[FFmpeg-devel] [PATCH 3/3] lavc/cbrt_tablegen: unroll table generation loop

[Ganesh Ajjanagadde]

On Thu, Dec 31, 2015 at 5:08 PM, Ronald S. Bultje <rsbultje at gmail.com> wrote:
> Hi,
>
> On Thu, Dec 31, 2015 at 6:53 PM, Ganesh Ajjanagadde <gajjanagadde at gmail.com>
> wrote:
>>
>> BTW, it would also help me if you or an AAC maintainer can come up
>> with a number below which dynamic initialization can always be done.
>
>
> I think the answer is "never", since 0 is always faster than any number. But
> that's not an absolute veto or anything.

It would be useful to know how long aac decoding takes in general, e.g
for n seconds of audio, what is the cycle count. Cycle counts of
initialization can get easily amortized in that. And that is really
what should determine one's heuristics for when to
statically/dynamically init.

I personally don't mind either way. All of this work grew out of a
remark by wm4 some weeks back:
https://ffmpeg.org/pipermail/ffmpeg-devel/2015-November/184018.html,
something I agree with mostly. The reason I ask is to prioritize
effort. For instance, if 300k cycles is too much to always dynamically
init, I will not bother shaving 20k cycles more: I don't want to spend
time on last mile optimizations unless it can result in help towards
wm4's goal, i.e I don't want an immediate nack to removal of the
hardcoded tables stuff here.

>
> Ronald