[x265] [PATCH] analysis: Intra picture estimation information sharing

Steve Borho steve at borho.org
Mon Sep 15 12:40:08 CEST 2014


On 09/15, gopu at multicorewareinc.com wrote:
> # HG changeset patch
> # User Gopu Govindaswamy <gopu at multicorewareinc.com>
> # Date 1410770251 -19800
> #      Mon Sep 15 14:07:31 2014 +0530
> # Node ID 9db768fa41ad927c66c1dc4ae446953862052ce4
> # Parent  184e56afa951815f4e295b4fcce094ee03361a2e
> analysis: Intra picture estimation information sharing
> 
> when --analysis-mode=save - the encoder runs a full encode and dump the
> best split and mode decisions into x265_analysis.dat(default file name if file
> name is not provided) file
> when --analysis-mode=load - the encoder reads the best split and mode decisions
> from x265_analysis.dat and bypass the actual split and mode decisions, and
> therefore perform a much faster encode
> 
> diff -r 184e56afa951 -r 9db768fa41ad source/Lib/TLibCommon/TComRom.cpp
> --- a/source/Lib/TLibCommon/TComRom.cpp	Fri Sep 12 12:02:46 2014 +0530
> +++ b/source/Lib/TLibCommon/TComRom.cpp	Mon Sep 15 14:07:31 2014 +0530
> @@ -505,5 +505,19 @@
>      0x38, 
>  };
>  
> +    /* Contains how much to increment shared depth buffer for different ctu sizes to get next best depth.
> +     * here,
> +     * depth 0 = 64x64, depth 1 = 32x32, depth 2 = 16x16 and depth 3 = 8x8
> +     * if ctu = 64, depth buffer size is 256 combination of depth values 0, 1, 2, 3.
> +     * if ctu = 32, depth buffer size is 64 combination of depth values 1, 2, 3.
> +     * if ctu = 16, depth buffer size is 16 combination of depth values 2, 3 */

the comment should be w/s aligned with the array, lines 2&3 should be
combined

> +const uint32_t g_depthInc[3][4] =
> +{
> +    { 16,  4,  0, 0},
> +    { 64, 16,  4, 1},
> +    {256, 64, 16, 4}
> +};
> +
>  }
>  //! \}
> diff -r 184e56afa951 -r 9db768fa41ad source/Lib/TLibCommon/TComRom.h
> --- a/source/Lib/TLibCommon/TComRom.h	Fri Sep 12 12:02:46 2014 +0530
> +++ b/source/Lib/TLibCommon/TComRom.h	Mon Sep 15 14:07:31 2014 +0530
> @@ -155,6 +155,8 @@
>  // Intra tables
>  extern const uint8_t g_intraFilterFlags[35];
>  
> +extern const uint32_t g_depthInc[3][4];
> +
>  }
>  
>  #endif  //ifndef X265_TCOMROM_H
> diff -r 184e56afa951 -r 9db768fa41ad source/encoder/analysis.cpp
> --- a/source/encoder/analysis.cpp	Fri Sep 12 12:02:46 2014 +0530
> +++ b/source/encoder/analysis.cpp	Mon Sep 15 14:07:31 2014 +0530
> @@ -311,14 +311,24 @@
>      uint32_t numPartition = cu->getTotalNumPart();
>      if (m_bestCU[0]->m_slice->m_sliceType == I_SLICE)
>      {
> -        compressIntraCU(m_bestCU[0], m_tempCU[0], false, cu, cu->m_CULocalData);
> -        if (m_param->analysisMode == 1)
> +        if (m_param->analysisMode == 2)

our code should always use the X265_ANALYSIS_LOAD|SAVE macros,
except when checking != 0.

>          {
> -            memcpy(&m_bestCU[0]->m_pic->m_intraData->depth[cu->getAddr() * cu->m_numPartitions], m_bestCU[0]->getDepth(), sizeof(uint8_t) * cu->getTotalNumPart());
> -            memcpy(&m_bestCU[0]->m_pic->m_intraData->modes[cu->getAddr() * cu->m_numPartitions], m_bestCU[0]->getLumaIntraDir(), sizeof(uint8_t) * cu->getTotalNumPart());
> -            memcpy(&m_bestCU[0]->m_pic->m_intraData->partSizes[cu->getAddr() * cu->m_numPartitions], m_bestCU[0]->getPartitionSize(), sizeof(char) * cu->getTotalNumPart());
> -            m_bestCU[0]->m_pic->m_intraData->cuAddr[cu->getAddr()] = cu->getAddr();
> -            m_bestCU[0]->m_pic->m_intraData->poc[cu->getAddr()]    = cu->m_pic->m_POC;
> +            sharedCompressIntraCU(m_bestCU[0], m_tempCU[0], false, cu, cu->m_CULocalData, 
> +                &m_bestCU[0]->m_pic->m_intraData->depth[cu->getAddr() * cu->m_numPartitions],
> +                &m_bestCU[0]->m_pic->m_intraData->partSizes[cu->getAddr() * cu->m_numPartitions],
> +                &m_bestCU[0]->m_pic->m_intraData->modes[cu->getAddr() * cu->m_numPartitions]);

Pointer checking needs to be done at some point, probably at the frame
level. If the user doesn't allocate a buffer, we shouldn't crash.

We should probably also be setting the analysis pointers to NULL in the
input picture structure prior to returning from x265_encoder_encode() so
they do not accidentally re-use the same buffers for more than one
picture.  In short, we need to be a lot more defensive about API abuses.

> +        }
> +        else
> +        {
> +            compressIntraCU(m_bestCU[0], m_tempCU[0], false, cu, cu->m_CULocalData);
> +            if (m_param->analysisMode == 1)
> +            {
> +                memcpy(&m_bestCU[0]->m_pic->m_intraData->depth[cu->getAddr() * cu->m_numPartitions], m_bestCU[0]->getDepth(), sizeof(uint8_t) * cu->getTotalNumPart());
> +                memcpy(&m_bestCU[0]->m_pic->m_intraData->modes[cu->getAddr() * cu->m_numPartitions], m_bestCU[0]->getLumaIntraDir(), sizeof(uint8_t) * cu->getTotalNumPart());
> +                memcpy(&m_bestCU[0]->m_pic->m_intraData->partSizes[cu->getAddr() * cu->m_numPartitions], m_bestCU[0]->getPartitionSize(), sizeof(char) * cu->getTotalNumPart());
> +                m_bestCU[0]->m_pic->m_intraData->cuAddr[cu->getAddr()] = cu->getAddr();
> +                m_bestCU[0]->m_pic->m_intraData->poc[cu->getAddr()]    = cu->m_pic->m_POC;
> +            }
>          }
>          if (m_param->bLogCuStats || m_param->rc.bStatWrite)
>          {
> @@ -533,7 +543,142 @@
>  #endif
>  }
>  
> -void Analysis::checkIntra(TComDataCU*& outBestCU, TComDataCU*& outTempCU, PartSize partSize, CU *cu)
> +void Analysis::sharedCompressIntraCU(TComDataCU*& outBestCU, TComDataCU*& outTempCU, uint32_t depth, TComDataCU* cuPicsym, CU *cu, uint8_t* sharedDepth, char* sharedPartSizes, uint8_t* sharedModes)
> +{
> +    Frame* pic = outBestCU->m_pic;
> +
> +    // if current depth == shared depth then skip further splitting.
> +    bool bSubBranch = true;
> +
> +    if (depth == 0)

!depth

> +    {
> +        // offset to next best depth in sharedDepth buffer
> +        m_zorder = 0;
> +
> +        // index to g_depthInc array to increment m_zorder offset to next depth
> +        m_ctuToDepthIndex = m_param->maxCUSize / 22;

this math is pretty magical. my guess is there's already a table
somewhere that does this more cleanly? Does this code work with
--ctu 16?

> +        // get original YUV data from picture
> +        m_origYuv[depth]->copyFromPicYuv(pic->getPicYuvOrg(), outBestCU->getAddr(), outBestCU->getZorderIdxInCU());
> +    }
> +    else
> +        m_origYuv[0]->copyPartToYuv(m_origYuv[depth], outBestCU->getZorderIdxInCU());
> +
> +    Slice* slice = outTempCU->m_slice;
> +    int32_t cu_split_flag = !(cu->flags & CU::LEAF);
> +    int32_t cu_unsplit_flag = !(cu->flags & CU::SPLIT_MANDATORY);

It looks like this function is recursively encoding the entire I slice
CTU. If that is the case the name should reflect that, perhaps
compressSharedIntraCTU.

> +    if (cu_unsplit_flag && ((m_zorder == outBestCU->getZorderIdxInCU()) && (depth == sharedDepth[m_zorder])))
> +    {
> +        m_quant.setQPforQuant(outTempCU);
> +        checkIntra(outBestCU, outTempCU, (PartSize)sharedPartSizes[m_zorder], cu, &sharedModes[m_zorder]);
> +
> +        if (!(depth == g_maxCUDepth))
> +        {
> +            m_entropyCoder->resetBits();
> +            m_entropyCoder->codeSplitFlag(outBestCU, 0, depth);
> +            outBestCU->m_totalBits += m_entropyCoder->getNumberOfWrittenBits();
> +        }
> +        if (m_rdCost.m_psyRd)
> +            outBestCU->m_totalPsyCost = m_rdCost.calcPsyRdCost(outBestCU->m_totalDistortion, outBestCU->m_totalBits, outBestCU->m_psyEnergy);
> +        else
> +            outBestCU->m_totalRDCost  = m_rdCost.calcRdCost(outBestCU->m_totalDistortion, outBestCU->m_totalBits);

How applicable is psy-rd for I slices in the shared re-use case?  Does
it influence splits or something? If it's not being used, we should
save the cycles

Should we be measuring cost at all in the reuse case?

> +        bSubBranch = false;
> +
> +        // increment m_zorder offset to point to next best depth in sharedDepth buffer
> +        m_zorder += g_depthInc[m_ctuToDepthIndex][sharedDepth[m_zorder]];
> +    }
> +
> +    // copy original YUV samples in lossless mode
> +    if (outBestCU->isLosslessCoded(0))
> +        fillOrigYUVBuffer(outBestCU, m_origYuv[depth]);
> +
> +    // further split
> +    if (cu_split_flag && bSubBranch)
> +    {
> +        uint32_t    nextDepth     = depth + 1;
> +        TComDataCU* subBestPartCU = m_bestCU[nextDepth];
> +        TComDataCU* subTempPartCU = m_tempCU[nextDepth];
> +        for (uint32_t partUnitIdx = 0; partUnitIdx < 4; partUnitIdx++)
> +        {
> +            CU *child_cu = cuPicsym->m_CULocalData + cu->childIdx + partUnitIdx;
> +
> +            if (child_cu->flags & CU::PRESENT)
> +            {
> +                int32_t qp = outTempCU->getQP(0);
> +                subBestPartCU->initSubCU(outTempCU, partUnitIdx, nextDepth, qp); // clear sub partition datas or init.
> +                subTempPartCU->initSubCU(outTempCU, partUnitIdx, nextDepth, qp); // clear sub partition datas or init.
> +                if (0 == partUnitIdx) //initialize RD with previous depth buffer
> +                    m_rdEntropyCoders[nextDepth][CI_CURR_BEST].load(m_rdEntropyCoders[depth][CI_CURR_BEST]);
> +                else
> +                    m_rdEntropyCoders[nextDepth][CI_CURR_BEST].load(m_rdEntropyCoders[nextDepth][CI_NEXT_BEST]);

we normally code this as:

if (partUnitIdx) // initialize RD with previous depth buffer
    m_rdEntropyCoders[nextDepth][CI_CURR_BEST].load(m_rdEntropyCoders[nextDepth][CI_NEXT_BEST]);
else
    m_rdEntropyCoders[nextDepth][CI_CURR_BEST].load(m_rdEntropyCoders[depth][CI_CURR_BEST]);

> +
> +                sharedCompressIntraCU(subBestPartCU, subTempPartCU, nextDepth, cuPicsym, child_cu, sharedDepth, sharedPartSizes, sharedModes);
> +                outTempCU->copyPartFrom(subBestPartCU, partUnitIdx, nextDepth); // Keep best part data to current temporary data.
> +
> +                // check if cost ==  MAX_INT64 then current depth != sharedDepth so, current CU is not best CU
> +                // set the cost to MAX_INT64 - 1 to mark it as not best CU
> +                if (m_rdCost.m_psyRd && subBestPartCU->m_totalPsyCost == MAX_INT64)
> +                    outTempCU->m_totalPsyCost = MAX_INT64 - 1;
> +                else if(subBestPartCU->m_totalRDCost == MAX_INT64)
> +                    outTempCU->m_totalRDCost = MAX_INT64 - 1;
> +
> +                copyYuv2Tmp(subBestPartCU->getTotalNumPart() * partUnitIdx, nextDepth);
> +            }
> +            else
> +            {
> +                subBestPartCU->copyToPic(nextDepth);
> +                outTempCU->copyPartFrom(subBestPartCU, partUnitIdx, nextDepth);
> +
> +                // increment m_zorder offset to point to next best depth in sharedDepth buffer
> +                m_zorder += g_depthInc[m_ctuToDepthIndex][sharedDepth[m_zorder]];
> +            }
> +        }
> +        if (cu->flags & CU::PRESENT)
> +        {
> +            m_entropyCoder->resetBits();
> +            m_entropyCoder->codeSplitFlag(outTempCU, 0, depth);
> +            outTempCU->m_totalBits += m_entropyCoder->getNumberOfWrittenBits(); // split bits
> +        }
> +
> +        // check if cost is greater than (MAX_INT64 - 1)
> +        if (m_rdCost.m_psyRd && outTempCU->m_totalPsyCost >= MAX_INT64)
> +            outTempCU->m_totalPsyCost = m_rdCost.calcPsyRdCost(outTempCU->m_totalDistortion, outTempCU->m_totalBits, outTempCU->m_psyEnergy);
> +        else if (outTempCU->m_totalRDCost >= MAX_INT64)
> +            outTempCU->m_totalRDCost = m_rdCost.calcRdCost(outTempCU->m_totalDistortion, outTempCU->m_totalBits);

Unrelated to this patch, but now that psy-rd is stable, we can do away
with the separate cost variables. We should always be measuring rd cost
or psy-rd cost, there's no reason to keep both variables.

> +        if (depth == slice->m_pps->maxCuDQPDepth && slice->m_pps->bUseDQP)
> +        {
> +            bool hasResidual = false;
> +            for (uint32_t blkIdx = 0; blkIdx < outTempCU->getTotalNumPart(); blkIdx++)
> +            {
> +                if (outTempCU->getCbf(blkIdx, TEXT_LUMA) || outTempCU->getCbf(blkIdx, TEXT_CHROMA_U) ||
> +                    outTempCU->getCbf(blkIdx, TEXT_CHROMA_V))
> +                {
> +                    hasResidual = true;
> +                    break;
> +                }
> +            }
> +
> +            uint32_t targetPartIdx = 0;
> +            if (hasResidual)
> +            {
> +                bool foundNonZeroCbf = false;
> +                outTempCU->setQPSubCUs(outTempCU->getRefQP(targetPartIdx), outTempCU, 0, depth, foundNonZeroCbf);
> +                X265_CHECK(foundNonZeroCbf, "expected to find non-zero CBF\n");
> +            }
> +            else
> +                outTempCU->setQPSubParts(outTempCU->getRefQP(targetPartIdx), 0, depth); // set QP to default QP
> +        }
> +        m_rdEntropyCoders[nextDepth][CI_NEXT_BEST].store(m_rdEntropyCoders[depth][CI_TEMP_BEST]);
> +        checkBestMode(outBestCU, outTempCU, depth);
> +    }
> +    outBestCU->copyToPic(depth);
> +    copyYuv2Pic(pic, outBestCU->getAddr(), outBestCU->getZorderIdxInCU(), depth);
> +}
> +
> +void Analysis::checkIntra(TComDataCU*& outBestCU, TComDataCU*& outTempCU, PartSize partSize, CU *cu, uint8_t* sharedModes)
>  {
>      //PPAScopeEvent(CheckRDCostIntra + depth);
>      uint32_t depth = g_log2Size[m_param->maxCUSize] - cu->log2CUSize;
> @@ -544,7 +689,10 @@
>      uint32_t tuDepthRange[2];
>      outTempCU->getQuadtreeTULog2MinSizeInCU(tuDepthRange, 0);
>  
> -    estIntraPredQT(outTempCU, m_origYuv[depth], m_tmpPredYuv[depth], m_tmpResiYuv[depth], m_tmpRecoYuv[depth], tuDepthRange);
> +    if (sharedModes)
> +        sharedIntraPredQT(outTempCU, m_origYuv[depth], m_tmpPredYuv[depth], m_tmpResiYuv[depth], m_tmpRecoYuv[depth], tuDepthRange, sharedModes);
> +    else
> +        estIntraPredQT(outTempCU, m_origYuv[depth], m_tmpPredYuv[depth], m_tmpResiYuv[depth], m_tmpRecoYuv[depth], tuDepthRange);
>  
>      estIntraPredChromaQT(outTempCU, m_origYuv[depth], m_tmpPredYuv[depth], m_tmpResiYuv[depth], m_tmpRecoYuv[depth]);
>  
> diff -r 184e56afa951 -r 9db768fa41ad source/encoder/analysis.h
> --- a/source/encoder/analysis.h	Fri Sep 12 12:02:46 2014 +0530
> +++ b/source/encoder/analysis.h	Mon Sep 15 14:07:31 2014 +0530
> @@ -100,6 +100,9 @@
>      StatisticLog  m_sliceTypeLog[3];
>      StatisticLog* m_log;
>  
> +    uint32_t      m_zorder;
> +    uint32_t      m_ctuToDepthIndex;

it seems like these should be derivable from existing CU fields, or
passed on the stack to sharedCompressIntraCU()

> +
>      Analysis();
>      bool create(uint32_t totalDepth, uint32_t maxWidth);
>      void destroy();
> @@ -110,7 +113,8 @@
>  
>      /* Warning: The interface for these functions will undergo significant changes as a major refactor is under progress */
>      void compressIntraCU(TComDataCU*& outBestCU, TComDataCU*& outTempCU, uint32_t depth, TComDataCU* cuPicsym, CU *cu);
> -    void checkIntra(TComDataCU*& outBestCU, TComDataCU*& outTempCU, PartSize partSize, CU *cu);
> +    void checkIntra(TComDataCU*& outBestCU, TComDataCU*& outTempCU, PartSize partSize, CU *cu, uint8_t* sharedModes=NULL);

I don't generally like default args. please update all callers instead

> +    void sharedCompressIntraCU(TComDataCU*& outBestCU, TComDataCU*& outTempCU, uint32_t depth, TComDataCU* cuPicsym, CU *cu, uint8_t* sharedDepth, char* sharedPartSizes, uint8_t* sharedModes);
>  
>      void compressInterCU_rd0_4(TComDataCU*& outBestCU, TComDataCU*& outTempCU, TComDataCU* cu, uint32_t depth, TComDataCU* cuPicsym, CU *cu_t,
>                                 int bInsidePicture, uint32_t partitionIndex, uint32_t minDepth);
> diff -r 184e56afa951 -r 9db768fa41ad source/encoder/search.cpp
> --- a/source/encoder/search.cpp	Fri Sep 12 12:02:46 2014 +0530
> +++ b/source/encoder/search.cpp	Mon Sep 15 14:07:31 2014 +0530
> @@ -1484,6 +1484,75 @@
>      x265_emms();
>  }
>  
> +void Search::sharedIntraPredQT(TComDataCU* cu, TComYuv* fencYuv, TComYuv* predYuv, ShortYuv* resiYuv, TComYuv* reconYuv, uint32_t depthRange[2], uint8_t* sharedModes)
> +{
> +    uint32_t depth        = cu->getDepth(0);
> +    uint32_t initTrDepth  = cu->getPartitionSize(0) == SIZE_2Nx2N ? 0 : 1;
> +    uint32_t numPU        = 1 << (2 * initTrDepth);
> +    uint32_t log2TrSize   = cu->getLog2CUSize(0) - initTrDepth;
> +    uint32_t qNumParts    = cu->getTotalNumPart() >> 2;
> +    uint32_t overallDistY = 0;
> +    static const uint8_t intraModeNumFast[] = { 8, 8, 3, 3, 3 }; // 4x4, 8x8, 16x16, 32x32, 64x64

this array is unused

> +
> +    // loop over partitions
> +    uint32_t partOffset = 0;
> +    uint32_t puDistY;
> +    uint64_t puCost;
> +    for (uint32_t pu = 0; pu < numPU; pu++, partOffset += qNumParts)
> +    {
> +        uint32_t bestPUMode = sharedModes[pu];
> +        uint32_t bestPUDistY = 0;

these two variables both seem a bit redundant

> +        cu->setLumaIntraDirSubParts(bestPUMode, partOffset, depth + initTrDepth);
> +
> +        // set context models
> +        m_entropyCoder->load(m_rdEntropyCoders[depth][CI_CURR_BEST]);
> +
> +        // determine residual for partition
> +        puCost = 0;
> +        puDistY = xRecurIntraCodingQT(cu, initTrDepth, partOffset, fencYuv, predYuv, resiYuv, true, puCost, depthRange);
> +
> +        bestPUDistY = puDistY;
> +        xSetIntraResultQT(cu, initTrDepth, partOffset, reconYuv);
> +
> +        // update overall distortion
> +        overallDistY += bestPUDistY;
> +
> +        if (pu != numPU - 1)
> +        {
> +            uint32_t zorder      = cu->getZorderIdxInCU() + partOffset;
> +            pixel*   dst         = cu->m_pic->getPicYuvRec()->getLumaAddr(cu->getAddr(), zorder);
> +            uint32_t dststride   = cu->m_pic->getPicYuvRec()->getStride();
> +            pixel*   src         = reconYuv->getLumaAddr(partOffset);
> +            uint32_t srcstride   = reconYuv->getStride();
> +            primitives.luma_copy_pp[log2TrSize - 2](dst, dststride, src, srcstride);
> +        }
> +
> +        // update PU data
> +        cu->setLumaIntraDirSubParts(bestPUMode, partOffset, depth + initTrDepth);

is this call redundant?

> +        cu->copyToPic((uint8_t)depth, pu, initTrDepth);
> +    }
> +
> +    if (numPU > 1)
> +    {
> +        // set Cbf for all blocks
> +        uint32_t combCbfY = 0;
> +        uint32_t partIdx  = 0;
> +        for (uint32_t part = 0; part < 4; part++, partIdx += qNumParts)
> +            combCbfY |= cu->getCbf(partIdx, TEXT_LUMA,     1);
> +
> +        for (uint32_t offs = 0; offs < 4 * qNumParts; offs++)
> +            cu->getCbf(TEXT_LUMA)[offs] |= combCbfY;
> +

white-space

> +    }
> +
> +    // reset context models
> +    m_entropyCoder->load(m_rdEntropyCoders[depth][CI_CURR_BEST]);
> +
> +    // set distortion (rate and r-d costs are determined later)
> +    cu->m_totalDistortion = overallDistY;

cu->m_totalDistortion could be updated within the loop directly

> +}
> +
>  void Search::getBestIntraModeChroma(TComDataCU* cu, TComYuv* fencYuv, TComYuv* predYuv)
>  {
>      uint32_t depth   = cu->getDepth(0);
> diff -r 184e56afa951 -r 9db768fa41ad source/encoder/search.h
> --- a/source/encoder/search.h	Fri Sep 12 12:02:46 2014 +0530
> +++ b/source/encoder/search.h	Mon Sep 15 14:07:31 2014 +0530
> @@ -109,6 +109,7 @@
>      bool initSearch(x265_param *param, ScalingList& scalingList);
>  
>      void estIntraPredQT(TComDataCU* cu, TComYuv* fencYuv, TComYuv* predYuv, ShortYuv* resiYuv, TComYuv* reconYuv, uint32_t depthRange[2]);
> +    void sharedIntraPredQT(TComDataCU* cu, TComYuv* fencYuv, TComYuv* predYuv, ShortYuv* resiYuv, TComYuv* reconYuv, uint32_t depthRange[2], uint8_t* sharedModes);
>      void estIntraPredChromaQT(TComDataCU* cu, TComYuv* fencYuv, TComYuv* predYuv, ShortYuv* resiYuv, TComYuv* reconYuv);
>  
>      // estimation inter prediction (non-skip)
> _______________________________________________
> 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