[FFmpeg-devel] [PATCH] avcodec/mpegaudio_tablegen: speed up table generation

[Ganesh Ajjanagadde]

On Wed, Nov 25, 2015 at 9:22 PM, Timothy Gu <timothygu99 at gmail.com> wrote:
> On Wed, Nov 25, 2015 at 6:03 PM Timothy Gu <timothygu99 at gmail.com> wrote:
>
>> On Wed, Nov 25, 2015 at 05:17:29PM -0500, Ganesh Ajjanagadde wrote:
>> > +            double f = value * cbrt_lut[value] * pow(2, (exponent -
>> 400) * 0.25 + FRAC_BITS + 5) / IMDCT_SCALAR;
>>
>> While at it, you could change pow(2 to exp2(, which has a libm.h shim
>> and is easily 4 times faster than pow() on my machine (glibc 2.19,
>> Haswell).
>>
>
> In light of the comments made by Ronald and James in the other thread, it
> would probably be better if you look for instances of optimization
> opportunities in other files that matter rather than these tablegen files
> that will only be run once during the build process. A simple `git grep -E
> 'powf?\(2'` shows an awful lot.

That is not true, the default configure line generates these at
runtime (which was how I obtained the cycle numbers), translating to
cycles burnt. The Arch Linux package I know for a fact does not build
with "hardcoded tables". If it was just build time, I of course would
not have bothered.

Think of it this way: ~ 10^6 cycles is ~ 1 ms of running a CPU, and is
part of trivially avoidable latency. Why it is still being done at
runtime by default, I still don't understand why, and I requested for
an email thread/explanation.

I am working on all instances of pow, this was just one of them. For
me, the answer is simple: any runtime savings on the order of 10^5
cycles (and above) is worth it unless it causes a huge loss of
readability, even if it is in a init function that executes once.

>
> (BTW next time you send similar patches you c/should put them in a patch
> series so that it's easier to track on the ML. I missed the other thread
> before reading this one.)

Will keep in mind; I sent these as I did them.

>
> Timtohy
> _______________________________________________
> ffmpeg-devel mailing list
> ffmpeg-devel at ffmpeg.org
> http://ffmpeg.org/mailman/listinfo/ffmpeg-devel