[FFmpeg-devel] [PATCH 01/11] ffmpeg: use log10 instead of log()/log(10)
gajjanag at mit.edu
Thu Oct 29 12:50:10 CET 2015
On Thu, Oct 29, 2015 at 7:36 AM, wm4 <nfxjfg at googlemail.com> wrote:
> On Thu, 29 Oct 2015 07:27:20 -0400
> Ganesh Ajjanagadde <gajjanag at mit.edu> wrote:
>> On Thu, Oct 29, 2015 at 5:43 AM, wm4 <nfxjfg at googlemail.com> wrote:
>> > On Thu, 29 Oct 2015 00:19:59 -0400
>> > Ganesh Ajjanagadde <gajjanagadde at gmail.com> wrote:
>> >> This is more concise and conveys the intent better.
>> >> Furthermore, it is likely more precise as well due to lack of floating
>> >> point division.
>> >> Signed-off-by: Ganesh Ajjanagadde <gajjanagadde at gmail.com>
>> >> ---
>> > These patches are all pretty similar. And likely tedious to check for
>> > correctness. So low gain, while some potential for regressions.
>> In the time it took you to write that comment, you could have easily
>> reviewed a couple. log10 is already being used in the codebase, why
>> not make it consistent and also make it more precise?
>> There is a good reason why libc has a log10 function.
>> Maybe you don't care about such things, but I do: it is (roughly)
>> analogous to using double instead of float for filters/resampling etc
>> - noise floor should not be increased unless there is a clear benefit.
>> Here there is none from using log as opposed to log10.
> Not going to play this superficially-review-mass-patches-with-little-to-
> with-regression-fallout game.
> You're giving the ML more patches to review than it can deal with it, so
> my reaction is to flat out reject such patch-spam for things which seem to
> have little benefit and are not 100% trivial.
How? I send fewer patches than many other developers here.
> (This reminds me of mass cosmetics from Libav...)
Comparing this with cosmetics is not accurate - these are NOT
cosmetics. They affect the bit accuracy of the results. To prove to
the likes of wm4 that this is not cosmetic, here are some simple
computing 10*log10(x) vs 10*log10(x)/log(10.0), top is first, bottom
is second expr:
x = 1e-12:
x = DBL_MIN:
x = 5*1e32:
These are differences obtained within the first 15 digits. I am pretty
confident that by finding the worst case, one can get even starker
> ffmpeg-devel mailing list
> ffmpeg-devel at ffmpeg.org
More information about the ffmpeg-devel