[x265] [PATCH RFC] psyrd: bug fix in use of psyrdcost for PU/TU decision in inter
Steve Borho
steve at borho.org
Thu Jun 12 07:32:32 CEST 2014
On Wed, Jun 11, 2014 at 6:13 AM, <sumalatha at multicorewareinc.com> wrote:
> # HG changeset patch
> # User Sumalatha Polureddy<sumalatha at multicorewareinc.com>
> # Date 1402483864 -19800
> # Node ID a407bf9959c957d8a3c355594a060abab8a182fc
> # Parent dbe573edb3459f57ab058318e4227f6b19045c2d
> psyrd: bug fix in use of psyrdcost for PU/TU decision in inter
please fold these two patches together and re-submit
> diff -r dbe573edb345 -r a407bf9959c9 source/Lib/TLibEncoder/TEncSearch.cpp
> --- a/source/Lib/TLibEncoder/TEncSearch.cpp Tue Jun 10 17:47:15 2014 +0530
> +++ b/source/Lib/TLibEncoder/TEncSearch.cpp Wed Jun 11 16:21:04 2014 +0530
> @@ -2777,7 +2777,18 @@
> m_entropyCoder->resetBits();
> m_entropyCoder->encodeQtRootCbfZero(cu);
> uint32_t zeroResiBits = m_entropyCoder->getNumberOfWrittenBits();
> - uint64_t zeroCost = m_rdCost->calcRdCost(zeroDistortion, zeroResiBits);
> + uint64_t zeroCost = 0;
> + uint32_t zeroPsyEnergyY = 0;
> + if (m_rdCost->psyRdEnabled())
> + {
> + int size = g_convertToBit[cuSize];
> + zeroPsyEnergyY = m_rdCost->psyCost(size, fencYuv->getLumaAddr(), fencYuv->getStride(),
> + (pixel*)RDCost::zeroPel, MAX_CU_SIZE); // need to check whether zero distortion is similar to psyenergy of fenc
> + zeroCost = m_rdCost->calcPsyRdCost(zeroDistortion, zeroResiBits, zeroPsyEnergyY);
> + }
> + else
> + zeroCost = m_rdCost->calcRdCost(zeroDistortion, zeroResiBits);
> +
> if (cu->isLosslessCoded(0))
> {
> zeroCost = cost + 1;
> @@ -2785,6 +2796,7 @@
> if (zeroCost < cost)
> {
> distortion = zeroDistortion;
> + cu->m_psyEnergy = zeroPsyEnergyY;
>
> const uint32_t qpartnum = cu->getPic()->getNumPartInCU() >> (depth << 1);
> ::memset(cu->getTransformIdx(), 0, qpartnum * sizeof(uint8_t));
> @@ -2807,7 +2819,10 @@
>
> bits = xSymbolBitsInter(cu);
>
> - cost = m_rdCost->calcRdCost(distortion, bits);
> + if (m_rdCost->psyRdEnabled())
> + cost = m_rdCost->calcPsyRdCost(distortion, bits, cu->m_psyEnergy);
> + else
> + cost = m_rdCost->calcRdCost(distortion, bits);
>
> if (cost < bestCost)
> {
> @@ -3258,6 +3273,7 @@
> m_trQuant->invtransformNxN(cu->getCUTransquantBypass(absPartIdx), REG_DCT, curResiY, strideResiY, coeffCurY, trSize, scalingListType, false, lastPos[TEXT_LUMA][0]); //this is for inter mode only
>
> const uint32_t nonZeroDistY = primitives.sse_ss[partSize](resiYuv->getLumaAddr(absPartIdx), resiYuv->m_width, curResiY, strideResiY);
> + uint32_t nonZeroPsyEnergyY = 0;
> if (m_rdCost->psyRdEnabled())
> {
> pixel* pred = predYuv->getLumaAddr(absPartIdx);
> @@ -3267,22 +3283,20 @@
> uint32_t stride = fencYuv->getStride();
> //===== reconstruction =====
> primitives.luma_add_ps[partSize](reconIPred, reconIPredStride, pred, curResiY, stride, strideResiY);
> + int size = g_convertToBit[trSize];
> + nonZeroPsyEnergyY = m_rdCost->psyCost(size, fencYuv->getLumaAddr(absPartIdx), fencYuv->getStride(),
> + cu->getPic()->getPicYuvRec()->getLumaAddr(cu->getAddr(), zorder), cu->getPic()->getPicYuvRec()->getStride());
> }
> if (cu->isLosslessCoded(0))
> {
> distY = nonZeroDistY;
> + psyEnergyY = nonZeroPsyEnergyY;
> }
> else
> {
> uint64_t singleCostY = 0;
> if (m_rdCost->psyRdEnabled())
> - {
> - int size = g_convertToBit[trSize];
> - uint32_t zorder = cu->getZorderIdxInCU() + absPartIdx;
> - psyEnergyY = m_rdCost->psyCost(size, fencYuv->getLumaAddr(absPartIdx), fencYuv->getStride(),
> - cu->getPic()->getPicYuvRec()->getLumaAddr(cu->getAddr(), zorder), cu->getPic()->getPicYuvRec()->getStride());
> - singleCostY = m_rdCost->calcPsyRdCost(nonZeroDistY, singleBitsComp[TEXT_LUMA][0], psyEnergyY);
> - }
> + singleCostY = m_rdCost->calcPsyRdCost(nonZeroDistY, singleBitsComp[TEXT_LUMA][0], nonZeroPsyEnergyY);
> else
> singleCostY = m_rdCost->calcRdCost(nonZeroDistY, singleBitsComp[TEXT_LUMA][0]);
> m_entropyCoder->resetBits();
> @@ -3307,6 +3321,7 @@
> else
> {
> distY = nonZeroDistY;
> + psyEnergyY = nonZeroPsyEnergyY;
> if (checkTransformSkipY)
> {
> minCost[TEXT_LUMA][0] = singleCostY;
> @@ -3370,6 +3385,7 @@
> uint32_t dist = primitives.sse_ss[partSizeC](resiYuv->getCbAddr(absPartIdxC), resiYuv->m_cwidth,
> curResiU, strideResiC);
> const uint32_t nonZeroDistU = m_rdCost->scaleChromaDistCb(dist);
> + uint32_t nonZeroPsyEnergyU = 0;
> if (m_rdCost->psyRdEnabled())
> {
> pixel* pred = predYuv->getCbAddr(absPartIdxC);
> @@ -3379,21 +3395,21 @@
> uint32_t stride = fencYuv->getCStride();
> //===== reconstruction =====
> primitives.luma_add_ps[partSizeC](reconIPred, reconIPredStride, pred, curResiU, stride, strideResiC);
> + int size = g_convertToBit[trSizeC];
> + nonZeroPsyEnergyU = m_rdCost->psyCost(size, fencYuv->getCbAddr(absPartIdxC), fencYuv->getCStride(),
> + cu->getPic()->getPicYuvRec()->getCbAddr(cu->getAddr(), zorder), cu->getPic()->getPicYuvRec()->getCStride());
> }
> if (cu->isLosslessCoded(0))
> {
> distU = nonZeroDistU;
> + psyEnergyU = nonZeroPsyEnergyU;
> }
> else
> {
> uint64_t singleCostU = 0;
> if (m_rdCost->psyRdEnabled())
> {
> - int size = g_convertToBit[trSizeC];
> - uint32_t zorder = cu->getZorderIdxInCU() + absPartIdxC;
> - psyEnergyU = m_rdCost->psyCost(size, fencYuv->getCbAddr(absPartIdxC), fencYuv->getCStride(),
> - cu->getPic()->getPicYuvRec()->getCbAddr(cu->getAddr(), zorder), cu->getPic()->getPicYuvRec()->getCStride());
> - singleCostU = m_rdCost->calcPsyRdCost(nonZeroDistU, singleBitsComp[TEXT_CHROMA_U][0], psyEnergyU);
> + singleCostU = m_rdCost->calcPsyRdCost(nonZeroDistU, singleBitsComp[TEXT_CHROMA_U][0], nonZeroPsyEnergyU);
> }
> else
> singleCostU = m_rdCost->calcRdCost(nonZeroDistU, singleBitsComp[TEXT_CHROMA_U][0]);
> @@ -3419,6 +3435,7 @@
> else
> {
> distU = nonZeroDistU;
> + psyEnergyU = nonZeroPsyEnergyU;
> if (checkTransformSkipUV)
> {
> minCost[TEXT_CHROMA_U][tuIterator.m_section] = singleCostU;
> @@ -3462,6 +3479,7 @@
> uint32_t dist = primitives.sse_ss[partSizeC](resiYuv->getCrAddr(absPartIdxC), resiYuv->m_cwidth,
> curResiV, strideResiC);
> const uint32_t nonZeroDistV = m_rdCost->scaleChromaDistCr(dist);
> + uint32_t nonZeroPsyEnergyV = 0;
>
> if (m_rdCost->psyRdEnabled())
> {
> @@ -3472,21 +3490,21 @@
> uint32_t stride = fencYuv->getCStride();
> //===== reconstruction =====
> primitives.luma_add_ps[partSizeC](reconIPred, reconIPredStride, pred, curResiV, stride, strideResiC);
> + int size = g_convertToBit[trSizeC];
> + nonZeroPsyEnergyV = m_rdCost->psyCost(size, fencYuv->getCrAddr(absPartIdxC), fencYuv->getCStride(),
> + cu->getPic()->getPicYuvRec()->getCrAddr(cu->getAddr(), zorder), cu->getPic()->getPicYuvRec()->getCStride());
> }
> if (cu->isLosslessCoded(0))
> {
> distV = nonZeroDistV;
> + psyEnergyV = nonZeroPsyEnergyV;
> }
> else
> {
> uint64_t singleCostV = 0;
> if (m_rdCost->psyRdEnabled())
> {
> - int size = g_convertToBit[trSizeC];
> - uint32_t zorder = cu->getZorderIdxInCU() + absPartIdxC;
> - psyEnergyV = m_rdCost->psyCost(size, fencYuv->getCrAddr(absPartIdxC), fencYuv->getCStride(),
> - cu->getPic()->getPicYuvRec()->getCrAddr(cu->getAddr(), zorder), cu->getPic()->getPicYuvRec()->getCStride());
> - singleCostV = m_rdCost->calcPsyRdCost(nonZeroDistV, singleBitsComp[TEXT_CHROMA_V][tuIterator.m_section], psyEnergyV);
> + singleCostV = m_rdCost->calcPsyRdCost(nonZeroDistV, singleBitsComp[TEXT_CHROMA_V][tuIterator.m_section], nonZeroPsyEnergyV);
> }
> else
> singleCostV = m_rdCost->calcRdCost(nonZeroDistV, singleBitsComp[TEXT_CHROMA_V][tuIterator.m_section]);
> @@ -3512,6 +3530,7 @@
> else
> {
> distV = nonZeroDistV;
> + psyEnergyV = nonZeroPsyEnergyV;
> if (checkTransformSkipUV)
> {
> minCost[TEXT_CHROMA_V][tuIterator.m_section] = singleCostV;
> @@ -3548,6 +3567,7 @@
> if (checkTransformSkipY)
> {
> uint32_t nonZeroDistY = 0, absSumTransformSkipY;
> + uint32_t nonZeroPsyEnergyY = 0;
> uint64_t singleCostY = MAX_INT64;
>
> coeff_t bestCoeffY[MAX_TS_SIZE * MAX_TS_SIZE];
> @@ -3599,9 +3619,9 @@
> //===== reconstruction =====
> primitives.luma_add_ps[partSize](reconIPred, reconIPredStride, pred, curResiY, stride, strideResiY);
> int size = g_convertToBit[trSize];
> - psyEnergyY = m_rdCost->psyCost(size, fencYuv->getLumaAddr(absPartIdx), fencYuv->getStride(),
> + nonZeroPsyEnergyY = m_rdCost->psyCost(size, fencYuv->getLumaAddr(absPartIdx), fencYuv->getStride(),
> cu->getPic()->getPicYuvRec()->getLumaAddr(cu->getAddr(), zorder), cu->getPic()->getPicYuvRec()->getStride());
> - singleCostY = m_rdCost->calcPsyRdCost(nonZeroDistY, skipSingleBitsY, psyEnergyY);
> + singleCostY = m_rdCost->calcPsyRdCost(nonZeroDistY, skipSingleBitsY, nonZeroPsyEnergyY);
> }
> else
> singleCostY = m_rdCost->calcRdCost(nonZeroDistY, skipSingleBitsY);
> @@ -3616,7 +3636,7 @@
> else
> {
> singleDistComp[TEXT_LUMA][0] = nonZeroDistY;
> - singlePsyEnergyComp[TEXT_LUMA][0] = psyEnergyY;
> + singlePsyEnergyComp[TEXT_LUMA][0] = nonZeroPsyEnergyY;
> absSum[TEXT_LUMA][0] = absSumTransformSkipY;
> bestTransformMode[TEXT_LUMA][0] = 1;
> }
> @@ -3627,6 +3647,7 @@
> if (bCodeChroma && checkTransformSkipUV)
> {
> uint32_t nonZeroDistU = 0, nonZeroDistV = 0, absSumTransformSkipU, absSumTransformSkipV;
> + uint32_t nonZeroPsyEnergyU = 0, nonZeroPsyEnergyV = 0;
> uint64_t singleCostU = MAX_INT64;
> uint64_t singleCostV = MAX_INT64;
>
> @@ -3704,9 +3725,9 @@
> //===== reconstruction =====
> primitives.luma_add_ps[partSizeC](reconIPred, reconIPredStride, pred, curResiU, stride, strideResiC);
> int size = g_convertToBit[trSizeC];
> - psyEnergyU = m_rdCost->psyCost(size, fencYuv->getCbAddr(absPartIdxC), fencYuv->getCStride(),
> + nonZeroPsyEnergyU = m_rdCost->psyCost(size, fencYuv->getCbAddr(absPartIdxC), fencYuv->getCStride(),
> cu->getPic()->getPicYuvRec()->getCbAddr(cu->getAddr(), zorder), cu->getPic()->getPicYuvRec()->getCStride());
> - singleCostU = m_rdCost->calcPsyRdCost(nonZeroDistU, singleBitsComp[TEXT_CHROMA_U][tuIterator.m_section], psyEnergyU);
> + singleCostU = m_rdCost->calcPsyRdCost(nonZeroDistU, singleBitsComp[TEXT_CHROMA_U][tuIterator.m_section], nonZeroPsyEnergyU);
> }
> else
> singleCostU = m_rdCost->calcRdCost(nonZeroDistU, singleBitsComp[TEXT_CHROMA_U][tuIterator.m_section]);
> @@ -3722,7 +3743,7 @@
> else
> {
> singleDistComp[TEXT_CHROMA_U][tuIterator.m_section] = nonZeroDistU;
> - singlePsyEnergyComp[TEXT_CHROMA_U][tuIterator.m_section] = psyEnergyU;
> + singlePsyEnergyComp[TEXT_CHROMA_U][tuIterator.m_section] = nonZeroPsyEnergyU;
> absSum[TEXT_CHROMA_U][tuIterator.m_section] = absSumTransformSkipU;
> bestTransformMode[TEXT_CHROMA_U][tuIterator.m_section] = 1;
> }
> @@ -3743,7 +3764,6 @@
> uint32_t dist = primitives.sse_ss[partSizeC](resiYuv->getCrAddr(absPartIdxC), resiYuv->m_cwidth,
> curResiV, strideResiC);
> nonZeroDistV = m_rdCost->scaleChromaDistCr(dist);
> - singleCostV = m_rdCost->calcRdCost(nonZeroDistV, singleBitsComp[TEXT_CHROMA_V][tuIterator.m_section]);
> if (m_rdCost->psyRdEnabled())
> {
> pixel* pred = predYuv->getCrAddr(absPartIdxC);
> @@ -3754,9 +3774,9 @@
> //===== reconstruction =====
> primitives.luma_add_ps[partSizeC](reconIPred, reconIPredStride, pred, curResiV, stride, strideResiC);
> int size = g_convertToBit[trSizeC];
> - psyEnergyV = m_rdCost->psyCost(size, fencYuv->getCrAddr(absPartIdxC), fencYuv->getCStride(),
> + nonZeroPsyEnergyV = m_rdCost->psyCost(size, fencYuv->getCrAddr(absPartIdxC), fencYuv->getCStride(),
> cu->getPic()->getPicYuvRec()->getCrAddr(cu->getAddr(), zorder), cu->getPic()->getPicYuvRec()->getCStride());
> - singleCostV = m_rdCost->calcPsyRdCost(nonZeroDistV, singleBitsComp[TEXT_CHROMA_V][tuIterator.m_section], psyEnergyV);
> + singleCostV = m_rdCost->calcPsyRdCost(nonZeroDistV, singleBitsComp[TEXT_CHROMA_V][tuIterator.m_section], nonZeroPsyEnergyV);
> }
> else
> singleCostV = m_rdCost->calcRdCost(nonZeroDistV, singleBitsComp[TEXT_CHROMA_V][tuIterator.m_section]);
> @@ -3771,7 +3791,7 @@
> else
> {
> singleDistComp[TEXT_CHROMA_V][tuIterator.m_section] = nonZeroDistV;
> - singlePsyEnergyComp[TEXT_CHROMA_V][tuIterator.m_section] = psyEnergyV;
> + singlePsyEnergyComp[TEXT_CHROMA_V][tuIterator.m_section] = nonZeroPsyEnergyV;
> absSum[TEXT_CHROMA_V][tuIterator.m_section] = absSumTransformSkipV;
> bestTransformMode[TEXT_CHROMA_V][tuIterator.m_section] = 1;
> }
> _______________________________________________
> x265-devel mailing list
> x265-devel at videolan.org
> https://mailman.videolan.org/listinfo/x265-devel
--
Steve Borho
More information about the x265-devel
mailing list