[FFmpeg-devel] [PATCHv2] lavc/aacenc_utils: replace powf(x, y) by expf(logf(x), y)

[Ganesh Ajjanagadde]

On Sun, Mar 13, 2016 at 7:51 AM, Ronald S. Bultje <rsbultje at gmail.com> wrote:
> Hi,
>
> On Sat, Mar 12, 2016 at 11:40 AM, Ganesh Ajjanagadde <gajjanag at gmail.com>
> wrote:
>>
>> diff --git a/libavutil/internal.h b/libavutil/internal.h
>> index da76ca2..aa43754 100644
>> --- a/libavutil/internal.h
>> +++ b/libavutil/internal.h
>> @@ -315,6 +315,22 @@ static av_always_inline float ff_exp10f(float x)
>>  }
>>
>>  /**
>> + * Compute x^y for floating point x, y. Note: this function is faster
>> than the
>> + * libm variant due to mainly 2 reasons:
>> + * 1. It does not handle any edge cases. In particular, this is only
>> guaranteed
>> + * to work correctly for x > 0.
>> + * 2. It is not as accurate as a standard nearly "correctly rounded" libm
>> variant.
>> + * @param x base
>> + * @param y exponent
>> + * @return x^y
>> + */
>> +static av_always_inline float ff_fast_pow(float x, float y)
>> +{
>> +    return expf(logf(x) * y);
>> +}
>
>
> Thanks, mostly OK. Small comments:
>
> - I wonder if this should move to a separate file, e.g. it seems more
> fitting in mathematics.h or libm.h. internal.h seems like a strange choice.
> I don't know which is better, I'd personally probably go for libm.h but I
> can see why some people wouldn't like it since it slightly changes the
> meaning of that file.

I don't like moving it to libm either for this reason.

I chose lavu/internal for 2 reasons:
1. simply to group with other similar things, (ff_exp10, ff_exp10f)
that also went into lavu/internal.
2. more fundamentally, it appears lavu/mathematics.h is a public
header. I am quite strongly against making ff_exp10, ff_fast_powf etc
public.

> - it should be called ff_fast_powf since it's float-based, not double-based
> (compare pow vs. powf)

Oops, changed.

>
> (I also noticed there are some huge huge huge functions in libm.h that
> probably shouldn't be in a header but move to a .c file instead, if anyone
> is looking for a cleanup target - e.g. hypoth and erf.)

Although I would like to do it myself, I won't simply because I can't
test effectively as I lack a broken libm. Note that lavu/libm is not
just bloated but also broken in other ways as well; recall discussion
on how some of the things there should not really be macros but
functions.

>
> Ronald