncSearch::xTZSearch( TComDataCU* pcCU, TComPattern* pcPatternKey, Pel* piRefY, Int iRefStride, TComMv* pcMvSrchRngLT, TComMv* pcMvSrchRngRB, TComMv& rcMv, UInt& ruiSAD ) 4727 4371 { … … 4918 4562 pcPatternKey->getROIYHeight(), 4919 4563 iRefStride, 4920 0, 0 , 0, 0);4564 0, 0 ); 4921 4565 4922 4566 // Half-pel refinement … … 4936 4580 rcMvQter += rcMvHalf; rcMvQter <<= 1; 4937 4581 ruiCost = xPatternRefinement( pcPatternKey, baseRefMv, 1, rcMvQter ); 4582 } 4583 4584 /** encode residual and calculate rate-distortion for a CU block 4585 * \param pcCU 4586 * \param pcYuvOrg 4587 * \param pcYuvPred 4588 * \param rpcYuvResi 4589 * \param rpcYuvResiBest 4590 * \param rpcYuvRec 4591 * \param bSkipRes 4592 * \returns Void 4593 */ 4594 Void TEncSearch::encodeResAndCalcRdInterCU( TComDataCU* pcCU, TComYuv* pcYuvOrg, TComYuv* pcYuvPred, TComYuv*& rpcYuvResi, TComYuv*& rpcYuvResiBest, TComYuv*& rpcYuvRec, Bool bSkipRes ) 4595 { 4596 #if NO_RESIDUAL_FLAG_FOR_BLPRED 4597 if ( pcCU->isIntra(0) && !pcCU->isIntraBL(0)) 4598 #else 4599 if ( pcCU->isIntra(0) ) 4600 #endif 4601 { 4602 return; 4603 } 4604 4605 Bool bHighPass = pcCU->getSlice()->getDepth() ? true : false; 4606 UInt uiBits = 0, uiBitsBest = 0; 4607 UInt uiDistortion = 0, uiDistortionBest = 0; 4608 4609 UInt uiWidth = pcCU->getWidth ( 0 ); 4610 UInt uiHeight = pcCU->getHeight( 0 ); 4611 4612 // No residual coding : SKIP mode 4613 if ( bSkipRes ) 4614 { 4615 pcCU->setSkipFlagSubParts( true, 0, pcCU->getDepth(0) ); 4616 4617 rpcYuvResi->clear(); 4618 4619 pcYuvPred->copyToPartYuv( rpcYuvRec, 0 ); 4620 4621 #if WEIGHTED_CHROMA_DISTORTION 4622 uiDistortion = m_pcRdCost->getDistPart(g_bitDepthY, rpcYuvRec->getLumaAddr(), rpcYuvRec->getStride(), pcYuvOrg->getLumaAddr(), pcYuvOrg->getStride(), uiWidth, uiHeight ) 4623 + m_pcRdCost->getDistPart(g_bitDepthC, rpcYuvRec->getCbAddr(), rpcYuvRec->getCStride(), pcYuvOrg->getCbAddr(), pcYuvOrg->getCStride(), uiWidth >> 1, uiHeight >> 1, TEXT_CHROMA_U ) 4624 + m_pcRdCost->getDistPart(g_bitDepthC, rpcYuvRec->getCrAddr(), rpcYuvRec->getCStride(), pcYuvOrg->getCrAddr(), pcYuvOrg->getCStride(), uiWidth >> 1, uiHeight >> 1, TEXT_CHROMA_V ); 4625 #else 4626 uiDistortion = m_pcRdCost->getDistPart(g_bitDepthY, rpcYuvRec->getLumaAddr(), rpcYuvRec->getStride(), pcYuvOrg->getLumaAddr(), pcYuvOrg->getStride(), uiWidth, uiHeight ) 4627 + m_pcRdCost->getDistPart(g_bitDepthC, rpcYuvRec->getCbAddr(), rpcYuvRec->getCStride(), pcYuvOrg->getCbAddr(), pcYuvOrg->getCStride(), uiWidth >> 1, uiHeight >> 1 ) 4628 + m_pcRdCost->getDistPart(g_bitDepthC, rpcYuvRec->getCrAddr(), rpcYuvRec->getCStride(), pcYuvOrg->getCrAddr(), pcYuvOrg->getCStride(), uiWidth >> 1, uiHeight >> 1 ); 4629 #endif 4630 4631 if( m_bUseSBACRD ) 4632 m_pcRDGoOnSbacCoder->load(m_pppcRDSbacCoder[pcCU->getDepth(0)][CI_CURR_BEST]); 4633 4634 m_pcEntropyCoder->resetBits(); 4635 if (pcCU->getSlice()->getPPS()->getTransquantBypassEnableFlag()) 4636 { 4637 m_pcEntropyCoder->encodeCUTransquantBypassFlag(pcCU, 0, true); 4638 } 4639 m_pcEntropyCoder->encodeSkipFlag(pcCU, 0, true); 4640 m_pcEntropyCoder->encodeMergeIndex( pcCU, 0, true ); 4641 4642 uiBits = m_pcEntropyCoder->getNumberOfWrittenBits(); 4643 pcCU->getTotalBits() = uiBits; 4644 pcCU->getTotalDistortion() = uiDistortion; 4645 pcCU->getTotalCost() = m_pcRdCost->calcRdCost( uiBits, uiDistortion ); 4646 4647 if( m_bUseSBACRD ) 4648 m_pcRDGoOnSbacCoder->store(m_pppcRDSbacCoder[pcCU->getDepth(0)][CI_TEMP_BEST]); 4649 4650 pcCU->setCbfSubParts( 0, 0, 0, 0, pcCU->getDepth( 0 ) ); 4651 pcCU->setTrIdxSubParts( 0, 0, pcCU->getDepth(0) ); 4652 4653 return; 4654 } 4655 4656 // Residual coding. 4657 Int qp, qpBest = 0, qpMin, qpMax; 4658 Double dCost, dCostBest = MAX_DOUBLE; 4659 4660 UInt uiTrLevel = 0; 4661 if( (pcCU->getWidth(0) > pcCU->getSlice()->getSPS()->getMaxTrSize()) ) 4662 { 4663 while( pcCU->getWidth(0) > (pcCU->getSlice()->getSPS()->getMaxTrSize()<<uiTrLevel) ) uiTrLevel++; 4664 } 4665 UInt uiMaxTrMode = pcCU->getSlice()->getSPS()->getMaxTrDepth() + uiTrLevel; 4666 4667 while((uiWidth>>uiMaxTrMode) < (g_uiMaxCUWidth>>g_uiMaxCUDepth)) uiMaxTrMode--; 4668 4669 qpMin = bHighPass ? Clip3( -pcCU->getSlice()->getSPS()->getQpBDOffsetY(), MAX_QP, pcCU->getQP(0) - m_iMaxDeltaQP ) : pcCU->getQP( 0 ); 4670 qpMax = bHighPass ? Clip3( -pcCU->getSlice()->getSPS()->getQpBDOffsetY(), MAX_QP, pcCU->getQP(0) + m_iMaxDeltaQP ) : pcCU->getQP( 0 ); 4671 4672 rpcYuvResi->subtract( pcYuvOrg, pcYuvPred, 0, uiWidth ); 4673 4674 for ( qp = qpMin; qp <= qpMax; qp++ ) 4675 { 4676 dCost = 0.; 4677 uiBits = 0; 4678 uiDistortion = 0; 4679 if( m_bUseSBACRD ) 4680 { 4681 m_pcRDGoOnSbacCoder->load( m_pppcRDSbacCoder[ pcCU->getDepth( 0 ) ][ CI_CURR_BEST ] ); 4682 } 4683 4684 UInt uiZeroDistortion = 0; 4685 xEstimateResidualQT( pcCU, 0, 0, 0, rpcYuvResi, pcCU->getDepth(0), dCost, uiBits, uiDistortion, &uiZeroDistortion ); 4686 4687 m_pcEntropyCoder->resetBits(); 4688 m_pcEntropyCoder->encodeQtRootCbfZero( pcCU ); 4689 UInt zeroResiBits = m_pcEntropyCoder->getNumberOfWrittenBits(); 4690 Double dZeroCost = m_pcRdCost->calcRdCost( zeroResiBits, uiZeroDistortion ); 4691 if(pcCU->isLosslessCoded( 0 )) 4692 { 4693 dZeroCost = dCost + 1; 4694 } 4695 if ( dZeroCost < dCost ) 4696 { 4697 dCost = dZeroCost; 4698 uiBits = 0; 4699 uiDistortion = uiZeroDistortion; 4700 4701 const UInt uiQPartNum = pcCU->getPic()->getNumPartInCU() >> (pcCU->getDepth(0) << 1); 4702 ::memset( pcCU->getTransformIdx() , 0, uiQPartNum * sizeof(UChar) ); 4703 ::memset( pcCU->getCbf( TEXT_LUMA ) , 0, uiQPartNum * sizeof(UChar) ); 4704 ::memset( pcCU->getCbf( TEXT_CHROMA_U ), 0, uiQPartNum * sizeof(UChar) ); 4705 ::memset( pcCU->getCbf( TEXT_CHROMA_V ), 0, uiQPartNum * sizeof(UChar) ); 4706 ::memset( pcCU->getCoeffY() , 0, uiWidth * uiHeight * sizeof( TCoeff ) ); 4707 ::memset( pcCU->getCoeffCb() , 0, uiWidth * uiHeight * sizeof( TCoeff ) >> 2 ); 4708 ::memset( pcCU->getCoeffCr() , 0, uiWidth * uiHeight * sizeof( TCoeff ) >> 2 ); 4709 pcCU->setTransformSkipSubParts ( 0, 0, 0, 0, pcCU->getDepth(0) ); 4710 } 4711 #if NO_RESIDUAL_FLAG_FOR_BLPRED 4712 else if(pcCU->getLayerId() > 0 && pcCU->isIntraBL(0) && uiZeroDistortion == uiDistortion) // all zeros 4713 { 4714 const UInt uiQPartNum = pcCU->getPic()->getNumPartInCU() >> (pcCU->getDepth(0) << 1); 4715 ::memset( pcCU->getTransformIdx() , 0, uiQPartNum * sizeof(UChar) ); 4716 ::memset( pcCU->getCbf( TEXT_LUMA ) , 0, uiQPartNum * sizeof(UChar) ); 4717 ::memset( pcCU->getCbf( TEXT_CHROMA_U ), 0, uiQPartNum * sizeof(UChar) ); 4718 ::memset( pcCU->getCbf( TEXT_CHROMA_V ), 0, uiQPartNum * sizeof(UChar) ); 4719 ::memset( pcCU->getCoeffY() , 0, uiWidth * uiHeight * sizeof( TCoeff ) ); 4720 ::memset( pcCU->getCoeffCb() , 0, uiWidth * uiHeight * sizeof( TCoeff ) >> 2 ); 4721 ::memset( pcCU->getCoeffCr() , 0, uiWidth * uiHeight * sizeof( TCoeff ) >> 2 ); 4722 pcCU->setTransformSkipSubParts ( 0, 0, 0, 0, pcCU->getDepth(0) ); 4723 } 4724 #endif 4725 else 4726 { 4727 xSetResidualQTData( pcCU, 0, 0, 0, NULL, pcCU->getDepth(0), false ); 4728 } 4729 4730 if( m_bUseSBACRD ) 4731 { 4732 m_pcRDGoOnSbacCoder->load( m_pppcRDSbacCoder[pcCU->getDepth(0)][CI_CURR_BEST] ); 4733 } 4734 #if 0 // check 4735 { 4736 m_pcEntropyCoder->resetBits(); 4737 m_pcEntropyCoder->encodeCoeff( pcCU, 0, pcCU->getDepth(0), pcCU->getWidth(0), pcCU->getHeight(0) ); 4738 const UInt uiBitsForCoeff = m_pcEntropyCoder->getNumberOfWrittenBits(); 4739 if( m_bUseSBACRD ) 4740 { 4741 m_pcRDGoOnSbacCoder->load( m_pppcRDSbacCoder[pcCU->getDepth(0)][CI_CURR_BEST] ); 4742 } 4743 if( uiBitsForCoeff != uiBits ) 4744 assert( 0 ); 4745 } 4746 #endif 4747 uiBits = 0; 4748 { 4749 TComYuv *pDummy = NULL; 4750 xAddSymbolBitsInter( pcCU, 0, 0, uiBits, pDummy, NULL, pDummy ); 4751 } 4752 4753 4754 Double dExactCost = m_pcRdCost->calcRdCost( uiBits, uiDistortion ); 4755 dCost = dExactCost; 4756 4757 if ( dCost < dCostBest ) 4758 { 4759 if ( !pcCU->getQtRootCbf( 0 ) ) 4760 { 4761 rpcYuvResiBest->clear(); 4762 } 4763 else 4764 { 4765 xSetResidualQTData( pcCU, 0, 0, 0, rpcYuvResiBest, pcCU->getDepth(0), true ); 4766 } 4767 4768 if( qpMin != qpMax && qp != qpMax ) 4769 { 4770 const UInt uiQPartNum = pcCU->getPic()->getNumPartInCU() >> (pcCU->getDepth(0) << 1); 4771 ::memcpy( m_puhQTTempTrIdx, pcCU->getTransformIdx(), uiQPartNum * sizeof(UChar) ); 4772 ::memcpy( m_puhQTTempCbf[0], pcCU->getCbf( TEXT_LUMA ), uiQPartNum * sizeof(UChar) ); 4773 ::memcpy( m_puhQTTempCbf[1], pcCU->getCbf( TEXT_CHROMA_U ), uiQPartNum * sizeof(UChar) ); 4774 ::memcpy( m_puhQTTempCbf[2], pcCU->getCbf( TEXT_CHROMA_V ), uiQPartNum * sizeof(UChar) ); 4775 ::memcpy( m_pcQTTempCoeffY, pcCU->getCoeffY(), uiWidth * uiHeight * sizeof( TCoeff ) ); 4776 ::memcpy( m_pcQTTempCoeffCb, pcCU->getCoeffCb(), uiWidth * uiHeight * sizeof( TCoeff ) >> 2 ); 4777 ::memcpy( m_pcQTTempCoeffCr, pcCU->getCoeffCr(), uiWidth * uiHeight * sizeof( TCoeff ) >> 2 ); 4778 #if ADAPTIVE_QP_SELECTION 4779 ::memcpy( m_pcQTTempArlCoeffY, pcCU->getArlCoeffY(), uiWidth * uiHeight * sizeof( Int ) ); 4780 ::memcpy( m_pcQTTempArlCoeffCb, pcCU->getArlCoeffCb(), uiWidth * uiHeight * sizeof( Int ) >> 2 ); 4781 ::memcpy( m_pcQTTempArlCoeffCr, pcCU->getArlCoeffCr(), uiWidth * uiHeight * sizeof( Int ) >> 2 ); 4782 #endif 4783 ::memcpy( m_puhQTTempTransformSkipFlag[0], pcCU->getTransformSkip(TEXT_LUMA), uiQPartNum * sizeof( UChar ) ); 4784 ::memcpy( m_puhQTTempTransformSkipFlag[1], pcCU->getTransformSkip(TEXT_CHROMA_U), uiQPartNum * sizeof( UChar ) ); 4785 ::memcpy( m_puhQTTempTransformSkipFlag[2], pcCU->getTransformSkip(TEXT_CHROMA_V), uiQPartNum * sizeof( UChar ) ); 4786 } 4787 uiBitsBest = uiBits; 4788 uiDistortionBest = uiDistortion; 4789 dCostBest = dCost; 4790 qpBest = qp; 4791 if( m_bUseSBACRD ) 4792 { 4793 m_pcRDGoOnSbacCoder->store( m_pppcRDSbacCoder[ pcCU->getDepth( 0 ) ][ CI_TEMP_BEST ] ); 4794 } 4795 } 4796 } 4797 4798 assert ( dCostBest != MAX_DOUBLE ); 4799 4800 if( qpMin != qpMax && qpBest != qpMax ) 4801 { 4802 if( m_bUseSBACRD ) 4803 { 4804 assert( 0 ); // check 4805 m_pcRDGoOnSbacCoder->load( m_pppcRDSbacCoder[ pcCU->getDepth( 0 ) ][ CI_TEMP_BEST ] ); 4806 } 4807 // copy best cbf and trIdx to pcCU 4808 const UInt uiQPartNum = pcCU->getPic()->getNumPartInCU() >> (pcCU->getDepth(0) << 1); 4809 ::memcpy( pcCU->getTransformIdx(), m_puhQTTempTrIdx, uiQPartNum * sizeof(UChar) ); 4810 ::memcpy( pcCU->getCbf( TEXT_LUMA ), m_puhQTTempCbf[0], uiQPartNum * sizeof(UChar) ); 4811 ::memcpy( pcCU->getCbf( TEXT_CHROMA_U ), m_puhQTTempCbf[1], uiQPartNum * sizeof(UChar) ); 4812 ::memcpy( pcCU->getCbf( TEXT_CHROMA_V ), m_puhQTTempCbf[2], uiQPartNum * sizeof(UChar) ); 4813 ::memcpy( pcCU->getCoeffY(), m_pcQTTempCoeffY, uiWidth * uiHeight * sizeof( TCoeff ) ); 4814 ::memcpy( pcCU->getCoeffCb(), m_pcQTTempCoeffCb, uiWidth * uiHeight * sizeof( TCoeff ) >> 2 ); 4815 ::memcpy( pcCU->getCoeffCr(), m_pcQTTempCoeffCr, uiWidth * uiHeight * sizeof( TCoeff ) >> 2 ); 4816 #if ADAPTIVE_QP_SELECTION 4817 ::memcpy( pcCU->getArlCoeffY(), m_pcQTTempArlCoeffY, uiWidth * uiHeight * sizeof( Int ) ); 4818 ::memcpy( pcCU->getArlCoeffCb(), m_pcQTTempArlCoeffCb, uiWidth * uiHeight * sizeof( Int ) >> 2 ); 4819 ::memcpy( pcCU->getArlCoeffCr(), m_pcQTTempArlCoeffCr, uiWidth * uiHeight * sizeof( Int ) >> 2 ); 4820 #endif 4821 ::memcpy( pcCU->getTransformSkip(TEXT_LUMA), m_puhQTTempTransformSkipFlag[0], uiQPartNum * sizeof( UChar ) ); 4822 ::memcpy( pcCU->getTransformSkip(TEXT_CHROMA_U), m_puhQTTempTransformSkipFlag[1], uiQPartNum * sizeof( UChar ) ); 4823 ::memcpy( pcCU->getTransformSkip(TEXT_CHROMA_V), m_puhQTTempTransformSkipFlag[2], uiQPartNum * sizeof( UChar ) ); 4824 } 4825 rpcYuvRec->addClip ( pcYuvPred, rpcYuvResiBest, 0, uiWidth ); 4826 4827 // update with clipped distortion and cost (qp estimation loop uses unclipped values) 4828 #if WEIGHTED_CHROMA_DISTORTION 4829 uiDistortionBest = m_pcRdCost->getDistPart(g_bitDepthY, rpcYuvRec->getLumaAddr(), rpcYuvRec->getStride(), pcYuvOrg->getLumaAddr(), pcYuvOrg->getStride(), uiWidth, uiHeight ) 4830 + m_pcRdCost->getDistPart(g_bitDepthC, rpcYuvRec->getCbAddr(), rpcYuvRec->getCStride(), pcYuvOrg->getCbAddr(), pcYuvOrg->getCStride(), uiWidth >> 1, uiHeight >> 1, TEXT_CHROMA_U ) 4831 + m_pcRdCost->getDistPart(g_bitDepthC, rpcYuvRec->getCrAddr(), rpcYuvRec->getCStride(), pcYuvOrg->getCrAddr(), pcYuvOrg->getCStride(), uiWidth >> 1, uiHeight >> 1, TEXT_CHROMA_V ); 4832 #else 4833 uiDistortionBest = m_pcRdCost->getDistPart(g_bitDepthY, rpcYuvRec->getLumaAddr(), rpcYuvRec->getStride(), pcYuvOrg->getLumaAddr(), pcYuvOrg->getStride(), uiWidth, uiHeight ) 4834 + m_pcRdCost->getDistPart(g_bitDepthC, rpcYuvRec->getCbAddr(), rpcYuvRec->getCStride(), pcYuvOrg->getCbAddr(), pcYuvOrg->getCStride(), uiWidth >> 1, uiHeight >> 1 ) 4835 + m_pcRdCost->getDistPart(g_bitDepthC, rpcYuvRec->getCrAddr(), rpcYuvRec->getCStride(), pcYuvOrg->getCrAddr(), pcYuvOrg->getCStride(), uiWidth >> 1, uiHeight >> 1 ); 4836 #endif 4837 dCostBest = m_pcRdCost->calcRdCost( uiBitsBest, uiDistortionBest ); 4838 4839 pcCU->getTotalBits() = uiBitsBest; 4840 pcCU->getTotalDistortion() = uiDistortionBest; 4841 pcCU->getTotalCost() = dCostBest; 4842 4843 if ( pcCU->isSkipped(0) ) 4844 { 4845 pcCU->setCbfSubParts( 0, 0, 0, 0, pcCU->getDepth( 0 ) ); 4846 } 4847 4848 pcCU->setQPSubParts( qpBest, 0, pcCU->getDepth(0) ); 4849 } 4850 4851 Void TEncSearch::xEstimateResidualQT( TComDataCU* pcCU, UInt uiQuadrant, UInt uiAbsPartIdx, UInt absTUPartIdx, TComYuv* pcResi, const UInt uiDepth, Double &rdCost, UInt &ruiBits, UInt &ruiDist, UInt *puiZeroDist ) 4852 { 4853 const UInt uiTrMode = uiDepth - pcCU->getDepth( 0 ); 4854 4855 assert( pcCU->getDepth( 0 ) == pcCU->getDepth( uiAbsPartIdx ) ); 4856 const UInt uiLog2TrSize = g_aucConvertToBit[pcCU->getSlice()->getSPS()->getMaxCUWidth() >> uiDepth]+2; 4857 4858 UInt SplitFlag = ((pcCU->getSlice()->getSPS()->getQuadtreeTUMaxDepthInter() == 1) && pcCU->getPredictionMode(uiAbsPartIdx) == MODE_INTER && ( pcCU->getPartitionSize(uiAbsPartIdx) != SIZE_2Nx2N )); 4859 Bool bCheckFull; 4860 if ( SplitFlag && uiDepth == pcCU->getDepth(uiAbsPartIdx) && ( uiLog2TrSize > pcCU->getQuadtreeTULog2MinSizeInCU(uiAbsPartIdx) ) ) 4861 bCheckFull = false; 4862 else 4863 bCheckFull = ( uiLog2TrSize <= pcCU->getSlice()->getSPS()->getQuadtreeTULog2MaxSize() ); 4864 4865 const Bool bCheckSplit = ( uiLog2TrSize > pcCU->getQuadtreeTULog2MinSizeInCU(uiAbsPartIdx) ); 4866 4867 assert( bCheckFull || bCheckSplit ); 4868 4869 Bool bCodeChroma = true; 4870 UInt uiTrModeC = uiTrMode; 4871 UInt uiLog2TrSizeC = uiLog2TrSize-1; 4872 if( uiLog2TrSize == 2 ) 4873 { 4874 uiLog2TrSizeC++; 4875 uiTrModeC --; 4876 UInt uiQPDiv = pcCU->getPic()->getNumPartInCU() >> ( ( pcCU->getDepth( 0 ) + uiTrModeC ) << 1 ); 4877 bCodeChroma = ( ( uiAbsPartIdx % uiQPDiv ) == 0 ); 4878 } 4879 4880 const UInt uiSetCbf = 1 << uiTrMode; 4881 // code full block 4882 Double dSingleCost = MAX_DOUBLE; 4883 UInt uiSingleBits = 0; 4884 UInt uiSingleDist = 0; 4885 UInt uiAbsSumY = 0, uiAbsSumU = 0, uiAbsSumV = 0; 4886 UInt uiBestTransformMode[3] = {0}; 4887 4888 if( m_bUseSBACRD ) 4889 { 4890 m_pcRDGoOnSbacCoder->store( m_pppcRDSbacCoder[ uiDepth ][ CI_QT_TRAFO_ROOT ] ); 4891 } 4892 4893 if( bCheckFull ) 4894 { 4895 const UInt uiNumCoeffPerAbsPartIdxIncrement = pcCU->getSlice()->getSPS()->getMaxCUWidth() * pcCU->getSlice()->getSPS()->getMaxCUHeight() >> ( pcCU->getSlice()->getSPS()->getMaxCUDepth() << 1 ); 4896 const UInt uiQTTempAccessLayer = pcCU->getSlice()->getSPS()->getQuadtreeTULog2MaxSize() - uiLog2TrSize; 4897 TCoeff *pcCoeffCurrY = m_ppcQTTempCoeffY [uiQTTempAccessLayer] + uiNumCoeffPerAbsPartIdxIncrement * uiAbsPartIdx; 4898 TCoeff *pcCoeffCurrU = m_ppcQTTempCoeffCb[uiQTTempAccessLayer] + (uiNumCoeffPerAbsPartIdxIncrement * uiAbsPartIdx>>2); 4899 TCoeff *pcCoeffCurrV = m_ppcQTTempCoeffCr[uiQTTempAccessLayer] + (uiNumCoeffPerAbsPartIdxIncrement * uiAbsPartIdx>>2); 4900 #if ADAPTIVE_QP_SELECTION 4901 Int *pcArlCoeffCurrY = m_ppcQTTempArlCoeffY [uiQTTempAccessLayer] + uiNumCoeffPerAbsPartIdxIncrement * uiAbsPartIdx; 4902 Int *pcArlCoeffCurrU = m_ppcQTTempArlCoeffCb[uiQTTempAccessLayer] + (uiNumCoeffPerAbsPartIdxIncrement * uiAbsPartIdx>>2); 4903 Int *pcArlCoeffCurrV = m_ppcQTTempArlCoeffCr[uiQTTempAccessLayer] + (uiNumCoeffPerAbsPartIdxIncrement * uiAbsPartIdx>>2); 4904 #endif 4905 4906 Int trWidth = 0, trHeight = 0, trWidthC = 0, trHeightC = 0; 4907 UInt absTUPartIdxC = uiAbsPartIdx; 4908 4909 trWidth = trHeight = 1 << uiLog2TrSize; 4910 trWidthC = trHeightC = 1 <<uiLog2TrSizeC; 4911 pcCU->setTrIdxSubParts( uiDepth - pcCU->getDepth( 0 ), uiAbsPartIdx, uiDepth ); 4912 Double minCostY = MAX_DOUBLE; 4913 Double minCostU = MAX_DOUBLE; 4914 Double minCostV = MAX_DOUBLE; 4915 Bool checkTransformSkipY = pcCU->getSlice()->getPPS()->getUseTransformSkip() && trWidth == 4 && trHeight == 4; 4916 Bool checkTransformSkipUV = pcCU->getSlice()->getPPS()->getUseTransformSkip() && trWidthC == 4 && trHeightC == 4; 4917 4918 checkTransformSkipY &= (!pcCU->isLosslessCoded(0)); 4919 checkTransformSkipUV &= (!pcCU->isLosslessCoded(0)); 4920 4921 pcCU->setTransformSkipSubParts ( 0, TEXT_LUMA, uiAbsPartIdx, uiDepth ); 4922 if( bCodeChroma ) 4923 { 4924 pcCU->setTransformSkipSubParts ( 0, TEXT_CHROMA_U, uiAbsPartIdx, pcCU->getDepth(0)+uiTrModeC ); 4925 pcCU->setTransformSkipSubParts ( 0, TEXT_CHROMA_V, uiAbsPartIdx, pcCU->getDepth(0)+uiTrModeC ); 4926 } 4927 4928 if (m_pcEncCfg->getUseRDOQ()) 4929 { 4930 m_pcEntropyCoder->estimateBit(m_pcTrQuant->m_pcEstBitsSbac, trWidth, trHeight, TEXT_LUMA ); 4931 } 4932 4933 m_pcTrQuant->setQPforQuant( pcCU->getQP( 0 ), TEXT_LUMA, pcCU->getSlice()->getSPS()->getQpBDOffsetY(), 0 ); 4934 4935 #if RDOQ_CHROMA_LAMBDA 4936 m_pcTrQuant->selectLambda(TEXT_LUMA); 4937 #endif 4938 m_pcTrQuant->transformNxN( pcCU, pcResi->getLumaAddr( absTUPartIdx ), pcResi->getStride (), pcCoeffCurrY, 4939 #if ADAPTIVE_QP_SELECTION 4940 pcArlCoeffCurrY, 4941 #endif 4942 trWidth, trHeight, uiAbsSumY, TEXT_LUMA, uiAbsPartIdx ); 4943 4944 pcCU->setCbfSubParts( uiAbsSumY ? uiSetCbf : 0, TEXT_LUMA, uiAbsPartIdx, uiDepth ); 4945 4946 if( bCodeChroma ) 4947 { 4948 if (m_pcEncCfg->getUseRDOQ()) 4949 { 4950 m_pcEntropyCoder->estimateBit(m_pcTrQuant->m_pcEstBitsSbac, trWidthC, trHeightC, TEXT_CHROMA ); 4951 } 4952 4953 Int curChromaQpOffset = pcCU->getSlice()->getPPS()->getChromaCbQpOffset() + pcCU->getSlice()->getSliceQpDeltaCb(); 4954 m_pcTrQuant->setQPforQuant( pcCU->getQP( 0 ), TEXT_CHROMA, pcCU->getSlice()->getSPS()->getQpBDOffsetC(), curChromaQpOffset ); 4955 4956 #if RDOQ_CHROMA_LAMBDA 4957 m_pcTrQuant->selectLambda(TEXT_CHROMA); 4958 #endif 4959 4960 m_pcTrQuant->transformNxN( pcCU, pcResi->getCbAddr(absTUPartIdxC), pcResi->getCStride(), pcCoeffCurrU, 4961 #if ADAPTIVE_QP_SELECTION 4962 pcArlCoeffCurrU, 4963 #endif 4964 trWidthC, trHeightC, uiAbsSumU, TEXT_CHROMA_U, uiAbsPartIdx ); 4965 4966 curChromaQpOffset = pcCU->getSlice()->getPPS()->getChromaCrQpOffset() + pcCU->getSlice()->getSliceQpDeltaCr(); 4967 m_pcTrQuant->setQPforQuant( pcCU->getQP( 0 ), TEXT_CHROMA, pcCU->getSlice()->getSPS()->getQpBDOffsetC(), curChromaQpOffset ); 4968 m_pcTrQuant->transformNxN( pcCU, pcResi->getCrAddr(absTUPartIdxC), pcResi->getCStride(), pcCoeffCurrV, 4969 #if ADAPTIVE_QP_SELECTION 4970 pcArlCoeffCurrV, 4971 #endif 4972 trWidthC, trHeightC, uiAbsSumV, TEXT_CHROMA_V, uiAbsPartIdx ); 4973 4974 pcCU->setCbfSubParts( uiAbsSumU ? uiSetCbf : 0, TEXT_CHROMA_U, uiAbsPartIdx, pcCU->getDepth(0)+uiTrModeC ); 4975 pcCU->setCbfSubParts( uiAbsSumV ? uiSetCbf : 0, TEXT_CHROMA_V, uiAbsPartIdx, pcCU->getDepth(0)+uiTrModeC ); 4976 } 4977 4978 m_pcEntropyCoder->resetBits(); 4979 4980 { 4981 m_pcEntropyCoder->encodeQtCbf( pcCU, uiAbsPartIdx, TEXT_LUMA, uiTrMode ); 4982 } 4983 4984 m_pcEntropyCoder->encodeCoeffNxN( pcCU, pcCoeffCurrY, uiAbsPartIdx, trWidth, trHeight, uiDepth, TEXT_LUMA ); 4985 const UInt uiSingleBitsY = m_pcEntropyCoder->getNumberOfWrittenBits(); 4986 4987 UInt uiSingleBitsU = 0; 4988 UInt uiSingleBitsV = 0; 4989 if( bCodeChroma ) 4990 { 4991 { 4992 m_pcEntropyCoder->encodeQtCbf ( pcCU, uiAbsPartIdx, TEXT_CHROMA_U, uiTrMode ); 4993 } 4994 m_pcEntropyCoder->encodeCoeffNxN( pcCU, pcCoeffCurrU, uiAbsPartIdx, trWidthC, trHeightC, uiDepth, TEXT_CHROMA_U ); 4995 uiSingleBitsU = m_pcEntropyCoder->getNumberOfWrittenBits() - uiSingleBitsY; 4996 4997 { 4998 m_pcEntropyCoder->encodeQtCbf ( pcCU, uiAbsPartIdx, TEXT_CHROMA_V, uiTrMode ); 4999 } 5000 m_pcEntropyCoder->encodeCoeffNxN( pcCU, pcCoeffCurrV, uiAbsPartIdx, trWidthC, trHeightC, uiDepth, TEXT_CHROMA_V ); 5001 uiSingleBitsV = m_pcEntropyCoder->getNumberOfWrittenBits() - ( uiSingleBitsY + uiSingleBitsU ); 5002 } 5003 5004 const UInt uiNumSamplesLuma = 1 << (uiLog2TrSize<<1); 5005 const UInt uiNumSamplesChro = 1 << (uiLog2TrSizeC<<1); 5006 5007 ::memset( m_pTempPel, 0, sizeof( Pel ) * uiNumSamplesLuma ); // not necessary needed for inside of recursion (only at the beginning) 5008 5009 UInt uiDistY = m_pcRdCost->getDistPart(g_bitDepthY, m_pTempPel, trWidth, pcResi->getLumaAddr( absTUPartIdx ), pcResi->getStride(), trWidth, trHeight ); // initialized with zero residual destortion 5010 5011 if ( puiZeroDist ) 5012 { 5013 *puiZeroDist += uiDistY; 5014 } 5015 if( uiAbsSumY ) 5016 { 5017 Pel *pcResiCurrY = m_pcQTTempTComYuv[ uiQTTempAccessLayer ].getLumaAddr( absTUPartIdx ); 5018 5019 m_pcTrQuant->setQPforQuant( pcCU->getQP( 0 ), TEXT_LUMA, pcCU->getSlice()->getSPS()->getQpBDOffsetY(), 0 ); 5020 5021 Int scalingListType = 3 + g_eTTable[(Int)TEXT_LUMA]; 5022 assert(scalingListType < 6); 5023 #if NO_RESIDUAL_FLAG_FOR_BLPRED 5024 if(pcCU->isIntraBL(uiAbsPartIdx) ) 5025 { 5026 m_pcTrQuant->invtransformNxN( pcCU->getCUTransquantBypass(uiAbsPartIdx), TEXT_LUMA,DC_IDX, pcResiCurrY, m_pcQTTempTComYuv[uiQTTempAccessLayer].getStride(), pcCoeffCurrY, trWidth, trHeight, scalingListType );//this is for inter mode only 5027 } 5028 else 5029 { 5030 m_pcTrQuant->invtransformNxN( pcCU->getCUTransquantBypass(uiAbsPartIdx), TEXT_LUMA,REG_DCT, pcResiCurrY, m_pcQTTempTComYuv[uiQTTempAccessLayer].getStride(), pcCoeffCurrY, trWidth, trHeight, scalingListType );//this is for inter mode only 5031 } 5032 #else 5033 m_pcTrQuant->invtransformNxN( pcCU->getCUTransquantBypass(uiAbsPartIdx), TEXT_LUMA,REG_DCT, pcResiCurrY, m_pcQTTempTComYuv[uiQTTempAccessLayer].getStride(), pcCoeffCurrY, trWidth, trHeight, scalingListType );//this is for inter mode only 5034 #endif 5035 5036 const UInt uiNonzeroDistY = m_pcRdCost->getDistPart(g_bitDepthY, m_pcQTTempTComYuv[uiQTTempAccessLayer].getLumaAddr( absTUPartIdx ), m_pcQTTempTComYuv[uiQTTempAccessLayer].getStride(), 5037 pcResi->getLumaAddr( absTUPartIdx ), pcResi->getStride(), trWidth,trHeight ); 5038 if (pcCU->isLosslessCoded(0)) 5039 { 5040 uiDistY = uiNonzeroDistY; 5041 } 5042 else 5043 { 5044 const Double singleCostY = m_pcRdCost->calcRdCost( uiSingleBitsY, uiNonzeroDistY ); 5045 m_pcEntropyCoder->resetBits(); 5046 m_pcEntropyCoder->encodeQtCbfZero( pcCU, TEXT_LUMA, uiTrMode ); 5047 const UInt uiNullBitsY = m_pcEntropyCoder->getNumberOfWrittenBits(); 5048 const Double nullCostY = m_pcRdCost->calcRdCost( uiNullBitsY, uiDistY ); 5049 if( nullCostY < singleCostY ) 5050 { 5051 uiAbsSumY = 0; 5052 ::memset( pcCoeffCurrY, 0, sizeof( TCoeff ) * uiNumSamplesLuma ); 5053 if( checkTransformSkipY ) 5054 { 5055 minCostY = nullCostY; 5056 } 5057 } 5058 else 5059 { 5060 uiDistY = uiNonzeroDistY; 5061 if( checkTransformSkipY ) 5062 { 5063 minCostY = singleCostY; 5064 } 5065 } 5066 } 5067 } 5068 else if( checkTransformSkipY ) 5069 { 5070 m_pcEntropyCoder->resetBits(); 5071 m_pcEntropyCoder->encodeQtCbfZero( pcCU, TEXT_LUMA, uiTrMode ); 5072 const UInt uiNullBitsY = m_pcEntropyCoder->getNumberOfWrittenBits(); 5073 minCostY = m_pcRdCost->calcRdCost( uiNullBitsY, uiDistY ); 5074 } 5075 5076 if( !uiAbsSumY ) 5077 { 5078 Pel *pcPtr = m_pcQTTempTComYuv[uiQTTempAccessLayer].getLumaAddr( absTUPartIdx ); 5079 const UInt uiStride = m_pcQTTempTComYuv[uiQTTempAccessLayer].getStride(); 5080 for( UInt uiY = 0; uiY < trHeight; ++uiY ) 5081 { 5082 ::memset( pcPtr, 0, sizeof( Pel ) * trWidth ); 5083 pcPtr += uiStride; 5084 } 5085 } 5086 5087 UInt uiDistU = 0; 5088 UInt uiDistV = 0; 5089 if( bCodeChroma ) 5090 { 5091 uiDistU = m_pcRdCost->getDistPart(g_bitDepthC, m_pTempPel, trWidthC, pcResi->getCbAddr( absTUPartIdxC ), pcResi->getCStride(), trWidthC, trHeightC 5092 #if WEIGHTED_CHROMA_DISTORTION 5093 , TEXT_CHROMA_U 5094 #endif 5095 ); // initialized with zero residual destortion 5096 if ( puiZeroDist ) 5097 { 5098 *puiZeroDist += uiDistU; 5099 } 5100 if( uiAbsSumU ) 5101 { 5102 Pel *pcResiCurrU = m_pcQTTempTComYuv[uiQTTempAccessLayer].getCbAddr( absTUPartIdxC ); 5103 5104 Int curChromaQpOffset = pcCU->getSlice()->getPPS()->getChromaCbQpOffset() + pcCU->getSlice()->getSliceQpDeltaCb(); 5105 m_pcTrQuant->setQPforQuant( pcCU->getQP( 0 ), TEXT_CHROMA, pcCU->getSlice()->getSPS()->getQpBDOffsetC(), curChromaQpOffset ); 5106 5107 Int scalingListType = 3 + g_eTTable[(Int)TEXT_CHROMA_U]; 5108 assert(scalingListType < 6); 5109 m_pcTrQuant->invtransformNxN( pcCU->getCUTransquantBypass(uiAbsPartIdx), TEXT_CHROMA,REG_DCT, pcResiCurrU, m_pcQTTempTComYuv[uiQTTempAccessLayer].getCStride(), pcCoeffCurrU, trWidthC, trHeightC, scalingListType ); 5110 5111 const UInt uiNonzeroDistU = m_pcRdCost->getDistPart(g_bitDepthC, m_pcQTTempTComYuv[uiQTTempAccessLayer].getCbAddr( absTUPartIdxC), m_pcQTTempTComYuv[uiQTTempAccessLayer].getCStride(), 5112 pcResi->getCbAddr( absTUPartIdxC), pcResi->getCStride(), trWidthC, trHeightC 5113 #if WEIGHTED_CHROMA_DISTORTION 5114 , TEXT_CHROMA_U 5115 #endif 5116 ); 5117 5118 if(pcCU->isLosslessCoded(0)) 5119 { 5120 uiDistU = uiNonzeroDistU; 5121 } 5122 else 5123 { 5124 const Double dSingleCostU = m_pcRdCost->calcRdCost( uiSingleBitsU, uiNonzeroDistU ); 5125 m_pcEntropyCoder->resetBits(); 5126 m_pcEntropyCoder->encodeQtCbfZero( pcCU, TEXT_CHROMA_U, uiTrMode ); 5127 const UInt uiNullBitsU = m_pcEntropyCoder->getNumberOfWrittenBits(); 5128 const Double dNullCostU = m_pcRdCost->calcRdCost( uiNullBitsU, uiDistU ); 5129 if( dNullCostU < dSingleCostU ) 5130 { 5131 uiAbsSumU = 0; 5132 ::memset( pcCoeffCurrU, 0, sizeof( TCoeff ) * uiNumSamplesChro ); 5133 if( checkTransformSkipUV ) 5134 { 5135 minCostU = dNullCostU; 5136 } 5137 } 5138 else 5139 { 5140 uiDistU = uiNonzeroDistU; 5141 if( checkTransformSkipUV ) 5142 { 5143 minCostU = dSingleCostU; 5144 } 5145 } 5146 } 5147 } 5148 else if( checkTransformSkipUV ) 5149 { 5150 m_pcEntropyCoder->resetBits(); 5151 m_pcEntropyCoder->encodeQtCbfZero( pcCU, TEXT_CHROMA_U, uiTrModeC ); 5152 const UInt uiNullBitsU = m_pcEntropyCoder->getNumberOfWrittenBits(); 5153 minCostU = m_pcRdCost->calcRdCost( uiNullBitsU, uiDistU ); 5154 } 5155 if( !uiAbsSumU ) 5156 { 5157 Pel *pcPtr = m_pcQTTempTComYuv[uiQTTempAccessLayer].getCbAddr( absTUPartIdxC ); 5158 const UInt uiStride = m_pcQTTempTComYuv[uiQTTempAccessLayer].getCStride(); 5159 for( UInt uiY = 0; uiY < trHeightC; ++uiY ) 5160 { 5161 ::memset( pcPtr, 0, sizeof(Pel) * trWidthC ); 5162 pcPtr += uiStride; 5163 } 5164 } 5165 5166 uiDistV = m_pcRdCost->getDistPart(g_bitDepthC, m_pTempPel, trWidthC, pcResi->getCrAddr( absTUPartIdxC), pcResi->getCStride(), trWidthC, trHeightC 5167 #if WEIGHTED_CHROMA_DISTORTION 5168 , TEXT_CHROMA_V 5169 #endif 5170 ); // initialized with zero residual destortion 5171 if ( puiZeroDist ) 5172 { 5173 *puiZeroDist += uiDistV; 5174 } 5175 if( uiAbsSumV ) 5176 { 5177 Pel *pcResiCurrV = m_pcQTTempTComYuv[uiQTTempAccessLayer].getCrAddr( absTUPartIdxC ); 5178 Int curChromaQpOffset = pcCU->getSlice()->getPPS()->getChromaCrQpOffset() + pcCU->getSlice()->getSliceQpDeltaCr(); 5179 m_pcTrQuant->setQPforQuant( pcCU->getQP( 0 ), TEXT_CHROMA, pcCU->getSlice()->getSPS()->getQpBDOffsetC(), curChromaQpOffset ); 5180 5181 Int scalingListType = 3 + g_eTTable[(Int)TEXT_CHROMA_V]; 5182 assert(scalingListType < 6); 5183 m_pcTrQuant->invtransformNxN( pcCU->getCUTransquantBypass(uiAbsPartIdx), TEXT_CHROMA,REG_DCT, pcResiCurrV, m_pcQTTempTComYuv[uiQTTempAccessLayer].getCStride(), pcCoeffCurrV, trWidthC, trHeightC, scalingListType ); 5184 5185 const UInt uiNonzeroDistV = m_pcRdCost->getDistPart(g_bitDepthC, m_pcQTTempTComYuv[uiQTTempAccessLayer].getCrAddr( absTUPartIdxC ), m_pcQTTempTComYuv[uiQTTempAccessLayer].getCStride(), 5186 pcResi->getCrAddr( absTUPartIdxC ), pcResi->getCStride(), trWidthC, trHeightC 5187 #if WEIGHTED_CHROMA_DISTORTION 5188 , TEXT_CHROMA_V 5189 #endif 5190 ); 5191 if (pcCU->isLosslessCoded(0)) 5192 { 5193 uiDistV = uiNonzeroDistV; 5194 } 5195 else 5196 { 5197 const Double dSingleCostV = m_pcRdCost->calcRdCost( uiSingleBitsV, uiNonzeroDistV ); 5198 m_pcEntropyCoder->resetBits(); 5199 m_pcEntropyCoder->encodeQtCbfZero( pcCU, TEXT_CHROMA_V, uiTrMode ); 5200 const UInt uiNullBitsV = m_pcEntropyCoder->getNumberOfWrittenBits(); 5201 const Double dNullCostV = m_pcRdCost->calcRdCost( uiNullBitsV, uiDistV ); 5202 if( dNullCostV < dSingleCostV ) 5203 { 5204 uiAbsSumV = 0; 5205 ::memset( pcCoeffCurrV, 0, sizeof( TCoeff ) * uiNumSamplesChro ); 5206 if( checkTransformSkipUV ) 5207 { 5208 minCostV = dNullCostV; 5209 } 5210 } 5211 else 5212 { 5213 uiDistV = uiNonzeroDistV; 5214 if( checkTransformSkipUV ) 5215 { 5216 minCostV = dSingleCostV; 5217 } 5218 } 5219 } 5220 } 5221 else if( checkTransformSkipUV ) 5222 { 5223 m_pcEntropyCoder->resetBits(); 5224 m_pcEntropyCoder->encodeQtCbfZero( pcCU, TEXT_CHROMA_V, uiTrModeC ); 5225 const UInt uiNullBitsV = m_pcEntropyCoder->getNumberOfWrittenBits(); 5226 minCostV = m_pcRdCost->calcRdCost( uiNullBitsV, uiDistV ); 5227 } 5228 if( !uiAbsSumV ) 5229 { 5230 Pel *pcPtr = m_pcQTTempTComYuv[uiQTTempAccessLayer].getCrAddr( absTUPartIdxC ); 5231 const UInt uiStride = m_pcQTTempTComYuv[uiQTTempAccessLayer].getCStride(); 5232 for( UInt uiY = 0; uiY < trHeightC; ++uiY ) 5233 { 5234 ::memset( pcPtr, 0, sizeof(Pel) * trWidthC ); 5235 pcPtr += uiStride; 5236 } 5237 } 5238 } 5239 pcCU->setCbfSubParts( uiAbsSumY ? uiSetCbf : 0, TEXT_LUMA, uiAbsPartIdx, uiDepth ); 5240 if( bCodeChroma ) 5241 { 5242 pcCU->setCbfSubParts( uiAbsSumU ? uiSetCbf : 0, TEXT_CHROMA_U, uiAbsPartIdx, pcCU->getDepth(0)+uiTrModeC ); 5243 pcCU->setCbfSubParts( uiAbsSumV ? uiSetCbf : 0, TEXT_CHROMA_V, uiAbsPartIdx, pcCU->getDepth(0)+uiTrModeC ); 5244 } 5245 5246 if( checkTransformSkipY ) 5247 { 5248 UInt uiNonzeroDistY, uiAbsSumTransformSkipY; 5249 Double dSingleCostY; 5250 5251 Pel *pcResiCurrY = m_pcQTTempTComYuv[ uiQTTempAccessLayer ].getLumaAddr( absTUPartIdx ); 5252 UInt resiYStride = m_pcQTTempTComYuv[ uiQTTempAccessLayer ].getStride(); 5253 5254 TCoeff bestCoeffY[32*32]; 5255 memcpy( bestCoeffY, pcCoeffCurrY, sizeof(TCoeff) * uiNumSamplesLuma ); 5256 5257 #if ADAPTIVE_QP_SELECTION 5258 TCoeff bestArlCoeffY[32*32]; 5259 memcpy( bestArlCoeffY, pcArlCoeffCurrY, sizeof(TCoeff) * uiNumSamplesLuma ); 5260 #endif 5261 5262 Pel bestResiY[32*32]; 5263 for ( Int i = 0; i < trHeight; ++i ) 5264 { 5265 memcpy( &bestResiY[i*trWidth], pcResiCurrY+i*resiYStride, sizeof(Pel) * trWidth ); 5266 } 5267 5268 if( m_bUseSBACRD ) 5269 { 5270 m_pcRDGoOnSbacCoder->load( m_pppcRDSbacCoder[ uiDepth ][ CI_QT_TRAFO_ROOT ] ); 5271 } 5272 5273 pcCU->setTransformSkipSubParts ( 1, TEXT_LUMA, uiAbsPartIdx, uiDepth ); 5274 5275 if (m_pcEncCfg->getUseRDOQTS()) 5276 { 5277 m_pcEntropyCoder->estimateBit( m_pcTrQuant->m_pcEstBitsSbac, trWidth, trHeight, TEXT_LUMA ); 5278 } 5279 5280 m_pcTrQuant->setQPforQuant( pcCU->getQP( 0 ), TEXT_LUMA, pcCU->getSlice()->getSPS()->getQpBDOffsetY(), 0 ); 5281 5282 #if RDOQ_CHROMA_LAMBDA 5283 m_pcTrQuant->selectLambda(TEXT_LUMA); 5284 #endif 5285 m_pcTrQuant->transformNxN( pcCU, pcResi->getLumaAddr( absTUPartIdx ), pcResi->getStride (), pcCoeffCurrY, 5286 #if ADAPTIVE_QP_SELECTION 5287 pcArlCoeffCurrY, 5288 #endif 5289 trWidth, trHeight, uiAbsSumTransformSkipY, TEXT_LUMA, uiAbsPartIdx, true ); 5290 pcCU->setCbfSubParts( uiAbsSumTransformSkipY ? uiSetCbf : 0, TEXT_LUMA, uiAbsPartIdx, uiDepth ); 5291 5292 if( uiAbsSumTransformSkipY != 0 ) 5293 { 5294 m_pcEntropyCoder->resetBits(); 5295 m_pcEntropyCoder->encodeQtCbf( pcCU, uiAbsPartIdx, TEXT_LUMA, uiTrMode ); 5296 m_pcEntropyCoder->encodeCoeffNxN( pcCU, pcCoeffCurrY, uiAbsPartIdx, trWidth, trHeight, uiDepth, TEXT_LUMA ); 5297 const UInt uiTsSingleBitsY = m_pcEntropyCoder->getNumberOfWrittenBits(); 5298 5299 m_pcTrQuant->setQPforQuant( pcCU->getQP( 0 ), TEXT_LUMA, pcCU->getSlice()->getSPS()->getQpBDOffsetY(), 0 ); 5300 5301 Int scalingListType = 3 + g_eTTable[(Int)TEXT_LUMA]; 5302 assert(scalingListType < 6); 5303 5304 #if NO_RESIDUAL_FLAG_FOR_BLPRED 5305 if(pcCU->isIntraBL(uiAbsPartIdx) ) 5306 { 5307 m_pcTrQuant->invtransformNxN( pcCU->getCUTransquantBypass(uiAbsPartIdx), TEXT_LUMA,DC_IDX, pcResiCurrY, m_pcQTTempTComYuv[uiQTTempAccessLayer].getStride(), pcCoeffCurrY, trWidth, trHeight, scalingListType, true ); 5308 } 5309 else 5310 { 5311 m_pcTrQuant->invtransformNxN( pcCU->getCUTransquantBypass(uiAbsPartIdx), TEXT_LUMA,REG_DCT, pcResiCurrY, m_pcQTTempTComYuv[uiQTTempAccessLayer].getStride(), pcCoeffCurrY, trWidth, trHeight, scalingListType, true ); 5312 } 5313 #else 5314 m_pcTrQuant->invtransformNxN( pcCU->getCUTransquantBypass(uiAbsPartIdx), TEXT_LUMA,REG_DCT, pcResiCurrY, m_pcQTTempTComYuv[uiQTTempAccessLayer].getStride(), pcCoeffCurrY, trWidth, trHeight, scalingListType, true ); 5315 #endif 5316 5317 uiNonzeroDistY = m_pcRdCost->getDistPart(g_bitDepthY, m_pcQTTempTComYuv[uiQTTempAccessLayer].getLumaAddr( absTUPartIdx ), m_pcQTTempTComYuv[uiQTTempAccessLayer].getStride(), 5318 pcResi->getLumaAddr( absTUPartIdx ), pcResi->getStride(), trWidth, trHeight ); 5319 5320 dSingleCostY = m_pcRdCost->calcRdCost( uiTsSingleBitsY, uiNonzeroDistY ); 5321 } 5322 5323 if( !uiAbsSumTransformSkipY || minCostY < dSingleCostY ) 5324 { 5325 pcCU->setTransformSkipSubParts ( 0, TEXT_LUMA, uiAbsPartIdx, uiDepth ); 5326 memcpy( pcCoeffCurrY, bestCoeffY, sizeof(TCoeff) * uiNumSamplesLuma ); 5327 #if ADAPTIVE_QP_SELECTION 5328 memcpy( pcArlCoeffCurrY, bestArlCoeffY, sizeof(TCoeff) * uiNumSamplesLuma ); 5329 #endif 5330 for( Int i = 0; i < trHeight; ++i ) 5331 { 5332 memcpy( pcResiCurrY+i*resiYStride, &bestResiY[i*trWidth], sizeof(Pel) * trWidth ); 5333 } 5334 } 5335 else 5336 { 5337 uiDistY = uiNonzeroDistY; 5338 uiAbsSumY = uiAbsSumTransformSkipY; 5339 uiBestTransformMode[0] = 1; 5340 } 5341 5342 pcCU->setCbfSubParts( uiAbsSumY ? uiSetCbf : 0, TEXT_LUMA, uiAbsPartIdx, uiDepth ); 5343 } 5344 5345 if( bCodeChroma && checkTransformSkipUV ) 5346 { 5347 UInt uiNonzeroDistU, uiNonzeroDistV, uiAbsSumTransformSkipU, uiAbsSumTransformSkipV; 5348 Double dSingleCostU, dSingleCostV; 5349 5350 Pel *pcResiCurrU = m_pcQTTempTComYuv[uiQTTempAccessLayer].getCbAddr( absTUPartIdxC ); 5351 Pel *pcResiCurrV = m_pcQTTempTComYuv[uiQTTempAccessLayer].getCrAddr( absTUPartIdxC ); 5352 UInt resiCStride = m_pcQTTempTComYuv[uiQTTempAccessLayer].getCStride(); 5353 5354 TCoeff bestCoeffU[32*32], bestCoeffV[32*32]; 5355 memcpy( bestCoeffU, pcCoeffCurrU, sizeof(TCoeff) * uiNumSamplesChro ); 5356 memcpy( bestCoeffV, pcCoeffCurrV, sizeof(TCoeff) * uiNumSamplesChro ); 5357 5358 #if ADAPTIVE_QP_SELECTION 5359 TCoeff bestArlCoeffU[32*32], bestArlCoeffV[32*32]; 5360 memcpy( bestArlCoeffU, pcArlCoeffCurrU, sizeof(TCoeff) * uiNumSamplesChro ); 5361 memcpy( bestArlCoeffV, pcArlCoeffCurrV, sizeof(TCoeff) * uiNumSamplesChro ); 5362 #endif 5363 5364 Pel bestResiU[32*32], bestResiV[32*32]; 5365 for (Int i = 0; i < trHeightC; ++i ) 5366 { 5367 memcpy( &bestResiU[i*trWidthC], pcResiCurrU+i*resiCStride, sizeof(Pel) * trWidthC ); 5368 memcpy( &bestResiV[i*trWidthC], pcResiCurrV+i*resiCStride, sizeof(Pel) * trWidthC ); 5369 } 5370 5371 if( m_bUseSBACRD ) 5372 { 5373 m_pcRDGoOnSbacCoder->load( m_pppcRDSbacCoder[ uiDepth ][ CI_QT_TRAFO_ROOT ] ); 5374 } 5375 5376 pcCU->setTransformSkipSubParts ( 1, TEXT_CHROMA_U, uiAbsPartIdx, pcCU->getDepth(0)+uiTrModeC ); 5377 pcCU->setTransformSkipSubParts ( 1, TEXT_CHROMA_V, uiAbsPartIdx, pcCU->getDepth(0)+uiTrModeC ); 5378 5379 if (m_pcEncCfg->getUseRDOQTS()) 5380 { 5381 m_pcEntropyCoder->estimateBit(m_pcTrQuant->m_pcEstBitsSbac, trWidthC, trHeightC, TEXT_CHROMA ); 5382 } 5383 5384 Int curChromaQpOffset = pcCU->getSlice()->getPPS()->getChromaCbQpOffset() + pcCU->getSlice()->getSliceQpDeltaCb(); 5385 m_pcTrQuant->setQPforQuant( pcCU->getQP( 0 ), TEXT_CHROMA, pcCU->getSlice()->getSPS()->getQpBDOffsetC(), curChromaQpOffset ); 5386 5387 #if RDOQ_CHROMA_LAMBDA 5388 m_pcTrQuant->selectLambda(TEXT_CHROMA); 5389 #endif 5390 5391 m_pcTrQuant->transformNxN( pcCU, pcResi->getCbAddr(absTUPartIdxC), pcResi->getCStride(), pcCoeffCurrU, 5392 #if ADAPTIVE_QP_SELECTION 5393 pcArlCoeffCurrU, 5394 #endif 5395 trWidthC, trHeightC, uiAbsSumTransformSkipU, TEXT_CHROMA_U, uiAbsPartIdx, true ); 5396 curChromaQpOffset = pcCU->getSlice()->getPPS()->getChromaCrQpOffset() + pcCU->getSlice()->getSliceQpDeltaCr(); 5397 m_pcTrQuant->setQPforQuant( pcCU->getQP( 0 ), TEXT_CHROMA, pcCU->getSlice()->getSPS()->getQpBDOffsetC(), curChromaQpOffset ); 5398 m_pcTrQuant->transformNxN( pcCU, pcResi->getCrAddr(absTUPartIdxC), pcResi->getCStride(), pcCoeffCurrV, 5399 #if ADAPTIVE_QP_SELECTION 5400 pcArlCoeffCurrV, 5401 #endif 5402 trWidthC, trHeightC, uiAbsSumTransformSkipV, TEXT_CHROMA_V, uiAbsPartIdx, true ); 5403 5404 pcCU->setCbfSubParts( uiAbsSumTransformSkipU ? uiSetCbf : 0, TEXT_CHROMA_U, uiAbsPartIdx, pcCU->getDepth(0)+uiTrModeC ); 5405 pcCU->setCbfSubParts( uiAbsSumTransformSkipV ? uiSetCbf : 0, TEXT_CHROMA_V, uiAbsPartIdx, pcCU->getDepth(0)+uiTrModeC ); 5406 5407 m_pcEntropyCoder->resetBits(); 5408 uiSingleBitsU = 0; 5409 uiSingleBitsV = 0; 5410 5411 if( uiAbsSumTransformSkipU ) 5412 { 5413 m_pcEntropyCoder->encodeQtCbf ( pcCU, uiAbsPartIdx, TEXT_CHROMA_U, uiTrMode ); 5414 m_pcEntropyCoder->encodeCoeffNxN( pcCU, pcCoeffCurrU, uiAbsPartIdx, trWidthC, trHeightC, uiDepth, TEXT_CHROMA_U ); 5415 uiSingleBitsU = m_pcEntropyCoder->getNumberOfWrittenBits(); 5416 5417 curChromaQpOffset = pcCU->getSlice()->getPPS()->getChromaCbQpOffset() + pcCU->getSlice()->getSliceQpDeltaCb(); 5418 m_pcTrQuant->setQPforQuant( pcCU->getQP( 0 ), TEXT_CHROMA, pcCU->getSlice()->getSPS()->getQpBDOffsetC(), curChromaQpOffset ); 5419 5420 Int scalingListType = 3 + g_eTTable[(Int)TEXT_CHROMA_U]; 5421 assert(scalingListType < 6); 5422 5423 m_pcTrQuant->invtransformNxN( pcCU->getCUTransquantBypass(uiAbsPartIdx), TEXT_CHROMA,REG_DCT, pcResiCurrU, m_pcQTTempTComYuv[uiQTTempAccessLayer].getCStride(), pcCoeffCurrU, trWidthC, trHeightC, scalingListType, true ); 5424 5425 uiNonzeroDistU = m_pcRdCost->getDistPart(g_bitDepthC, m_pcQTTempTComYuv[uiQTTempAccessLayer].getCbAddr( absTUPartIdxC), m_pcQTTempTComYuv[uiQTTempAccessLayer].getCStride(), 5426 pcResi->getCbAddr( absTUPartIdxC), pcResi->getCStride(), trWidthC, trHeightC 5427 #if WEIGHTED_CHROMA_DISTORTION 5428 , TEXT_CHROMA_U 5429 #endif 5430 ); 5431 5432 dSingleCostU = m_pcRdCost->calcRdCost( uiSingleBitsU, uiNonzeroDistU ); 5433 } 5434 5435 if( !uiAbsSumTransformSkipU || minCostU < dSingleCostU ) 5436 { 5437 pcCU->setTransformSkipSubParts ( 0, TEXT_CHROMA_U, uiAbsPartIdx, pcCU->getDepth(0)+uiTrModeC ); 5438 5439 memcpy( pcCoeffCurrU, bestCoeffU, sizeof (TCoeff) * uiNumSamplesChro ); 5440 #if ADAPTIVE_QP_SELECTION 5441 memcpy( pcArlCoeffCurrU, bestArlCoeffU, sizeof (TCoeff) * uiNumSamplesChro ); 5442 #endif 5443 for( Int i = 0; i < trHeightC; ++i ) 5444 { 5445 memcpy( pcResiCurrU+i*resiCStride, &bestResiU[i*trWidthC], sizeof(Pel) * trWidthC ); 5446 } 5447 } 5448 else 5449 { 5450 uiDistU = uiNonzeroDistU; 5451 uiAbsSumU = uiAbsSumTransformSkipU; 5452 uiBestTransformMode[1] = 1; 5453 } 5454 5455 if( uiAbsSumTransformSkipV ) 5456 { 5457 m_pcEntropyCoder->encodeQtCbf ( pcCU, uiAbsPartIdx, TEXT_CHROMA_V, uiTrMode ); 5458 m_pcEntropyCoder->encodeCoeffNxN( pcCU, pcCoeffCurrV, uiAbsPartIdx, trWidthC, trHeightC, uiDepth, TEXT_CHROMA_V ); 5459 uiSingleBitsV = m_pcEntropyCoder->getNumberOfWrittenBits() - uiSingleBitsU; 5460 5461 curChromaQpOffset = pcCU->getSlice()->getPPS()->getChromaCrQpOffset() + pcCU->getSlice()->getSliceQpDeltaCr(); 5462 m_pcTrQuant->setQPforQuant( pcCU->getQP( 0 ), TEXT_CHROMA, pcCU->getSlice()->getSPS()->getQpBDOffsetC(), curChromaQpOffset ); 5463 5464 Int scalingListType = 3 + g_eTTable[(Int)TEXT_CHROMA_V]; 5465 assert(scalingListType < 6); 5466 5467 m_pcTrQuant->invtransformNxN( pcCU->getCUTransquantBypass(uiAbsPartIdx), TEXT_CHROMA,REG_DCT, pcResiCurrV, m_pcQTTempTComYuv[uiQTTempAccessLayer].getCStride(), pcCoeffCurrV, trWidthC, trHeightC, scalingListType, true ); 5468 5469 uiNonzeroDistV = m_pcRdCost->getDistPart(g_bitDepthC, m_pcQTTempTComYuv[uiQTTempAccessLayer].getCrAddr( absTUPartIdxC ), m_pcQTTempTComYuv[uiQTTempAccessLayer].getCStride(), 5470 pcResi->getCrAddr( absTUPartIdxC ), pcResi->getCStride(), trWidthC, trHeightC 5471 #if WEIGHTED_CHROMA_DISTORTION 5472 , TEXT_CHROMA_V 5473 #endif 5474 ); 5475 5476 dSingleCostV = m_pcRdCost->calcRdCost( uiSingleBitsV, uiNonzeroDistV ); 5477 } 5478 5479 if( !uiAbsSumTransformSkipV || minCostV < dSingleCostV ) 5480 { 5481 pcCU->setTransformSkipSubParts ( 0, TEXT_CHROMA_V, uiAbsPartIdx, pcCU->getDepth(0)+uiTrModeC ); 5482 5483 memcpy( pcCoeffCurrV, bestCoeffV, sizeof(TCoeff) * uiNumSamplesChro ); 5484 #if ADAPTIVE_QP_SELECTION 5485 memcpy( pcArlCoeffCurrV, bestArlCoeffV, sizeof(TCoeff) * uiNumSamplesChro ); 5486 #endif 5487 for( Int i = 0; i < trHeightC; ++i ) 5488 { 5489 memcpy( pcResiCurrV+i*resiCStride, &bestResiV[i*trWidthC], sizeof(Pel) * trWidthC ); 5490 } 5491 } 5492 else 5493 { 5494 uiDistV = uiNonzeroDistV; 5495 uiAbsSumV = uiAbsSumTransformSkipV; 5496 uiBestTransformMode[2] = 1; 5497 } 5498 5499 pcCU->setCbfSubParts( uiAbsSumU ? uiSetCbf : 0, TEXT_CHROMA_U, uiAbsPartIdx, pcCU->getDepth(0)+uiTrModeC ); 5500 pcCU->setCbfSubParts( uiAbsSumV ? uiSetCbf : 0, TEXT_CHROMA_V, uiAbsPartIdx, pcCU->getDepth(0)+uiTrModeC ); 5501 } 5502 5503 if( m_bUseSBACRD ) 5504 { 5505 m_pcRDGoOnSbacCoder->load( m_pppcRDSbacCoder[ uiDepth ][ CI_QT_TRAFO_ROOT ] ); 5506 } 5507 5508 m_pcEntropyCoder->resetBits(); 5509 5510 { 5511 if( uiLog2TrSize > pcCU->getQuadtreeTULog2MinSizeInCU(uiAbsPartIdx) ) 5512 { 5513 m_pcEntropyCoder->encodeTransformSubdivFlag( 0, 5 - uiLog2TrSize ); 5514 } 5515 } 5516 5517 { 5518 if( bCodeChroma ) 5519 { 5520 m_pcEntropyCoder->encodeQtCbf( pcCU, uiAbsPartIdx, TEXT_CHROMA_U, uiTrMode ); 5521 m_pcEntropyCoder->encodeQtCbf( pcCU, uiAbsPartIdx, TEXT_CHROMA_V, uiTrMode ); 5522 } 5523 5524 m_pcEntropyCoder->encodeQtCbf( pcCU, uiAbsPartIdx, TEXT_LUMA, uiTrMode ); 5525 } 5526 5527 m_pcEntropyCoder->encodeCoeffNxN( pcCU, pcCoeffCurrY, uiAbsPartIdx, trWidth, trHeight, uiDepth, TEXT_LUMA ); 5528 5529 if( bCodeChroma ) 5530 { 5531 m_pcEntropyCoder->encodeCoeffNxN( pcCU, pcCoeffCurrU, uiAbsPartIdx, trWidthC, trHeightC, uiDepth, TEXT_CHROMA_U ); 5532 m_pcEntropyCoder->encodeCoeffNxN( pcCU, pcCoeffCurrV, uiAbsPartIdx, trWidthC, trHeightC, uiDepth, TEXT_CHROMA_V ); 5533 } 5534 5535 uiSingleBits = m_pcEntropyCoder->getNumberOfWrittenBits(); 5536 5537 uiSingleDist = uiDistY + uiDistU + uiDistV; 5538 dSingleCost = m_pcRdCost->calcRdCost( uiSingleBits, uiSingleDist ); 5539 } 5540 5541 // code sub-blocks 5542 if( bCheckSplit ) 5543 { 5544 if( m_bUseSBACRD && bCheckFull ) 5545 { 5546 m_pcRDGoOnSbacCoder->store( m_pppcRDSbacCoder[ uiDepth ][ CI_QT_TRAFO_TEST ] ); 5547 m_pcRDGoOnSbacCoder->load ( m_pppcRDSbacCoder[ uiDepth ][ CI_QT_TRAFO_ROOT ] ); 5548 } 5549 UInt uiSubdivDist = 0; 5550 UInt uiSubdivBits = 0; 5551 Double dSubdivCost = 0.0; 5552 5553 const UInt uiQPartNumSubdiv = pcCU->getPic()->getNumPartInCU() >> ((uiDepth + 1 ) << 1); 5554 for( UInt ui = 0; ui < 4; ++ui ) 5555 { 5556 UInt nsAddr = uiAbsPartIdx + ui * uiQPartNumSubdiv; 5557 xEstimateResidualQT( pcCU, ui, uiAbsPartIdx + ui * uiQPartNumSubdiv, nsAddr, pcResi, uiDepth + 1, dSubdivCost, uiSubdivBits, uiSubdivDist, bCheckFull ? NULL : puiZeroDist ); 5558 } 5559 5560 UInt uiYCbf = 0; 5561 UInt uiUCbf = 0; 5562 UInt uiVCbf = 0; 5563 for( UInt ui = 0; ui < 4; ++ui ) 5564 { 5565 uiYCbf |= pcCU->getCbf( uiAbsPartIdx + ui * uiQPartNumSubdiv, TEXT_LUMA, uiTrMode + 1 ); 5566 uiUCbf |= pcCU->getCbf( uiAbsPartIdx + ui * uiQPartNumSubdiv, TEXT_CHROMA_U, uiTrMode + 1 ); 5567 uiVCbf |= pcCU->getCbf( uiAbsPartIdx + ui * uiQPartNumSubdiv, TEXT_CHROMA_V, uiTrMode + 1 ); 5568 } 5569 for( UInt ui = 0; ui < 4 * uiQPartNumSubdiv; ++ui ) 5570 { 5571 pcCU->getCbf( TEXT_LUMA )[uiAbsPartIdx + ui] |= uiYCbf << uiTrMode; 5572 pcCU->getCbf( TEXT_CHROMA_U )[uiAbsPartIdx + ui] |= uiUCbf << uiTrMode; 5573 pcCU->getCbf( TEXT_CHROMA_V )[uiAbsPartIdx + ui] |= uiVCbf << uiTrMode; 5574 } 5575 5576 if( m_bUseSBACRD ) 5577 { 5578 m_pcRDGoOnSbacCoder->load( m_pppcRDSbacCoder[ uiDepth ][ CI_QT_TRAFO_ROOT ] ); 5579 } 5580 m_pcEntropyCoder->resetBits(); 5581 5582 { 5583 xEncodeResidualQT( pcCU, uiAbsPartIdx, uiDepth, true, TEXT_LUMA ); 5584 xEncodeResidualQT( pcCU, uiAbsPartIdx, uiDepth, false, TEXT_LUMA ); 5585 xEncodeResidualQT( pcCU, uiAbsPartIdx, uiDepth, false, TEXT_CHROMA_U ); 5586 xEncodeResidualQT( pcCU, uiAbsPartIdx, uiDepth, false, TEXT_CHROMA_V ); 5587 } 5588 5589 uiSubdivBits = m_pcEntropyCoder->getNumberOfWrittenBits(); 5590 dSubdivCost = m_pcRdCost->calcRdCost( uiSubdivBits, uiSubdivDist ); 5591 5592 if( uiYCbf || uiUCbf || uiVCbf || !bCheckFull ) 5593 { 5594 if( dSubdivCost < dSingleCost ) 5595 { 5596 rdCost += dSubdivCost; 5597 ruiBits += uiSubdivBits; 5598 ruiDist += uiSubdivDist; 5599 return; 5600 } 5601 } 5602 pcCU->setTransformSkipSubParts ( uiBestTransformMode[0], TEXT_LUMA, uiAbsPartIdx, uiDepth ); 5603 if(bCodeChroma) 5604 { 5605 pcCU->setTransformSkipSubParts ( uiBestTransformMode[1], TEXT_CHROMA_U, uiAbsPartIdx, pcCU->getDepth(0)+uiTrModeC ); 5606 pcCU->setTransformSkipSubParts ( uiBestTransformMode[2], TEXT_CHROMA_V, uiAbsPartIdx, pcCU->getDepth(0)+uiTrModeC ); 5607 } 5608 assert( bCheckFull ); 5609 if( m_bUseSBACRD ) 5610 { 5611 m_pcRDGoOnSbacCoder->load( m_pppcRDSbacCoder[ uiDepth ][ CI_QT_TRAFO_TEST ] ); 5612 } 5613 } 5614 rdCost += dSingleCost; 5615 ruiBits += uiSingleBits; 5616 ruiDist += uiSingleDist; 5617 5618 pcCU->setTrIdxSubParts( uiTrMode, uiAbsPartIdx, uiDepth ); 5619 5620 pcCU->setCbfSubParts( uiAbsSumY ? uiSetCbf : 0, TEXT_LUMA, uiAbsPartIdx, uiDepth ); 5621 if( bCodeChroma ) 5622 { 5623 pcCU->setCbfSubParts( uiAbsSumU ? uiSetCbf : 0, TEXT_CHROMA_U, uiAbsPartIdx, pcCU->getDepth(0)+uiTrModeC ); 5624 pcCU->setCbfSubParts( uiAbsSumV ? uiSetCbf : 0, TEXT_CHROMA_V, uiAbsPartIdx, pcCU->getDepth(0)+uiTrModeC ); 5625 } 5626 } 5627 5628 Void TEncSearch::xEncodeResidualQT( TComDataCU* pcCU, UInt uiAbsPartIdx, const UInt uiDepth, Bool bSubdivAndCbf, TextType eType ) 5629 { 5630 assert( pcCU->getDepth( 0 ) == pcCU->getDepth( uiAbsPartIdx ) ); 5631 const UInt uiCurrTrMode = uiDepth - pcCU->getDepth( 0 ); 5632 const UInt uiTrMode = pcCU->getTransformIdx( uiAbsPartIdx ); 5633 5634 const Bool bSubdiv = uiCurrTrMode != uiTrMode; 5635 5636 const UInt uiLog2TrSize = g_aucConvertToBit[pcCU->getSlice()->getSPS()->getMaxCUWidth() >> uiDepth]+2; 5637 5638 { 5639 if( bSubdivAndCbf && uiLog2TrSize <= pcCU->getSlice()->getSPS()->getQuadtreeTULog2MaxSize() && uiLog2TrSize > pcCU->getQuadtreeTULog2MinSizeInCU(uiAbsPartIdx) ) 5640 { 5641 m_pcEntropyCoder->encodeTransformSubdivFlag( bSubdiv, 5 - uiLog2TrSize ); 5642 } 5643 } 5644 5645 { 5646 #if INTRA_BL 5647 #if NO_RESIDUAL_FLAG_FOR_BLPRED 5648 assert( !pcCU->isIntra(uiAbsPartIdx) || pcCU->isIntraBL(uiAbsPartIdx)); 5649 #else 5650 assert( !pcCU->isIntra(uiAbsPartIdx) ); 5651 #endif 5652 #else 5653 assert( pcCU->getPredictionMode(uiAbsPartIdx) != MODE_INTRA ); 5654 #endif 5655 if( bSubdivAndCbf ) 5656 { 5657 const Bool bFirstCbfOfCU = uiCurrTrMode == 0; 5658 if( bFirstCbfOfCU || uiLog2TrSize > 2 ) 5659 { 5660 if( bFirstCbfOfCU || pcCU->getCbf( uiAbsPartIdx, TEXT_CHROMA_U, uiCurrTrMode - 1 ) ) 5661 { 5662 m_pcEntropyCoder->encodeQtCbf( pcCU, uiAbsPartIdx, TEXT_CHROMA_U, uiCurrTrMode ); 5663 } 5664 if( bFirstCbfOfCU || pcCU->getCbf( uiAbsPartIdx, TEXT_CHROMA_V, uiCurrTrMode - 1 ) ) 5665 { 5666 m_pcEntropyCoder->encodeQtCbf( pcCU, uiAbsPartIdx, TEXT_CHROMA_V, uiCurrTrMode ); 5667 } 5668 } 5669 else if( uiLog2TrSize == 2 ) 5670 { 5671 assert( pcCU->getCbf( uiAbsPartIdx, TEXT_CHROMA_U, uiCurrTrMode ) == pcCU->getCbf( uiAbsPartIdx, TEXT_CHROMA_U, uiCurrTrMode - 1 ) ); 5672 assert( pcCU->getCbf( uiAbsPartIdx, TEXT_CHROMA_V, uiCurrTrMode ) == pcCU->getCbf( uiAbsPartIdx, TEXT_CHROMA_V, uiCurrTrMode - 1 ) ); 5673 } 5674 } 5675 } 5676 5677 if( !bSubdiv ) 5678 { 5679 const UInt uiNumCoeffPerAbsPartIdxIncrement = pcCU->getSlice()->getSPS()->getMaxCUWidth() * pcCU->getSlice()->getSPS()->getMaxCUHeight() >> ( pcCU->getSlice()->getSPS()->getMaxCUDepth() << 1 ); 5680 //assert( 16 == uiNumCoeffPerAbsPartIdxIncrement ); // check 5681 const UInt uiQTTempAccessLayer = pcCU->getSlice()->getSPS()->getQuadtreeTULog2MaxSize() - uiLog2TrSize; 5682 TCoeff *pcCoeffCurrY = m_ppcQTTempCoeffY [uiQTTempAccessLayer] + uiNumCoeffPerAbsPartIdxIncrement * uiAbsPartIdx; 5683 TCoeff *pcCoeffCurrU = m_ppcQTTempCoeffCb[uiQTTempAccessLayer] + (uiNumCoeffPerAbsPartIdxIncrement * uiAbsPartIdx>>2); 5684 TCoeff *pcCoeffCurrV = m_ppcQTTempCoeffCr[uiQTTempAccessLayer] + (uiNumCoeffPerAbsPartIdxIncrement * uiAbsPartIdx>>2); 5685 5686 Bool bCodeChroma = true; 5687 UInt uiTrModeC = uiTrMode; 5688 UInt uiLog2TrSizeC = uiLog2TrSize-1; 5689 if( uiLog2TrSize == 2 ) 5690 { 5691 uiLog2TrSizeC++; 5692 uiTrModeC --; 5693 UInt uiQPDiv = pcCU->getPic()->getNumPartInCU() >> ( ( pcCU->getDepth( 0 ) + uiTrModeC ) << 1 ); 5694 bCodeChroma = ( ( uiAbsPartIdx % uiQPDiv ) == 0 ); 5695 } 5696 5697 if( bSubdivAndCbf ) 5698 { 5699 { 5700 m_pcEntropyCoder->encodeQtCbf( pcCU, uiAbsPartIdx, TEXT_LUMA, uiTrMode ); 5701 } 5702 } 5703 else 5704 { 5705 if( eType == TEXT_LUMA && pcCU->getCbf( uiAbsPartIdx, TEXT_LUMA, uiTrMode ) ) 5706 { 5707 Int trWidth = 1 << uiLog2TrSize; 5708 Int trHeight = 1 << uiLog2TrSize; 5709 m_pcEntropyCoder->encodeCoeffNxN( pcCU, pcCoeffCurrY, uiAbsPartIdx, trWidth, trHeight, uiDepth, TEXT_LUMA ); 5710 } 5711 if( bCodeChroma ) 5712 { 5713 Int trWidth = 1 << uiLog2TrSizeC; 5714 Int trHeight = 1 << uiLog2TrSizeC; 5715 if( eType == TEXT_CHROMA_U && pcCU->getCbf( uiAbsPartIdx, TEXT_CHROMA_U, uiTrMode ) ) 5716 { 5717 m_pcEntropyCoder->encodeCoeffNxN( pcCU, pcCoeffCurrU, uiAbsPartIdx, trWidth, trHeight, uiDepth, TEXT_CHROMA_U ); 5718 } 5719 if( eType == TEXT_CHROMA_V && pcCU->getCbf( uiAbsPartIdx, TEXT_CHROMA_V, uiTrMode ) ) 5720 { 5721 m_pcEntropyCoder->encodeCoeffNxN( pcCU, pcCoeffCurrV, uiAbsPartIdx, trWidth, trHeight, uiDepth, TEXT_CHROMA_V ); 5722 } 5723 } 5724 } 5725 } 5726 else 5727 { 5728 if( bSubdivAndCbf || pcCU->getCbf( uiAbsPartIdx, eType, uiCurrTrMode ) ) 5729 { 5730 const UInt uiQPartNumSubdiv = pcCU->getPic()->getNumPartInCU() >> ((uiDepth + 1 ) << 1); 5731 for( UInt ui = 0; ui < 4; ++ui ) 5732 { 5733 xEncodeResidualQT( pcCU, uiAbsPartIdx + ui * uiQPartNumSubdiv, uiDepth + 1, bSubdivAndCbf, eType ); 5734 } 5735 } 5736 } 5737 } 5738 5739 Void TEncSearch::xSetResidualQTData( TComDataCU* pcCU, UInt uiQuadrant, UInt uiAbsPartIdx, UInt absTUPartIdx, TComYuv* pcResi, UInt uiDepth, Bool bSpatial ) 5740 { 5741 assert( pcCU->getDepth( 0 ) == pcCU->getDepth( uiAbsPartIdx ) ); 5742 const UInt uiCurrTrMode = uiDepth - pcCU->getDepth( 0 ); 5743 const UInt uiTrMode = pcCU->getTransformIdx( uiAbsPartIdx ); 5744 5745 if( uiCurrTrMode == uiTrMode ) 5746 { 5747 const UInt uiLog2TrSize = g_aucConvertToBit[pcCU->getSlice()->getSPS()->getMaxCUWidth() >> uiDepth]+2; 5748 const UInt uiQTTempAccessLayer = pcCU->getSlice()->getSPS()->getQuadtreeTULog2MaxSize() - uiLog2TrSize; 5749 5750 Bool bCodeChroma = true; 5751 UInt uiTrModeC = uiTrMode; 5752 UInt uiLog2TrSizeC = uiLog2TrSize-1; 5753 if( uiLog2TrSize == 2 ) 5754 { 5755 uiLog2TrSizeC++; 5756 uiTrModeC --; 5757 UInt uiQPDiv = pcCU->getPic()->getNumPartInCU() >> ( ( pcCU->getDepth( 0 ) + uiTrModeC ) << 1 ); 5758 bCodeChroma = ( ( uiAbsPartIdx % uiQPDiv ) == 0 ); 5759 } 5760 5761 if( bSpatial ) 5762 { 5763 Int trWidth = 1 << uiLog2TrSize; 5764 Int trHeight = 1 << uiLog2TrSize; 5765 m_pcQTTempTComYuv[uiQTTempAccessLayer].copyPartToPartLuma ( pcResi, absTUPartIdx, trWidth , trHeight ); 5766 5767 if( bCodeChroma ) 5768 { 5769 { 5770 m_pcQTTempTComYuv[uiQTTempAccessLayer].copyPartToPartChroma( pcResi, uiAbsPartIdx, 1 << uiLog2TrSizeC, 1 << uiLog2TrSizeC ); 5771 } 5772 } 5773 } 5774 else 5775 { 5776 UInt uiNumCoeffPerAbsPartIdxIncrement = pcCU->getSlice()->getSPS()->getMaxCUWidth() * pcCU->getSlice()->getSPS()->getMaxCUHeight() >> ( pcCU->getSlice()->getSPS()->getMaxCUDepth() << 1 ); 5777 UInt uiNumCoeffY = ( 1 << ( uiLog2TrSize << 1 ) ); 5778 TCoeff* pcCoeffSrcY = m_ppcQTTempCoeffY [uiQTTempAccessLayer] + uiNumCoeffPerAbsPartIdxIncrement * uiAbsPartIdx; 5779 TCoeff* pcCoeffDstY = pcCU->getCoeffY() + uiNumCoeffPerAbsPartIdxIncrement * uiAbsPartIdx; 5780 ::memcpy( pcCoeffDstY, pcCoeffSrcY, sizeof( TCoeff ) * uiNumCoeffY ); 5781 #if ADAPTIVE_QP_SELECTION 5782 Int* pcArlCoeffSrcY = m_ppcQTTempArlCoeffY [uiQTTempAccessLayer] + uiNumCoeffPerAbsPartIdxIncrement * uiAbsPartIdx; 5783 Int* pcArlCoeffDstY = pcCU->getArlCoeffY() + uiNumCoeffPerAbsPartIdxIncrement * uiAbsPartIdx; 5784 ::memcpy( pcArlCoeffDstY, pcArlCoeffSrcY, sizeof( Int ) * uiNumCoeffY ); 5785 #endif 5786 if( bCodeChroma ) 5787 { 5788 UInt uiNumCoeffC = ( 1 << ( uiLog2TrSizeC << 1 ) ); 5789 TCoeff* pcCoeffSrcU = m_ppcQTTempCoeffCb[uiQTTempAccessLayer] + (uiNumCoeffPerAbsPartIdxIncrement * uiAbsPartIdx>>2); 5790 TCoeff* pcCoeffSrcV = m_ppcQTTempCoeffCr[uiQTTempAccessLayer] + (uiNumCoeffPerAbsPartIdxIncrement * uiAbsPartIdx>>2); 5791 TCoeff* pcCoeffDstU = pcCU->getCoeffCb() + (uiNumCoeffPerAbsPartIdxIncrement * uiAbsPartIdx>>2); 5792 TCoeff* pcCoeffDstV = pcCU->getCoeffCr() + (uiNumCoeffPerAbsPartIdxIncrement * uiAbsPartIdx>>2); 5793 ::memcpy( pcCoeffDstU, pcCoeffSrcU, sizeof( TCoeff ) * uiNumCoeffC ); 5794 ::memcpy( pcCoeffDstV, pcCoeffSrcV, sizeof( TCoeff ) * uiNumCoeffC ); 5795 #if ADAPTIVE_QP_SELECTION 5796 Int* pcArlCoeffSrcU = m_ppcQTTempArlCoeffCb[uiQTTempAccessLayer] + (uiNumCoeffPerAbsPartIdxIncrement * uiAbsPartIdx>>2); 5797 Int* pcArlCoeffSrcV = m_ppcQTTempArlCoeffCr[uiQTTempAccessLayer] + (uiNumCoeffPerAbsPartIdxIncrement * uiAbsPartIdx>>2); 5798 Int* pcArlCoeffDstU = pcCU->getArlCoeffCb() + (uiNumCoeffPerAbsPartIdxIncrement * uiAbsPartIdx>>2); 5799 Int* pcArlCoeffDstV = pcCU->getArlCoeffCr() + (uiNumCoeffPerAbsPartIdxIncrement * uiAbsPartIdx>>2); 5800 ::memcpy( pcArlCoeffDstU, pcArlCoeffSrcU, sizeof( Int ) * uiNumCoeffC ); 5801 ::memcpy( pcArlCoeffDstV, pcArlCoeffSrcV, sizeof( Int ) * uiNumCoeffC ); 5802 #endif 5803 } 5804 } 5805 } 5806 else 5807 { 5808 const UInt uiQPartNumSubdiv = pcCU->getPic()->getNumPartInCU() >> ((uiDepth + 1 ) << 1); 5809 for( UInt ui = 0; ui < 4; ++ui ) 5810 { 5811 UInt nsAddr = uiAbsPartIdx + ui * uiQPartNumSubdiv; 5812 xSetResidualQTData( pcCU, ui, uiAbsPartIdx + ui * uiQPartNumSubdiv, nsAddr, pcResi, uiDepth + 1, bSpatial ); 5813 } 5814 } 5815 } 5816 5817 UInt TEncSearch::xModeBitsIntra( TComDataCU* pcCU, UInt uiMode, UInt uiPU, UInt uiPartOffset, UInt uiDepth, UInt uiInitTrDepth ) 5818 { 5819 if( m_bUseSBACRD ) 5820 { 5821 // Reload only contexts required for coding intra mode information 5822 m_pcRDGoOnSbacCoder->loadIntraDirModeLuma( m_pppcRDSbacCoder[uiDepth][CI_CURR_BEST] ); 5823 } 5824 5825 pcCU->setLumaIntraDirSubParts ( uiMode, uiPartOffset, uiDepth + uiInitTrDepth ); 5826 5827 m_pcEntropyCoder->resetBits(); 5828 m_pcEntropyCoder->encodeIntraDirModeLuma ( pcCU, uiPartOffset); 5829 5830 return m_pcEntropyCoder->getNumberOfWrittenBits(); 5831 } 5832 5833 UInt TEncSearch::xUpdateCandList( UInt uiMode, Double uiCost, UInt uiFastCandNum, UInt * CandModeList, Double * CandCostList ) 5834 { 5835 UInt i; 5836 UInt shift=0; 5837 5838 while ( shift<uiFastCandNum && uiCost<CandCostList[ uiFastCandNum-1-shift ] ) shift++; 5839 5840 if( shift!=0 ) 5841 { 5842 for(i=1; i<shift; i++) 5843 { 5844 CandModeList[ uiFastCandNum-i ] = CandModeList[ uiFastCandNum-1-i ]; 5845 CandCostList[ uiFastCandNum-i ] = CandCostList[ uiFastCandNum-1-i ]; 5846 } 5847 CandModeList[ uiFastCandNum-shift ] = uiMode; 5848 CandCostList[ uiFastCandNum-shift ] = uiCost; 5849 return 1; 5850 } 5851 5852 return 0; 5853 } 5854 5855 /** add inter-prediction syntax elements for a CU block 5856 * \param pcCU 5857 * \param uiQp 5858 * \param uiTrMode 5859 * \param ruiBits 5860 * \param rpcYuvRec 5861 * \param pcYuvPred 5862 * \param rpcYuvResi 5863 * \returns Void 5864 */ 5865 Void TEncSearch::xAddSymbolBitsInter( TComDataCU* pcCU, UInt uiQp, UInt uiTrMode, UInt& ruiBits, TComYuv*& rpcYuvRec, TComYuv*pcYuvPred, TComYuv*& rpcYuvResi ) 5866 { 5867 if(pcCU->getMergeFlag( 0 ) && pcCU->getPartitionSize( 0 ) == SIZE_2Nx2N && !pcCU->getQtRootCbf( 0 )) 5868 { 5869 pcCU->setSkipFlagSubParts( true, 0, pcCU->getDepth(0) ); 5870 5871 m_pcEntropyCoder->resetBits(); 5872 if(pcCU->getSlice()->getPPS()->getTransquantBypassEnableFlag()) 5873 { 5874 m_pcEntropyCoder->encodeCUTransquantBypassFlag(pcCU, 0, true); 5875 } 5876 m_pcEntropyCoder->encodeSkipFlag(pcCU, 0, true); 5877 m_pcEntropyCoder->encodeMergeIndex(pcCU, 0, true); 5878 ruiBits += m_pcEntropyCoder->getNumberOfWrittenBits(); 5879 } 5880 else 5881 { 5882 m_pcEntropyCoder->resetBits(); 5883 if(pcCU->getSlice()->getPPS()->getTransquantBypassEnableFlag()) 5884 { 5885 m_pcEntropyCoder->encodeCUTransquantBypassFlag(pcCU, 0, true); 5886 } 5887 m_pcEntropyCoder->encodeSkipFlag ( pcCU, 0, true ); 5888 #if INTRA_BL 5889 if(m_pcEncCfg->getLayerId()) 5890 { 5891 m_pcEntropyCoder->encodeIntraBLFlag(pcCU, 0, true); 5892 #if !NO_RESIDUAL_FLAG_FOR_BLPRED 5893 assert( pcCU->isIntraBL( 0 ) == false ); 5894 #endif 5895 } 5896 #endif 5897 #if NO_RESIDUAL_FLAG_FOR_BLPRED 5898 if( !pcCU->isIntraBL(0)) 5899 { 5900 #endif 5901 m_pcEntropyCoder->encodePredMode( pcCU, 0, true ); 5902 m_pcEntropyCoder->encodePartSize( pcCU, 0, pcCU->getDepth(0), true ); 5903 m_pcEntropyCoder->encodePredInfo( pcCU, 0, true ); 5904 #if NO_RESIDUAL_FLAG_FOR_BLPRED 5905 } 5906 #endif 5907 Bool bDummy = false; 5908 m_pcEntropyCoder->encodeCoeff ( pcCU, 0, pcCU->getDepth(0), pcCU->getWidth(0), pcCU->getHeight(0), bDummy ); 5909 5910 ruiBits += m_pcEntropyCoder->getNumberOfWrittenBits(); 5911 } 5912 } 5913 5914 /** 5915 * \brief Generate half-sample interpolated block 5916 * 5917 * \param pattern Reference picture ROI 5918 * \param biPred Flag indicating whether block is for biprediction 5919 */ 5920 Void TEncSearch::xExtDIFUpSamplingH( TComPattern* pattern, Bool biPred ) 5921 { 5922 Int width = pattern->getROIYWidth(); 5923 Int height = pattern->getROIYHeight(); 5924 Int srcStride = pattern->getPatternLStride(); 5925 5926 Int intStride = m_filteredBlockTmp[0].getStride(); 5927 Int dstStride = m_filteredBlock[0][0].getStride(); 5928 Short *intPtr; 5929 Short *dstPtr; 5930 Int filterSize = NTAPS_LUMA; 5931 Int halfFilterSize = (filterSize>>1); 5932 Pel *srcPtr = pattern->getROIY() - halfFilterSize*srcStride - 1; 5933 5934 m_if.filterHorLuma(srcPtr, srcStride, m_filteredBlockTmp[0].getLumaAddr(), intStride, width+1, height+filterSize, 0, false); 5935 m_if.filterHorLuma(srcPtr, srcStride, m_filteredBlockTmp[2].getLumaAddr(), intStride, width+1, height+filterSize, 2, false); 5936 5937 intPtr = m_filteredBlockTmp[0].getLumaAddr() + halfFilterSize * intStride + 1; 5938 dstPtr = m_filteredBlock[0][0].getLumaAddr(); 5939 m_if.filterVerLuma(intPtr, intStride, dstPtr, dstStride, width+0, height+0, 0, false, true); 5940 5941 intPtr = m_filteredBlockTmp[0].getLumaAddr() + (halfFilterSize-1) * intStride + 1; 5942 dstPtr = m_filteredBlock[2][0].getLumaAddr(); 5943 m_if.filterVerLuma(intPtr, intStride, dstPtr, dstStride, width+0, height+1, 2, false, true); 5944 5945 intPtr = m_filteredBlockTmp[2].getLumaAddr() + halfFilterSize * intStride; 5946 dstPtr = m_filteredBlock[0][2].getLumaAddr(); 5947 m_if.filterVerLuma(intPtr, intStride, dstPtr, dstStride, width+1, height+0, 0, false, true); 5948 5949 intPtr = m_filteredBlockTmp[2].getLumaAddr() + (halfFilterSize-1) * intStride; 5950 dstPtr = m_filteredBlock[2][2].getLumaAddr(); 5951 m_if.filterVerLuma(intPtr, intStride, dstPtr, dstStride, width+1, height+1, 2, false, true); 5952 } 5953 5954 /** 5955 * \brief Generate quarter-sample interpolated blocks 5956 * 5957 * \param pattern Reference picture ROI 5958 * \param halfPelRef Half-pel mv 5959 * \param biPred Flag indicating whether block is for biprediction 5960 */ 5961 Void TEncSearch::xExtDIFUpSamplingQ( TComPattern* pattern, TComMv halfPelRef, Bool biPred ) 5962 { 5963 Int width = pattern->getROIYWidth(); 5964 Int height = pattern->getROIYHeight(); 5965 Int srcStride = pattern->getPatternLStride(); 5966 5967 Pel *srcPtr; 5968 Int intStride = m_filteredBlockTmp[0].getStride(); 5969 Int dstStride = m_filteredBlock[0][0].getStride(); 5970 Short *intPtr; 5971 Short *dstPtr; 5972 Int filterSize = NTAPS_LUMA; 5973 5974 Int halfFilterSize = (filterSize>>1); 5975 5976 Int extHeight = (halfPelRef.getVer() == 0) ? height + filterSize : height + filterSize-1; 5977 5978 // Horizontal filter 1/4 5979 srcPtr = pattern->getROIY() - halfFilterSize * srcStride - 1; 5980 intPtr = m_filteredBlockTmp[1].getLumaAddr(); 5981 if (halfPelRef.getVer() > 0) 5982 { 5983 srcPtr += srcStride; 5984 } 5985 if (halfPelRef.getHor() >= 0) 5986 { 5987 srcPtr += 1; 5988 } 5989 m_if.filterHorLuma(srcPtr, srcStride, intPtr, intStride, width, extHeight, 1, false); 5990 5991 // Horizontal filter 3/4 5992 srcPtr = pattern->getROIY() - halfFilterSize*srcStride - 1; 5993 intPtr = m_filteredBlockTmp[3].getLumaAddr(); 5994 if (halfPelRef.getVer() > 0) 5995 { 5996 srcPtr += srcStride; 5997 } 5998 if (halfPelRef.getHor() > 0) 5999 { 6000 srcPtr += 1; 6001 } 6002 m_if.filterHorLuma(srcPtr, srcStride, intPtr, intStride, width, extHeight, 3, false); 6003 6004 // Generate @ 1,1 6005 intPtr = m_filteredBlockTmp[1].getLumaAddr() + (halfFilterSize-1) * intStride; 6006 dstPtr = m_filteredBlock[1][1].getLumaAddr(); 6007 if (halfPelRef.getVer() == 0) 6008 { 6009 intPtr += intStride; 6010 } 6011 m_if.filterVerLuma(intPtr, intStride, dstPtr, dstStride, width, height, 1, false, true); 6012 6013 // Generate @ 3,1 6014 intPtr = m_filteredBlockTmp[1].getLumaAddr() + (halfFilterSize-1) * intStride; 6015 dstPtr = m_filteredBlock[3][1].getLumaAddr(); 6016 m_if.filterVerLuma(intPtr, intStride, dstPtr, dstStride, width, height, 3, false, true); 6017 6018 if (halfPelRef.getVer() != 0) 6019 { 6020 // Generate @ 2,1 6021 intPtr = m_filteredBlockTmp[1].getLumaAddr() + (halfFilterSize-1) * intStride; 6022 dstPtr = m_filteredBlock[2][1].getLumaAddr(); 6023 if (halfPelRef.getVer() == 0) 6024 { 6025 intPtr += intStride; 6026 } 6027 m_if.filterVerLuma(intPtr, intStride, dstPtr, dstStride, width, height, 2, false, true); 6028 6029 // Generate @ 2,3 6030 intPtr = m_filteredBlockTmp[3].getLumaAddr() + (halfFilterSize-1) * intStride; 6031 dstPtr = m_filteredBlock[2][3].getLumaAddr(); 6032 if (halfPelRef.getVer() == 0) 6033 { 6034 intPtr += intStride; 6035 } 6036 m_if.filterVerLuma(intPtr, intStride, dstPtr, dstStride, width, height, 2, false, true); 6037 } 6038 else 6039 { 6040 // Generate @ 0,1 6041 intPtr = m_filteredBlockTmp[1].getLumaAddr() + halfFilterSize * intStride; 6042 dstPtr = m_filteredBlock[0][1].getLumaAddr(); 6043 m_if.filterVerLuma(intPtr, intStride, dstPtr, dstStride, width, height, 0, false, true); 6044 6045 // Generate @ 0,3 6046 intPtr = m_filteredBlockTmp[3].getLumaAddr() + halfFilterSize * intStride; 6047 dstPtr = m_filteredBlock[0][3].getLumaAddr(); 6048 m_if.filterVerLuma(intPtr, intStride, dstPtr, dstStride, width, height, 0, false, true); 6049 } 6050 6051 if (halfPelRef.getHor() != 0) 6052 { 6053 // Generate @ 1,2 6054 intPtr = m_filteredBlockTmp[2].getLumaAddr() + (halfFilterSize-1) * intStride; 6055 dstPtr = m_filteredBlock[1][2].getLumaAddr(); 6056 if (halfPelRef.getHor() > 0) 6057 { 6058 intPtr += 1; 6059 } 6060 if (halfPelRef.getVer() >= 0) 6061 { 6062 intPtr += intStride; 6063 } 6064 m_if.filterVerLuma(intPtr, intStride, dstPtr, dstStride, width, height, 1, false, true); 6065 6066 // Generate @ 3,2 6067 intPtr = m_filteredBlockTmp[2].getLumaAddr() + (halfFilterSize-1) * intStride; 6068 dstPtr = m_filteredBlock[3][2].getLumaAddr(); 6069 if (halfPelRef.getHor() > 0) 6070 { 6071 intPtr += 1; 6072 } 6073 if (halfPelRef.getVer() > 0) 6074 { 6075 intPtr += intStride; 6076 } 6077 m_if.filterVerLuma(intPtr, intStride, dstPtr, dstStride, width, height, 3, false, true); 6078 } 6079 else 6080 { 6081 // Generate @ 1,0 6082 intPtr = m_filteredBlockTmp[0].getLumaAddr() + (halfFilterSize-1) * intStride + 1; 6083 dstPtr = m_filteredBlock[1][0].getLumaAddr(); 6084 if (halfPelRef.getVer() >= 0) 6085 { 6086 intPtr += intStride; 6087 } 6088 m_if.filterVerLuma(intPtr, intStride, dstPtr, dstStride, width, height, 1, false, true); 6089 6090 // Generate @ 3,0 6091 intPtr = m_filteredBlockTmp[0].getLumaAddr() + (halfFilterSize-1) * intStride + 1; 6092 dstPtr = m_filteredBlock[3][0].getLumaAddr(); 6093 if (halfPelRef.getVer() > 0) 6094 { 6095 intPtr += intStride; 6096 } 6097 m_if.filterVerLuma(intPtr, intStride, dstPtr, dstStride, width, height, 3, false, true); 6098 } 6099 6100 // Generate @ 1,3 6101 intPtr = m_filteredBlockTmp[3].getLumaAddr() + (halfFilterSize-1) * intStride; 6102 dstPtr = m_filteredBlock[1][3].getLumaAddr(); 6103 if (halfPelRef.getVer() == 0) 6104 { 6105 intPtr += intStride; 6106 } 6107 m_if.filterVerLuma(intPtr, intStride, dstPtr, dstStride, width, height, 1, false, true); 6108 6109 // Generate @ 3,3 6110 intPtr = m_filteredBlockTmp[3].getLumaAddr() + (halfFilterSize-1) * intStride; 6111 dstPtr = m_filteredBlock[3][3].getLumaAddr(); 6112 m_if.filterVerLuma(intPtr, intStride, dstPtr, dstStride, width, height, 3, false, true); 6113 } 6114 6115 /** set wp tables 6116 * \param TComDataCU* pcCU 6117 * \param iRefIdx 6118 * \param eRefPicListCur 6119 * \returns Void 6120 */ 6121 Void TEncSearch::setWpScalingDistParam( TComDataCU* pcCU, Int iRefIdx, RefPicList eRefPicListCur ) 6122 { 6123 if ( iRefIdx<0 ) 6124 { 6125 m_cDistParam.bApplyWeight = false; 6126 return; 6127 } 6128 6129 TComSlice *pcSlice = pcCU->getSlice(); 6130 TComPPS *pps = pcCU->getSlice()->getPPS(); 6131 wpScalingParam *wp0 , *wp1; 6132 m_cDistParam.bApplyWeight = ( pcSlice->getSliceType()==P_SLICE && pps->getUseWP() ) || ( pcSlice->getSliceType()==B_SLICE && pps->getWPBiPred() ) ; 6133 if ( !m_cDistParam.bApplyWeight ) return; 6134 6135 Int iRefIdx0 = ( eRefPicListCur == REF_PIC_LIST_0 ) ? iRefIdx : (-1); 6136 Int iRefIdx1 = ( eRefPicListCur == REF_PIC_LIST_1 ) ? iRefIdx : (-1); 6137 6138 getWpScaling( pcCU, iRefIdx0, iRefIdx1, wp0 , wp1 ); 6139 6140 if ( iRefIdx0 < 0 ) wp0 = NULL; 6141 if ( iRefIdx1 < 0 ) wp1 = NULL; 6142 6143 m_cDistParam.wpCur = NULL; 6144 6145 if ( eRefPicListCur == REF_PIC_LIST_0 ) 6146 { 6147 m_cDistParam.wpCur = wp0; 6148 } 6149 else 6150 { 6151 m_cDistParam.wpCur = wp1; 6152 } 4938 6153 } 4939 6154 … … 4961 6176 } 4962 6177 #endif 4963 /** encode residual and calculate rate-distortion for a CU block 4964 * \param pcCU 4965 * \param pcYuvOrg 4966 * \param pcYuvPred 4967 * \param rpcYuvResi 4968 * \param rpcYuvResiBest 4969 * \param rpcYuvRec 4970 * \param bSkipRes 4971 * \returns Void 4972 */ 4973 Void TEncSearch::encodeResAndCalcRdInterCU( TComDataCU* pcCU, TComYuv* pcYuvOrg, TComYuv* pcYuvPred, TComYuv*& rpcYuvResi, TComYuv*& rpcYuvResiBest, TComYuv*& rpcYuvRec, Bool bSkipRes ) 6178 6179 #if ENCODER_FAST_MODE 6180 Bool TEncSearch::predInterSearchILRUni( TComDataCU* pcCU, TComYuv* pcOrgYuv, TComYuv*& rpcPredYuv, TComYuv*& rpcResiYuv, TComYuv*& rpcRecoYuv ) 4974 6181 { 4975 #if NO_RESIDUAL_FLAG_FOR_BLPRED 4976 if ( pcCU->isIntra(0) && !pcCU->isIntraBL(0)) 4977 #else 4978 if ( pcCU->isIntra(0) ) 4979 #endif 4980 { 4981 return; 4982 } 4983 4984 Bool bHighPass = pcCU->getSlice()->getDepth() ? true : false; 4985 UInt uiBits = 0, uiBitsBest = 0; 4986 UInt uiDistortion = 0, uiDistortionBest = 0; 4987 4988 UInt uiWidth = pcCU->getWidth ( 0 ); 4989 UInt uiHeight = pcCU->getHeight( 0 ); 4990 4991 // No residual coding : SKIP mode 4992 if ( bSkipRes ) 4993 { 4994 #if SKIP_FLAG 4995 pcCU->setSkipFlagSubParts( true, 0, pcCU->getDepth(0) ); 4996 #endif 4997 4998 rpcYuvResi->clear(); 4999 5000 pcYuvPred->copyToPartYuv( rpcYuvRec, 0 ); 5001 5002 #if WEIGHTED_CHROMA_DISTORTION 5003 uiDistortion = m_pcRdCost->getDistPart( rpcYuvRec->getLumaAddr(), rpcYuvRec->getStride(), pcYuvOrg->getLumaAddr(), pcYuvOrg->getStride(), uiWidth, uiHeight ) 5004 + m_pcRdCost->getDistPart( rpcYuvRec->getCbAddr(), rpcYuvRec->getCStride(), pcYuvOrg->getCbAddr(), pcYuvOrg->getCStride(), uiWidth >> 1, uiHeight >> 1, true ) 5005 + m_pcRdCost->getDistPart( rpcYuvRec->getCrAddr(), rpcYuvRec->getCStride(), pcYuvOrg->getCrAddr(), pcYuvOrg->getCStride(), uiWidth >> 1, uiHeight >> 1, true ); 5006 #else 5007 uiDistortion = m_pcRdCost->getDistPart( rpcYuvRec->getLumaAddr(), rpcYuvRec->getStride(), pcYuvOrg->getLumaAddr(), pcYuvOrg->getStride(), uiWidth, uiHeight ) 5008 + m_pcRdCost->getDistPart( rpcYuvRec->getCbAddr(), rpcYuvRec->getCStride(), pcYuvOrg->getCbAddr(), pcYuvOrg->getCStride(), uiWidth >> 1, uiHeight >> 1 ) 5009 + m_pcRdCost->getDistPart( rpcYuvRec->getCrAddr(), rpcYuvRec->getCStride(), pcYuvOrg->getCrAddr(), pcYuvOrg->getCStride(), uiWidth >> 1, uiHeight >> 1 ); 5010 #endif 5011 5012 if( m_bUseSBACRD ) 5013 m_pcRDGoOnSbacCoder->load(m_pppcRDSbacCoder[pcCU->getDepth(0)][CI_CURR_BEST]); 5014 5015 m_pcEntropyCoder->resetBits(); 5016 if (pcCU->getSlice()->getPPS()->getTransquantBypassEnableFlag()) 5017 { 5018 m_pcEntropyCoder->encodeCUTransquantBypassFlag(pcCU, 0, true); 5019 } 5020 m_pcEntropyCoder->encodeSkipFlag(pcCU, 0, true); 5021 m_pcEntropyCoder->encodeMergeIndex( pcCU, 0, 0, true ); 5022 5023 uiBits = m_pcEntropyCoder->getNumberOfWrittenBits(); 5024 pcCU->getTotalBits() = uiBits; 5025 pcCU->getTotalDistortion() = uiDistortion; 5026 pcCU->getTotalCost() = m_pcRdCost->calcRdCost( uiBits, uiDistortion ); 5027 5028 if( m_bUseSBACRD ) 5029 m_pcRDGoOnSbacCoder->store(m_pppcRDSbacCoder[pcCU->getDepth(0)][CI_TEMP_BEST]); 5030 5031 pcCU->setCbfSubParts( 0, 0, 0, 0, pcCU->getDepth( 0 ) ); 5032 pcCU->setTrIdxSubParts( 0, 0, pcCU->getDepth(0) ); 5033 5034 return; 5035 } 5036 5037 // Residual coding. 5038 Int qp, qpBest = 0, qpMin, qpMax; 5039 Double dCost, dCostBest = MAX_DOUBLE; 5040 5041 UInt uiTrLevel = 0; 5042 if( (pcCU->getWidth(0) > pcCU->getSlice()->getSPS()->getMaxTrSize()) ) 5043 { 5044 while( pcCU->getWidth(0) > (pcCU->getSlice()->getSPS()->getMaxTrSize()<<uiTrLevel) ) uiTrLevel++; 5045 } 5046 UInt uiMaxTrMode = pcCU->getSlice()->getSPS()->getMaxTrDepth() + uiTrLevel; 5047 5048 while((uiWidth>>uiMaxTrMode) < (g_uiMaxCUWidth>>g_uiMaxCUDepth)) uiMaxTrMode--; 5049 5050 qpMin = bHighPass ? Clip3( -pcCU->getSlice()->getSPS()->getQpBDOffsetY(), MAX_QP, pcCU->getQP(0) - m_iMaxDeltaQP ) : pcCU->getQP( 0 ); 5051 qpMax = bHighPass ? Clip3( -pcCU->getSlice()->getSPS()->getQpBDOffsetY(), MAX_QP, pcCU->getQP(0) + m_iMaxDeltaQP ) : pcCU->getQP( 0 ); 5052 5053 rpcYuvResi->subtract( pcYuvOrg, pcYuvPred, 0, uiWidth ); 5054 5055 for ( qp = qpMin; qp <= qpMax; qp++ ) 5056 { 5057 dCost = 0.; 5058 uiBits = 0; 5059 uiDistortion = 0; 5060 if( m_bUseSBACRD ) 5061 { 5062 m_pcRDGoOnSbacCoder->load( m_pppcRDSbacCoder[ pcCU->getDepth( 0 ) ][ CI_CURR_BEST ] ); 5063 } 5064 5065 UInt uiZeroDistortion = 0; 5066 #if IBDI_DISTORTION 5067 xEstimateResidualQT( pcCU, 0, 0, 0, pcYuvOrg, pcYuvPred, rpcYuvResi, pcCU->getDepth(0), dCost, uiBits, uiDistortion, &uiZeroDistortion ); 5068 #else 5069 xEstimateResidualQT( pcCU, 0, 0, 0, rpcYuvResi, pcCU->getDepth(0), dCost, uiBits, uiDistortion, &uiZeroDistortion ); 5070 #endif 5071 5072 #if TU_ZERO_CBF_RDO 5073 m_pcEntropyCoder->resetBits(); 5074 m_pcEntropyCoder->encodeQtRootCbfZero( pcCU, 0 ); 5075 UInt zeroResiBits = m_pcEntropyCoder->getNumberOfWrittenBits(); 5076 double dZeroCost = m_pcRdCost->calcRdCost( zeroResiBits, uiZeroDistortion ); 5077 #else 5078 double dZeroCost = m_pcRdCost->calcRdCost( 0, uiZeroDistortion ); 5079 #endif 5080 if(pcCU->isLosslessCoded( 0 )) 5081 { 5082 dZeroCost = dCost + 1; 5083 } 5084 if ( dZeroCost < dCost ) 5085 { 5086 dCost = dZeroCost; 5087 uiBits = 0; 5088 uiDistortion = uiZeroDistortion; 5089 5090 const UInt uiQPartNum = pcCU->getPic()->getNumPartInCU() >> (pcCU->getDepth(0) << 1); 5091 ::memset( pcCU->getTransformIdx() , 0, uiQPartNum * sizeof(UChar) ); 5092 ::memset( pcCU->getCbf( TEXT_LUMA ) , 0, uiQPartNum * sizeof(UChar) ); 5093 ::memset( pcCU->getCbf( TEXT_CHROMA_U ), 0, uiQPartNum * sizeof(UChar) ); 5094 ::memset( pcCU->getCbf( TEXT_CHROMA_V ), 0, uiQPartNum * sizeof(UChar) ); 5095 ::memset( pcCU->getCoeffY() , 0, uiWidth * uiHeight * sizeof( TCoeff ) ); 5096 ::memset( pcCU->getCoeffCb() , 0, uiWidth * uiHeight * sizeof( TCoeff ) >> 2 ); 5097 ::memset( pcCU->getCoeffCr() , 0, uiWidth * uiHeight * sizeof( TCoeff ) >> 2 ); 5098 #if INTER_TRANSFORMSKIP 5099 pcCU->setTransformSkipSubParts ( 0, 0, 0, 0, pcCU->getDepth(0) ); 5100 #endif 5101 } 5102 #if NO_RESIDUAL_FLAG_FOR_BLPRED 5103 else if(pcCU->getLayerId() > 0 && pcCU->isIntraBL(0) && uiZeroDistortion == uiDistortion) // all zeros 5104 { 5105 const UInt uiQPartNum = pcCU->getPic()->getNumPartInCU() >> (pcCU->getDepth(0) << 1); 5106 ::memset( pcCU->getTransformIdx() , 0, uiQPartNum * sizeof(UChar) ); 5107 ::memset( pcCU->getCbf( TEXT_LUMA ) , 0, uiQPartNum * sizeof(UChar) ); 5108 ::memset( pcCU->getCbf( TEXT_CHROMA_U ), 0, uiQPartNum * sizeof(UChar) ); 5109 ::memset( pcCU->getCbf( TEXT_CHROMA_V ), 0, uiQPartNum * sizeof(UChar) ); 5110 ::memset( pcCU->getCoeffY() , 0, uiWidth * uiHeight * sizeof( TCoeff ) ); 5111 ::memset( pcCU->getCoeffCb() , 0, uiWidth * uiHeight * sizeof( TCoeff ) >> 2 ); 5112 ::memset( pcCU->getCoeffCr() , 0, uiWidth * uiHeight * sizeof( TCoeff ) >> 2 ); 5113 #if INTER_TRANSFORMSKIP 5114 pcCU->setTransformSkipSubParts ( 0, 0, 0, 0, pcCU->getDepth(0) ); 5115 #endif 5116 } 5117 #endif 5118 else 5119 { 5120 xSetResidualQTData( pcCU, 0, 0, 0, NULL, pcCU->getDepth(0), false ); 5121 } 5122 5123 if( m_bUseSBACRD ) 5124 { 5125 m_pcRDGoOnSbacCoder->load( m_pppcRDSbacCoder[pcCU->getDepth(0)][CI_CURR_BEST] ); 5126 } 5127 #if 0 // check 5128 { 5129 m_pcEntropyCoder->resetBits(); 5130 m_pcEntropyCoder->encodeCoeff( pcCU, 0, pcCU->getDepth(0), pcCU->getWidth(0), pcCU->getHeight(0) ); 5131 const UInt uiBitsForCoeff = m_pcEntropyCoder->getNumberOfWrittenBits(); 5132 if( m_bUseSBACRD ) 5133 { 5134 m_pcRDGoOnSbacCoder->load( m_pppcRDSbacCoder[pcCU->getDepth(0)][CI_CURR_BEST] ); 5135 } 5136 if( uiBitsForCoeff != uiBits ) 5137 assert( 0 ); 5138 } 5139 #endif 5140 uiBits = 0; 5141 { 5142 TComYuv *pDummy = NULL; 5143 xAddSymbolBitsInter( pcCU, 0, 0, uiBits, pDummy, NULL, pDummy ); 5144 } 5145 5146 5147 Double dExactCost = m_pcRdCost->calcRdCost( uiBits, uiDistortion ); 5148 dCost = dExactCost; 5149 5150 if ( dCost < dCostBest ) 5151 { 5152 if ( !pcCU->getQtRootCbf( 0 ) ) 5153 { 5154 rpcYuvResiBest->clear(); 5155 } 5156 else 5157 { 5158 xSetResidualQTData( pcCU, 0, 0, 0, rpcYuvResiBest, pcCU->getDepth(0), true ); 5159 } 5160 5161 if( qpMin != qpMax && qp != qpMax ) 5162 { 5163 const UInt uiQPartNum = pcCU->getPic()->getNumPartInCU() >> (pcCU->getDepth(0) << 1); 5164 ::memcpy( m_puhQTTempTrIdx, pcCU->getTransformIdx(), uiQPartNum * sizeof(UChar) ); 5165 ::memcpy( m_puhQTTempCbf[0], pcCU->getCbf( TEXT_LUMA ), uiQPartNum * sizeof(UChar) ); 5166 ::memcpy( m_puhQTTempCbf[1], pcCU->getCbf( TEXT_CHROMA_U ), uiQPartNum * sizeof(UChar) ); 5167 ::memcpy( m_puhQTTempCbf[2], pcCU->getCbf( TEXT_CHROMA_V ), uiQPartNum * sizeof(UChar) ); 5168 ::memcpy( m_pcQTTempCoeffY, pcCU->getCoeffY(), uiWidth * uiHeight * sizeof( TCoeff ) ); 5169 ::memcpy( m_pcQTTempCoeffCb, pcCU->getCoeffCb(), uiWidth * uiHeight * sizeof( TCoeff ) >> 2 ); 5170 ::memcpy( m_pcQTTempCoeffCr, pcCU->getCoeffCr(), uiWidth * uiHeight * sizeof( TCoeff ) >> 2 ); 5171 #if ADAPTIVE_QP_SELECTION 5172 ::memcpy( m_pcQTTempArlCoeffY, pcCU->getArlCoeffY(), uiWidth * uiHeight * sizeof( Int ) ); 5173 ::memcpy( m_pcQTTempArlCoeffCb, pcCU->getArlCoeffCb(), uiWidth * uiHeight * sizeof( Int ) >> 2 ); 5174 ::memcpy( m_pcQTTempArlCoeffCr, pcCU->getArlCoeffCr(), uiWidth * uiHeight * sizeof( Int ) >> 2 ); 5175 #endif 5176 #if INTER_TRANSFORMSKIP 5177 ::memcpy( m_puhQTTempTransformSkipFlag[0], pcCU->getTransformSkip(TEXT_LUMA), uiQPartNum * sizeof( UChar ) ); 5178 ::memcpy( m_puhQTTempTransformSkipFlag[1], pcCU->getTransformSkip(TEXT_CHROMA_U), uiQPartNum * sizeof( UChar ) ); 5179 ::memcpy( m_puhQTTempTransformSkipFlag[2], pcCU->getTransformSkip(TEXT_CHROMA_V), uiQPartNum * sizeof( UChar ) ); 5180 #endif 5181 } 5182 uiBitsBest = uiBits; 5183 uiDistortionBest = uiDistortion; 5184 dCostBest = dCost; 5185 qpBest = qp; 5186 if( m_bUseSBACRD ) 5187 { 5188 m_pcRDGoOnSbacCoder->store( m_pppcRDSbacCoder[ pcCU->getDepth( 0 ) ][ CI_TEMP_BEST ] ); 5189 } 5190 } 5191 } 5192 5193 assert ( dCostBest != MAX_DOUBLE ); 5194 5195 if( qpMin != qpMax && qpBest != qpMax ) 5196 { 5197 if( m_bUseSBACRD ) 5198 { 5199 assert( 0 ); // check 5200 m_pcRDGoOnSbacCoder->load( m_pppcRDSbacCoder[ pcCU->getDepth( 0 ) ][ CI_TEMP_BEST ] ); 5201 } 5202 // copy best cbf and trIdx to pcCU 5203 const UInt uiQPartNum = pcCU->getPic()->getNumPartInCU() >> (pcCU->getDepth(0) << 1); 5204 ::memcpy( pcCU->getTransformIdx(), m_puhQTTempTrIdx, uiQPartNum * sizeof(UChar) ); 5205 ::memcpy( pcCU->getCbf( TEXT_LUMA ), m_puhQTTempCbf[0], uiQPartNum * sizeof(UChar) ); 5206 ::memcpy( pcCU->getCbf( TEXT_CHROMA_U ), m_puhQTTempCbf[1], uiQPartNum * sizeof(UChar) ); 5207 ::memcpy( pcCU->getCbf( TEXT_CHROMA_V ), m_puhQTTempCbf[2], uiQPartNum * sizeof(UChar) ); 5208 ::memcpy( pcCU->getCoeffY(), m_pcQTTempCoeffY, uiWidth * uiHeight * sizeof( TCoeff ) ); 5209 ::memcpy( pcCU->getCoeffCb(), m_pcQTTempCoeffCb, uiWidth * uiHeight * sizeof( TCoeff ) >> 2 ); 5210 ::memcpy( pcCU->getCoeffCr(), m_pcQTTempCoeffCr, uiWidth * uiHeight * sizeof( TCoeff ) >> 2 ); 5211 #if ADAPTIVE_QP_SELECTION 5212 ::memcpy( pcCU->getArlCoeffY(), m_pcQTTempArlCoeffY, uiWidth * uiHeight * sizeof( Int ) ); 5213 ::memcpy( pcCU->getArlCoeffCb(), m_pcQTTempArlCoeffCb, uiWidth * uiHeight * sizeof( Int ) >> 2 ); 5214 ::memcpy( pcCU->getArlCoeffCr(), m_pcQTTempArlCoeffCr, uiWidth * uiHeight * sizeof( Int ) >> 2 ); 5215 #endif 5216 #if INTER_TRANSFORMSKIP 5217 ::memcpy( pcCU->getTransformSkip(TEXT_LUMA), m_puhQTTempTransformSkipFlag[0], uiQPartNum * sizeof( UChar ) ); 5218 ::memcpy( pcCU->getTransformSkip(TEXT_CHROMA_U), m_puhQTTempTransformSkipFlag[1], uiQPartNum * sizeof( UChar ) ); 5219 ::memcpy( pcCU->getTransformSkip(TEXT_CHROMA_V), m_puhQTTempTransformSkipFlag[2], uiQPartNum * sizeof( UChar ) ); 5220 #endif 5221 } 5222 rpcYuvRec->addClip ( pcYuvPred, rpcYuvResiBest, 0, uiWidth ); 5223 5224 // update with clipped distortion and cost (qp estimation loop uses unclipped values) 5225 #if WEIGHTED_CHROMA_DISTORTION 5226 uiDistortionBest = m_pcRdCost->getDistPart( rpcYuvRec->getLumaAddr(), rpcYuvRec->getStride(), pcYuvOrg->getLumaAddr(), pcYuvOrg->getStride(), uiWidth, uiHeight ) 5227 + m_pcRdCost->getDistPart( rpcYuvRec->getCbAddr(), rpcYuvRec->getCStride(), pcYuvOrg->getCbAddr(), pcYuvOrg->getCStride(), uiWidth >> 1, uiHeight >> 1, true ) 5228 + m_pcRdCost->getDistPart( rpcYuvRec->getCrAddr(), rpcYuvRec->getCStride(), pcYuvOrg->getCrAddr(), pcYuvOrg->getCStride(), uiWidth >> 1, uiHeight >> 1, true ); 5229 #else 5230 uiDistortionBest = m_pcRdCost->getDistPart( rpcYuvRec->getLumaAddr(), rpcYuvRec->getStride(), pcYuvOrg->getLumaAddr(), pcYuvOrg->getStride(), uiWidth, uiHeight ) 5231 + m_pcRdCost->getDistPart( rpcYuvRec->getCbAddr(), rpcYuvRec->getCStride(), pcYuvOrg->getCbAddr(), pcYuvOrg->getCStride(), uiWidth >> 1, uiHeight >> 1 ) 5232 + m_pcRdCost->getDistPart( rpcYuvRec->getCrAddr(), rpcYuvRec->getCStride(), pcYuvOrg->getCrAddr(), pcYuvOrg->getCStride(), uiWidth >> 1, uiHeight >> 1 ); 5233 #endif 5234 dCostBest = m_pcRdCost->calcRdCost( uiBitsBest, uiDistortionBest ); 5235 5236 pcCU->getTotalBits() = uiBitsBest; 5237 pcCU->getTotalDistortion() = uiDistortionBest; 5238 pcCU->getTotalCost() = dCostBest; 5239 5240 if ( pcCU->isSkipped(0) ) 5241 { 5242 pcCU->setCbfSubParts( 0, 0, 0, 0, pcCU->getDepth( 0 ) ); 5243 } 5244 5245 pcCU->setQPSubParts( qpBest, 0, pcCU->getDepth(0) ); 6182 rpcPredYuv->clear(); 6183 rpcRecoYuv->clear(); 6184 6185 Int iNumPredDir = pcCU->getSlice()->isInterP() ? 1 : 2; 6186 6187 TComMv cMv[2]; 6188 TComMv cMvPred[2][33]; 6189 TComMv cMvTemp[2][33]; 6190 TComMv TempMv; 6191 6192 Int iRefIdx[2]={0,0}; 6193 6194 Int aaiMvpIdx[2][33]; 6195 Int aaiMvpNum[2][33]; 6196 6197 UInt uiMbBits[3] = {1, 1, 0}; 6198 UInt uiLastMode = 0; 6199 6200 UInt uiCost[2] = { MAX_UINT, MAX_UINT }; //uni, rdCost 6201 UInt uiCostTemp; 6202 UInt biPDistTemp = MAX_INT; 6203 UInt uiBitsTemp; 6204 6205 PartSize ePartSize = pcCU->getPartitionSize( 0 ); //2Nx2N 6206 Int iPartIdx = 0; //one PU in CU 6207 UInt uiPartAddr; 6208 Int iRoiWidth, iRoiHeight; 6209 6210 xGetBlkBits( ePartSize, pcCU->getSlice()->isInterP(), iPartIdx, uiLastMode, uiMbBits); 6211 pcCU->getPartIndexAndSize( iPartIdx, uiPartAddr, iRoiWidth, iRoiHeight ); 6212 6213 for( Int iRefList = 0; iRefList < iNumPredDir; iRefList++) //list 6214 { 6215 RefPicList eRefPicList = ( iRefList ? REF_PIC_LIST_1 : REF_PIC_LIST_0 ); 6216 6217 Int iRefIdxTemp = -1; 6218 Bool foundILR = false; 6219 for( Int refIdx = 0; refIdx < pcCU->getSlice()->getNumRefIdx(eRefPicList); refIdx++ ) 6220 { 6221 if( pcCU->getSlice()->getRefPic(eRefPicList, refIdx)->isILR() ) 6222 { 6223 iRefIdxTemp = refIdx; 6224 foundILR = true; 6225 break; 6226 } 6227 } 6228 6229 if(!foundILR) //no ILR in eRefPiclist 6230 { 6231 continue; 6232 } 6233 6234 uiBitsTemp = uiMbBits[iRefList]; 6235 if ( pcCU->getSlice()->getNumRefIdx(eRefPicList) > 1 ) 6236 { 6237 uiBitsTemp += iRefIdxTemp+1; 6238 if ( iRefIdxTemp == pcCU->getSlice()->getNumRefIdx(eRefPicList)-1 ) uiBitsTemp--; 6239 } 6240 6241 xEstimateMvPredAMVP( pcCU, pcOrgYuv, iPartIdx, eRefPicList, iRefIdxTemp, cMvPred[iRefList][iRefIdxTemp], false, &biPDistTemp); 6242 aaiMvpIdx[iRefList][iRefIdxTemp] = pcCU->getMVPIdx(eRefPicList, uiPartAddr); 6243 aaiMvpNum[iRefList][iRefIdxTemp] = pcCU->getMVPNum(eRefPicList, uiPartAddr); 6244 6245 uiBitsTemp += m_auiMVPIdxCost[aaiMvpIdx[iRefList][iRefIdxTemp]][AMVP_MAX_NUM_CANDS]; 6246 6247 xMotionEstimation ( pcCU, pcOrgYuv, iPartIdx, eRefPicList, &cMvPred[iRefList][iRefIdxTemp], iRefIdxTemp, cMvTemp[iRefList][iRefIdxTemp], uiBitsTemp, uiCostTemp ); 6248 xCheckBestMVP(pcCU, eRefPicList, cMvTemp[iRefList][iRefIdxTemp], cMvPred[iRefList][iRefIdxTemp], aaiMvpIdx[iRefList][iRefIdxTemp], uiBitsTemp, uiCostTemp); 6249 6250 if( uiCostTemp < uiCost[iRefList] ) 6251 { 6252 uiCost[iRefList] = uiCostTemp; 6253 6254 cMv[iRefList] = cMvTemp[iRefList][iRefIdxTemp]; 6255 iRefIdx[iRefList] = iRefIdxTemp; 6256 6257 pcCU->getCUMvField(eRefPicList)->setAllMv( cMv[iRefList], ePartSize, uiPartAddr, 0, iPartIdx ); 6258 pcCU->getCUMvField(eRefPicList)->setAllRefIdx( iRefIdx[iRefList], ePartSize, uiPartAddr, 0, iPartIdx ); 6259 } 6260 } 6261 6262 if( uiCost[0] == MAX_UINT && uiCost[1] == MAX_UINT ) //no ILR in both list0 and list1 6263 { 6264 return false; 6265 } 6266 6267 // Clear Motion Field 6268 pcCU->getCUMvField(REF_PIC_LIST_0)->setAllMvField( TComMvField(), ePartSize, uiPartAddr, 0, iPartIdx ); 6269 pcCU->getCUMvField(REF_PIC_LIST_1)->setAllMvField( TComMvField(), ePartSize, uiPartAddr, 0, iPartIdx ); 6270 pcCU->getCUMvField(REF_PIC_LIST_0)->setAllMvd ( TComMv(), ePartSize, uiPartAddr, 0, iPartIdx ); 6271 pcCU->getCUMvField(REF_PIC_LIST_1)->setAllMvd ( TComMv(), ePartSize, uiPartAddr, 0, iPartIdx ); 6272 6273 pcCU->setMVPIdxSubParts( -1, REF_PIC_LIST_0, uiPartAddr, iPartIdx, pcCU->getDepth(uiPartAddr)); 6274 pcCU->setMVPNumSubParts( -1, REF_PIC_LIST_0, uiPartAddr, iPartIdx, pcCU->getDepth(uiPartAddr)); 6275 pcCU->setMVPIdxSubParts( -1, REF_PIC_LIST_1, uiPartAddr, iPartIdx, pcCU->getDepth(uiPartAddr)); 6276 pcCU->setMVPNumSubParts( -1, REF_PIC_LIST_1, uiPartAddr, iPartIdx, pcCU->getDepth(uiPartAddr)); 6277 6278 if( uiCost[0] <= uiCost[1] ) //list0 ILR 6279 { 6280 pcCU->getCUMvField(REF_PIC_LIST_0)->setAllMv ( cMv[0], ePartSize, uiPartAddr, 0, iPartIdx ); 6281 pcCU->getCUMvField(REF_PIC_LIST_0)->setAllRefIdx( iRefIdx[0], ePartSize, uiPartAddr, 0, iPartIdx ); 6282 6283 TempMv = cMv[0] - cMvPred[0][iRefIdx[0]]; 6284 pcCU->getCUMvField(REF_PIC_LIST_0)->setAllMvd( TempMv, ePartSize, uiPartAddr, 0, iPartIdx ); 6285 6286 pcCU->setInterDirSubParts( 1, uiPartAddr, iPartIdx, pcCU->getDepth(0) ); 6287 6288 pcCU->setMVPIdxSubParts( aaiMvpIdx[0][iRefIdx[0]], REF_PIC_LIST_0, uiPartAddr, iPartIdx, pcCU->getDepth(uiPartAddr)); 6289 pcCU->setMVPNumSubParts( aaiMvpNum[0][iRefIdx[0]], REF_PIC_LIST_0, uiPartAddr, iPartIdx, pcCU->getDepth(uiPartAddr)); 6290 } 6291 else if( uiCost[1] < uiCost[0] ) //list1 ILR 6292 { 6293 pcCU->getCUMvField(REF_PIC_LIST_1)->setAllMv ( cMv[1], ePartSize, uiPartAddr, 0, iPartIdx ); 6294 pcCU->getCUMvField(REF_PIC_LIST_1)->setAllRefIdx( iRefIdx[1], ePartSize, uiPartAddr, 0, iPartIdx ); 6295 6296 TempMv = cMv[1] - cMvPred[1][iRefIdx[1]]; 6297 pcCU->getCUMvField(REF_PIC_LIST_1)->setAllMvd( TempMv, ePartSize, uiPartAddr, 0, iPartIdx ); 6298 6299 pcCU->setInterDirSubParts( 2, uiPartAddr, iPartIdx, pcCU->getDepth(0) ); 6300 6301 pcCU->setMVPIdxSubParts( aaiMvpIdx[1][iRefIdx[1]], REF_PIC_LIST_1, uiPartAddr, iPartIdx, pcCU->getDepth(uiPartAddr)); 6302 pcCU->setMVPNumSubParts( aaiMvpNum[1][iRefIdx[1]], REF_PIC_LIST_1, uiPartAddr, iPartIdx, pcCU->getDepth(uiPartAddr)); 6303 } 6304 else 6305 { 6306 assert(0); 6307 } 6308 6309 pcCU->setMergeFlagSubParts( false, uiPartAddr, iPartIdx, pcCU->getDepth( uiPartAddr ) ); 6310 6311 motionCompensation ( pcCU, rpcPredYuv, REF_PIC_LIST_X, iPartIdx ); 6312 6313 setWpScalingDistParam( pcCU, -1, REF_PIC_LIST_X ); 6314 6315 return true; 5246 6316 } 5247 6317 5248 #if IBDI_DISTORTION 5249 Void TEncSearch::xEstimateResidualQT( TComDataCU* pcCU, UInt uiQuadrant, UInt uiAbsPartIdx, UInt absTUPartIdx, TComYuv* pcOrg, TComYuv* pcPred, TComYuv* pcResi, const UInt uiDepth, Double &rdCost, UInt &ruiBits, UInt &ruiDist, UInt *puiZeroDist ) 5250 #else 5251 Void TEncSearch::xEstimateResidualQT( TComDataCU* pcCU, UInt uiQuadrant, UInt uiAbsPartIdx, UInt absTUPartIdx, TComYuv* pcResi, const UInt uiDepth, Double &rdCost, UInt &ruiBits, UInt &ruiDist, UInt *puiZeroDist ) 5252 #endif 6318 6319 #endif 6320 6321 #if INTRA_BL 6322 Void 6323 TEncSearch::estIntraBLPredQT( TComDataCU* pcCU, 6324 TComYuv* pcOrgYuv, 6325 TComYuv* pcPredYuv, 6326 TComYuv* pcResiYuv, 6327 TComYuv* pcRecoYuv ) 5253 6328 { 5254 const UInt uiTrMode = uiDepth - pcCU->getDepth( 0);5255 5256 assert( pcCU->getDepth( 0 ) == pcCU->getDepth( uiAbsPartIdx ) );5257 const UInt uiLog2TrSize = g_aucConvertToBit[pcCU->getSlice()->getSPS()->getMaxCUWidth() >> uiDepth]+2;5258 5259 UInt SplitFlag = ((pcCU->getSlice()->getSPS()->getQuadtreeTUMaxDepthInter() == 1) && pcCU->getPredictionMode(uiAbsPartIdx) == MODE_INTER && ( pcCU->getPartitionSize(uiAbsPartIdx) != SIZE_2Nx2N ));5260 Bool bCheckFull;5261 if ( SplitFlag && uiDepth == pcCU->getDepth(uiAbsPartIdx) && ( uiLog2TrSize > pcCU->getQuadtreeTULog2MinSizeInCU(uiAbsPartIdx) ) )5262 bCheckFull = false;6329 UInt uiDepth = pcCU->getDepth(0); 6330 UInt uiOverallDistY = 0; 6331 UInt uiOverallDistC = 0; 6332 6333 //===== set QP and clear Cbf ===== 6334 if ( pcCU->getSlice()->getPPS()->getUseDQP() == true) 6335 { 6336 pcCU->setQPSubParts( pcCU->getQP(0), 0, uiDepth ); 6337 } 5263 6338 else 5264 bCheckFull = ( uiLog2TrSize <= pcCU->getSlice()->getSPS()->getQuadtreeTULog2MaxSize() ); 5265 5266 const Bool bCheckSplit = ( uiLog2TrSize > pcCU->getQuadtreeTULog2MinSizeInCU(uiAbsPartIdx) ); 5267 5268 assert( bCheckFull || bCheckSplit ); 5269 5270 Bool bCodeChroma = true; 5271 UInt uiTrModeC = uiTrMode; 5272 UInt uiLog2TrSizeC = uiLog2TrSize-1; 5273 if( uiLog2TrSize == 2 ) 5274 { 5275 uiLog2TrSizeC++; 5276 uiTrModeC --; 5277 UInt uiQPDiv = pcCU->getPic()->getNumPartInCU() >> ( ( pcCU->getDepth( 0 ) + uiTrModeC ) << 1 ); 5278 bCodeChroma = ( ( uiAbsPartIdx % uiQPDiv ) == 0 ); 5279 } 5280 5281 const UInt uiSetCbf = 1 << uiTrMode; 5282 // code full block 5283 Double dSingleCost = MAX_DOUBLE; 5284 UInt uiSingleBits = 0; 5285 UInt uiSingleDist = 0; 5286 UInt uiAbsSumY = 0, uiAbsSumU = 0, uiAbsSumV = 0; 5287 #if INTER_TRANSFORMSKIP 5288 UInt uiBestTransformMode[3] = {0}; 5289 #endif 5290 6339 { 6340 pcCU->setQPSubParts( pcCU->getSlice()->getSliceQp(), 0, uiDepth ); 6341 } 6342 6343 //===== init pattern for luma prediction ===== 6344 Bool bAboveAvail = false; 6345 Bool bLeftAvail = false; 6346 pcCU->getPattern()->initPattern ( pcCU, 0, 0 ); 6347 pcCU->getPattern()->initAdiPattern( pcCU, 0, 0, m_piYuvExt, m_iYuvExtStride, m_iYuvExtHeight, bAboveAvail, bLeftAvail ); 6348 6349 pcCU->setLumaIntraDirSubParts ( DC_IDX, 0, uiDepth ); 6350 6351 // set context models 5291 6352 if( m_bUseSBACRD ) 5292 6353 { 5293 m_pcRDGoOnSbacCoder->store( m_pppcRDSbacCoder[ uiDepth ][ CI_QT_TRAFO_ROOT ] ); 5294 } 5295 5296 if( bCheckFull ) 5297 { 5298 const UInt uiNumCoeffPerAbsPartIdxIncrement = pcCU->getSlice()->getSPS()->getMaxCUWidth() * pcCU->getSlice()->getSPS()->getMaxCUHeight() >> ( pcCU->getSlice()->getSPS()->getMaxCUDepth() << 1 ); 5299 const UInt uiQTTempAccessLayer = pcCU->getSlice()->getSPS()->getQuadtreeTULog2MaxSize() - uiLog2TrSize; 5300 TCoeff *pcCoeffCurrY = m_ppcQTTempCoeffY [uiQTTempAccessLayer] + uiNumCoeffPerAbsPartIdxIncrement * uiAbsPartIdx; 5301 TCoeff *pcCoeffCurrU = m_ppcQTTempCoeffCb[uiQTTempAccessLayer] + (uiNumCoeffPerAbsPartIdxIncrement * uiAbsPartIdx>>2); 5302 TCoeff *pcCoeffCurrV = m_ppcQTTempCoeffCr[uiQTTempAccessLayer] + (uiNumCoeffPerAbsPartIdxIncrement * uiAbsPartIdx>>2); 5303 #if ADAPTIVE_QP_SELECTION 5304 Int *pcArlCoeffCurrY = m_ppcQTTempArlCoeffY [uiQTTempAccessLayer] + uiNumCoeffPerAbsPartIdxIncrement * uiAbsPartIdx; 5305 Int *pcArlCoeffCurrU = m_ppcQTTempArlCoeffCb[uiQTTempAccessLayer] + (uiNumCoeffPerAbsPartIdxIncrement * uiAbsPartIdx>>2); 5306 Int *pcArlCoeffCurrV = m_ppcQTTempArlCoeffCr[uiQTTempAccessLayer] + (uiNumCoeffPerAbsPartIdxIncrement * uiAbsPartIdx>>2); 5307 #endif 5308 5309 Int trWidth = 0, trHeight = 0, trWidthC = 0, trHeightC = 0; 5310 UInt absTUPartIdxC = uiAbsPartIdx; 5311 5312 trWidth = trHeight = 1 << uiLog2TrSize; 5313 trWidthC = trHeightC = 1 <<uiLog2TrSizeC; 5314 #if !REMOVE_NSQT 5315 pcCU->getNSQTSize ( uiTrMode, uiAbsPartIdx, trWidth, trHeight ); 5316 pcCU->getNSQTSize ( uiTrModeC, uiAbsPartIdx, trWidthC, trHeightC ); 5317 5318 if( bCodeChroma && pcCU->useNonSquareTrans( uiTrMode, uiAbsPartIdx ) && !( uiLog2TrSizeC == pcCU->getSlice()->getSPS()->getQuadtreeTULog2MinSize() && uiTrModeC == 1 ) ) 5319 { 5320 absTUPartIdxC = pcCU->getNSAddrChroma( uiLog2TrSizeC, uiTrModeC, uiQuadrant, absTUPartIdx ); 5321 } 5322 #endif 5323 pcCU->setTrIdxSubParts( uiDepth - pcCU->getDepth( 0 ), uiAbsPartIdx, uiDepth ); 5324 #if INTER_TRANSFORMSKIP 5325 Double minCostY = MAX_DOUBLE; 5326 Double minCostU = MAX_DOUBLE; 5327 Double minCostV = MAX_DOUBLE; 5328 #if PPS_TS_FLAG 5329 Bool checkTransformSkipY = pcCU->getSlice()->getPPS()->getUseTransformSkip() && trWidth == 4 && trHeight == 4; 5330 Bool checkTransformSkipUV = pcCU->getSlice()->getPPS()->getUseTransformSkip() && trWidthC == 4 && trHeightC == 4; 5331 #else 5332 Bool checkTransformSkipY = pcCU->getSlice()->getSPS()->getUseTransformSkip() && trWidth == 4 && trHeight == 4; 5333 Bool checkTransformSkipUV = pcCU->getSlice()->getSPS()->getUseTransformSkip() && trWidthC == 4 && trHeightC == 4; 5334 #endif 5335 5336 checkTransformSkipY &= (!pcCU->isLosslessCoded(0)); 5337 checkTransformSkipUV &= (!pcCU->isLosslessCoded(0)); 5338 5339 pcCU->setTransformSkipSubParts ( 0, TEXT_LUMA, uiAbsPartIdx, uiDepth ); 5340 if( bCodeChroma ) 5341 { 5342 pcCU->setTransformSkipSubParts ( 0, TEXT_CHROMA_U, uiAbsPartIdx, pcCU->getDepth(0)+uiTrModeC ); 5343 pcCU->setTransformSkipSubParts ( 0, TEXT_CHROMA_V, uiAbsPartIdx, pcCU->getDepth(0)+uiTrModeC ); 5344 } 5345 #endif 5346 5347 if (m_pcEncCfg->getUseRDOQ()) 5348 { 5349 m_pcEntropyCoder->estimateBit(m_pcTrQuant->m_pcEstBitsSbac, trWidth, trHeight, TEXT_LUMA ); 5350 } 5351 5352 m_pcTrQuant->setQPforQuant( pcCU->getQP( 0 ), TEXT_LUMA, pcCU->getSlice()->getSPS()->getQpBDOffsetY(), 0 ); 5353 5354 #if RDOQ_CHROMA_LAMBDA 5355 m_pcTrQuant->selectLambda(TEXT_LUMA); 5356 #endif 5357 m_pcTrQuant->transformNxN( pcCU, pcResi->getLumaAddr( absTUPartIdx ), pcResi->getStride (), pcCoeffCurrY, 5358 #if ADAPTIVE_QP_SELECTION 5359 pcArlCoeffCurrY, 5360 #endif 5361 trWidth, trHeight, uiAbsSumY, TEXT_LUMA, uiAbsPartIdx ); 5362 5363 pcCU->setCbfSubParts( uiAbsSumY ? uiSetCbf : 0, TEXT_LUMA, uiAbsPartIdx, uiDepth ); 5364 5365 if( bCodeChroma ) 5366 { 5367 if (m_pcEncCfg->getUseRDOQ()) 5368 { 5369 m_pcEntropyCoder->estimateBit(m_pcTrQuant->m_pcEstBitsSbac, trWidthC, trHeightC, TEXT_CHROMA ); 5370 } 5371 5372 #if CHROMA_QP_EXTENSION 5373 Int curChromaQpOffset = pcCU->getSlice()->getPPS()->getChromaCbQpOffset() + pcCU->getSlice()->getSliceQpDeltaCb(); 5374 m_pcTrQuant->setQPforQuant( pcCU->getQP( 0 ), TEXT_CHROMA, pcCU->getSlice()->getSPS()->getQpBDOffsetC(), curChromaQpOffset ); 5375 #else 5376 m_pcTrQuant->setQPforQuant( pcCU->getQP( 0 ), TEXT_CHROMA, pcCU->getSlice()->getSPS()->getQpBDOffsetC(), pcCU->getSlice()->getPPS()->getChromaCbQpOffset() ); 5377 #endif 5378 5379 #if RDOQ_CHROMA_LAMBDA 5380 m_pcTrQuant->selectLambda(TEXT_CHROMA); 5381 #endif 5382 5383 m_pcTrQuant->transformNxN( pcCU, pcResi->getCbAddr(absTUPartIdxC), pcResi->getCStride(), pcCoeffCurrU, 5384 #if ADAPTIVE_QP_SELECTION 5385 pcArlCoeffCurrU, 5386 #endif 5387 trWidthC, trHeightC, uiAbsSumU, TEXT_CHROMA_U, uiAbsPartIdx ); 5388 5389 #if CHROMA_QP_EXTENSION 5390 curChromaQpOffset = pcCU->getSlice()->getPPS()->getChromaCrQpOffset() + pcCU->getSlice()->getSliceQpDeltaCr(); 5391 m_pcTrQuant->setQPforQuant( pcCU->getQP( 0 ), TEXT_CHROMA, pcCU->getSlice()->getSPS()->getQpBDOffsetC(), curChromaQpOffset ); 5392 #else 5393 m_pcTrQuant->setQPforQuant( pcCU->getQP( 0 ), TEXT_CHROMA, pcCU->getSlice()->getSPS()->getQpBDOffsetC(), pcCU->getSlice()->getPPS()->getChromaCrQpOffset() ); 5394 #endif 5395 m_pcTrQuant->transformNxN( pcCU, pcResi->getCrAddr(absTUPartIdxC), pcResi->getCStride(), pcCoeffCurrV, 5396 #if ADAPTIVE_QP_SELECTION 5397 pcArlCoeffCurrV, 5398 #endif 5399 trWidthC, trHeightC, uiAbsSumV, TEXT_CHROMA_V, uiAbsPartIdx ); 5400 5401 pcCU->setCbfSubParts( uiAbsSumU ? uiSetCbf : 0, TEXT_CHROMA_U, uiAbsPartIdx, pcCU->getDepth(0)+uiTrModeC ); 5402 pcCU->setCbfSubParts( uiAbsSumV ? uiSetCbf : 0, TEXT_CHROMA_V, uiAbsPartIdx, pcCU->getDepth(0)+uiTrModeC ); 5403 } 5404 5405 m_pcEntropyCoder->resetBits(); 5406 5407 { 5408 m_pcEntropyCoder->encodeQtCbf( pcCU, uiAbsPartIdx, TEXT_LUMA, uiTrMode ); 5409 } 5410 5411 m_pcEntropyCoder->encodeCoeffNxN( pcCU, pcCoeffCurrY, uiAbsPartIdx, trWidth, trHeight, uiDepth, TEXT_LUMA ); 5412 const UInt uiSingleBitsY = m_pcEntropyCoder->getNumberOfWrittenBits(); 5413 5414 UInt uiSingleBitsU = 0; 5415 UInt uiSingleBitsV = 0; 5416 if( bCodeChroma ) 5417 { 5418 { 5419 m_pcEntropyCoder->encodeQtCbf ( pcCU, uiAbsPartIdx, TEXT_CHROMA_U, uiTrMode ); 5420 } 5421 m_pcEntropyCoder->encodeCoeffNxN( pcCU, pcCoeffCurrU, uiAbsPartIdx, trWidthC, trHeightC, uiDepth, TEXT_CHROMA_U ); 5422 uiSingleBitsU = m_pcEntropyCoder->getNumberOfWrittenBits() - uiSingleBitsY; 5423 5424 { 5425 m_pcEntropyCoder->encodeQtCbf ( pcCU, uiAbsPartIdx, TEXT_CHROMA_V, uiTrMode ); 5426 } 5427 m_pcEntropyCoder->encodeCoeffNxN( pcCU, pcCoeffCurrV, uiAbsPartIdx, trWidthC, trHeightC, uiDepth, TEXT_CHROMA_V ); 5428 uiSingleBitsV = m_pcEntropyCoder->getNumberOfWrittenBits() - ( uiSingleBitsY + uiSingleBitsU ); 5429 } 5430 5431 const UInt uiNumSamplesLuma = 1 << (uiLog2TrSize<<1); 5432 const UInt uiNumSamplesChro = 1 << (uiLog2TrSizeC<<1); 5433 5434 ::memset( m_pTempPel, 0, sizeof( Pel ) * uiNumSamplesLuma ); // not necessary needed for inside of recursion (only at the beginning) 5435 5436 #if IBDI_DISTORTION 5437 UInt uiDistY = m_pcRdCost->getDistPart( pcPred->getLumaAddr( absTUPartIdx ), pcPred->getStride(), pcOrg->getLumaAddr( absTUPartIdx), pcOrg->getStride(), trWidth, trHeight); 5438 #else 5439 UInt uiDistY = m_pcRdCost->getDistPart( m_pTempPel, trWidth, pcResi->getLumaAddr( absTUPartIdx ), pcResi->getStride(), trWidth, trHeight ); // initialized with zero residual destortion 5440 #endif 5441 if ( puiZeroDist ) 5442 { 5443 *puiZeroDist += uiDistY; 5444 } 5445 if( uiAbsSumY ) 5446 { 5447 Pel *pcResiCurrY = m_pcQTTempTComYuv[ uiQTTempAccessLayer ].getLumaAddr( absTUPartIdx ); 5448 5449 m_pcTrQuant->setQPforQuant( pcCU->getQP( 0 ), TEXT_LUMA, pcCU->getSlice()->getSPS()->getQpBDOffsetY(), 0 ); 5450 5451 Int scalingListType = 3 + g_eTTable[(Int)TEXT_LUMA]; 5452 assert(scalingListType < 6); 5453 #if NO_RESIDUAL_FLAG_FOR_BLPRED 5454 if(pcCU->isIntraBL(uiAbsPartIdx) ) 5455 { 5456 m_pcTrQuant->invtransformNxN( pcCU->getCUTransquantBypass(uiAbsPartIdx), TEXT_LUMA,DC_IDX, pcResiCurrY, m_pcQTTempTComYuv[uiQTTempAccessLayer].getStride(), pcCoeffCurrY, trWidth, trHeight, scalingListType );//this is for inter mode only 5457 } 5458 else 5459 { 5460 m_pcTrQuant->invtransformNxN( pcCU->getCUTransquantBypass(uiAbsPartIdx), TEXT_LUMA,REG_DCT, pcResiCurrY, m_pcQTTempTComYuv[uiQTTempAccessLayer].getStride(), pcCoeffCurrY, trWidth, trHeight, scalingListType );//this is for inter mode only 5461 } 5462 #else 5463 m_pcTrQuant->invtransformNxN( pcCU->getCUTransquantBypass(uiAbsPartIdx), TEXT_LUMA,REG_DCT, pcResiCurrY, m_pcQTTempTComYuv[uiQTTempAccessLayer].getStride(), pcCoeffCurrY, trWidth, trHeight, scalingListType );//this is for inter mode only 5464 #endif 5465 5466 const UInt uiNonzeroDistY = m_pcRdCost->getDistPart( m_pcQTTempTComYuv[uiQTTempAccessLayer].getLumaAddr( absTUPartIdx ), m_pcQTTempTComYuv[uiQTTempAccessLayer].getStride(), 5467 pcResi->getLumaAddr( absTUPartIdx ), pcResi->getStride(), trWidth,trHeight ); 5468 if (pcCU->isLosslessCoded(0)) 5469 { 5470 uiDistY = uiNonzeroDistY; 5471 } 5472 else 5473 { 5474 const Double singleCostY = m_pcRdCost->calcRdCost( uiSingleBitsY, uiNonzeroDistY ); 5475 #if TU_ZERO_CBF_RDO 5476 m_pcEntropyCoder->resetBits(); 5477 m_pcEntropyCoder->encodeQtCbfZero( pcCU, uiAbsPartIdx, TEXT_LUMA, uiTrMode ); 5478 const UInt uiNullBitsY = m_pcEntropyCoder->getNumberOfWrittenBits(); 5479 const Double nullCostY = m_pcRdCost->calcRdCost( uiNullBitsY, uiDistY ); 5480 #else 5481 const Double nullCostY = m_pcRdCost->calcRdCost( 0, uiDistY ); 5482 #endif 5483 if( nullCostY < singleCostY ) 5484 { 5485 uiAbsSumY = 0; 5486 ::memset( pcCoeffCurrY, 0, sizeof( TCoeff ) * uiNumSamplesLuma ); 5487 #if INTER_TRANSFORMSKIP 5488 if( checkTransformSkipY ) 5489 { 5490 minCostY = nullCostY; 5491 } 5492 #endif 5493 } 5494 else 5495 { 5496 uiDistY = uiNonzeroDistY; 5497 #if INTER_TRANSFORMSKIP 5498 if( checkTransformSkipY ) 5499 { 5500 minCostY = singleCostY; 5501 } 5502 #endif 5503 } 5504 } 5505 } 5506 #if INTER_TRANSFORMSKIP 5507 else if( checkTransformSkipY ) 5508 { 5509 #if TU_ZERO_CBF_RDO 5510 m_pcEntropyCoder->resetBits(); 5511 m_pcEntropyCoder->encodeQtCbfZero( pcCU, uiAbsPartIdx, TEXT_LUMA, uiTrMode ); 5512 const UInt uiNullBitsY = m_pcEntropyCoder->getNumberOfWrittenBits(); 5513 minCostY = m_pcRdCost->calcRdCost( uiNullBitsY, uiDistY ); 5514 #else 5515 minCostY = m_pcRdCost->calcRdCost( 0, uiDistY ); 5516 #endif 5517 } 5518 #endif 5519 5520 if( !uiAbsSumY ) 5521 { 5522 Pel *pcPtr = m_pcQTTempTComYuv[uiQTTempAccessLayer].getLumaAddr( absTUPartIdx ); 5523 const UInt uiStride = m_pcQTTempTComYuv[uiQTTempAccessLayer].getStride(); 5524 for( UInt uiY = 0; uiY < trHeight; ++uiY ) 5525 { 5526 ::memset( pcPtr, 0, sizeof( Pel ) * trWidth ); 5527 pcPtr += uiStride; 5528 } 5529 } 5530 5531 UInt uiDistU = 0; 5532 UInt uiDistV = 0; 5533 if( bCodeChroma ) 5534 { 5535 #if IBDI_DISTORTION 5536 uiDistU = m_pcRdCost->getDistPart( pcPred->getCbAddr( absTUPartIdxC ), pcPred->getCStride(), pcOrg->getCbAddr( absTUPartIdxC ), pcOrg->getCStride(), trWidthC, trHeightC 5537 #if WEIGHTED_CHROMA_DISTORTION 5538 , true 5539 #endif 5540 ); 5541 #else 5542 uiDistU = m_pcRdCost->getDistPart( m_pTempPel, trWidthC, pcResi->getCbAddr( absTUPartIdxC ), pcResi->getCStride(), trWidthC, trHeightC 5543 #if WEIGHTED_CHROMA_DISTORTION 5544 , true 5545 #endif 5546 ); // initialized with zero residual destortion 5547 #endif 5548 if ( puiZeroDist ) 5549 { 5550 *puiZeroDist += uiDistU; 5551 } 5552 if( uiAbsSumU ) 5553 { 5554 Pel *pcResiCurrU = m_pcQTTempTComYuv[uiQTTempAccessLayer].getCbAddr( absTUPartIdxC ); 5555 5556 #if CHROMA_QP_EXTENSION 5557 Int curChromaQpOffset = pcCU->getSlice()->getPPS()->getChromaCbQpOffset() + pcCU->getSlice()->getSliceQpDeltaCb(); 5558 m_pcTrQuant->setQPforQuant( pcCU->getQP( 0 ), TEXT_CHROMA, pcCU->getSlice()->getSPS()->getQpBDOffsetC(), curChromaQpOffset ); 5559 #else 5560 m_pcTrQuant->setQPforQuant( pcCU->getQP( 0 ), TEXT_CHROMA, pcCU->getSlice()->getSPS()->getQpBDOffsetC(), pcCU->getSlice()->getPPS()->getChromaCbQpOffset() ); 5561 #endif 5562 5563 Int scalingListType = 3 + g_eTTable[(Int)TEXT_CHROMA_U]; 5564 assert(scalingListType < 6); 5565 m_pcTrQuant->invtransformNxN( pcCU->getCUTransquantBypass(uiAbsPartIdx), TEXT_CHROMA,REG_DCT, pcResiCurrU, m_pcQTTempTComYuv[uiQTTempAccessLayer].getCStride(), pcCoeffCurrU, trWidthC, trHeightC, scalingListType ); 5566 5567 const UInt uiNonzeroDistU = m_pcRdCost->getDistPart( m_pcQTTempTComYuv[uiQTTempAccessLayer].getCbAddr( absTUPartIdxC), m_pcQTTempTComYuv[uiQTTempAccessLayer].getCStride(), 5568 pcResi->getCbAddr( absTUPartIdxC), pcResi->getCStride(), trWidthC, trHeightC 5569 #if WEIGHTED_CHROMA_DISTORTION 5570 , true 5571 #endif 5572 ); 5573 5574 if(pcCU->isLosslessCoded(0)) 5575 { 5576 uiDistU = uiNonzeroDistU; 5577 } 5578 else 5579 { 5580 const Double dSingleCostU = m_pcRdCost->calcRdCost( uiSingleBitsU, uiNonzeroDistU ); 5581 #if TU_ZERO_CBF_RDO 5582 m_pcEntropyCoder->resetBits(); 5583 m_pcEntropyCoder->encodeQtCbfZero( pcCU, uiAbsPartIdx, TEXT_CHROMA_U, uiTrMode ); 5584 const UInt uiNullBitsU = m_pcEntropyCoder->getNumberOfWrittenBits(); 5585 const Double dNullCostU = m_pcRdCost->calcRdCost( uiNullBitsU, uiDistU ); 5586 #else 5587 const Double dNullCostU = m_pcRdCost->calcRdCost( 0, uiDistU ); 5588 #endif 5589 if( dNullCostU < dSingleCostU ) 5590 { 5591 uiAbsSumU = 0; 5592 ::memset( pcCoeffCurrU, 0, sizeof( TCoeff ) * uiNumSamplesChro ); 5593 #if INTER_TRANSFORMSKIP 5594 if( checkTransformSkipUV ) 5595 { 5596 minCostU = dNullCostU; 5597 } 5598 #endif 5599 } 5600 else 5601 { 5602 uiDistU = uiNonzeroDistU; 5603 #if INTER_TRANSFORMSKIP 5604 if( checkTransformSkipUV ) 5605 { 5606 minCostU = dSingleCostU; 5607 } 5608 #endif 5609 } 5610 } 5611 } 5612 #if INTER_TRANSFORMSKIP 5613 else if( checkTransformSkipUV ) 5614 { 5615 #if TU_ZERO_CBF_RDO 5616 m_pcEntropyCoder->resetBits(); 5617 m_pcEntropyCoder->encodeQtCbfZero( pcCU, uiAbsPartIdx, TEXT_CHROMA_U, uiTrModeC ); 5618 const UInt uiNullBitsU = m_pcEntropyCoder->getNumberOfWrittenBits(); 5619 minCostU = m_pcRdCost->calcRdCost( uiNullBitsU, uiDistU ); 5620 #else 5621 minCostU = m_pcRdCost->calcRdCost( 0, uiDistU ); 5622 #endif 5623 } 5624 #endif 5625 if( !uiAbsSumU ) 5626 { 5627 Pel *pcPtr = m_pcQTTempTComYuv[uiQTTempAccessLayer].getCbAddr( absTUPartIdxC ); 5628 const UInt uiStride = m_pcQTTempTComYuv[uiQTTempAccessLayer].getCStride(); 5629 for( UInt uiY = 0; uiY < trHeightC; ++uiY ) 5630 { 5631 ::memset( pcPtr, 0, sizeof(Pel) * trWidthC ); 5632 pcPtr += uiStride; 5633 } 5634 } 5635 5636 #if IBDI_DISTORTION 5637 uiDistV = m_pcRdCost->getDistPart( pcPred->getCrAddr( absTUPartIdxC ), pcPred->getCStride(), pcOrg->getCrAddr( absTUPartIdxC ), pcOrg->getCStride(), trWidthC, trHeightC 5638 #if WEIGHTED_CHROMA_DISTORTION 5639 , true 5640 #endif 5641 ); 5642 #else 5643 uiDistV = m_pcRdCost->getDistPart( m_pTempPel, trWidthC, pcResi->getCrAddr( absTUPartIdxC), pcResi->getCStride(), trWidthC, trHeightC 5644 #if WEIGHTED_CHROMA_DISTORTION 5645 , true 5646 #endif 5647 ); // initialized with zero residual destortion 5648 #endif 5649 if ( puiZeroDist ) 5650 { 5651 *puiZeroDist += uiDistV; 5652 } 5653 if( uiAbsSumV ) 5654 { 5655 Pel *pcResiCurrV = m_pcQTTempTComYuv[uiQTTempAccessLayer].getCrAddr( absTUPartIdxC ); 5656 if( !uiAbsSumU ) 5657 { 5658 #if CHROMA_QP_EXTENSION 5659 Int curChromaQpOffset = pcCU->getSlice()->getPPS()->getChromaCrQpOffset() + pcCU->getSlice()->getSliceQpDeltaCr(); 5660 m_pcTrQuant->setQPforQuant( pcCU->getQP( 0 ), TEXT_CHROMA, pcCU->getSlice()->getSPS()->getQpBDOffsetC(), curChromaQpOffset ); 5661 #else 5662 m_pcTrQuant->setQPforQuant( pcCU->getQP( 0 ), TEXT_CHROMA, pcCU->getSlice()->getSPS()->getQpBDOffsetC(), pcCU->getSlice()->getPPS()->getChromaCrQpOffset() ); 5663 #endif 5664 } 5665 Int scalingListType = 3 + g_eTTable[(Int)TEXT_CHROMA_V]; 5666 assert(scalingListType < 6); 5667 m_pcTrQuant->invtransformNxN( pcCU->getCUTransquantBypass(uiAbsPartIdx), TEXT_CHROMA,REG_DCT, pcResiCurrV, m_pcQTTempTComYuv[uiQTTempAccessLayer].getCStride(), pcCoeffCurrV, trWidthC, trHeightC, scalingListType ); 5668 5669 const UInt uiNonzeroDistV = m_pcRdCost->getDistPart( m_pcQTTempTComYuv[uiQTTempAccessLayer].getCrAddr( absTUPartIdxC ), m_pcQTTempTComYuv[uiQTTempAccessLayer].getCStride(), 5670 pcResi->getCrAddr( absTUPartIdxC ), pcResi->getCStride(), trWidthC, trHeightC 5671 #if WEIGHTED_CHROMA_DISTORTION 5672 , true 5673 #endif 5674 ); 5675 if (pcCU->isLosslessCoded(0)) 5676 { 5677 uiDistV = uiNonzeroDistV; 5678 } 5679 else 5680 { 5681 const Double dSingleCostV = m_pcRdCost->calcRdCost( uiSingleBitsV, uiNonzeroDistV ); 5682 #if TU_ZERO_CBF_RDO 5683 m_pcEntropyCoder->resetBits(); 5684 m_pcEntropyCoder->encodeQtCbfZero( pcCU, uiAbsPartIdx, TEXT_CHROMA_V, uiTrMode ); 5685 const UInt uiNullBitsV = m_pcEntropyCoder->getNumberOfWrittenBits(); 5686 const Double dNullCostV = m_pcRdCost->calcRdCost( uiNullBitsV, uiDistV ); 5687 #else 5688 const Double dNullCostV = m_pcRdCost->calcRdCost( 0, uiDistV ); 5689 #endif 5690 if( dNullCostV < dSingleCostV ) 5691 { 5692 uiAbsSumV = 0; 5693 ::memset( pcCoeffCurrV, 0, sizeof( TCoeff ) * uiNumSamplesChro ); 5694 #if INTER_TRANSFORMSKIP 5695 if( checkTransformSkipUV ) 5696 { 5697 minCostV = dNullCostV; 5698 } 5699 #endif 5700 } 5701 else 5702 { 5703 uiDistV = uiNonzeroDistV; 5704 #if INTER_TRANSFORMSKIP 5705 if( checkTransformSkipUV ) 5706 { 5707 minCostV = dSingleCostV; 5708 } 5709 #endif 5710 } 5711 } 5712 } 5713 #if INTER_TRANSFORMSKIP 5714 else if( checkTransformSkipUV ) 5715 { 5716 #if TU_ZERO_CBF_RDO 5717 m_pcEntropyCoder->resetBits(); 5718 m_pcEntropyCoder->encodeQtCbfZero( pcCU, uiAbsPartIdx, TEXT_CHROMA_V, uiTrModeC ); 5719 const UInt uiNullBitsV = m_pcEntropyCoder->getNumberOfWrittenBits(); 5720 minCostV = m_pcRdCost->calcRdCost( uiNullBitsV, uiDistV ); 5721 #else 5722 minCostV = m_pcRdCost->calcRdCost( 0, uiDistV ); 5723 #endif 5724 } 5725 #endif 5726 if( !uiAbsSumV ) 5727 { 5728 Pel *pcPtr = m_pcQTTempTComYuv[uiQTTempAccessLayer].getCrAddr( absTUPartIdxC ); 5729 const UInt uiStride = m_pcQTTempTComYuv[uiQTTempAccessLayer].getCStride(); 5730 for( UInt uiY = 0; uiY < trHeightC; ++uiY ) 5731 { 5732 ::memset( pcPtr, 0, sizeof(Pel) * trWidthC ); 5733 pcPtr += uiStride; 5734 } 5735 } 5736 } 5737 pcCU->setCbfSubParts( uiAbsSumY ? uiSetCbf : 0, TEXT_LUMA, uiAbsPartIdx, uiDepth ); 5738 if( bCodeChroma ) 5739 { 5740 pcCU->setCbfSubParts( uiAbsSumU ? uiSetCbf : 0, TEXT_CHROMA_U, uiAbsPartIdx, pcCU->getDepth(0)+uiTrModeC ); 5741 pcCU->setCbfSubParts( uiAbsSumV ? uiSetCbf : 0, TEXT_CHROMA_V, uiAbsPartIdx, pcCU->getDepth(0)+uiTrModeC ); 5742 } 5743 5744 #if INTER_TRANSFORMSKIP 5745 if( checkTransformSkipY ) 5746 { 5747 UInt uiNonzeroDistY, uiAbsSumTransformSkipY; 5748 Double dSingleCostY; 5749 5750 Pel *pcResiCurrY = m_pcQTTempTComYuv[ uiQTTempAccessLayer ].getLumaAddr( absTUPartIdx ); 5751 UInt resiYStride = m_pcQTTempTComYuv[ uiQTTempAccessLayer ].getStride(); 5752 5753 TCoeff bestCoeffY[32*32]; 5754 memcpy( bestCoeffY, pcCoeffCurrY, sizeof(TCoeff) * uiNumSamplesLuma ); 5755 5756 #if ADAPTIVE_QP_SELECTION 5757 TCoeff bestArlCoeffY[32*32]; 5758 memcpy( bestArlCoeffY, pcArlCoeffCurrY, sizeof(TCoeff) * uiNumSamplesLuma ); 5759 #endif 5760 5761 Pel bestResiY[32*32]; 5762 for ( Int i = 0; i < trHeight; ++i ) 5763 { 5764 memcpy( &bestResiY[i*trWidth], pcResiCurrY+i*resiYStride, sizeof(Pel) * trWidth ); 5765 } 5766 5767 if( m_bUseSBACRD ) 5768 { 5769 m_pcRDGoOnSbacCoder->load( m_pppcRDSbacCoder[ uiDepth ][ CI_QT_TRAFO_ROOT ] ); 5770 } 5771 5772 pcCU->setTransformSkipSubParts ( 1, TEXT_LUMA, uiAbsPartIdx, uiDepth ); 5773 5774 if (m_pcEncCfg->getUseRDOQ()) 5775 { 5776 m_pcEntropyCoder->estimateBit( m_pcTrQuant->m_pcEstBitsSbac, trWidth, trHeight, TEXT_LUMA ); 5777 } 5778 5779 m_pcTrQuant->setQPforQuant( pcCU->getQP( 0 ), TEXT_LUMA, pcCU->getSlice()->getSPS()->getQpBDOffsetY(), 0 ); 5780 5781 #if RDOQ_CHROMA_LAMBDA 5782 m_pcTrQuant->selectLambda(TEXT_LUMA); 5783 #endif 5784 m_pcTrQuant->transformNxN( pcCU, pcResi->getLumaAddr( absTUPartIdx ), pcResi->getStride (), pcCoeffCurrY, 5785 #if ADAPTIVE_QP_SELECTION 5786 pcArlCoeffCurrY, 5787 #endif 5788 trWidth, trHeight, uiAbsSumTransformSkipY, TEXT_LUMA, uiAbsPartIdx, true ); 5789 pcCU->setCbfSubParts( uiAbsSumTransformSkipY ? uiSetCbf : 0, TEXT_LUMA, uiAbsPartIdx, uiDepth ); 5790 5791 if( uiAbsSumTransformSkipY != 0 ) 5792 { 5793 m_pcEntropyCoder->resetBits(); 5794 m_pcEntropyCoder->encodeQtCbf( pcCU, uiAbsPartIdx, TEXT_LUMA, uiTrMode ); 5795 m_pcEntropyCoder->encodeCoeffNxN( pcCU, pcCoeffCurrY, uiAbsPartIdx, trWidth, trHeight, uiDepth, TEXT_LUMA ); 5796 const UInt uiTsSingleBitsY = m_pcEntropyCoder->getNumberOfWrittenBits(); 5797 5798 m_pcTrQuant->setQPforQuant( pcCU->getQP( 0 ), TEXT_LUMA, pcCU->getSlice()->getSPS()->getQpBDOffsetY(), 0 ); 5799 5800 Int scalingListType = 3 + g_eTTable[(Int)TEXT_LUMA]; 5801 assert(scalingListType < 6); 5802 5803 #if NO_RESIDUAL_FLAG_FOR_BLPRED 5804 if(pcCU->isIntraBL(uiAbsPartIdx) ) 5805 { 5806 m_pcTrQuant->invtransformNxN( pcCU->getCUTransquantBypass(uiAbsPartIdx), TEXT_LUMA,DC_IDX, pcResiCurrY, m_pcQTTempTComYuv[uiQTTempAccessLayer].getStride(), pcCoeffCurrY, trWidth, trHeight, scalingListType, true ); 5807 } 5808 else 5809 { 5810 m_pcTrQuant->invtransformNxN( pcCU->getCUTransquantBypass(uiAbsPartIdx), TEXT_LUMA,REG_DCT, pcResiCurrY, m_pcQTTempTComYuv[uiQTTempAccessLayer].getStride(), pcCoeffCurrY, trWidth, trHeight, scalingListType, true ); 5811 } 5812 #else 5813 m_pcTrQuant->invtransformNxN( pcCU->getCUTransquantBypass(uiAbsPartIdx), TEXT_LUMA,REG_DCT, pcResiCurrY, m_pcQTTempTComYuv[uiQTTempAccessLayer].getStride(), pcCoeffCurrY, trWidth, trHeight, scalingListType, true ); 5814 #endif 5815 5816 uiNonzeroDistY = m_pcRdCost->getDistPart( m_pcQTTempTComYuv[uiQTTempAccessLayer].getLumaAddr( absTUPartIdx ), m_pcQTTempTComYuv[uiQTTempAccessLayer].getStride(), 5817 pcResi->getLumaAddr( absTUPartIdx ), pcResi->getStride(), trWidth, trHeight ); 5818 5819 dSingleCostY = m_pcRdCost->calcRdCost( uiTsSingleBitsY, uiNonzeroDistY ); 5820 } 5821 5822 if( !uiAbsSumTransformSkipY || minCostY < dSingleCostY ) 5823 { 5824 pcCU->setTransformSkipSubParts ( 0, TEXT_LUMA, uiAbsPartIdx, uiDepth ); 5825 memcpy( pcCoeffCurrY, bestCoeffY, sizeof(TCoeff) * uiNumSamplesLuma ); 5826 #if ADAPTIVE_QP_SELECTION 5827 memcpy( pcArlCoeffCurrY, bestArlCoeffY, sizeof(TCoeff) * uiNumSamplesLuma ); 5828 #endif 5829 for( Int i = 0; i < trHeight; ++i ) 5830 { 5831 memcpy( pcResiCurrY+i*resiYStride, &bestResiY[i*trWidth], sizeof(Pel) * trWidth ); 5832 } 5833 } 5834 else 5835 { 5836 uiDistY = uiNonzeroDistY; 5837 uiAbsSumY = uiAbsSumTransformSkipY; 5838 uiBestTransformMode[0] = 1; 5839 } 5840 5841 pcCU->setCbfSubParts( uiAbsSumY ? uiSetCbf : 0, TEXT_LUMA, uiAbsPartIdx, uiDepth ); 5842 } 5843 5844 if( bCodeChroma && checkTransformSkipUV ) 5845 { 5846 UInt uiNonzeroDistU, uiNonzeroDistV, uiAbsSumTransformSkipU, uiAbsSumTransformSkipV; 5847 Double dSingleCostU, dSingleCostV; 5848 5849 Pel *pcResiCurrU = m_pcQTTempTComYuv[uiQTTempAccessLayer].getCbAddr( absTUPartIdxC ); 5850 Pel *pcResiCurrV = m_pcQTTempTComYuv[uiQTTempAccessLayer].getCrAddr( absTUPartIdxC ); 5851 UInt resiCStride = m_pcQTTempTComYuv[uiQTTempAccessLayer].getCStride(); 5852 5853 TCoeff bestCoeffU[32*32], bestCoeffV[32*32]; 5854 memcpy( bestCoeffU, pcCoeffCurrU, sizeof(TCoeff) * uiNumSamplesChro ); 5855 memcpy( bestCoeffV, pcCoeffCurrV, sizeof(TCoeff) * uiNumSamplesChro ); 5856 5857 #if ADAPTIVE_QP_SELECTION 5858 TCoeff bestArlCoeffU[32*32], bestArlCoeffV[32*32]; 5859 memcpy( bestArlCoeffU, pcArlCoeffCurrU, sizeof(TCoeff) * uiNumSamplesChro ); 5860 memcpy( bestArlCoeffV, pcArlCoeffCurrV, sizeof(TCoeff) * uiNumSamplesChro ); 5861 #endif 5862 5863 Pel bestResiU[32*32], bestResiV[32*32]; 5864 for (Int i = 0; i < trHeightC; ++i ) 5865 { 5866 memcpy( &bestResiU[i*trWidthC], pcResiCurrU+i*resiCStride, sizeof(Pel) * trWidthC ); 5867 memcpy( &bestResiV[i*trWidthC], pcResiCurrV+i*resiCStride, sizeof(Pel) * trWidthC ); 5868 } 5869 5870 if( m_bUseSBACRD ) 5871 { 5872 m_pcRDGoOnSbacCoder->load( m_pppcRDSbacCoder[ uiDepth ][ CI_QT_TRAFO_ROOT ] ); 5873 } 5874 5875 pcCU->setTransformSkipSubParts ( 1, TEXT_CHROMA_U, uiAbsPartIdx, pcCU->getDepth(0)+uiTrModeC ); 5876 pcCU->setTransformSkipSubParts ( 1, TEXT_CHROMA_V, uiAbsPartIdx, pcCU->getDepth(0)+uiTrModeC ); 5877 5878 if (m_pcEncCfg->getUseRDOQ()) 5879 { 5880 m_pcEntropyCoder->estimateBit(m_pcTrQuant->m_pcEstBitsSbac, trWidthC, trHeightC, TEXT_CHROMA ); 5881 } 5882 5883 #if CHROMA_QP_EXTENSION 5884 Int curChromaQpOffset = pcCU->getSlice()->getPPS()->getChromaCbQpOffset() + pcCU->getSlice()->getSliceQpDeltaCb(); 5885 m_pcTrQuant->setQPforQuant( pcCU->getQP( 0 ), TEXT_CHROMA, pcCU->getSlice()->getSPS()->getQpBDOffsetC(), curChromaQpOffset ); 5886 #else 5887 m_pcTrQuant->setQPforQuant( pcCU->getQP( 0 ), TEXT_CHROMA, pcCU->getSlice()->getSPS()->getQpBDOffsetC(), pcCU->getSlice()->getPPS()->getChromaCbQpOffset() ); 5888 #endif 5889 5890 #if RDOQ_CHROMA_LAMBDA 5891 m_pcTrQuant->selectLambda(TEXT_CHROMA); 5892 #endif 5893 5894 m_pcTrQuant->transformNxN( pcCU, pcResi->getCbAddr(absTUPartIdxC), pcResi->getCStride(), pcCoeffCurrU, 5895 #if ADAPTIVE_QP_SELECTION 5896 pcArlCoeffCurrU, 5897 #endif 5898 trWidthC, trHeightC, uiAbsSumTransformSkipU, TEXT_CHROMA_U, uiAbsPartIdx, true ); 5899 #if CHROMA_QP_EXTENSION 5900 curChromaQpOffset = pcCU->getSlice()->getPPS()->getChromaCrQpOffset() + pcCU->getSlice()->getSliceQpDeltaCr(); 5901 m_pcTrQuant->setQPforQuant( pcCU->getQP( 0 ), TEXT_CHROMA, pcCU->getSlice()->getSPS()->getQpBDOffsetC(), curChromaQpOffset ); 5902 #else 5903 m_pcTrQuant->setQPforQuant( pcCU->getQP( 0 ), TEXT_CHROMA, pcCU->getSlice()->getSPS()->getQpBDOffsetC(), pcCU->getSlice()->getPPS()->getChromaCrQpOffset() ); 5904 #endif 5905 m_pcTrQuant->transformNxN( pcCU, pcResi->getCrAddr(absTUPartIdxC), pcResi->getCStride(), pcCoeffCurrV, 5906 #if ADAPTIVE_QP_SELECTION 5907 pcArlCoeffCurrV, 5908 #endif 5909 trWidthC, trHeightC, uiAbsSumTransformSkipV, TEXT_CHROMA_V, uiAbsPartIdx, true ); 5910 5911 pcCU->setCbfSubParts( uiAbsSumTransformSkipU ? uiSetCbf : 0, TEXT_CHROMA_U, uiAbsPartIdx, pcCU->getDepth(0)+uiTrModeC ); 5912 pcCU->setCbfSubParts( uiAbsSumTransformSkipV ? uiSetCbf : 0, TEXT_CHROMA_V, uiAbsPartIdx, pcCU->getDepth(0)+uiTrModeC ); 5913 5914 m_pcEntropyCoder->resetBits(); 5915 uiSingleBitsU = 0; 5916 uiSingleBitsV = 0; 5917 5918 if( uiAbsSumTransformSkipU ) 5919 { 5920 m_pcEntropyCoder->encodeQtCbf ( pcCU, uiAbsPartIdx, TEXT_CHROMA_U, uiTrMode ); 5921 m_pcEntropyCoder->encodeCoeffNxN( pcCU, pcCoeffCurrU, uiAbsPartIdx, trWidthC, trHeightC, uiDepth, TEXT_CHROMA_U ); 5922 uiSingleBitsU = m_pcEntropyCoder->getNumberOfWrittenBits(); 5923 5924 #if CHROMA_QP_EXTENSION 5925 curChromaQpOffset = pcCU->getSlice()->getPPS()->getChromaCbQpOffset() + pcCU->getSlice()->getSliceQpDeltaCb(); 5926 m_pcTrQuant->setQPforQuant( pcCU->getQP( 0 ), TEXT_CHROMA, pcCU->getSlice()->getSPS()->getQpBDOffsetC(), curChromaQpOffset ); 5927 #else 5928 m_pcTrQuant->setQPforQuant( pcCU->getQP( 0 ), TEXT_CHROMA, pcCU->getSlice()->getSPS()->getQpBDOffsetC(), pcCU->getSlice()->getPPS()->getChromaCbQpOffset() ); 5929 #endif 5930 5931 Int scalingListType = 3 + g_eTTable[(Int)TEXT_CHROMA_U]; 5932 assert(scalingListType < 6); 5933 5934 m_pcTrQuant->invtransformNxN( pcCU->getCUTransquantBypass(uiAbsPartIdx), TEXT_CHROMA,REG_DCT, pcResiCurrU, m_pcQTTempTComYuv[uiQTTempAccessLayer].getCStride(), pcCoeffCurrU, trWidthC, trHeightC, scalingListType, true ); 5935 5936 uiNonzeroDistU = m_pcRdCost->getDistPart( m_pcQTTempTComYuv[uiQTTempAccessLayer].getCbAddr( absTUPartIdxC), m_pcQTTempTComYuv[uiQTTempAccessLayer].getCStride(), 5937 pcResi->getCbAddr( absTUPartIdxC), pcResi->getCStride(), trWidthC, trHeightC 5938 #if WEIGHTED_CHROMA_DISTORTION 5939 , true 5940 #endif 5941 ); 5942 5943 dSingleCostU = m_pcRdCost->calcRdCost( uiSingleBitsU, uiNonzeroDistU ); 5944 } 5945 5946 if( !uiAbsSumTransformSkipU || minCostU < dSingleCostU ) 5947 { 5948 pcCU->setTransformSkipSubParts ( 0, TEXT_CHROMA_U, uiAbsPartIdx, pcCU->getDepth(0)+uiTrModeC ); 5949 5950 memcpy( pcCoeffCurrU, bestCoeffU, sizeof (TCoeff) * uiNumSamplesChro ); 5951 #if ADAPTIVE_QP_SELECTION 5952 memcpy( pcArlCoeffCurrU, bestArlCoeffU, sizeof (TCoeff) * uiNumSamplesChro ); 5953 #endif 5954 for( Int i = 0; i < trHeightC; ++i ) 5955 { 5956 memcpy( pcResiCurrU+i*resiCStride, &bestResiU[i*trWidthC], sizeof(Pel) * trWidthC ); 5957 } 5958 } 5959 else 5960 { 5961 uiDistU = uiNonzeroDistU; 5962 uiAbsSumU = uiAbsSumTransformSkipU; 5963 uiBestTransformMode[1] = 1; 5964 } 5965 5966 if( uiAbsSumTransformSkipV ) 5967 { 5968 m_pcEntropyCoder->encodeQtCbf ( pcCU, uiAbsPartIdx, TEXT_CHROMA_V, uiTrMode ); 5969 m_pcEntropyCoder->encodeCoeffNxN( pcCU, pcCoeffCurrV, uiAbsPartIdx, trWidthC, trHeightC, uiDepth, TEXT_CHROMA_V ); 5970 uiSingleBitsV = m_pcEntropyCoder->getNumberOfWrittenBits() - uiSingleBitsU; 5971 5972 #if CHROMA_QP_EXTENSION 5973 curChromaQpOffset = pcCU->getSlice()->getPPS()->getChromaCrQpOffset() + pcCU->getSlice()->getSliceQpDeltaCr(); 5974 m_pcTrQuant->setQPforQuant( pcCU->getQP( 0 ), TEXT_CHROMA, pcCU->getSlice()->getSPS()->getQpBDOffsetC(), curChromaQpOffset ); 5975 #else 5976 m_pcTrQuant->setQPforQuant( pcCU->getQP( 0 ), TEXT_CHROMA, pcCU->getSlice()->getSPS()->getQpBDOffsetC(), pcCU->getSlice()->getPPS()->getChromaCrQpOffset() ); 5977 #endif 5978 5979 Int scalingListType = 3 + g_eTTable[(Int)TEXT_CHROMA_V]; 5980 assert(scalingListType < 6); 5981 5982 m_pcTrQuant->invtransformNxN( pcCU->getCUTransquantBypass(uiAbsPartIdx), TEXT_CHROMA,REG_DCT, pcResiCurrV, m_pcQTTempTComYuv[uiQTTempAccessLayer].getCStride(), pcCoeffCurrV, trWidthC, trHeightC, scalingListType, true ); 5983 5984 uiNonzeroDistV = m_pcRdCost->getDistPart( m_pcQTTempTComYuv[uiQTTempAccessLayer].getCrAddr( absTUPartIdxC ), m_pcQTTempTComYuv[uiQTTempAccessLayer].getCStride(), 5985 pcResi->getCrAddr( absTUPartIdxC ), pcResi->getCStride(), trWidthC, trHeightC 5986 #if WEIGHTED_CHROMA_DISTORTION 5987 , true 5988 #endif 5989 ); 5990 5991 dSingleCostV = m_pcRdCost->calcRdCost( uiSingleBitsV, uiNonzeroDistV ); 5992 } 5993 5994 if( !uiAbsSumTransformSkipV || minCostV < dSingleCostV ) 5995 { 5996 pcCU->setTransformSkipSubParts ( 0, TEXT_CHROMA_V, uiAbsPartIdx, pcCU->getDepth(0)+uiTrModeC ); 5997 5998 memcpy( pcCoeffCurrV, bestCoeffV, sizeof(TCoeff) * uiNumSamplesChro ); 5999 #if ADAPTIVE_QP_SELECTION 6000 memcpy( pcArlCoeffCurrV, bestArlCoeffV, sizeof(TCoeff) * uiNumSamplesChro ); 6001 #endif 6002 for( Int i = 0; i < trHeightC; ++i ) 6003 { 6004 memcpy( pcResiCurrV+i*resiCStride, &bestResiV[i*trWidthC], sizeof(Pel) * trWidthC ); 6005 } 6006 } 6007 else 6008 { 6009 uiDistV = uiNonzeroDistV; 6010 uiAbsSumV = uiAbsSumTransformSkipV; 6011 uiBestTransformMode[2] = 1; 6012 } 6013 6014 pcCU->setCbfSubParts( uiAbsSumU ? uiSetCbf : 0, TEXT_CHROMA_U, uiAbsPartIdx, pcCU->getDepth(0)+uiTrModeC ); 6015 pcCU->setCbfSubParts( uiAbsSumV ? uiSetCbf : 0, TEXT_CHROMA_V, uiAbsPartIdx, pcCU->getDepth(0)+uiTrModeC ); 6016 } 6017 #endif 6018 6019 if( m_bUseSBACRD ) 6020 { 6021 m_pcRDGoOnSbacCoder->load( m_pppcRDSbacCoder[ uiDepth ][ CI_QT_TRAFO_ROOT ] ); 6022 } 6023 6024 m_pcEntropyCoder->resetBits(); 6025 6026 { 6027 if( uiLog2TrSize > pcCU->getQuadtreeTULog2MinSizeInCU(uiAbsPartIdx) ) 6028 { 6029 #if TRANS_SPLIT_FLAG_CTX_REDUCTION 6030 m_pcEntropyCoder->encodeTransformSubdivFlag( 0, 5 - uiLog2TrSize ); 6031 #else 6032 m_pcEntropyCoder->encodeTransformSubdivFlag( 0, uiDepth ); 6033 #endif 6034 } 6035 } 6036 6037 { 6038 if( bCodeChroma ) 6039 { 6040 m_pcEntropyCoder->encodeQtCbf( pcCU, uiAbsPartIdx, TEXT_CHROMA_U, uiTrMode ); 6041 m_pcEntropyCoder->encodeQtCbf( pcCU, uiAbsPartIdx, TEXT_CHROMA_V, uiTrMode ); 6042 } 6043 6044 m_pcEntropyCoder->encodeQtCbf( pcCU, uiAbsPartIdx, TEXT_LUMA, uiTrMode ); 6045 } 6046 6047 m_pcEntropyCoder->encodeCoeffNxN( pcCU, pcCoeffCurrY, uiAbsPartIdx, trWidth, trHeight, uiDepth, TEXT_LUMA ); 6048 6049 if( bCodeChroma ) 6050 { 6051 m_pcEntropyCoder->encodeCoeffNxN( pcCU, pcCoeffCurrU, uiAbsPartIdx, trWidthC, trHeightC, uiDepth, TEXT_CHROMA_U ); 6052 m_pcEntropyCoder->encodeCoeffNxN( pcCU, pcCoeffCurrV, uiAbsPartIdx, trWidthC, trHeightC, uiDepth, TEXT_CHROMA_V ); 6053 } 6054 6055 uiSingleBits = m_pcEntropyCoder->getNumberOfWrittenBits(); 6056 6057 uiSingleDist = uiDistY + uiDistU + uiDistV; 6058 dSingleCost = m_pcRdCost->calcRdCost( uiSingleBits, uiSingleDist ); 6059 } 6060 6061 // code sub-blocks 6062 if( bCheckSplit ) 6063 { 6064 if( m_bUseSBACRD && bCheckFull ) 6065 { 6066 m_pcRDGoOnSbacCoder->store( m_pppcRDSbacCoder[ uiDepth ][ CI_QT_TRAFO_TEST ] ); 6067 m_pcRDGoOnSbacCoder->load ( m_pppcRDSbacCoder[ uiDepth ][ CI_QT_TRAFO_ROOT ] ); 6068 } 6069 UInt uiSubdivDist = 0; 6070 UInt uiSubdivBits = 0; 6071 Double dSubdivCost = 0.0; 6072 6073 const UInt uiQPartNumSubdiv = pcCU->getPic()->getNumPartInCU() >> ((uiDepth + 1 ) << 1); 6074 for( UInt ui = 0; ui < 4; ++ui ) 6075 { 6076 #if REMOVE_NSQT 6077 UInt nsAddr = uiAbsPartIdx + ui * uiQPartNumSubdiv; 6078 #else 6079 UInt nsAddr = 0; 6080 nsAddr = pcCU->getNSAbsPartIdx( uiLog2TrSize - 1, uiAbsPartIdx + ui * uiQPartNumSubdiv, absTUPartIdx, ui, uiTrMode + 1 ); 6081 #endif 6082 #if IBDI_DISTORTION 6083 xEstimateResidualQT( pcCU, ui, uiAbsPartIdx + ui * uiQPartNumSubdiv, nsAddr, pcOrg, pcPred, pcResi, uiDepth + 1, dSubdivCost, uiSubdivBits, uiSubdivDist, bCheckFull ? NULL : puiZeroDist ); 6084 #else 6085 xEstimateResidualQT( pcCU, ui, uiAbsPartIdx + ui * uiQPartNumSubdiv, nsAddr, pcResi, uiDepth + 1, dSubdivCost, uiSubdivBits, uiSubdivDist, bCheckFull ? NULL : puiZeroDist ); 6086 #endif 6087 } 6088 6089 UInt uiYCbf = 0; 6090 UInt uiUCbf = 0; 6091 UInt uiVCbf = 0; 6092 for( UInt ui = 0; ui < 4; ++ui ) 6093 { 6094 uiYCbf |= pcCU->getCbf( uiAbsPartIdx + ui * uiQPartNumSubdiv, TEXT_LUMA, uiTrMode + 1 ); 6095 uiUCbf |= pcCU->getCbf( uiAbsPartIdx + ui * uiQPartNumSubdiv, TEXT_CHROMA_U, uiTrMode + 1 ); 6096 uiVCbf |= pcCU->getCbf( uiAbsPartIdx + ui * uiQPartNumSubdiv, TEXT_CHROMA_V, uiTrMode + 1 ); 6097 } 6098 for( UInt ui = 0; ui < 4 * uiQPartNumSubdiv; ++ui ) 6099 { 6100 pcCU->getCbf( TEXT_LUMA )[uiAbsPartIdx + ui] |= uiYCbf << uiTrMode; 6101 pcCU->getCbf( TEXT_CHROMA_U )[uiAbsPartIdx + ui] |= uiUCbf << uiTrMode; 6102 pcCU->getCbf( TEXT_CHROMA_V )[uiAbsPartIdx + ui] |= uiVCbf << uiTrMode; 6103 } 6104 6105 if( m_bUseSBACRD ) 6106 { 6107 m_pcRDGoOnSbacCoder->load( m_pppcRDSbacCoder[ uiDepth ][ CI_QT_TRAFO_ROOT ] ); 6108 } 6109 m_pcEntropyCoder->resetBits(); 6110 6111 { 6112 xEncodeResidualQT( pcCU, uiAbsPartIdx, uiDepth, true, TEXT_LUMA ); 6113 xEncodeResidualQT( pcCU, uiAbsPartIdx, uiDepth, false, TEXT_LUMA ); 6114 xEncodeResidualQT( pcCU, uiAbsPartIdx, uiDepth, false, TEXT_CHROMA_U ); 6115 xEncodeResidualQT( pcCU, uiAbsPartIdx, uiDepth, false, TEXT_CHROMA_V ); 6116 } 6117 6118 uiSubdivBits = m_pcEntropyCoder->getNumberOfWrittenBits(); 6119 dSubdivCost = m_pcRdCost->calcRdCost( uiSubdivBits, uiSubdivDist ); 6120 6121 if( uiYCbf || uiUCbf || uiVCbf || !bCheckFull ) 6122 { 6123 if( dSubdivCost < dSingleCost ) 6124 { 6125 rdCost += dSubdivCost; 6126 ruiBits += uiSubdivBits; 6127 ruiDist += uiSubdivDist; 6128 return; 6129 } 6130 } 6131 #if INTER_TRANSFORMSKIP 6132 pcCU->setTransformSkipSubParts ( uiBestTransformMode[0], TEXT_LUMA, uiAbsPartIdx, uiDepth ); 6133 if(bCodeChroma) 6134 { 6135 pcCU->setTransformSkipSubParts ( uiBestTransformMode[1], TEXT_CHROMA_U, uiAbsPartIdx, pcCU->getDepth(0)+uiTrModeC ); 6136 pcCU->setTransformSkipSubParts ( uiBestTransformMode[2], TEXT_CHROMA_V, uiAbsPartIdx, pcCU->getDepth(0)+uiTrModeC ); 6137 } 6138 #endif 6139 assert( bCheckFull ); 6140 if( m_bUseSBACRD ) 6141 { 6142 m_pcRDGoOnSbacCoder->load( m_pppcRDSbacCoder[ uiDepth ][ CI_QT_TRAFO_TEST ] ); 6143 } 6144 } 6145 rdCost += dSingleCost; 6146 ruiBits += uiSingleBits; 6147 ruiDist += uiSingleDist; 6148 6149 pcCU->setTrIdxSubParts( uiTrMode, uiAbsPartIdx, uiDepth ); 6150 6151 pcCU->setCbfSubParts( uiAbsSumY ? uiSetCbf : 0, TEXT_LUMA, uiAbsPartIdx, uiDepth ); 6152 if( bCodeChroma ) 6153 { 6154 pcCU->setCbfSubParts( uiAbsSumU ? uiSetCbf : 0, TEXT_CHROMA_U, uiAbsPartIdx, pcCU->getDepth(0)+uiTrModeC ); 6155 pcCU->setCbfSubParts( uiAbsSumV ? uiSetCbf : 0, TEXT_CHROMA_V, uiAbsPartIdx, pcCU->getDepth(0)+uiTrModeC ); 6156 } 6354 m_pcRDGoOnSbacCoder->load( m_pppcRDSbacCoder[uiDepth][CI_CURR_BEST] ); 6355 } 6356 6357 // determine residual for partition 6358 Double dPUCost = 0.0; 6359 xRecurIntraCodingQT( pcCU, 0, 0, false, pcOrgYuv, pcPredYuv, pcResiYuv, uiOverallDistY, uiOverallDistC, false, dPUCost ); 6360 xSetIntraResultQT( pcCU, 0, 0, false, pcRecoYuv ); 6361 6362 //=== update PU data ==== 6363 pcCU->copyToPic( uiDepth, 0, 0 ); 6364 6365 //===== reset context models ===== 6366 if(m_bUseSBACRD) 6367 { 6368 m_pcRDGoOnSbacCoder->load(m_pppcRDSbacCoder[uiDepth][CI_CURR_BEST]); 6369 } 6370 6371 //===== set distortion (rate and r-d costs are determined later) ===== 6372 pcCU->getTotalDistortion() = uiOverallDistY + uiOverallDistC; 6157 6373 } 6158 6374 6159 Void TEncSearch::xEncodeResidualQT( TComDataCU* pcCU, UInt uiAbsPartIdx, const UInt uiDepth, Bool bSubdivAndCbf, TextType eType ) 6160 { 6161 assert( pcCU->getDepth( 0 ) == pcCU->getDepth( uiAbsPartIdx ) ); 6162 const UInt uiCurrTrMode = uiDepth - pcCU->getDepth( 0 ); 6163 const UInt uiTrMode = pcCU->getTransformIdx( uiAbsPartIdx ); 6164 6165 const Bool bSubdiv = uiCurrTrMode != uiTrMode; 6166 6167 const UInt uiLog2TrSize = g_aucConvertToBit[pcCU->getSlice()->getSPS()->getMaxCUWidth() >> uiDepth]+2; 6168 6169 { 6170 if( bSubdivAndCbf && uiLog2TrSize <= pcCU->getSlice()->getSPS()->getQuadtreeTULog2MaxSize() && uiLog2TrSize > pcCU->getQuadtreeTULog2MinSizeInCU(uiAbsPartIdx) ) 6171 { 6172 #if TRANS_SPLIT_FLAG_CTX_REDUCTION 6173 m_pcEntropyCoder->encodeTransformSubdivFlag( bSubdiv, 5 - uiLog2TrSize ); 6174 #else 6175 m_pcEntropyCoder->encodeTransformSubdivFlag( bSubdiv, uiDepth ); 6176 #endif 6177 } 6178 } 6179 6180 { 6181 #if INTRA_BL 6182 #if NO_RESIDUAL_FLAG_FOR_BLPRED 6183 assert( !pcCU->isIntra(uiAbsPartIdx) || pcCU->isIntraBL(uiAbsPartIdx)); 6184 #else 6185 assert( !pcCU->isIntra(uiAbsPartIdx) ); 6186 #endif 6187 #else 6188 assert( pcCU->getPredictionMode(uiAbsPartIdx) != MODE_INTRA ); 6189 #endif 6190 if( bSubdivAndCbf ) 6191 { 6192 const Bool bFirstCbfOfCU = uiCurrTrMode == 0; 6193 if( bFirstCbfOfCU || uiLog2TrSize > 2 ) 6194 { 6195 if( bFirstCbfOfCU || pcCU->getCbf( uiAbsPartIdx, TEXT_CHROMA_U, uiCurrTrMode - 1 ) ) 6196 { 6197 m_pcEntropyCoder->encodeQtCbf( pcCU, uiAbsPartIdx, TEXT_CHROMA_U, uiCurrTrMode ); 6198 } 6199 if( bFirstCbfOfCU || pcCU->getCbf( uiAbsPartIdx, TEXT_CHROMA_V, uiCurrTrMode - 1 ) ) 6200 { 6201 m_pcEntropyCoder->encodeQtCbf( pcCU, uiAbsPartIdx, TEXT_CHROMA_V, uiCurrTrMode ); 6202 } 6203 } 6204 else if( uiLog2TrSize == 2 ) 6205 { 6206 assert( pcCU->getCbf( uiAbsPartIdx, TEXT_CHROMA_U, uiCurrTrMode ) == pcCU->getCbf( uiAbsPartIdx, TEXT_CHROMA_U, uiCurrTrMode - 1 ) ); 6207 assert( pcCU->getCbf( uiAbsPartIdx, TEXT_CHROMA_V, uiCurrTrMode ) == pcCU->getCbf( uiAbsPartIdx, TEXT_CHROMA_V, uiCurrTrMode - 1 ) ); 6208 } 6209 } 6210 } 6211 6212 if( !bSubdiv ) 6213 { 6214 const UInt uiNumCoeffPerAbsPartIdxIncrement = pcCU->getSlice()->getSPS()->getMaxCUWidth() * pcCU->getSlice()->getSPS()->getMaxCUHeight() >> ( pcCU->getSlice()->getSPS()->getMaxCUDepth() << 1 ); 6215 //assert( 16 == uiNumCoeffPerAbsPartIdxIncrement ); // check 6216 const UInt uiQTTempAccessLayer = pcCU->getSlice()->getSPS()->getQuadtreeTULog2MaxSize() - uiLog2TrSize; 6217 TCoeff *pcCoeffCurrY = m_ppcQTTempCoeffY [uiQTTempAccessLayer] + uiNumCoeffPerAbsPartIdxIncrement * uiAbsPartIdx; 6218 TCoeff *pcCoeffCurrU = m_ppcQTTempCoeffCb[uiQTTempAccessLayer] + (uiNumCoeffPerAbsPartIdxIncrement * uiAbsPartIdx>>2); 6219 TCoeff *pcCoeffCurrV = m_ppcQTTempCoeffCr[uiQTTempAccessLayer] + (uiNumCoeffPerAbsPartIdxIncrement * uiAbsPartIdx>>2); 6220 6221 Bool bCodeChroma = true; 6222 UInt uiTrModeC = uiTrMode; 6223 UInt uiLog2TrSizeC = uiLog2TrSize-1; 6224 if( uiLog2TrSize == 2 ) 6225 { 6226 uiLog2TrSizeC++; 6227 uiTrModeC --; 6228 UInt uiQPDiv = pcCU->getPic()->getNumPartInCU() >> ( ( pcCU->getDepth( 0 ) + uiTrModeC ) << 1 ); 6229 bCodeChroma = ( ( uiAbsPartIdx % uiQPDiv ) == 0 ); 6230 } 6231 6232 if( bSubdivAndCbf ) 6233 { 6234 { 6235 m_pcEntropyCoder->encodeQtCbf( pcCU, uiAbsPartIdx, TEXT_LUMA, uiTrMode ); 6236 } 6237 } 6238 else 6239 { 6240 if( eType == TEXT_LUMA && pcCU->getCbf( uiAbsPartIdx, TEXT_LUMA, uiTrMode ) ) 6241 { 6242 Int trWidth = 1 << uiLog2TrSize; 6243 Int trHeight = 1 << uiLog2TrSize; 6244 #if !REMOVE_NSQT 6245 pcCU->getNSQTSize( uiTrMode, uiAbsPartIdx, trWidth, trHeight ); 6246 #endif 6247 m_pcEntropyCoder->encodeCoeffNxN( pcCU, pcCoeffCurrY, uiAbsPartIdx, trWidth, trHeight, uiDepth, TEXT_LUMA ); 6248 } 6249 if( bCodeChroma ) 6250 { 6251 Int trWidth = 1 << uiLog2TrSizeC; 6252 Int trHeight = 1 << uiLog2TrSizeC; 6253 #if !REMOVE_NSQT 6254 pcCU->getNSQTSize( uiTrMode, uiAbsPartIdx, trWidth, trHeight ); 6255 #endif 6256 if( eType == TEXT_CHROMA_U && pcCU->getCbf( uiAbsPartIdx, TEXT_CHROMA_U, uiTrMode ) ) 6257 { 6258 m_pcEntropyCoder->encodeCoeffNxN( pcCU, pcCoeffCurrU, uiAbsPartIdx, trWidth, trHeight, uiDepth, TEXT_CHROMA_U ); 6259 } 6260 if( eType == TEXT_CHROMA_V && pcCU->getCbf( uiAbsPartIdx, TEXT_CHROMA_V, uiTrMode ) ) 6261 { 6262 m_pcEntropyCoder->encodeCoeffNxN( pcCU, pcCoeffCurrV, uiAbsPartIdx, trWidth, trHeight, uiDepth, TEXT_CHROMA_V ); 6263 } 6264 } 6265 } 6266 } 6267 else 6268 { 6269 if( bSubdivAndCbf || pcCU->getCbf( uiAbsPartIdx, eType, uiCurrTrMode ) ) 6270 { 6271 const UInt uiQPartNumSubdiv = pcCU->getPic()->getNumPartInCU() >> ((uiDepth + 1 ) << 1); 6272 for( UInt ui = 0; ui < 4; ++ui ) 6273 { 6274 xEncodeResidualQT( pcCU, uiAbsPartIdx + ui * uiQPartNumSubdiv, uiDepth + 1, bSubdivAndCbf, eType ); 6275 } 6276 } 6277 } 6278 } 6279 6280 Void TEncSearch::xSetResidualQTData( TComDataCU* pcCU, UInt uiQuadrant, UInt uiAbsPartIdx, UInt absTUPartIdx, TComYuv* pcResi, UInt uiDepth, Bool bSpatial ) 6281 { 6282 assert( pcCU->getDepth( 0 ) == pcCU->getDepth( uiAbsPartIdx ) ); 6283 const UInt uiCurrTrMode = uiDepth - pcCU->getDepth( 0 ); 6284 const UInt uiTrMode = pcCU->getTransformIdx( uiAbsPartIdx ); 6285 6286 if( uiCurrTrMode == uiTrMode ) 6287 { 6288 const UInt uiLog2TrSize = g_aucConvertToBit[pcCU->getSlice()->getSPS()->getMaxCUWidth() >> uiDepth]+2; 6289 const UInt uiQTTempAccessLayer = pcCU->getSlice()->getSPS()->getQuadtreeTULog2MaxSize() - uiLog2TrSize; 6290 6291 Bool bCodeChroma = true; 6292 UInt uiTrModeC = uiTrMode; 6293 UInt uiLog2TrSizeC = uiLog2TrSize-1; 6294 if( uiLog2TrSize == 2 ) 6295 { 6296 uiLog2TrSizeC++; 6297 uiTrModeC --; 6298 UInt uiQPDiv = pcCU->getPic()->getNumPartInCU() >> ( ( pcCU->getDepth( 0 ) + uiTrModeC ) << 1 ); 6299 bCodeChroma = ( ( uiAbsPartIdx % uiQPDiv ) == 0 ); 6300 } 6301 6302 if( bSpatial ) 6303 { 6304 Int trWidth = 1 << uiLog2TrSize; 6305 Int trHeight = 1 << uiLog2TrSize; 6306 #if !REMOVE_NSQT 6307 pcCU->getNSQTSize( uiTrMode, uiAbsPartIdx, trWidth, trHeight ); 6308 #endif 6309 m_pcQTTempTComYuv[uiQTTempAccessLayer].copyPartToPartLuma ( pcResi, absTUPartIdx, trWidth , trHeight ); 6310 6311 if( bCodeChroma ) 6312 { 6313 #if !REMOVE_NSQT 6314 Int trWidthC = 1 << uiLog2TrSizeC; 6315 Int trHeightC = 1 << uiLog2TrSizeC; 6316 UInt absTUPartIdxC = absTUPartIdx; 6317 pcCU->getNSQTSize( uiTrModeC, uiAbsPartIdx, trWidthC, trHeightC ); 6318 6319 if( pcCU->useNonSquareTrans( uiTrModeC, uiAbsPartIdx ) && !( uiLog2TrSizeC == pcCU->getSlice()->getSPS()->getQuadtreeTULog2MinSize() && uiTrModeC == 1 ) ) 6320 { 6321 absTUPartIdxC = pcCU->getNSAddrChroma( uiLog2TrSizeC, uiTrModeC, uiQuadrant, absTUPartIdx ); 6322 m_pcQTTempTComYuv[uiQTTempAccessLayer].copyPartToPartChroma( pcResi, absTUPartIdxC, trWidthC, trHeightC ); 6323 } 6324 else 6325 #endif 6326 { 6327 m_pcQTTempTComYuv[uiQTTempAccessLayer].copyPartToPartChroma( pcResi, uiAbsPartIdx, 1 << uiLog2TrSizeC, 1 << uiLog2TrSizeC ); 6328 } 6329 } 6330 } 6331 else 6332 { 6333 UInt uiNumCoeffPerAbsPartIdxIncrement = pcCU->getSlice()->getSPS()->getMaxCUWidth() * pcCU->getSlice()->getSPS()->getMaxCUHeight() >> ( pcCU->getSlice()->getSPS()->getMaxCUDepth() << 1 ); 6334 UInt uiNumCoeffY = ( 1 << ( uiLog2TrSize << 1 ) ); 6335 TCoeff* pcCoeffSrcY = m_ppcQTTempCoeffY [uiQTTempAccessLayer] + uiNumCoeffPerAbsPartIdxIncrement * uiAbsPartIdx; 6336 TCoeff* pcCoeffDstY = pcCU->getCoeffY() + uiNumCoeffPerAbsPartIdxIncrement * uiAbsPartIdx; 6337 ::memcpy( pcCoeffDstY, pcCoeffSrcY, sizeof( TCoeff ) * uiNumCoeffY ); 6338 #if ADAPTIVE_QP_SELECTION 6339 Int* pcArlCoeffSrcY = m_ppcQTTempArlCoeffY [uiQTTempAccessLayer] + uiNumCoeffPerAbsPartIdxIncrement * uiAbsPartIdx; 6340 Int* pcArlCoeffDstY = pcCU->getArlCoeffY() + uiNumCoeffPerAbsPartIdxIncrement * uiAbsPartIdx; 6341 ::memcpy( pcArlCoeffDstY, pcArlCoeffSrcY, sizeof( Int ) * uiNumCoeffY ); 6342 #endif 6343 if( bCodeChroma ) 6344 { 6345 UInt uiNumCoeffC = ( 1 << ( uiLog2TrSizeC << 1 ) ); 6346 TCoeff* pcCoeffSrcU = m_ppcQTTempCoeffCb[uiQTTempAccessLayer] + (uiNumCoeffPerAbsPartIdxIncrement * uiAbsPartIdx>>2); 6347 TCoeff* pcCoeffSrcV = m_ppcQTTempCoeffCr[uiQTTempAccessLayer] + (uiNumCoeffPerAbsPartIdxIncrement * uiAbsPartIdx>>2); 6348 TCoeff* pcCoeffDstU = pcCU->getCoeffCb() + (uiNumCoeffPerAbsPartIdxIncrement * uiAbsPartIdx>>2); 6349 TCoeff* pcCoeffDstV = pcCU->getCoeffCr() + (uiNumCoeffPerAbsPartIdxIncrement * uiAbsPartIdx>>2); 6350 ::memcpy( pcCoeffDstU, pcCoeffSrcU, sizeof( TCoeff ) * uiNumCoeffC ); 6351 ::memcpy( pcCoeffDstV, pcCoeffSrcV, sizeof( TCoeff ) * uiNumCoeffC ); 6352 #if ADAPTIVE_QP_SELECTION 6353 Int* pcArlCoeffSrcU = m_ppcQTTempArlCoeffCb[uiQTTempAccessLayer] + (uiNumCoeffPerAbsPartIdxIncrement * uiAbsPartIdx>>2); 6354 Int* pcArlCoeffSrcV = m_ppcQTTempArlCoeffCr[uiQTTempAccessLayer] + (uiNumCoeffPerAbsPartIdxIncrement * uiAbsPartIdx>>2); 6355 Int* pcArlCoeffDstU = pcCU->getArlCoeffCb() + (uiNumCoeffPerAbsPartIdxIncrement * uiAbsPartIdx>>2); 6356 Int* pcArlCoeffDstV = pcCU->getArlCoeffCr() + (uiNumCoeffPerAbsPartIdxIncrement * uiAbsPartIdx>>2); 6357 ::memcpy( pcArlCoeffDstU, pcArlCoeffSrcU, sizeof( Int ) * uiNumCoeffC ); 6358 ::memcpy( pcArlCoeffDstV, pcArlCoeffSrcV, sizeof( Int ) * uiNumCoeffC ); 6359 #endif 6360 } 6361 } 6362 } 6363 else 6364 { 6365 const UInt uiQPartNumSubdiv = pcCU->getPic()->getNumPartInCU() >> ((uiDepth + 1 ) << 1); 6366 #if !REMOVE_NSQT 6367 const UInt uiLog2TrSize = g_aucConvertToBit[pcCU->getSlice()->getSPS()->getMaxCUWidth() >> uiDepth] + 2; 6368 #endif 6369 for( UInt ui = 0; ui < 4; ++ui ) 6370 { 6371 #if REMOVE_NSQT 6372 UInt nsAddr = uiAbsPartIdx + ui * uiQPartNumSubdiv; 6373 #else 6374 UInt nsAddr = 0; 6375 nsAddr = pcCU->getNSAbsPartIdx( uiLog2TrSize-1, uiAbsPartIdx + ui * uiQPartNumSubdiv, absTUPartIdx, ui, uiCurrTrMode + 1); 6376 #endif 6377 xSetResidualQTData( pcCU, ui, uiAbsPartIdx + ui * uiQPartNumSubdiv, nsAddr, pcResi, uiDepth + 1, bSpatial ); 6378 } 6379 } 6380 } 6381 6382 UInt TEncSearch::xModeBitsIntra( TComDataCU* pcCU, UInt uiMode, UInt uiPU, UInt uiPartOffset, UInt uiDepth, UInt uiInitTrDepth ) 6383 { 6384 if( m_bUseSBACRD ) 6385 { 6386 // Reload only contexts required for coding intra mode information 6387 m_pcRDGoOnSbacCoder->loadIntraDirModeLuma( m_pppcRDSbacCoder[uiDepth][CI_CURR_BEST] ); 6388 } 6389 6390 pcCU->setLumaIntraDirSubParts ( uiMode, uiPartOffset, uiDepth + uiInitTrDepth ); 6391 6392 m_pcEntropyCoder->resetBits(); 6393 m_pcEntropyCoder->encodeIntraDirModeLuma ( pcCU, uiPartOffset); 6394 6395 return m_pcEntropyCoder->getNumberOfWrittenBits(); 6396 } 6397 6398 UInt TEncSearch::xUpdateCandList( UInt uiMode, Double uiCost, UInt uiFastCandNum, UInt * CandModeList, Double * CandCostList ) 6399 { 6400 UInt i; 6401 UInt shift=0; 6402 6403 while ( shift<uiFastCandNum && uiCost<CandCostList[ uiFastCandNum-1-shift ] ) shift++; 6404 6405 if( shift!=0 ) 6406 { 6407 for(i=1; i<shift; i++) 6408 { 6409 CandModeList[ uiFastCandNum-i ] = CandModeList[ uiFastCandNum-1-i ]; 6410 CandCostList[ uiFastCandNum-i ] = CandCostList[ uiFastCandNum-1-i ]; 6411 } 6412 CandModeList[ uiFastCandNum-shift ] = uiMode; 6413 CandCostList[ uiFastCandNum-shift ] = uiCost; 6414 return 1; 6415 } 6416 6417 return 0; 6418 } 6419 6420 /** add inter-prediction syntax elements for a CU block 6421 * \param pcCU 6422 * \param uiQp 6423 * \param uiTrMode 6424 * \param ruiBits 6425 * \param rpcYuvRec 6426 * \param pcYuvPred 6427 * \param rpcYuvResi 6428 * \returns Void 6429 */ 6430 Void TEncSearch::xAddSymbolBitsInter( TComDataCU* pcCU, UInt uiQp, UInt uiTrMode, UInt& ruiBits, TComYuv*& rpcYuvRec, TComYuv*pcYuvPred, TComYuv*& rpcYuvResi ) 6431 { 6432 #if SKIP_FLAG 6433 if(pcCU->getMergeFlag( 0 ) && pcCU->getPartitionSize( 0 ) == SIZE_2Nx2N && !pcCU->getQtRootCbf( 0 )) 6434 #else 6435 if ( pcCU->isSkipped( 0 ) ) 6436 #endif 6437 { 6438 #if SKIP_FLAG 6439 pcCU->setSkipFlagSubParts( true, 0, pcCU->getDepth(0) ); 6440 #endif 6441 6442 m_pcEntropyCoder->resetBits(); 6443 if(pcCU->getSlice()->getPPS()->getTransquantBypassEnableFlag()) 6444 { 6445 m_pcEntropyCoder->encodeCUTransquantBypassFlag(pcCU, 0, true); 6446 } 6447 m_pcEntropyCoder->encodeSkipFlag(pcCU, 0, true); 6448 m_pcEntropyCoder->encodeMergeIndex(pcCU, 0, 0, true); 6449 ruiBits += m_pcEntropyCoder->getNumberOfWrittenBits(); 6450 } 6451 else 6452 { 6453 m_pcEntropyCoder->resetBits(); 6454 if(pcCU->getSlice()->getPPS()->getTransquantBypassEnableFlag()) 6455 { 6456 m_pcEntropyCoder->encodeCUTransquantBypassFlag(pcCU, 0, true); 6457 } 6458 m_pcEntropyCoder->encodeSkipFlag ( pcCU, 0, true ); 6459 #if INTRA_BL 6460 if(m_pcEncCfg->getLayerId()) 6461 { 6462 m_pcEntropyCoder->encodeIntraBLFlag(pcCU, 0, true); 6463 #if !NO_RESIDUAL_FLAG_FOR_BLPRED 6464 assert( pcCU->isIntraBL( 0 ) == false ); 6465 #endif 6466 } 6467 #endif 6468 #if NO_RESIDUAL_FLAG_FOR_BLPRED 6469 if( !pcCU->isIntraBL(0)) 6470 { 6471 #endif 6472 m_pcEntropyCoder->encodePredMode( pcCU, 0, true ); 6473 m_pcEntropyCoder->encodePartSize( pcCU, 0, pcCU->getDepth(0), true ); 6474 m_pcEntropyCoder->encodePredInfo( pcCU, 0, true ); 6475 #if NO_RESIDUAL_FLAG_FOR_BLPRED 6476 } 6477 #endif 6478 Bool bDummy = false; 6479 m_pcEntropyCoder->encodeCoeff ( pcCU, 0, pcCU->getDepth(0), pcCU->getWidth(0), pcCU->getHeight(0), bDummy ); 6480 6481 ruiBits += m_pcEntropyCoder->getNumberOfWrittenBits(); 6482 } 6483 } 6484 6485 /** 6486 * \brief Generate half-sample interpolated block 6487 * 6488 * \param pattern Reference picture ROI 6489 * \param biPred Flag indicating whether block is for biprediction 6490 */ 6491 Void TEncSearch::xExtDIFUpSamplingH( TComPattern* pattern, Bool biPred ) 6492 { 6493 Int width = pattern->getROIYWidth(); 6494 Int height = pattern->getROIYHeight(); 6495 Int srcStride = pattern->getPatternLStride(); 6496 6497 Int intStride = m_filteredBlockTmp[0].getStride(); 6498 Int dstStride = m_filteredBlock[0][0].getStride(); 6499 Short *intPtr; 6500 Short *dstPtr; 6501 Int filterSize = NTAPS_LUMA; 6502 Int halfFilterSize = (filterSize>>1); 6503 Pel *srcPtr = pattern->getROIY() - halfFilterSize*srcStride - 1; 6504 6505 m_if.filterHorLuma(srcPtr, srcStride, m_filteredBlockTmp[0].getLumaAddr(), intStride, width+1, height+filterSize, 0, false); 6506 m_if.filterHorLuma(srcPtr, srcStride, m_filteredBlockTmp[2].getLumaAddr(), intStride, width+1, height+filterSize, 2, false); 6507 6508 intPtr = m_filteredBlockTmp[0].getLumaAddr() + halfFilterSize * intStride + 1; 6509 dstPtr = m_filteredBlock[0][0].getLumaAddr(); 6510 m_if.filterVerLuma(intPtr, intStride, dstPtr, dstStride, width+0, height+0, 0, false, true); 6511 6512 intPtr = m_filteredBlockTmp[0].getLumaAddr() + (halfFilterSize-1) * intStride + 1; 6513 dstPtr = m_filteredBlock[2][0].getLumaAddr(); 6514 m_if.filterVerLuma(intPtr, intStride, dstPtr, dstStride, width+0, height+1, 2, false, true); 6515 6516 intPtr = m_filteredBlockTmp[2].getLumaAddr() + halfFilterSize * intStride; 6517 dstPtr = m_filteredBlock[0][2].getLumaAddr(); 6518 m_if.filterVerLuma(intPtr, intStride, dstPtr, dstStride, width+1, height+0, 0, false, true); 6519 6520 intPtr = m_filteredBlockTmp[2].getLumaAddr() + (halfFilterSize-1) * intStride; 6521 dstPtr = m_filteredBlock[2][2].getLumaAddr(); 6522 m_if.filterVerLuma(intPtr, intStride, dstPtr, dstStride, width+1, height+1, 2, false, true); 6523 } 6524 6525 /** 6526 * \brief Generate quarter-sample interpolated blocks 6527 * 6528 * \param pattern Reference picture ROI 6529 * \param halfPelRef Half-pel mv 6530 * \param biPred Flag indicating whether block is for biprediction 6531 */ 6532 Void TEncSearch::xExtDIFUpSamplingQ( TComPattern* pattern, TComMv halfPelRef, Bool biPred ) 6533 { 6534 Int width = pattern->getROIYWidth(); 6535 Int height = pattern->getROIYHeight(); 6536 Int srcStride = pattern->getPatternLStride(); 6537 6538 Pel *srcPtr; 6539 Int intStride = m_filteredBlockTmp[0].getStride(); 6540 Int dstStride = m_filteredBlock[0][0].getStride(); 6541 Short *intPtr; 6542 Short *dstPtr; 6543 Int filterSize = NTAPS_LUMA; 6544 6545 Int halfFilterSize = (filterSize>>1); 6546 6547 Int extHeight = (halfPelRef.getVer() == 0) ? height + filterSize : height + filterSize-1; 6548 6549 // Horizontal filter 1/4 6550 srcPtr = pattern->getROIY() - halfFilterSize * srcStride - 1; 6551 intPtr = m_filteredBlockTmp[1].getLumaAddr(); 6552 if (halfPelRef.getVer() > 0) 6553 { 6554 srcPtr += srcStride; 6555 } 6556 if (halfPelRef.getHor() >= 0) 6557 { 6558 srcPtr += 1; 6559 } 6560 m_if.filterHorLuma(srcPtr, srcStride, intPtr, intStride, width, extHeight, 1, false); 6561 6562 // Horizontal filter 3/4 6563 srcPtr = pattern->getROIY() - halfFilterSize*srcStride - 1; 6564 intPtr = m_filteredBlockTmp[3].getLumaAddr(); 6565 if (halfPelRef.getVer() > 0) 6566 { 6567 srcPtr += srcStride; 6568 } 6569 if (halfPelRef.getHor() > 0) 6570 { 6571 srcPtr += 1; 6572 } 6573 m_if.filterHorLuma(srcPtr, srcStride, intPtr, intStride, width, extHeight, 3, false); 6574 6575 // Generate @ 1,1 6576 intPtr = m_filteredBlockTmp[1].getLumaAddr() + (halfFilterSize-1) * intStride; 6577 dstPtr = m_filteredBlock[1][1].getLumaAddr(); 6578 if (halfPelRef.getVer() == 0) 6579 { 6580 intPtr += intStride; 6581 } 6582 m_if.filterVerLuma(intPtr, intStride, dstPtr, dstStride, width, height, 1, false, true); 6583 6584 // Generate @ 3,1 6585 intPtr = m_filteredBlockTmp[1].getLumaAddr() + (halfFilterSize-1) * intStride; 6586 dstPtr = m_filteredBlock[3][1].getLumaAddr(); 6587 m_if.filterVerLuma(intPtr, intStride, dstPtr, dstStride, width, height, 3, false, true); 6588 6589 if (halfPelRef.getVer() != 0) 6590 { 6591 // Generate @ 2,1 6592 intPtr = m_filteredBlockTmp[1].getLumaAddr() + (halfFilterSize-1) * intStride; 6593 dstPtr = m_filteredBlock[2][1].getLumaAddr(); 6594 if (halfPelRef.getVer() == 0) 6595 { 6596 intPtr += intStride; 6597 } 6598 m_if.filterVerLuma(intPtr, intStride, dstPtr, dstStride, width, height, 2, false, true); 6599 6600 // Generate @ 2,3 6601 intPtr = m_filteredBlockTmp[3].getLumaAddr() + (halfFilterSize-1) * intStride; 6602 dstPtr = m_filteredBlock[2][3].getLumaAddr(); 6603 if (halfPelRef.getVer() == 0) 6604 { 6605 intPtr += intStride; 6606 } 6607 m_if.filterVerLuma(intPtr, intStride, dstPtr, dstStride, width, height, 2, false, true); 6608 } 6609 else 6610 { 6611 // Generate @ 0,1 6612 intPtr = m_filteredBlockTmp[1].getLumaAddr() + halfFilterSize * intStride; 6613 dstPtr = m_filteredBlock[0][1].getLumaAddr(); 6614 m_if.filterVerLuma(intPtr, intStride, dstPtr, dstStride, width, height, 0, false, true); 6615 6616 // Generate @ 0,3 6617 intPtr = m_filteredBlockTmp[3].getLumaAddr() + halfFilterSize * intStride; 6618 dstPtr = m_filteredBlock[0][3].getLumaAddr(); 6619 m_if.filterVerLuma(intPtr, intStride, dstPtr, dstStride, width, height, 0, false, true); 6620 } 6621 6622 if (halfPelRef.getHor() != 0) 6623 { 6624 // Generate @ 1,2 6625 intPtr = m_filteredBlockTmp[2].getLumaAddr() + (halfFilterSize-1) * intStride; 6626 dstPtr = m_filteredBlock[1][2].getLumaAddr(); 6627 if (halfPelRef.getHor() > 0) 6628 { 6629 intPtr += 1; 6630 } 6631 if (halfPelRef.getVer() >= 0) 6632 { 6633 intPtr += intStride; 6634 } 6635 m_if.filterVerLuma(intPtr, intStride, dstPtr, dstStride, width, height, 1, false, true); 6636 6637 // Generate @ 3,2 6638 intPtr = m_filteredBlockTmp[2].getLumaAddr() + (halfFilterSize-1) * intStride; 6639 dstPtr = m_filteredBlock[3][2].getLumaAddr(); 6640 if (halfPelRef.getHor() > 0) 6641 { 6642 intPtr += 1; 6643 } 6644 if (halfPelRef.getVer() > 0) 6645 { 6646 intPtr += intStride; 6647 } 6648 m_if.filterVerLuma(intPtr, intStride, dstPtr, dstStride, width, height, 3, false, true); 6649 } 6650 else 6651 { 6652 // Generate @ 1,0 6653 intPtr = m_filteredBlockTmp[0].getLumaAddr() + (halfFilterSize-1) * intStride + 1; 6654 dstPtr = m_filteredBlock[1][0].getLumaAddr(); 6655 if (halfPelRef.getVer() >= 0) 6656 { 6657 intPtr += intStride; 6658 } 6659 m_if.filterVerLuma(intPtr, intStride, dstPtr, dstStride, width, height, 1, false, true); 6660 6661 // Generate @ 3,0 6662 intPtr = m_filteredBlockTmp[0].getLumaAddr() + (halfFilterSize-1) * intStride + 1; 6663 dstPtr = m_filteredBlock[3][0].getLumaAddr(); 6664 if (halfPelRef.getVer() > 0) 6665 { 6666 intPtr += intStride; 6667 } 6668 m_if.filterVerLuma(intPtr, intStride, dstPtr, dstStride, width, height, 3, false, true); 6669 } 6670 6671 // Generate @ 1,3 6672 intPtr = m_filteredBlockTmp[3].getLumaAddr() + (halfFilterSize-1) * intStride; 6673 dstPtr = m_filteredBlock[1][3].getLumaAddr(); 6674 if (halfPelRef.getVer() == 0) 6675 { 6676 intPtr += intStride; 6677 } 6678 m_if.filterVerLuma(intPtr, intStride, dstPtr, dstStride, width, height, 1, false, true); 6679 6680 // Generate @ 3,3 6681 intPtr = m_filteredBlockTmp[3].getLumaAddr() + (halfFilterSize-1) * intStride; 6682 dstPtr = m_filteredBlock[3][3].getLumaAddr(); 6683 m_if.filterVerLuma(intPtr, intStride, dstPtr, dstStride, width, height, 3, false, true); 6684 } 6685 6686 /** set wp tables 6687 * \param TComDataCU* pcCU 6688 * \param iRefIdx 6689 * \param eRefPicListCur 6690 * \returns Void 6691 */ 6692 Void TEncSearch::setWpScalingDistParam( TComDataCU* pcCU, Int iRefIdx, RefPicList eRefPicListCur ) 6693 { 6694 if ( iRefIdx<0 ) 6695 { 6696 m_cDistParam.bApplyWeight = false; 6697 return; 6698 } 6699 6700 TComSlice *pcSlice = pcCU->getSlice(); 6701 TComPPS *pps = pcCU->getSlice()->getPPS(); 6702 wpScalingParam *wp0 , *wp1; 6703 m_cDistParam.bApplyWeight = ( pcSlice->getSliceType()==P_SLICE && pps->getUseWP() ) || ( pcSlice->getSliceType()==B_SLICE && pps->getWPBiPred() ) ; 6704 if ( !m_cDistParam.bApplyWeight ) return; 6705 6706 Int iRefIdx0 = ( eRefPicListCur == REF_PIC_LIST_0 ) ? iRefIdx : (-1); 6707 Int iRefIdx1 = ( eRefPicListCur == REF_PIC_LIST_1 ) ? iRefIdx : (-1); 6708 6709 getWpScaling( pcCU, iRefIdx0, iRefIdx1, wp0 , wp1 ); 6710 6711 if ( iRefIdx0 < 0 ) wp0 = NULL; 6712 if ( iRefIdx1 < 0 ) wp1 = NULL; 6713 6714 m_cDistParam.wpCur = NULL; 6715 6716 if ( eRefPicListCur == REF_PIC_LIST_0 ) 6717 { 6718 m_cDistParam.wpCur = wp0; 6719 } 6720 else 6721 { 6722 m_cDistParam.wpCur = wp1; 6723 } 6724 } 6725 6375 #endif 6726 6376 //! \}
Note: See TracChangeset for help on using the changeset viewer.
