[FFmpeg-devel] [PATCH 1/5] mlpenc: fix lossless check error in number_sbits

[Jai Luthra]

On Wed, Jul 10, 2019 at 12:14:56AM +0200, Lynne wrote:
>Jul 9, 2019, 9:18 PM by me at jailuthra.in:
>
>> we need two bits instead of one bit to represent -1 in bitstream
>>
>> Signed-off-by: Jai Luthra <me at jailuthra.in>
>> ---
>>  libavcodec/mlpenc.c | 2 +-
>>  1 file changed, 1 insertion(+), 1 deletion(-)
>>
>> diff --git a/libavcodec/mlpenc.c b/libavcodec/mlpenc.c
>> index deb171645c..f4948451f1 100644
>> --- a/libavcodec/mlpenc.c
>> +++ b/libavcodec/mlpenc.c
>> @@ -466,7 +466,7 @@ static void default_decoding_params(MLPEncodeContext *ctx,
>>  */
>>  static int inline number_sbits(int number)
>>  {
>> -    if (number < 0)
>> +    if (number < -1)
>>  number++;
>>
>
>This is different from the first patch's version. Sure its correct now?

Yep. Previous patch produced valid bitstream too, but this provides better 
compression [1] by representing numbers of the form -2^x with one less bit for 
x >= 1.

This makes more sense, as we can represent -2 in two-bit twos-complement 
notation as `10` so output should be 2 bits (instead of 3 by previous patch). 
(similarly for -4, -8, -16, ...)

The lossless errors were being caused when a block of samples were all either 
-1 or 0. This function implied all samples could be represented as single bit 
each, but down the pipeline after huff vlc calculations, the encoder pushed 
<nothing> on the bitstream for all samples, which was always interpreted as 0 
by the decoder and never -1.

NB: One can argue -1 and 0 in fact can be represented in a single bit as two's 
complement, 0 being `0` and -1 being `1`. But imho, single bit two's 
complement is a weird boundary case, and not considering it solves the issue 
here. If someone has a better idea pls suggest.

[1]: tested using both patches on 
https://samples.ffmpeg.org/flac/When%20I%20Grow%20Up.flac. Previous patch 
compressed it to 28788216 byte MLP stream, this one compresses it to 28787834 
byte MLP stream. Both streams are valid and decode to lossless bit-exact 
output.

thx

--
jlut