[FFmpeg-devel] [PATCH] lavc/aacenc_utils: replace powf(x, y) by expf(logf(x), y)

[Reimar Döffinger]

On 10.03.2016, at 03:06, Ganesh Ajjanagadde <gajjanag at gmail.com> wrote:

> On Wed, Mar 9, 2016 at 2:16 AM, Reimar Döffinger
> <Reimar.Doeffinger at gmx.de> wrote:
>> On 08.03.2016, at 04:48, Ganesh Ajjanagadde <gajjanag at gmail.com> wrote:
>> 
>>> +                    nzl += expf(logf(s / ethresh) * nzslope);
>> 
>> Shouldn't log2f/exp2f be faster?
>> log2f at least has CPU support on x86 AFAICT.
> 
> I had tested this, and no, though it is still faster than powf.
> 
> It still seems to rely on libm, note that we don't use -ffast-math and
> a look at https://github.com/lattera/glibc/tree/master/sysdeps/x86_64/fpu
> as well seems to say no. Problem is, GNU people like to prioritize
> "correctly rounded" behavior over fast, reasonably accurate code,
> sometimes to ludicruous degrees.
> 
> Personally, I don't know why we don't use -ffast-math, not many seem
> to care that heavily on strict IEEE semantics. Maybe it leads to too
> much variation across platforms?

You lose some guarantees. In particular, the compiler will assume NaNs do not happen and you cannot predict which code path (after a comparison for example) they take.
But some code for either security or correctness reasons needs them to be handled a certain way.
I guess in theory you could try to make sure fisnan is used in all those cases, but then you need to find them, and I think if you take -ffast-math description literally there is no guarantee that even fisnan continues to work... I am also not sure none of the code relies on order of operations to get the precision it needs.
So it is simply too dangerous.
Some more specific options might be possible to use though (but I think even full -ffast-math gains you almost nothing? Does it even help here?).