[FFmpeg-devel] [PATCH]HE-AACv1 try 3 (all missing functionality added)

[Måns Rullgård]

Michael Niedermayer <michaelni at gmx.at> writes:

> On Sat, Feb 13, 2010 at 05:31:59PM +0100, Vitor Sessak wrote:
> [...]
>>> +
>>> +static void vector_fmul_scalar_fadd_scalar_c(float *dst, const float 
>>> *src,
>>> +                                             float mul, float add, int 
>>> len)
>>> +{
>>> +    int i;
>>> +    for (i = 0; i < len; i++)
>>> +        dst[i] = src[i] * mul + add;
>>> +}
>>> +
>>
>> Did you benchmark if hand-unrolling these loops give any benefit? You know 
>> that len is a multiple of four, but the compiler don't.
>
> assert(len>0 && len%4==0);
> then the compiler knows it too :)

Only if you compile with asserts enabled, which will slow things down
immensely in code like this.

> and after you looked at what gcc generates if forced to unroll
> (theres a flag for that IIRC)
> then post a code pessimization bug to the gcc-devels
> -> only way to see gcc get better at these things

What we need is a "soft assert" whereby we can tell the compiler
things we know to be true without any code being generated for it.
Some compilers have that already.

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