[pulseaudio-discuss] [Ubuntu-audio-dev] Preferred resampler method on armhf.

[Sjoerd Simons]

On Thu, 2014-11-20 at 08:35 +0100, David Henningsson wrote:
> Crossposting to Debian and upstream lists.
> 
> Apparently Debian has a patch that uss16 speex-fixed-3 587738 604034 874023 421630 543090
es fixed point by default on armhf, 
> so I'm just echoing Luke's question here: Has anybody performed any 
> testing or benchmarks across armhf hardware, w r t fixed point vs 
> floating point resampling with speex and PulseAudio?

Yes, obviously i did at the time I made that patch otherwise i would
have been pulling facts out of thin air :p. I do remember discussion it
with Arun on various occasions, probably on IRC.

Unfortunately i can no longer find my numbers, i do remember testing it
on a Feroceon 88FR131 (QNAP TS-119, softfloat), Cortex A8 (i.mx53,
hardfloat) and Cortex A9 (i.mx6, hardfloat). The softfloat results
should be obvious, for hardfloat on one of them the numbers were about
equal while on the other integer was quite a bit faster (but i really
don't remember the way round, it's been a while).

If you're looking at this now i would suggest to redoing the test on
currentish boards (Some tests on cortex A15 and cortex A7 machines would
be great). In case you're happy to rig together a small benchmark i'm
more then happy to run it on my small collection of boards here?

Oh as a side-note, there are NEON patches for speex around but i'm not
sure if those have been merged upstream. I can't see any signs of NEON
support in the Debain speex package on a quick skim at least.

> On 2014-11-18 00:25, Luke Yelavich wrote:
> > Hey folks.
> > I'm starting to work on merging Ubuntu and Debian packaging for pulseaudio. One of the patches Debian currently has is to use fixed point on all arm flavours they support. The argument given is that float resampling doesn't necessarily perform better than fixed point resampling, even on hard float capable hardware, and sometimes it is worse. here is the blurb taken from the patch:
> >
> > Switch to use the speex fixed point resampler method by default on ARM.
> > Traditionately there wasn't any standard floating point hardware, so for
> > those machines it's an obvious choise. On machine using the hardfloat
> > ABI floating point still isn't an obvious win over integer performance,
> > it's of equal performance on some machines, but slower on others.
> >
> > Has anybody performed any testing or benchmarks on the hardware we support? Should we just drop this and continue to use floating point resampler code?
> >
> > Luke
> >
> > _______________________________________________
> > Mailing list: https://launchpad.net/~ubuntu-audio-dev
> > Post to     : ubuntu-audio-dev at lists.launchpad.net
> > Unsubscribe : https://launchpad.net/~ubuntu-audio-dev
> > More help   : https://help.launchpad.net/ListHelp
> >
>