[FFmpeg-devel] [PATCH] add fixed coeff LPC mode to alac encoder

[Jai Menon]

On Sun, Oct 04, 2009 at 08:53:03PM -0400, Justin Ruggles wrote:
> Michael Niedermayer wrote:
> 
> > On Thu, Oct 01, 2009 at 08:03:09PM -0400, Justin Ruggles wrote:
> >> Hi,
> >>
> >> We currently do not have a regression test for the ALAC encoder like we
> >> do for FLAC.  The ALAC encoder has 2 compression levels, one
> >> uncompressed, which does not cover much of the code, and one which uses
> >> floating point when calculating LPC coefficients, which does not lend
> >> itself to reliable regression tests across all test platforms.
> >>
> >> The way we do regression testing with FLAC is to use the fixed LPC
> >> coefficient mode, which skips the floating-point parts but still covers
> >> prediction, entropy coding, and other parts in the regression test.
> >>
> >> Here is a patch to the ALAC encoder to add a fixed-coeff LPC mode using
> >> compression_level == 1.  This will allow us to add a round-trip
> >> encode/decode regression test.  The default compression level is changed
> >> to 2 and keeps the same behavior.  The fixed-coeff mode is about 20%
> >> faster but has about 6% worse compression (about the same as FLAC with
> >> compression level 2).
> >>
> >> Thanks,
> >> Justin
> >>
> > 
> >>  alacenc.c |   13 +++++++++++--
> >>  1 file changed, 11 insertions(+), 2 deletions(-)
> >> f1f4235083438379b9b828846668c463de9addf2  alac_compr1.patch
> >> Index: libavcodec/alacenc.c
> >> ===================================================================
> >> --- libavcodec/alacenc.c	(revision 20134)
> >> +++ libavcodec/alacenc.c	(working copy)
> >> @@ -132,12 +132,21 @@
> >>      int shift[MAX_LPC_ORDER];
> >>      int opt_order;
> >>  
> >> +    if (s->avctx->compression_level > 1) {
> >>      opt_order = ff_lpc_calc_coefs(&s->dspctx, s->sample_buf[ch], s->avctx->frame_size, s->min_prediction_order, s->max_prediction_order,
> >>                                     ALAC_MAX_LPC_PRECISION, coefs, shift, 1, ORDER_METHOD_EST, ALAC_MAX_LPC_SHIFT, 1);
> >>  
> >>      s->lpc[ch].lpc_order = opt_order;
> >>      s->lpc[ch].lpc_quant = shift[opt_order-1];
> >>      memcpy(s->lpc[ch].lpc_coeff, coefs[opt_order-1], opt_order*sizeof(int));
> >> +    } else {
> >> +        s->lpc[ch].lpc_order = 4;
> >> +        s->lpc[ch].lpc_quant = 3;
> >> +        s->lpc[ch].lpc_coeff[0] = -4;
> >> +        s->lpc[ch].lpc_coeff[0] =  6;
> >> +        s->lpc[ch].lpc_coeff[0] = -4;
> >> +        s->lpc[ch].lpc_coeff[0] =  1;
> > 
> > these coefficients do not look optimal, did you generate them using the LPC
> > code over a large number of audio samples? You should IMHO!
> 
> you're right.  i did some more testing and found coeffs that come within
> about 1% of compression level 2 with most samples tested.
> 
> > besides, are you sure you want to assign them all to lpc_coeff[0] ?
> 
> oops... yeah, that would help. :)
> 
> new patch attached.

Feel free to apply anytime :) Sorry for the late reply, I was away for
quite sometime..

-- 
Jai Menon