[FFmpeg-devel] Review request - ra288.{c,h} ra144.{c,h}

[Justin Ruggles]

Justin Ruggles wrote:
> Vitor Sessak wrote:
>> Michael Niedermayer wrote:
>>> On Tue, Jul 29, 2008 at 08:20:45PM +0200, Vitor Sessak wrote:
>>>> Hi,
>>>>
>>>> Those four files never passed a review. I've just finished cleaning them 
>>>> up, so if anyone wants to review them (Michael already said he will), 
>>>> now is time.
>>> heres 288
>>>
>>> [...]
>>>> static void decode(RA288Context *ractx, float gain, int cb_coef)
>> [...]
>>
>>>> /**
>>>>  * Converts autocorrelation coefficients to LPC coefficients using the
>>>>  * Levinson-Durbin algorithm. See blocks 37 and 50 of the G.728 specification.
>>>>  *
>>>>  * @return 0 if success, -1 if fail
>>>>  */
>>>> static int eval_lpc_coeffs(const float *in, float *tgt, int n)
>>>> {
>>>>     int i, j;
>>>>     double f0, f1, f2;
>>>>
>>>>     if (in[n] == 0)
>>>>         return -1;
>>>>
>>>>     if ((f0 = *in) <= 0)
>>>>         return -1;
>>>>
>>>>     in--; // To avoid a -1 subtraction in the inner loop
>>>>
>>>>     for (i=1; i <= n; i++) {
>>>>         f1 = in[i+1];
>>>>
>>>>         for (j=0; j < i - 1; j++)
>>>>             f1 += in[i-j]*tgt[j];
>>>>
>>>>         tgt[i-1] = f2 = -f1/f0;
>>>>         for (j=0; j < i >> 1; j++) {
>>>>             float temp = tgt[j] + tgt[i-j-2]*f2;
>>>>             tgt[i-j-2] += tgt[j]*f2;
>>>>             tgt[j] = temp;
>>>>         }
>>>>         if ((f0 += f1*f2) < 0)
>>>>             return -1;
>>>>     }
>>>>
>>>>     return 0;
>>>> }
>>> duplicate of compute_lpc_coefs() ?
>> Yes, the two functions are practically identical, but 
>> compute_lpc_coefs() use doubles and ra288 uses floats.
>>
>> Justin: does FLAC really need double precision or that was just an 
>> arbitrary choice?
> 
> I know I ran some compression comparison tests at one point, and using
> floats vs. doubles did negatively affect compression.  How much, I don't
> remember.  I'll test again and post the results.  If it's not too bad,
> I'm ok with changing to floats.

test of 5 concatenated samples.
total playing time: 26m47s

system: athlon64 x2 6000+, linux, gcc 4.1.2

compression level 5:
double  -- 6.224s -- 171183921
no sse2 -- 6.600s -- 171183931
float   -- 6.452s -- 171839999 = 0.4% larger

compression level 8:
double  -- 13.385s -- 167634457
no sse2 -- 13.981s -- 167634492
float   -- 13.857s -- 168444919 = 0.4% larger


I'm guessing that Loren's optimized functions could be adapted to use
floats to give better results.

-Justin