[FFmpeg-devel] [RFC] remove table_4_3_value with CONFIG_SMALL

[Måns Rullgård]

Reimar D?ffinger <Reimar.Doeffinger at gmx.de> writes:

> On Tue, Oct 13, 2009 at 12:54:29PM +0200, Michael Niedermayer wrote:
>> On Tue, Oct 13, 2009 at 12:24:58PM +0200, Reimar D?ffinger wrote:
>> > Hello,
>> > at over 128kB this is one of the largest variables in .bss
>> > Since it is in .bss I'm not 100% sure it is worth avoiding, though
>> > it might help compiling on IA64 which currently can't handle .bss > 4 MB
>> > with gcc at least.
>> > Attached patch instead uses pow etc. if CONFIG_SMALL is set.
>> > Which on the other hand might be really silly because the systems where
>> > you'd use CONFIG_SMALL are those that are least likely to have a fast
>> > pow implementation...
>> > I also admit that the pow for the exponent could probably be trivially
>> > replaced by shift and a 4-entry multiplication table for the lower bits.
>> 
>> [...]
>> > @@ -218,6 +220,9 @@
>> >  /* compute value^(4/3) * 2^(exponent/4). It normalized to FRAC_BITS */
>> >  static inline int l3_unscale(int value, int exponent)
>> >  {
>> > +#if CONFIG_SMALL
>> > +    return pow(value, 4.0 / 3.0) * pow(2, exponent*0.25) + 0.5;
>> 
>> value * cbrtf(value) * exp2f(exponent*0.25) + 0.5
>
> Is exp2f going to be faster than
> (1 << (exponent >> 2)) * exptbl[exponent & 3]

As a rule of thumb, libm functions are very, very slow.

-- 
M?ns Rullg?rd
mans at mansr.com