On Mon, Mar 06, 2017 at 03:48:57PM +0100, Matthieu Bouron wrote:
> On Thu, Feb 23, 2017 at 04:59:16PM +0100, Matthieu Bouron wrote:
> > Hello,
> >
> > The following patchset add the ff_simple_idct function neon functions for
> > the
> > aarch64 platform. It's ported from armv7 simple_idct_neon with some
> > improvements:
> > * the source idct blocks are now loaded once and kept in v24-v31
> > * the source idct blocks are no longer overriden in idct_col4_top
> > * the destination is now written in one pass at the end of
> > ff_simple_idct{,_put,_add}_neon
> >
> > It is bitexact with the armv7 neon implementation.
> >
> > Here are some results (reported by {START,STOP}_TIMER) on an Odroid-C2
> > (Cortex
> > A53):
> >
> > Functions IDCT: simple IDCT: simpleneon
> > ff_simple_idct_put 9795 units 3170 units
> > ff_simple_idct_add 10227 units 3302 units
> >
>
> Ping.
I'd like to push the patch tomorrow if there is no objection.
If that helps, here is the output of mjpegdec with simple and simpleneon
idct methods.
Original: http://0x5c.me/idct/original.jpg
Simple: http://0x5c.me/idct/simplec.png
Simpleneon: http://0x5c.me/idct/simpleneon.png
The diff between simple and simpleneon shows off some off by 1
differences: http://0x5c.me/idct/diff.png (simpleneon aarch64 is bitexact
with its armv7 counterpart though).
Matthieu
_______________________________________________
ffmpeg-devel mailing list
ffmpeg-devel@ffmpeg.org
http://ffmpeg.org/mailman/listinfo/ffmpeg-devel
