From 76b64c4fc49b3911b09844346524fc61233fc1c6 Mon Sep 17 00:00:00 2001 From: Takeshi Tsukuba Date: Thu, 3 Oct 2019 19:23:19 +0900 Subject: [PATCH] changes for P0272 --- source/Lib/CommonLib/ChromaFormat.h | 1 - source/Lib/CommonLib/Quant.cpp | 31 +++++++++++++--- source/Lib/CommonLib/QuantRDOQ.cpp | 30 +++++++++++++--- source/Lib/CommonLib/QuantRDOQ.h | 4 +++ source/Lib/CommonLib/TrQuant.cpp | 41 ++++++++++++++++++++-- source/Lib/CommonLib/TypeDef.h | 2 ++ source/Lib/CommonLib/dtrace_codingstruct.h | 19 ++++++++++ 7 files changed, 115 insertions(+), 13 deletions(-) diff --git a/source/Lib/CommonLib/ChromaFormat.h b/source/Lib/CommonLib/ChromaFormat.h index d6930043..d643fd8f 100644 --- a/source/Lib/CommonLib/ChromaFormat.h +++ b/source/Lib/CommonLib/ChromaFormat.h @@ -113,7 +113,6 @@ static inline int getTransformShift(const int channelBitDepth, const Size size, return maxLog2TrDynamicRange - channelBitDepth - ( ( floorLog2(size.width) + floorLog2(size.height) ) >> 1 ); } - //------------------------------------------------ #if !JVET_O0650_SIGNAL_CHROMAQP_MAPPING_TABLE static inline int getScaledChromaQP(int unscaledChromaQP, const ChromaFormat chFmt) diff --git a/source/Lib/CommonLib/Quant.cpp b/source/Lib/CommonLib/Quant.cpp index a34a4faf..6a4b5ae5 100644 --- a/source/Lib/CommonLib/Quant.cpp +++ b/source/Lib/CommonLib/Quant.cpp @@ -467,7 +467,11 @@ void Quant::dequant(const TransformUnit &tu, const int QP_rem = cQP.rem; #endif +#if SONY_SIMPLE_TS + const int rightShift = (IQUANT_SHIFT - ( (isTransformSkip? 0 : iTransformShift) + QP_per)) + (enableScalingLists ? LOG2_SCALING_LIST_NEUTRAL_VALUE : 0); +#else const int rightShift = (IQUANT_SHIFT - (iTransformShift + QP_per)) + (enableScalingLists ? LOG2_SCALING_LIST_NEUTRAL_VALUE : 0); +#endif if(enableScalingLists) { @@ -531,6 +535,7 @@ void Quant::dequant(const TransformUnit &tu, const Intermediate_Int iCoeffQ = (Intermediate_Int(clipQCoef) * scale + iAdd) >> rightShift; piCoef[n] = TCoeff(Clip3(transformMinimum,transformMaximum,iCoeffQ)); + } } else @@ -543,6 +548,7 @@ void Quant::dequant(const TransformUnit &tu, const Intermediate_Int iCoeffQ = (Intermediate_Int(clipQCoef) * scale) << leftShift; piCoef[n] = TCoeff(Clip3(transformMinimum,transformMaximum,iCoeffQ)); + } } } @@ -1033,16 +1039,20 @@ void Quant::quant(TransformUnit &tu, const ComponentID &compID, const CCoeffBuf #else const int defaultQuantisationCoefficient = g_quantScales[needSqrtAdjustment?1:0][cQP.rem]; #endif - int iTransformShift = getTransformShift(channelBitDepth, rect.size(), maxLog2TrDynamicRange) + ( needSqrtAdjustment?-1:0); + // + int iTransformShift = getTransformShift(channelBitDepth, rect.size(), maxLog2TrDynamicRange) + (needSqrtAdjustment ? -1 : 0); if (useTransformSkip && sps.getSpsRangeExtension().getExtendedPrecisionProcessingFlag()) { iTransformShift = std::max(0, iTransformShift); } - #if JVET_O0919_TS_MIN_QP +#if SONY_SIMPLE_TS + const int iQBits = QUANT_SHIFT + cQP.per(useTransformSkip) + (useTransformSkip ? 0 : iTransformShift); +#else const int iQBits = QUANT_SHIFT + cQP.per(useTransformSkip) + iTransformShift; +#endif #else const int iQBits = QUANT_SHIFT + cQP.per + iTransformShift; #endif @@ -1072,6 +1082,7 @@ void Quant::quant(TransformUnit &tu, const ComponentID &compID, const CCoeffBuf const TCoeff quantisedCoefficient = quantisedMagnitude * iSign; piQCoef.buf[uiBlockPos] = Clip3( entropyCodingMinimum, entropyCodingMaximum, quantisedCoefficient ); + } // for n if( tu.cu->bdpcmMode && isLuma(compID) ) { @@ -1090,6 +1101,7 @@ void Quant::quant(TransformUnit &tu, const ComponentID &compID, const CCoeffBuf bool Quant::xNeedRDOQ(TransformUnit &tu, const ComponentID &compID, const CCoeffBuf &pSrc, const QpParam &cQP) { + const SPS &sps = *tu.cs->sps; const CompArea &rect = tu.blocks[compID]; const uint32_t uiWidth = rect.width; @@ -1126,7 +1138,6 @@ bool Quant::xNeedRDOQ(TransformUnit &tu, const ComponentID &compID, const CCoeff const int defaultQuantisationCoefficient = g_quantScales[needSqrtAdjustment?1:0][cQP.rem]; #endif int iTransformShift = getTransformShift(channelBitDepth, rect.size(), maxLog2TrDynamicRange) + (needSqrtAdjustment?-1:0); - if (useTransformSkip && sps.getSpsRangeExtension().getExtendedPrecisionProcessingFlag()) { iTransformShift = std::max(0, iTransformShift); @@ -1194,7 +1205,11 @@ void Quant::transformSkipQuantOneSample(TransformUnit &tu, const ComponentID &co */ #if JVET_O0919_TS_MIN_QP +#if SONY_SIMPLE_TS + const int iQBits = QUANT_SHIFT + cQP.per(useTransformSkip) + (useTransformSkip ? 0 : iTransformShift); +#else const int iQBits = QUANT_SHIFT + cQP.per(useTransformSkip) + iTransformShift; +#endif #else const int iQBits = QUANT_SHIFT + cQP.per + iTransformShift; #endif @@ -1249,7 +1264,12 @@ void Quant::invTrSkipDeQuantOneSample(TransformUnit &tu, const ComponentID &comp CHECK(scalingListType >= SCALING_LIST_NUM, "Invalid scaling list"); +#if SONY_SIMPLE_TS + const bool isTransformSkip = tu.mtsIdx == MTS_SKIP && isLuma(compID); + const int rightShift = (IQUANT_SHIFT - ((isTransformSkip? 0: iTransformShift) + QP_per)) + (enableScalingLists ? LOG2_SCALING_LIST_NEUTRAL_VALUE : 0); +#else const int rightShift = (IQUANT_SHIFT - (iTransformShift + QP_per)) + (enableScalingLists ? LOG2_SCALING_LIST_NEUTRAL_VALUE : 0); +#endif const TCoeff transformMinimum = -(1 << maxLog2TrDynamicRange); const TCoeff transformMaximum = (1 << maxLog2TrDynamicRange) - 1; @@ -1315,7 +1335,9 @@ void Quant::invTrSkipDeQuantOneSample(TransformUnit &tu, const ComponentID &comp } // Inverse transform-skip - +#if SONY_SIMPLE_TS + reconSample = Pel(dequantisedSample); +#else if (iTransformShift >= 0) { const TCoeff offset = iTransformShift == 0 ? 0 : (1 << (iTransformShift - 1)); @@ -1326,6 +1348,7 @@ void Quant::invTrSkipDeQuantOneSample(TransformUnit &tu, const ComponentID &comp const int iTrShiftNeg = -iTransformShift; reconSample = Pel(dequantisedSample << iTrShiftNeg); } +#endif } diff --git a/source/Lib/CommonLib/QuantRDOQ.cpp b/source/Lib/CommonLib/QuantRDOQ.cpp index 3cc44627..25c20522 100644 --- a/source/Lib/CommonLib/QuantRDOQ.cpp +++ b/source/Lib/CommonLib/QuantRDOQ.cpp @@ -378,8 +378,6 @@ void QuantRDOQ::setScalingList(ScalingList *scalingList, const int maxLog2TrDyna } } - - double QuantRDOQ::xGetErrScaleCoeff(const bool needsSqrt2, SizeType width, SizeType height, int qp, const int maxLog2TrDynamicRange, const int channelBitDepth) { const int iTransformShift = getTransformShift(channelBitDepth, Size(width, height), maxLog2TrDynamicRange); @@ -1286,9 +1284,18 @@ void QuantRDOQ::xRateDistOptQuantTS( TransformUnit &tu, const ComponentID &compI const bool needsSqrt2Scale = TU::needsSqrt2Scale( tu, compID ); // should always be false - transform-skipped blocks don't require sqrt(2) compensation. #if JVET_O0919_TS_MIN_QP const bool isTransformSkip = tu.mtsIdx==MTS_SKIP && isLuma(compID); +#if SONY_SIMPLE_TS + const int qBits = QUANT_SHIFT + qp.per(isTransformSkip) + (isTransformSkip ? 0 : transformShift) + (needsSqrt2Scale ? -1 : 0); // Right shift of non-RDOQ quantizer; level = (coeff*uiQ + offset)>>q_bits +#else const int qBits = QUANT_SHIFT + qp.per(isTransformSkip) + transformShift + ( needsSqrt2Scale ? -1 : 0 ); // Right shift of non-RDOQ quantizer; level = (coeff*uiQ + offset)>>q_bits +#endif const int quantisationCoefficient = g_quantScales[needsSqrt2Scale?1:0][qp.rem(isTransformSkip)]; +#if SONY_SIMPLE_TS + const double errorScale = xGetErrScaleCoeff(TU::needsSqrt2Scale(tu, compID), width, height, qp.rem(isTransformSkip), maxLog2TrDynamicRange, channelBitDepth) * pow(2, 2*transformShift); +#else const double errorScale = xGetErrScaleCoeff( TU::needsSqrt2Scale( tu, compID ), width, height, qp.rem(isTransformSkip), maxLog2TrDynamicRange, channelBitDepth ); +#endif + #else const int qBits = QUANT_SHIFT + qp.per + transformShift + (needsSqrt2Scale?-1:0); // Right shift of non-RDOQ quantizer; level = (coeff*uiQ + offset)>>q_bits const int quantisationCoefficient = g_quantScales[needsSqrt2Scale?1:0][qp.rem]; @@ -1343,10 +1350,10 @@ void QuantRDOQ::xRateDistOptQuantTS( TransformUnit &tu, const ComponentID &compI uint32_t upAbsLevel = std::min(uint32_t(entropyCodingMaximum), downAbsLevel + 1); m_testedLevels = 0; - coeffLevels[m_testedLevels++] = roundAbsLevel; + coeffLevels[m_testedLevels++] = roundAbsLevel; // 0: round if (minAbsLevel != roundAbsLevel) - coeffLevels[m_testedLevels++] = minAbsLevel; + coeffLevels[m_testedLevels++] = minAbsLevel; //1: min int rightPixel, belowPixel, predPixel; @@ -1354,7 +1361,7 @@ void QuantRDOQ::xRateDistOptQuantTS( TransformUnit &tu, const ComponentID &compI predPixel = cctx.deriveModCoeff(rightPixel, belowPixel, upAbsLevel, 0); if (upAbsLevel != roundAbsLevel && upAbsLevel != minAbsLevel && predPixel == 1) - coeffLevels[m_testedLevels++] = upAbsLevel; + coeffLevels[m_testedLevels++] = upAbsLevel; // 2: up double dErr = double(levelDouble); coeffLevelError[0] = dErr * dErr * errorScale; @@ -1541,9 +1548,18 @@ void QuantRDOQ::forwardRDPCM( TransformUnit &tu, const ComponentID &compID, cons const bool needsSqrt2Scale = TU::needsSqrt2Scale(tu, compID); // should always be false - transform-skipped blocks don't require sqrt(2) compensation. #if JVET_O0919_TS_MIN_QP const bool isTransformSkip = tu.mtsIdx==MTS_SKIP && isLuma(compID); +#if SONY_SIMPLE_TS + const int qBits = QUANT_SHIFT + qp.per(isTransformSkip) + (isTransformSkip? 0 : transformShift) + (needsSqrt2Scale ? -1 : 0); // Right shift of non-RDOQ quantizer; level = (coeff*uiQ + offset)>>q_bits +#else const int qBits = QUANT_SHIFT + qp.per(isTransformSkip) + transformShift + ( needsSqrt2Scale ? -1 : 0 ); // Right shift of non-RDOQ quantizer; level = (coeff*uiQ + offset)>>q_bits +#endif const int quantisationCoefficient = g_quantScales[needsSqrt2Scale ? 1 : 0][qp.rem(isTransformSkip)]; +#if SONY_SIMPLE_TS + const double errorScale = xGetErrScaleCoeff(TU::needsSqrt2Scale(tu, compID), width, height, qp.rem(isTransformSkip), maxLog2TrDynamicRange, channelBitDepth) * pow(2, 2 * transformShift); +#else const double errorScale = xGetErrScaleCoeff(TU::needsSqrt2Scale(tu, compID), width, height, qp.rem(isTransformSkip), maxLog2TrDynamicRange, channelBitDepth); +#endif + #else const int qBits = QUANT_SHIFT + qp.per + transformShift + (needsSqrt2Scale ? -1 : 0); // Right shift of non-RDOQ quantizer; level = (coeff*uiQ + offset)>>q_bits const int quantisationCoefficient = g_quantScales[needsSqrt2Scale ? 1 : 0][qp.rem]; @@ -1552,7 +1568,11 @@ void QuantRDOQ::forwardRDPCM( TransformUnit &tu, const ComponentID &compID, cons TrQuantParams trQuantParams; #if JVET_O0919_TS_MIN_QP +#if SONY_SIMPLE_TS + trQuantParams.rightShift = (IQUANT_SHIFT - ((isTransformSkip ? 0 : transformShift) + qp.per(isTransformSkip))); +#else trQuantParams.rightShift = (IQUANT_SHIFT - (transformShift + qp.per(isTransformSkip))); +#endif trQuantParams.qScale = g_invQuantScales[needsSqrt2Scale ? 1 : 0][qp.rem(isTransformSkip)]; #else trQuantParams.rightShift = (IQUANT_SHIFT - (transformShift + qp.per)); diff --git a/source/Lib/CommonLib/QuantRDOQ.h b/source/Lib/CommonLib/QuantRDOQ.h index 0010d40c..23083b1d 100644 --- a/source/Lib/CommonLib/QuantRDOQ.h +++ b/source/Lib/CommonLib/QuantRDOQ.h @@ -69,7 +69,11 @@ public: private: double* xGetErrScaleCoeffSL ( uint32_t list, uint32_t sizeX, uint32_t sizeY, int qp ) { return m_errScale[sizeX][sizeY][list][qp]; }; //!< get Error Scale Coefficent +#if SONY_SIMPLTE_TS_FIX_RDOQ + double xGetErrScaleCoeff ( const bool needsSqrt2, SizeType width, SizeType height, int qp, const int maxLog2TrDynamicRange, const int channelBitDepth, const bool isTrSkip); +#else double xGetErrScaleCoeff ( const bool needsSqrt2, SizeType width, SizeType height, int qp, const int maxLog2TrDynamicRange, const int channelBitDepth); +#endif double& xGetErrScaleCoeffNoScalingList ( uint32_t list, uint32_t sizeX, uint32_t sizeY, int qp ) { return m_errScaleNoScalingList[sizeX][sizeY][list][qp]; }; //!< get Error Scale Coefficent void xInitScalingList ( const QuantRDOQ* other ); void xDestroyScalingList (); diff --git a/source/Lib/CommonLib/TrQuant.cpp b/source/Lib/CommonLib/TrQuant.cpp index e498dd0f..f85e1f93 100644 --- a/source/Lib/CommonLib/TrQuant.cpp +++ b/source/Lib/CommonLib/TrQuant.cpp @@ -1009,11 +1009,22 @@ void TrQuant::xITransformSkip(const CCoeffBuf &pCoeff, { iTransformShift = std::max( 0, iTransformShift ); } - +#if !SONY_SIMPLE_TS int iWHScale = 1; - +#endif const bool rotateResidual = TU::isNonTransformedResidualRotated( tu, compID ); +#if SONY_SIMPLE_TS + + for (uint32_t y = 0; y < height; y++) + { + for (uint32_t x = 0; x < width; x++) + { + pResidual.at(x, y) = Pel((rotateResidual ? pCoeff.at(pCoeff.width - x - 1, pCoeff.height - y - 1) : pCoeff.at(x, y))); + } + } + +#else if( iTransformShift >= 0 ) { const TCoeff offset = iTransformShift == 0 ? 0 : ( 1 << ( iTransformShift - 1 ) ); @@ -1038,6 +1049,7 @@ void TrQuant::xITransformSkip(const CCoeffBuf &pCoeff, } } } +#endif } void TrQuant::xQuant(TransformUnit &tu, const ComponentID &compID, const CCoeffBuf &pSrc, TCoeff &uiAbsSum, const QpParam &cQP, const Ctx& ctx) @@ -1100,6 +1112,18 @@ void TrQuant::transformNxN( TransformUnit &tu, const ComponentID &compID, const { scaleSAD=1.0/1.414213562; // compensate for not scaling transform skip coefficients by 1/sqrt(2) } +#if SONY_SIMPLE_TS + if (tu.mtsIdx == MTS_SKIP) { + int trShift = getTransformShift(tu.cu->slice->getSPS()->getBitDepth(toChannelType(compID)), rect.size(), tu.cu->slice->getSPS()->getMaxLog2TrDynamicRange(toChannelType(compID))); + if(trShift>=0){ + scaleSAD *= (double)(1<( 0, iTransformShift ); } +#if !SONY_SIMPLE_TS int iWHScale = 1; - +#endif const bool rotateResidual = TU::isNonTransformedResidualRotated( tu, compID ); const uint32_t uiSizeMinus1 = ( width * height ) - 1; +#if SONY_SIMPLE_TS + for (uint32_t y = 0, coefficientIndex = 0; y < height; y++) + { + for (uint32_t x = 0; x < width; x++, coefficientIndex++) + { + psCoeff[rotateResidual ? uiSizeMinus1 - coefficientIndex : coefficientIndex] = TCoeff(resi.at(x, y)); + } + } +#else if( iTransformShift >= 0 ) { for( uint32_t y = 0, coefficientIndex = 0; y < height; y++ ) @@ -1435,6 +1469,7 @@ void TrQuant::xTransformSkip(const TransformUnit &tu, const ComponentID &compID, } } } +#endif } //! \} diff --git a/source/Lib/CommonLib/TypeDef.h b/source/Lib/CommonLib/TypeDef.h index 88063fe2..7b592fbe 100644 --- a/source/Lib/CommonLib/TypeDef.h +++ b/source/Lib/CommonLib/TypeDef.h @@ -50,6 +50,8 @@ #include #include +#define SONY_SIMPLE_TS 1 + #define JVET_O1164_RPR 1 // JVET-O1164: Reference picture resampling #if JVET_O1164_RPR #define JVET_O1164_PS 1 diff --git a/source/Lib/CommonLib/dtrace_codingstruct.h b/source/Lib/CommonLib/dtrace_codingstruct.h index eafc916e..306d4bc6 100644 --- a/source/Lib/CommonLib/dtrace_codingstruct.h +++ b/source/Lib/CommonLib/dtrace_codingstruct.h @@ -97,6 +97,24 @@ inline void dtraceModeCost(CodingStructure &cs, double lambda) intraModeC = 68; int imvVal = 0; imvVal = cs.cus[0]->imv; +#if SONY_DBG_FIX_MODE_COST + DTRACE(g_trace_ctx, D_MODE_COST, "ModeCost: %6lld %3d @(%4d,%4d) [%2dx%2d] %d (qp%d,pm%d,skip%d,mrg%d,fruc%d,obmc%d,ic%d,imv%d,affn%d,intraModeL%d,intraModeC%d, mtsIdx%d) tempCS = %lld (%d,%d)\n", + DTRACE_GET_COUNTER(g_trace_ctx, D_MODE_COST), + cs.slice->getPOC(), + cs.area.Y().x, cs.area.Y().y, + cs.area.Y().width, cs.area.Y().height, + cs.cus[0]->qtDepth, + cs.cus[0]->qp, + cs.cus[0]->predMode, + cs.cus[0]->skip, + cs.pus[0]->mergeFlag, + 0, 0, + imvVal, + 0, 0, + intraModeL, intraModeC, + cs.tus[0]->mtsIdx, + tempCost, tempBits, tempDist); +#else DTRACE( g_trace_ctx, D_MODE_COST, "ModeCost: %6lld %3d @(%4d,%4d) [%2dx%2d] %d (qp%d,pm%d,skip%d,mrg%d,fruc%d,obmc%d,ic%d,imv%d,affn%d,%d,%d) tempCS = %lld (%d,%d)\n", DTRACE_GET_COUNTER( g_trace_ctx, D_MODE_COST ), cs.slice->getPOC(), @@ -112,6 +130,7 @@ inline void dtraceModeCost(CodingStructure &cs, double lambda) 0, 0, intraModeL, intraModeC, tempCost, tempBits, tempDist ); +#endif } inline void dtraceBestMode(CodingStructure *&tempCS, CodingStructure *&bestCS, double lambda) -- 2.17.0.windows.1