[x265] Patch to restore performance prior to AQ mode 4 commit

[Aruna Matheswaran]

Thanks for the patch.
Pushed to Release_3.2 with indentation fixes.

On Tue, Oct 1, 2019 at 6:42 PM Joel Galí <joelgalid at gmail.com> wrote:

> Hi,
>
> AQ mode 4 commit (21db162) causes slowdown even if not used aq-mode
> 4. Some code related to AQ mode 4 is executed always.
>
> Slowdown is aprox. 10% !! on my system encoding 1080p 8-bit using
> --aq-mode 3 and preset medium.
>
> More info:
> https://forum.doom9.org/showthread.php?p=1885998#post1885998
>
> changeset:   12565:e39ec6f83fe5
> tag:         tip
> user:        joelgalid at gmail.com
> date:        Mon Sep 30 18:23:41 2019 +0200
> summary:     Fix: AQ mode 4 commit (21db162) introduces slowdown even is
> not used AQ mode 4.
>
> diff -r fdd69a766881 -r e39ec6f83fe5 source/encoder/slicetype.cpp
> --- a/source/encoder/slicetype.cpp Wed Sep 25 17:58:14 2019 +0530
> +++ b/source/encoder/slicetype.cpp Mon Sep 30 18:23:41 2019 +0200
> @@ -480,17 +480,24 @@
>              {
>  #define AQ_EDGE_BIAS 0.5
>  #define EDGE_INCLINATION 45
> -                uint32_t numCuInHeight = (maxRow + param->maxCUSize - 1)
> / param->maxCUSize;
> -                int maxHeight = numCuInHeight * param->maxCUSize;
> -                intptr_t stride = curFrame->m_fencPic->m_stride;
> -                pixel *edgePic = X265_MALLOC(pixel, stride * (maxHeight +
> (curFrame->m_fencPic->m_lumaMarginY * 2)));
> -                pixel *gaussianPic = X265_MALLOC(pixel, stride *
> (maxHeight + (curFrame->m_fencPic->m_lumaMarginY * 2)));
> -                pixel *thetaPic = X265_MALLOC(pixel, stride * (maxHeight
> + (curFrame->m_fencPic->m_lumaMarginY * 2)));
> -                memset(edgePic, 0, stride * (maxHeight +
> (curFrame->m_fencPic->m_lumaMarginY * 2)) * sizeof(pixel));
> -                memset(gaussianPic, 0, stride * (maxHeight +
> (curFrame->m_fencPic->m_lumaMarginY * 2)) * sizeof(pixel));
> -                memset(thetaPic, 0, stride * (maxHeight +
> (curFrame->m_fencPic->m_lumaMarginY * 2)) * sizeof(pixel));
> -                if (param->rc.aqMode == X265_AQ_EDGE)
> -                    edgeFilter(curFrame, edgePic, gaussianPic, thetaPic,
> stride, maxRow, maxCol);
> +
> + pixel *edgePic = NULL;
> + pixel *gaussianPic = NULL;
> + pixel *thetaPic = NULL;
> +
> + if (param->rc.aqMode == X265_AQ_EDGE)
> + {
> + uint32_t numCuInHeight = (maxRow + param->maxCUSize - 1) /
> param->maxCUSize;
> + int maxHeight = numCuInHeight * param->maxCUSize;
> + intptr_t stride = curFrame->m_fencPic->m_stride;
> + edgePic = X265_MALLOC(pixel, stride * (maxHeight +
> (curFrame->m_fencPic->m_lumaMarginY * 2)));
> + gaussianPic = X265_MALLOC(pixel, stride * (maxHeight +
> (curFrame->m_fencPic->m_lumaMarginY * 2)));
> + thetaPic = X265_MALLOC(pixel, stride * (maxHeight +
> (curFrame->m_fencPic->m_lumaMarginY * 2)));
> + memset(edgePic, 0, stride * (maxHeight +
> (curFrame->m_fencPic->m_lumaMarginY * 2)) * sizeof(pixel));
> + memset(gaussianPic, 0, stride * (maxHeight +
> (curFrame->m_fencPic->m_lumaMarginY * 2)) * sizeof(pixel));
> + memset(thetaPic, 0, stride * (maxHeight +
> (curFrame->m_fencPic->m_lumaMarginY * 2)) * sizeof(pixel));
> + edgeFilter(curFrame, edgePic, gaussianPic, thetaPic, stride, maxRow,
> maxCol);
> + }
>
>                  int blockXY = 0, inclinedEdge = 0;
>                  double avg_adj_pow2 = 0, avg_adj = 0, qp_adj = 0;
> @@ -507,8 +514,8 @@
>                              energy = acEnergyCu(curFrame, blockX, blockY,
> param->internalCsp, param->rc.qgSize);
>                              if (param->rc.aqMode == X265_AQ_EDGE)
>                              {
> -                                pixel *edgeImage = edgePic +
> curFrame->m_fencPic->m_lumaMarginY * stride +
> curFrame->m_fencPic->m_lumaMarginX;
> -                                pixel *edgeTheta = thetaPic +
> curFrame->m_fencPic->m_lumaMarginY * stride +
> curFrame->m_fencPic->m_lumaMarginX;
> +                                pixel *edgeImage = edgePic +
> curFrame->m_fencPic->m_lumaMarginY * curFrame->m_fencPic->m_stride +
> curFrame->m_fencPic->m_lumaMarginX;
> +                                pixel *edgeTheta = thetaPic +
> curFrame->m_fencPic->m_lumaMarginY * curFrame->m_fencPic->m_stride +
> curFrame->m_fencPic->m_lumaMarginX;
>                                  edgeDensity = edgeDensityCu(curFrame,
> edgeImage, edgeTheta, avgAngle, blockX, blockY, param->rc.qgSize);
>                                  if (edgeDensity)
>                                  {
> @@ -541,10 +548,14 @@
>                  }
>                  else
>                      strength = param->rc.aqStrength * 1.0397f;
> -
> -                X265_FREE(edgePic);
> -                X265_FREE(gaussianPic);
> -                X265_FREE(thetaPic);
> +
> + if (param->rc.aqMode == X265_AQ_EDGE)
> + {
> + X265_FREE(edgePic);
> + X265_FREE(gaussianPic);
> + X265_FREE(thetaPic);
> + }
> +
>                  blockXY = 0;
>                  for (int blockY = 0; blockY < maxRow; blockY += loopIncr)
>                  {
>
>
> Regards,
> Joel
> _______________________________________________
> x265-devel mailing list
> x265-devel at videolan.org
> https://mailman.videolan.org/listinfo/x265-devel
>


-- 
Regards,
Aruna
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.videolan.org/pipermail/x265-devel/attachments/20191003/a81beefb/attachment.html>