[FFmpeg-devel] AAC decoder round 9

[Robert Swain]

2008/8/20 Vitor Sessak <vitor1001 at gmail.com>:
> Robert Swain wrote:
>>
>> 2008/8/19 Robert Swain <robert.swain at gmail.com>:
>>>
>>> 2008/8/18 Michael Niedermayer <michaelni at gmx.at>:
>>>>
>>>> On Mon, Aug 18, 2008 at 08:42:53PM +0100, Robert Swain wrote:

[...]

>>>>> +            // tns_decode_coef
>>>>> +            lpc[0] = 1;
>>>>> +            for (m = 1; m <= order; m++) {
>>>>> +                lpc[m] = tns->coef[w][filt][m - 1];
>>>>> +                for (i = 1; i < m; i++)
>>>>> +                    b[i] = lpc[i] + lpc[m] * lpc[m-i];
>>>>> +                for (i = 1; i < m; i++)
>>>>> +                    lpc[i] = b[i];
>>>>> +            }
>>>>
>>>> This loop looks oddly similar to the end of compute_lpc_coefs()
>>>> and eval_lpc_coeffs() can something be factored out here?
>>>
>>> Indeed, the 3GPP ref source calls this code the conversion of PARCOR
>>> (PARtial autoCORrelation, according to a quick google) coefficients to
>>> LPC coefficients. It doesn't, however, say what algorithm is used.
>>>
>>> Short of fully unrolling the Levinson-Durbin algorithm functions you
>>> mentioned and comparing to the aac.c code, it's not easy to see if
>>> they're identical or not, nor which should be used. I'll try comparing
>>> the results of some conversions before I resort to manual unrolling
>>> with pen and paper. :)
>>
>> Hrm. I can't get eval_lpc_coeffs() to work with a coef input vector
>> populated from tns_tmp2_map.

[...]

>> It seems the various conditions on in[] and f0 cause it to exit with
>> an error with the coef[] vectors I've tried and as they're potentially
>> valid, that's disconcerting for using this code.
>>
>> compute_lpc_coefs() accepts a two dimensional lpc array as an
>> argument. I'm guessing this so that the coefficients are available for
>> various orders ready for testing to choose which is best or something
>> like that.
>>
>> Do you have any advice for how to proceed? I'm going to keep prodding
>> eval_lpc_coeffs() in the mean time.
>
> Can the following patch be modified to work with AAC too?

I don't like the patch as it is, but you've helped me quite a bit to
figure this out. :)

lpc[0] = 1 isn't used so I'd like to remove it and start the useful
coefficients from index 0. With this, the above-quoted loop can be
rewritten to use part of compute_lpc_coefs():

// tns_decode_coef
for (m = 0; m < order; m++) {
    float tmp;
    lpc[m] = tns->coef[w][filt][m];
    for (i = 0; i < m/2; i++) {
        tmp = lpc[i];
        lpc[i]     += lpc[m] * lpc[m-1-i];
        lpc[m-1-i] += lpc[m] * tmp;
    }
    if(m & 1)
        lpc[i]     += lpc[m] * lpc[i];
}

Doing it this way avoids the need for the b[] temporary array and
produces identical output I think. You should be able to see clearly
that this is the same as the main loops within compute_lpc_coefs().
So, now the question is how to refactor this code to satisfy everyone.

Michael - would you like a single generic macro like the one Vitor
proposes? Or maybe two macros, one for this inner set of loops and one
for the extra autocorrelation function return value normalisation to
convert to autocorrelation coefficients and checks that are needed for
the other implementations? Or something else?

Regards,
Rob