[x265] [PATCH 2 of 2] rc: tune initial predictor values for better frame size predictions in vbv lookahead
Steve Borho
steve at borho.org
Wed Apr 8 20:39:52 CEST 2015
On 04/08, aarthi at multicorewareinc.com wrote:
> # HG changeset patch
> # User Aarthi Thirumalai
> # Date 1426487738 -19800
> # Mon Mar 16 12:05:38 2015 +0530
> # Node ID b4646bdaf2202748d3a3b159b9bc8e3cfcd272de
> # Parent 6f159ec858f4eec484d722916c41911e3e4540f9
> rc: tune initial predictor values for better frame size predictions in vbv lookahead
>
> an overall improvement in ssim of around .05-.1 db can be seen.
> Also, improves visual quality at the start of the encode.
>
> SteamLocomotiveTrain_2560x1600_60_10bit_crop.yuv --bitrate 9000 --vbv-bufsize 9000 --strict-cbr
>
> Bitrate Y PSNR U PSNR V PSNR Global PSNR SSIM SSIM (dB)
> before 9168.43 39.689 45.096 45.152 41.048 0.93176 11.66
> after 9106.16 39.864 45.169 45.199 41.194 0.93518 11.883
do these pair of commits belong on stable?
> diff -r 6f159ec858f4 -r b4646bdaf220 source/encoder/ratecontrol.cpp
> --- a/source/encoder/ratecontrol.cpp Tue Mar 31 22:16:21 2015 +0530
> +++ b/source/encoder/ratecontrol.cpp Mon Mar 16 12:05:38 2015 +0530
> @@ -372,12 +372,17 @@
> /* Frame Predictors and Row predictors used in vbv */
> for (int i = 0; i < 4; i++)
> {
> - m_pred[i].coeff = 1.5;
> + m_pred[i].coeff = 1.0;
> m_pred[i].count = 1.0;
> m_pred[i].decay = 0.5;
> m_pred[i].offset = 0.0;
> }
> - m_pred[0].coeff = 1.0;
> + m_pred[0].coeff = m_pred[3].coeff = 0.75;
> + if (m_param->rc.qCompress >= 0.8) // when tuned for grain
> + {
> + m_pred[1].coeff = 0.75;
> + m_pred[0].coeff = m_pred[3].coeff = 0.50;
> + }
> if (!m_statFileOut && (m_param->rc.bStatWrite || m_param->rc.bStatRead))
> {
> /* If the user hasn't defined the stat filename, use the default value */
> @@ -1797,8 +1802,11 @@
> double pbits = predictSize(&m_pred[m_predType], q, (double)m_currentSatd);
> if (pbits > rce->frameSizeMaximum)
> q *= pbits / rce->frameSizeMaximum;
> -
> - if (!m_isCbr || (m_isAbr && m_currentSatd >= rce->movingAvgSum && q <= q0 / 2))
> + /* To detect frames that are more complex in SATD costs compared to prev window, yet
> + * lookahead vbv reduces its qscale by half its value. Be on safer side and avoid drastic
> + * qscale reductions for frames high in complexity */
> + bool mispredCheck = rce->movingAvgSum && m_currentSatd >= rce->movingAvgSum && q <= q0 / 2;
> + if (!m_isCbr || (m_isAbr && mispredCheck))
> q = X265_MAX(q0, q);
>
> if (m_rateFactorMaxIncrement)
> _______________________________________________
> 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