On 2014-11-20 13:55, Felipe Sateler wrote: > On Thu, Nov 20, 2014 at 8:30 AM, Sjoerd Simons <sjoerd at luon.net> wrote: >> On Thu, 2014-11-20 at 11:58 +0100, David Henningsson wrote: >>> >>> On 2014-11-20 08:40, Alexander E. Patrakov wrote: >>>> 20.11.2014 12:35, David Henningsson wrote: >>>>> Crossposting to Debian and upstream lists. >>>>> >>>>> Apparently Debian has a patch that uses 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? >>>> >>>> That patch has been superseded by this commit: >>>> >>>> http://cgit.freedesktop.org/pulseaudio/pulseaudio/commit/?id=ac984f59d36ef555bc5b0df9af1cd48193d0d14f >>>> >>>> >>>> So, if you prefer a fixed-point resampler on armhf (or, for that matter, >>>> on any other architecture), just compile speex with --enable-fixed-point >>>> there. >>> >>> Well, this is a more of an "upstream default" question rather than a >>> "why don't you recompile speex" question, and one where it might make >>> sense to come up with something reasonable across the board; either by >>> upstreaming "fixed point by default for armhf", or by Debian to drop its >>> patch. >>> >>> It does not look likely to me that PulseAudio upstream and Debian would >>> correspond to so different machine/user scenarios so that a Debian >>> specific patch would make sense here. >> >> Well with the patch mentioned here, i would probably drop that patch >> from Debian in any case as pulse would just follow speex (which is the >> right thing anyway).. Which makes the upstream choice in pulse a bit >> moot? >> >> That said i'd still be interesting on seeing speex resampling benchmark >> numbers on a wider set of recentish boards. > > In debian the speex maintainer has refused to enable floating point on > armhf [1][2], claiming benchmarks show that fixed point is better even > on armhf. I have not seen those benchmarks, so if it is possible to > corroborate or dispel this information it would be great. On ubuntu I > understand speex is built with floating point. > > I only have a raspberry pi to try things on, but I'd be glad to run a > benchmark there if that is of any use. > > [1] https://bugs.debian.org/cgi-bin/bugreport.cgi?bug=736752 > [2] https://bugs.debian.org/cgi-bin/bugreport.cgi?bug=700602 Right, thanks for the reminder. So, speex should be compiled with floating point, not because of performance, but because AGC is broken (or not even implemented!) when compiled with fixed point. So that's what we do on Ubuntu. So maybe we could stick with status quo then, for simplicity? If somebody cares enough it would be up to that person to provide benchmarks. -- David Henningsson, Canonical Ltd. https://launchpad.net/~diwic
