[FFmpeg-devel] Input requested on floating point decomposition for AAC Main

[Alex Converse]

On Tue, Nov 11, 2008 at 3:26 PM, M?ns Rullg?rd <mans at mansr.com> wrote:
> "Alex Converse" <alex.converse at gmail.com> writes:
>
>> On Tue, Nov 11, 2008 at 5:21 AM, Jason Garrett-Glaser
>> <darkshikari at gmail.com> wrote:
>>> On Mon, Nov 10, 2008 at 10:39 PM, Alex Converse <alex.converse at gmail.com> wrote:
>>>> To do 16-bit floating point rounding for AAC-Main, I need a function
>>>> that will decompose a float into a normalized scheme and it's
>>>> exponent. Conveniently there exists the x87 instruction FXTRACT for
>>>> this very purpose.
>>>
>>> Are you sure this is a good idea?  FXTRACT takes 170 clock cycles on
>>> Core 2, according to Agner, making it the single slowest floating
>>> point operation in x86 history, perhaps competing with FBSTP.
>>>
>>
>> Hmm, FXTRACT may be slow but it appears to not be as slow as frexpf.
>> On my Core2 Duo I'm getting:
>> 40092570 dezicycles in t_fxtract, 2 runs, 0 skips
>> 142385760 dezicycles in t_frexpf, 2 runs, 0 skips
>>
>> for this test code:
>> float t1() {
>>     float j, f, g;
>>     START_TIMER("t_fxtract");
>>     for (j = 1.0f; j <= 1048576.0f; j++)
>>         g = fxtract(j, &f);
>>     STOP_TIMER("t_fxtract");
>>     return g;
>> }
>>
>> float t2() {
>>     int i, j;
>>     float h;
>>     START_TIMER("t_frexpf");
>>     for (j = 1.0f; j <= 1048576.0f; j++)
>>         h = frexpf(j, &i);
>>     STOP_TIMER("t_frexpf");
>>     return h;
>> }
>
> What do frexpf() and fxtract() contain?
>

frexpf() is the standard libm function, see man frexp. fxtract() is:
float fxtract(float a, float* exp)
{
    float y, z;
    __asm__ (
             "fxtract\n\t"
             :"=t"(z), "=u"(y)
             :"0"(a)
    );
    *exp = y;
    return z;
}

--Alex