[x265] [PATCH] analysis: re-order analysis to do splits before ME or intra for pmode
Ashok Kumar Mishra
ashok at multicorewareinc.com
Fri Aug 14 16:39:37 CEST 2015
Below are the performance testing on Haswell with and without depth-search
applied on pmode.
*preset VERYSLOW*
*Before*
D:\ashok>x265_b.exe --input
\\HEVC-TEST-2\testsequences\ducks_take_off_1080p50.y4m --preset veryslow
--hash=1 --no-info --psnr --ssim -o test_b.hevc --pmode
encoded 500 frames in 901.20s (0.55 fps), 8813.94 kb/s, Avg QP:37.97,
Global PSNR: 30.369, SSIM Mean Y: 0.8159502 ( 7.351 dB)
*After*
D:\ashok>x265_a.exe --input
\\HEVC-TEST-2\testsequences\ducks_take_off_1080p50.y4m --preset veryslow
--hash=1 --no-info --psnr --ssim -o test_a.hevc --pmode --limit-refs 1
encoded 500 frames in 632.76s (0.79 fps), 8666.79 kb/s, Avg QP:37.90,
Global PSNR: 30.311, SSIM Mean Y: 0.8134978 ( 7.293 dB)
*preset SLOWER*
*Before*
D:\ashok>x265_b.exe --input
\\HEVC-TEST-2\testsequences\ducks_take_off_1080p50.y4m --preset slower
--hash=1 --no-info --psnr --ssim -o test_b.hevc --pmode
encoded 500 frames in 585.06s (0.85 fps), 8796.19 kb/s, Avg QP:38.03,
Global PSNR: 30.356, SSIM Mean Y: 0.8153866 ( 7.337 dB)
*After*
D:\ashok>x265_a.exe --input
\\HEVC-TEST-2\testsequences\ducks_take_off_1080p50.y4m --preset slower
--hash=1 --no-info --psnr --ssim -o test_a.hevc --pmode --limit-refs 1
encoded 500 frames in 439.82s (1.14 fps), 8740.25 kb/s, Avg QP:37.99,
Global PSNR: 30.334, SSIM Mean Y: 0.8144196 ( 7.315 dB)
*preset SLOW*
*Before*
D:\ashok>x265_b.exe --input
\\HEVC-TEST-2\testsequences\ducks_take_off_1080p50.y4m --preset slow
--hash=1 --no-info --psnr --ssim -o test_b.hevc --pmode
encoded 500 frames in 148.11s (3.38 fps), 8764.71 kb/s, Avg QP:38.08,
Global PSNR: 30.282, SSIM Mean Y: 0.8124724 ( 7.269 dB)
*After*
D:\ashok>x265_a.exe --input
\\HEVC-TEST-2\testsequences\ducks_take_off_1080p50.y4m --preset slow
--hash=1 --no-info --psnr --ssim -o test_a.hevc --pmode --limit-refs 1
encoded 500 frames in 110.45s (4.53 fps), 8712.04 kb/s, Avg QP:38.04,
Global PSNR: 30.265, SSIM Mean Y: 0.8117029 ( 7.252 dB)
*preset MEDIUM*
*Before*
D:\ashok>x265_b.exe --input
\\HEVC-TEST-2\testsequences\ducks_take_off_1080p50.y4m --preset medium
--hash=1 --no-info --psnr --ssim -o test_b.hevc --pmode
encoded 500 frames in 67.01s (7.46 fps), 8975.61 kb/s, Avg QP:37.97, Global
PSNR: 30.076, SSIM Mean Y: 0.8040911 ( 7.079 dB)
*After*
D:\ashok>x265_a.exe --input
\\HEVC-TEST-2\testsequences\ducks_take_off_1080p50.y4m --preset medium
--hash=1 --no-info --psnr --ssim -o test_a.hevc --pmode --limit-refs 1
encoded 500 frames in 37.56s (13.31 fps), 8954.61 kb/s, Avg QP:37.96,
Global PSNR: 30.070, SSIM Mean Y: 0.8038041 ( 7.073 dB)
On Fri, Aug 14, 2015 at 4:27 PM, Ashok Kumar Mishra <
ashok at multicorewareinc.com> wrote:
> Yes, performance has improved, will send after some time for all presets.
>
> On Fri, Aug 14, 2015 at 4:20 PM, Steve Borho <steve at borho.org> wrote:
>
>> On 08/14, ashok at multicorewareinc.com wrote:
>> > # HG changeset patch
>> > # User Ashok Kumar Mishra<ashok at multicorewareinc.com>
>> > # Date 1439540228 -19800
>> > # Fri Aug 14 13:47:08 2015 +0530
>> > # Node ID 9e26bef14543025908ed979b3d217417baf1ac8f
>> > # Parent d56b2466c04459205287e1581d8a36eebf372ba6
>> > analysis: re-order analysis to do splits before ME or intra for pmode
>>
>> I'm happy to see these patches. They look good, do you have any example
>> before/after performance and compression numbers?
>>
>> > diff -r d56b2466c044 -r 9e26bef14543 source/encoder/analysis.cpp
>> > --- a/source/encoder/analysis.cpp Wed Aug 12 18:12:20 2015 +0530
>> > +++ b/source/encoder/analysis.cpp Fri Aug 14 13:47:08 2015 +0530
>> > @@ -505,16 +505,82 @@
>> >
>> > X265_CHECK(m_param->rdLevel >= 2, "compressInterCU_dist does not
>> support RD 0 or 1\n");
>> >
>> > + PMODE pmode(*this, cuGeom);
>> > +
>> > + if (mightNotSplit && depth >= minDepth)
>> > + {
>> > + /* Initialize all prediction CUs based on parentCTU */
>> > + md.pred[PRED_MERGE].cu.initSubCU(parentCTU, cuGeom, qp);
>> > + md.pred[PRED_SKIP].cu.initSubCU(parentCTU, cuGeom, qp);
>> > +
>> > + if (m_param->rdLevel <= 4)
>> > + checkMerge2Nx2N_rd0_4(md.pred[PRED_SKIP],
>> md.pred[PRED_MERGE], cuGeom);
>> > + else
>> > + checkMerge2Nx2N_rd5_6(md.pred[PRED_SKIP],
>> md.pred[PRED_MERGE], cuGeom, false);
>> > + }
>> > +
>> > + bool bNoSplit = false;
>> > + if (md.bestMode)
>> > + {
>> > + bNoSplit = md.bestMode->cu.isSkipped(0);
>> > + if (mightSplit && depth && depth >= minDepth && !bNoSplit &&
>> m_param->rdLevel <= 4)
>> > + bNoSplit = recursionDepthCheck(parentCTU, cuGeom,
>> *md.bestMode);
>> > + }
>> > +
>> > + if (mightSplit && !bNoSplit)
>> > + {
>> > + Mode* splitPred = &md.pred[PRED_SPLIT];
>> > + splitPred->initCosts();
>> > + CUData* splitCU = &splitPred->cu;
>> > + splitCU->initSubCU(parentCTU, cuGeom, qp);
>> > +
>> > + uint32_t nextDepth = depth + 1;
>> > + ModeDepth& nd = m_modeDepth[nextDepth];
>> > + invalidateContexts(nextDepth);
>> > + Entropy* nextContext = &m_rqt[depth].cur;
>> > + int nextQP = qp;
>> > +
>> > + for (uint32_t subPartIdx = 0; subPartIdx < 4; subPartIdx++)
>> > + {
>> > + const CUGeom& childGeom = *(&cuGeom + cuGeom.childOffset +
>> subPartIdx);
>> > + if (childGeom.flags & CUGeom::PRESENT)
>> > + {
>> > + m_modeDepth[0].fencYuv.copyPartToYuv(nd.fencYuv,
>> childGeom.absPartIdx);
>> > + m_rqt[nextDepth].cur.load(*nextContext);
>> > +
>> > + if (m_slice->m_pps->bUseDQP && nextDepth <=
>> m_slice->m_pps->maxCuDQPDepth)
>> > + nextQP = setLambdaFromQP(parentCTU,
>> calculateQpforCuSize(parentCTU, childGeom));
>> > +
>> > + compressInterCU_dist(parentCTU, childGeom, nextQP);
>> > +
>> > + // Save best CU and pred data for this sub CU
>> > + splitCU->copyPartFrom(nd.bestMode->cu, childGeom,
>> subPartIdx);
>> > + splitPred->addSubCosts(*nd.bestMode);
>> > +
>> > +
>> nd.bestMode->reconYuv.copyToPartYuv(splitPred->reconYuv,
>> childGeom.numPartitions * subPartIdx);
>> > + nextContext = &nd.bestMode->contexts;
>> > + }
>> > + else
>> > + splitCU->setEmptyPart(childGeom, subPartIdx);
>> > + }
>> > + nextContext->store(splitPred->contexts);
>> > +
>> > + if (mightNotSplit)
>> > + addSplitFlagCost(*splitPred, cuGeom.depth);
>> > + else
>> > + updateModeCost(*splitPred);
>> > +
>> > + checkDQPForSplitPred(*splitPred, cuGeom);
>> > + }
>> > +
>> > if (mightNotSplit && depth >= minDepth)
>> > {
>> > int bTryAmp = m_slice->m_sps->maxAMPDepth > depth;
>> > int bTryIntra = m_slice->m_sliceType != B_SLICE ||
>> m_param->bIntraInBFrames;
>> >
>> > - PMODE pmode(*this, cuGeom);
>> > + if (m_slice->m_pps->bUseDQP && depth <=
>> m_slice->m_pps->maxCuDQPDepth && m_slice->m_pps->maxCuDQPDepth != 0)
>> > + setLambdaFromQP(parentCTU, qp);
>> >
>> > - /* Initialize all prediction CUs based on parentCTU */
>> > - md.pred[PRED_MERGE].cu.initSubCU(parentCTU, cuGeom, qp);
>> > - md.pred[PRED_SKIP].cu.initSubCU(parentCTU, cuGeom, qp);
>> > if (bTryIntra)
>> > {
>> > md.pred[PRED_INTRA].cu.initSubCU(parentCTU, cuGeom, qp);
>> > @@ -548,8 +614,6 @@
>> >
>> > if (m_param->rdLevel <= 4)
>> > {
>> > - checkMerge2Nx2N_rd0_4(md.pred[PRED_SKIP],
>> md.pred[PRED_MERGE], cuGeom);
>> > -
>> > {
>> > ProfileCUScope(parentCTU, pmodeBlockTime,
>> countPModeMasters);
>> > pmode.waitForExit();
>> > @@ -632,14 +696,13 @@
>> > }
>> > else
>> > {
>> > - checkMerge2Nx2N_rd5_6(md.pred[PRED_SKIP],
>> md.pred[PRED_MERGE], cuGeom, false);
>> > {
>> > ProfileCUScope(parentCTU, pmodeBlockTime,
>> countPModeMasters);
>> > pmode.waitForExit();
>> > }
>> >
>> > checkBestMode(md.pred[PRED_2Nx2N], depth);
>> > - if (m_slice->m_sliceType == B_SLICE)
>> > + if (m_slice->m_sliceType == B_SLICE &&
>> md.pred[PRED_BIDIR].sa8dCost < MAX_INT64)
>> > checkBestMode(md.pred[PRED_BIDIR], depth);
>> >
>> > if (m_param->bEnableRectInter)
>> > @@ -664,14 +727,6 @@
>> > }
>> > }
>> >
>> > - if (md.bestMode->rdCost == MAX_INT64 && !bTryIntra)
>> > - {
>> > - md.pred[PRED_INTRA].cu.initSubCU(parentCTU, cuGeom, qp);
>> > - checkIntraInInter(md.pred[PRED_INTRA], cuGeom);
>> > - encodeIntraInInter(md.pred[PRED_INTRA], cuGeom);
>> > - checkBestMode(md.pred[PRED_INTRA], depth);
>> > - }
>> > -
>> > if (m_bTryLossless)
>> > tryLossless(cuGeom);
>> >
>> > @@ -679,60 +734,9 @@
>> > addSplitFlagCost(*md.bestMode, cuGeom.depth);
>> > }
>> >
>> > - bool bNoSplit = false;
>> > - if (md.bestMode)
>> > - {
>> > - bNoSplit = md.bestMode->cu.isSkipped(0);
>> > - if (mightSplit && depth && depth >= minDepth && !bNoSplit &&
>> m_param->rdLevel <= 4)
>> > - bNoSplit = recursionDepthCheck(parentCTU, cuGeom,
>> *md.bestMode);
>> > - }
>> > -
>> > - if (mightSplit && !bNoSplit)
>> > - {
>> > - Mode* splitPred = &md.pred[PRED_SPLIT];
>> > - splitPred->initCosts();
>> > - CUData* splitCU = &splitPred->cu;
>> > - splitCU->initSubCU(parentCTU, cuGeom, qp);
>> > -
>> > - uint32_t nextDepth = depth + 1;
>> > - ModeDepth& nd = m_modeDepth[nextDepth];
>> > - invalidateContexts(nextDepth);
>> > - Entropy* nextContext = &m_rqt[depth].cur;
>> > - int nextQP = qp;
>> > -
>> > - for (uint32_t subPartIdx = 0; subPartIdx < 4; subPartIdx++)
>> > - {
>> > - const CUGeom& childGeom = *(&cuGeom + cuGeom.childOffset +
>> subPartIdx);
>> > - if (childGeom.flags & CUGeom::PRESENT)
>> > - {
>> > - m_modeDepth[0].fencYuv.copyPartToYuv(nd.fencYuv,
>> childGeom.absPartIdx);
>> > - m_rqt[nextDepth].cur.load(*nextContext);
>> > -
>> > - if (m_slice->m_pps->bUseDQP && nextDepth <=
>> m_slice->m_pps->maxCuDQPDepth)
>> > - nextQP = setLambdaFromQP(parentCTU,
>> calculateQpforCuSize(parentCTU, childGeom));
>> > -
>> > - compressInterCU_dist(parentCTU, childGeom, nextQP);
>> > -
>> > - // Save best CU and pred data for this sub CU
>> > - splitCU->copyPartFrom(nd.bestMode->cu, childGeom,
>> subPartIdx);
>> > - splitPred->addSubCosts(*nd.bestMode);
>> > -
>> > -
>> nd.bestMode->reconYuv.copyToPartYuv(splitPred->reconYuv,
>> childGeom.numPartitions * subPartIdx);
>> > - nextContext = &nd.bestMode->contexts;
>> > - }
>> > - else
>> > - splitCU->setEmptyPart(childGeom, subPartIdx);
>> > - }
>> > - nextContext->store(splitPred->contexts);
>> > -
>> > - if (mightNotSplit)
>> > - addSplitFlagCost(*splitPred, cuGeom.depth);
>> > - else
>> > - updateModeCost(*splitPred);
>> > -
>> > - checkDQPForSplitPred(*splitPred, cuGeom);
>> > - checkBestMode(*splitPred, depth);
>> > - }
>> > + /* compare split RD cost against best cost */
>> > + if (mightSplit && !bNoSplit)
>> > + checkBestMode(md.pred[PRED_SPLIT], depth);
>> >
>> > if (mightNotSplit)
>> > {
>> > @@ -746,8 +750,7 @@
>> >
>> > /* Copy best data to encData CTU and recon */
>> > md.bestMode->cu.copyToPic(depth);
>> > - if (md.bestMode != &md.pred[PRED_SPLIT])
>> > - md.bestMode->reconYuv.copyToPicYuv(*m_frame->m_reconPic,
>> cuAddr, cuGeom.absPartIdx);
>> > + md.bestMode->reconYuv.copyToPicYuv(*m_frame->m_reconPic, cuAddr,
>> cuGeom.absPartIdx);
>> > }
>> >
>> > uint32_t Analysis::compressInterCU_rd0_4(const CUData& parentCTU,
>> const CUGeom& cuGeom, int32_t qp)
>> > _______________________________________________
>> > x265-devel mailing list
>> > x265-devel at videolan.org
>> > https://mailman.videolan.org/listinfo/x265-devel
>>
>> --
>> Steve Borho
>> _______________________________________________
>> x265-devel mailing list
>> x265-devel at videolan.org
>> https://mailman.videolan.org/listinfo/x265-devel
>>
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.videolan.org/pipermail/x265-devel/attachments/20150814/3dbb7fa9/attachment-0001.html>
More information about the x265-devel
mailing list