On Mon, Mar 17, 2014 at 2:18 AM, <[email protected]> wrote:
> # HG changeset patch
> # User Kavitha Sampath <[email protected]>
> # Date 1395040372 -19800
> # Mon Mar 17 12:42:52 2014 +0530
> # Node ID 99551efb1b89f6dbb00761d4003ac5927fc760b7
> # Parent 8d5deb7cafd83ac554489deb8577f905a0cda6b3
> weightanalysis: fix stack corruption due to fabsf()
>
> diff -r 8d5deb7cafd8 -r 99551efb1b89 source/encoder/slicetype.cpp
> --- a/source/encoder/slicetype.cpp Mon Mar 17 00:47:24 2014 -0500
> +++ b/source/encoder/slicetype.cpp Mon Mar 17 12:42:52 2014 +0530
> @@ -1303,7 +1303,9 @@
> refMean = (float)ref->wp_sum[0] / (fenc->lines * fenc->width) / (1 <<
> (X265_DEPTH - 8));
>
> /* Early termination */
> - if (fabsf(refMean - fencMean) < 0.5f && fabsf(1.f - guessScale) <
> epsilon)
> + float meanDiff = refMean < fencMean ? fencMean - refMean : refMean -
> fencMean;
> + float guessVal = guessScale > 1.f ? guessScale - 1.f : 1.f - guessScale;
> + if (meanDiff < 0.5f && guessVal < epsilon)
this is seriously ugly. I think it would be better to understand why
fabsf() seems to experience problems on Win32 Debug builds and fix the
root cause. This is just papering over a deeper issue.
> return;
>
> int minoff = 0, minscale, mindenom;
> diff -r 8d5deb7cafd8 -r 99551efb1b89 source/encoder/weightPrediction.cpp
> --- a/source/encoder/weightPrediction.cpp Mon Mar 17 00:47:24 2014 -0500
> +++ b/source/encoder/weightPrediction.cpp Mon Mar 17 12:42:52 2014 +0530
> @@ -298,7 +298,9 @@
>
> /* Early termination */
> x265_emms();
> - if (fabsf(refMean[plane] - fencMean[plane]) < 0.5f && fabsf(1.f
> - guessScale[plane]) < epsilon)
> + float meanDiff = refMean[plane] < fencMean[plane] ?
> fencMean[plane] - refMean[plane] : refMean[plane] - fencMean[plane];
> + float guessVal = guessScale[plane] > 1.f ? guessScale[plane] -
> 1.f : 1.f - guessScale[plane];
> + if (meanDiff < 0.5f && guessVal < epsilon)
> {
> SET_WEIGHT(weights[plane], 0, 1, 0, 0);
> continue;
> _______________________________________________
> x265-devel mailing list
> [email protected]
> https://mailman.videolan.org/listinfo/x265-devel
--
Steve Borho
_______________________________________________
x265-devel mailing list
[email protected]
https://mailman.videolan.org/listinfo/x265-devel
