[FFmpeg-devel] [PATCH] avformat/avcodec: Add DTS-UHD demuxer and parser, movenc support.

[Roy Funderburk]

On 6/15/23 8:46 AM, Paul B Mahol wrote:
> get_vlc2 can be made for get_bits_var(), first table bits (that are still
> int and not uint8_t), the code that picks table index from which to take
> bits.
> 
> It is also possible to make it take both first index and rest of it and
> build bigger tables but that is very very advanced step for new
> contributors.
> 
> Use INIT_VLC_SPARSE_STATIC, there are myriad examples in libavcodec, one of
> them being imm4 decoder.
> 

I will change that table bits array to uint8_t.

I encountered an issue when trying to set up the VLC table with
INIT_VLC_SPARSE_STATIC.

The current get_bits_var:

    static int get_bits_var(GetBitContext *gb, const VarBits *var_bits)
    {
        static const uint8_t bits_used[8] = { 1, 1, 1, 1, 2, 2, 3, 3 };
        int code = show_bits(gb, 3); /* value range is [0, 7] */

        skip_bits(gb, bits_used[code]);
        if (var_bits->bits[code] == 0)
            return 0;
        return get_bits_long(gb, var_bits->bits[code]) + var_bits->add[code];
    }

Changed to use get_vlc2 would be:

    static int get_bits_var(GetBitContext *gb, const VarBits *var_bits)
    {
        int code = get_vlc2(&gb, vlc.table, 3, 1);
        if (var_bits->bits[code] == 0)
            return 0;
        return get_bits_long(gb, var_bits->bits[code]) + var_bits->add[code];
    }

The "vlc.table" that INIT_VLC_SPARSE_STATIC needs to output is:

    len: 1, 1, 1, 1, 2, 2, 3, 3
    sym: 4, 4, 4, 4, 8, 8, 16, 32

INIT_VLC_SPARSE_STATIC would get the same len and sym and this code table as input:
    code: 0, 1, 2, 3, 4, 5, 6, 7

INIT_VLC_SPARSE_STATIC rejects the "len" and "code" because the code
of "2" will not fit into one bit.  Regardless of the fact that the desired
output table is not a valid table according to VLC functions, that invalid
table is what is needed.

This goes back to how get_bits_var decodes the bit context.
get_bits_var looks at three bits to get an index in the range of 0-7.
But it may not skip all of those bits.  One or two of those bits may also be
read by the final get_bits_long.

    code = show_bits(gb, 3);
    skip_bits(gb, bits_used[code]);
    get_bits_long(gb,...);

The VLC functions to construct tables have validation in place that prevents
the construction of the table needed for get_vlc2() to work with
get_bits_var.

So the only way I can find to use get_vlc2() is to not use any of the vlc.h
initialization functions, and instead hard-code the tables to be used by
get_vlc2() like this:

VLCElem table[8] = { {4,1}, {4,1}, {4,1}, {4,1}, {8,2}, {8,2}, {16,3}, {32,3} };

Before I do that, I wanted to verify this would be acceptable. Would
this be a use of get_vlc2() that could lead to issues in the future
if the GET_VLC macro changed so that it would not skip fewer bits than the
code length?

What is your opinion on this?

Thanks,
-Roy