[FFmpeg-devel] [PATCH] Some ra144.c simplifications
Vitor Sessak
vitor1001
Sun Jun 29 13:28:14 CEST 2008
Michael Niedermayer wrote:
> On Sat, Jun 28, 2008 at 11:27:40PM +0200, Vitor Sessak wrote:
>> Michael Niedermayer wrote:
>>> On Tue, Jun 24, 2008 at 11:35:12PM +0200, Vitor Sessak wrote:
>>>> Michael Niedermayer wrote:
>>>>> On Sat, Jun 21, 2008 at 07:53:09AM +0200, Vitor Sessak wrote:
>>>>>> Michael Niedermayer wrote:
>>>>>>> On Wed, Jun 04, 2008 at 08:18:10PM +0200, Vitor Sessak wrote:
>>>>>>>> Michael Niedermayer wrote:
>>>>>>>>> On Wed, May 28, 2008 at 09:23:02PM +0200, Vitor Sessak wrote:
>>>>>>>>>> Michael Niedermayer wrote:
>>>>>>>>>>> On Wed, May 28, 2008 at 06:56:45PM +0200, Vitor Sessak wrote:
>>>>>>>>>>>> Michael Niedermayer wrote:
>>>>>>>>>>>>> On Tue, May 27, 2008 at 09:16:09PM +0200, Vitor Sessak wrote:
>>>>>>>>>>>>>> Michael Niedermayer wrote:
>>>>>>>>>>>>>>> On Sun, May 25, 2008 at 07:11:52PM +0200, Vitor Sessak wrote:
>>>>>>>>>>>>>>>> Michael Niedermayer wrote:
>>>>>>>>>>>>>>>>> On Sun, May 25, 2008 at 06:05:15PM +0200, Vitor Sessak
>>>>>>>>>>>>>>>>> wrote:
>>>>>>>>>>>>>>>>> [...]
>>>>>>>>>>>>>>>>>>>> ok
>>>>>>>>>>>>>>>>>>> One more...
>>>>>>>>>>>>>>>>>> ... and some more cleanup:
>>>>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>>>>> ra144_vector_add_wav.diff: Make add_wav() receive a vector
>>>>>>>>>>>>>>>>>> instead of three integers
>>>>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>>>>> ra144_params_dec2.diff: Do not calculate anything based in
>>>>>>>>>>>>>>>>>> l, it is unrolled in the loop anyway
>>>>>>>>>>>>>>>>> ok
>>>>>>>>>>>>>>>> Now s/(unsigned) short/(u)int16_t.
>>>>>>>>>>>>>>> ok
>>>>>>>>>>>>>> Next one. dec2() interpolates the block coefficients from the
>>>>>>>>>>>>>> previous one and fall back to a block-dependent schema if the
>>>>>>>>>>>>>> interpolation results in an unstable filter...
>>>>>>>>>>>>> [...]
>>>>>>>>>>>>>> + // Interpolate block coefficients from the this frame
>>>>>>>>>>>>>> forth block and
>>>>>>>>>>>>>> + // last frame forth block
>>>>>>>>>>>>>> for (x=0; x<30; x++)
>>>>>>>>>>>>>> - decsp[x] = (a * inp[x] + b * inp2[x]) >> 2;
>>>>>>>>>>>>>> + decsp[x] = (a * ractx->lpc_coef[x] + b *
>>>>>>>>>>>>>> ractx->lpc_coef_old[x])>> 2;
>>>>>>>>>>>>> ff_acelp_weighted_vector_sum()
>>>>>>>>>>>> Ok, but to do that I need to use int16_t. So I propose to apply
>>>>>>>>>>>> my original patch and then the attached one.
>>>>>>>>>>> hmm, ok
>>>>>>>>>> Done. Now remove the dec1() function (that was memcpy + 1 line of
>>>>>>>>>> code). As a side effect, it removes the need of a memcpy (the
>>>>>>>>>> dec1() call at decode_frame()).
>>>>>>>>> ok
>>>>>>>> Now the first patch (ra144_rescale_energy.diff) split the energy
>>>>>>>> rescaling out of the rms() function. The next patch remove *lpc_refl
>>>>>>>> from the context, since the only thing needed from the last frame is
>>>>>>>> the non rescaled output of rms().
>>>>>>> ok
>>>>>> Now, I'm almost finished with this. Two things remains:
>>>>>>
>>>>>> 1- When decoding a ra144 encoded file, ffmpeg produces lots of
>>>>>> "Multiple frames in a packet from stream 0" (see
>>>>>> http://fate.multimedia.cx/index.php?test_result=1911120 for an
>>>>>> example). This is because the decoder receives a 1000 byte sample and
>>>>>> decode only 20 bytes. The attached patch fix this (it decode all the 50
>>>>>> blocks).
>>>>> wrong solution, we need a AVParser that splits the 1000bytes in 20byte
>>>>> packets. A generic one that works based on block_align might be usefull
>>>>> for other cases as well ...
>>>> I'll see it later.
>>>>
>>>>>> 2- There are lots of unused table entries. Ok to remove them or do you
>>>>>> thing they can useful for anything (another codec?)?
>>>>> remove
>>>>>
>>>>>
>>>>>> Finally, if there is anything else you don't like about ra144.{c,h},
>>>>>> now is the time to say if you want me to have a look at it...
>>>>> 1st pass review of ra144.c is below :)
>>>>> (yes you regret now that you asked, i know ...)
>>>> It was my masochistic side that made that question =)
>>> good, please remind me to do another review when you are finished
>>> with that one :)
>>> With your help ra144.c will soon be pretty nice and clean, next comes
>>> ra288.c i assume :)
>> Why not? By the way, can I just clean it and you flame me in -cvslogs?
>
> yes, but id like to do a review of what remains when you run out of ideas ...
> (iam saying that so you can remind me as i will certainly forget ...)
>
>
>>> [...]
>>>>>> static void lpc_filter(const int16_t *lpc_coefs, const int16_t
>>>>>> *adapt_coef,
>>>>>> void *out, int *statbuf, int len)
>>>>>> {
>>>>>> int x, i;
>>>>>> uint16_t work[50];
>>>>>> int16_t *ptr = work;
>>>>>>
>>>>>> memcpy(work, statbuf,20);
>>>>> 10*sizeof(int16_t)
>>>>>
>>>>>
>>>>>> memcpy(work + 10, adapt_coef, len * 2);
>>>>>>
>>>>>> for (i=0; i<len; i++) {
>>>>>> int sum = 0;
>>>>>> int new_val;
>>>>>>
>>>>>> for(x=0; x<10; x++)
>>>>>> sum += lpc_coefs[9-x] * ptr[x];
>>>>>>
>>>>>> sum >>= 12;
>>>>>>
>>>>>> new_val = ptr[10] - sum;
>>>>>>
>>>>>> if (new_val < -32768 || new_val > 32767) {
>>>>>> memset(out, 0, len * 2);
>>>>>> memset(statbuf, 0, 20);
>>>>>> return;
>>>>>> }
>>>>>>
>>>>>> ptr[10] = new_val;
>>>>>> ptr++;
>>>>>> }
>>>>>>
>>>>>> memcpy(out, work+10, len * 2);
>>>>>> memcpy(statbuf, work + 40, 20);
>>>>>> }
>>>>> duplicate of ff_acelp_lp_synthesis_filter)(
>>>> No, they are slightly different (ff_acelp_lp_synthesis_filter use the
>>>> last n input values to predict out[n]
>>> no it does not, it uses filter_length which you could set to 10
>> Ok, I did it (patch attached). But there were two things I was not happy
>> with. First, ff_acelp_lp_synthesis_filter completely ignores
>> filter_coeffs[0] and uses filter_coeffs[1..filter_length]. I have no
>> idea why it is like that, but I think the correct inner loop would be
>>
>>> for(i=0; i<filter_length; i++)
>>> sum -= filter_coeffs[i] * out[n-i-1];
>> where now filter_length is 10 (and not 11!) for the 10th order filter.
To be more precise, are you ok with the following patch? Vectors in C
usually are written as filter_coeffs[0..length-1]...
>>
>> Also, I had to change the initial value of sum in
>> ff_acelp_lp_synthesis_filter(). I don't know if I can just change it or
>> if I should add a "round" boolean to the parameters of
>> ff_acelp_lp_synthesis_filter...
>
> an int rounder which is 0x800 / 0xFFF would make more sense
Done (didn't sent a patch before because there are not many ways to do
that).
>> One last thing, can I claim copyright on this file?
>
> which file?
See $subj ;-)
Btw, it should be ready for the next review cycle.
-Vitor
-------------- next part --------------
A non-text attachment was scrubbed...
Name: acelp2.diff
Type: text/x-patch
Size: 1911 bytes
Desc: not available
URL: <http://lists.mplayerhq.hu/pipermail/ffmpeg-devel/attachments/20080629/7997d3cf/attachment.bin>
More information about the ffmpeg-devel
mailing list