[FFmpeg-devel] [PATCH 1/2] VP4 video decoder

[James Almer]

On 5/14/2019 9:37 AM, Peter Ross wrote:
> On Sun, May 12, 2019 at 01:24:56PM +0200, Reimar Döffinger wrote:
>> On 12.05.2019, at 08:12, Peter Ross <pross at xvid.org> wrote:
>>> +static int read_mb_value(GetBitContext *gb)
>>> +{
>>> +    int v = 1;
>>> +    int size;
>>> +
>>> +    do {
>>> +        size = 0;
>>> +        if (!get_bits1(gb))
>>> +            break;
>>> +        v++;
>>> +        do {
>>> +            if (!get_bits1(gb))
>>> +                break;
>>> +            v += 1 << size++;
>>> +        } while (size < 8);
>>> +    } while (size == 8);
>>> +
>>> +    if (size)
>>> +        v += get_bits(gb, size);
>>> +
>>> +    return v;
>>> +}
>>
>> Maybe not worth it performance wise, but did you check if this could be simplified?
>> For example the get_bits1 cases that end up with size 0 could return directly.
>> Or it could peek ahead 9 bits in the bitstream and count the leading 1s to get v and size without looping (i.e. loop only for the 9 bits of 1s specifically).
>> Alternatively add a comment to clarify the encoding scheme it implements (like 9 consecutive 1s is a prefix encoding an offset of 257 etc).
> 
> thanks for these suggestions.
> 
> replacing get_bits() with OPEN_READER/UPDATE_CACHE/SHOW_UBITS/etc results in a
> consistent 0.50 % speedup.

You could also test the cached reader Paul wrote some time ago, and see
if it's faster. Just define CACHED_BITSTREAM_READER to 1 to enable it
and use the get_bits functions normally (don't use the above macros).

Make sure to test in both x86_32 and x86_64, since it uses a 64bit
variable as bit cache. If it's only faster in x86_64, you could enable
it only for those targets.

> 
> checking the initial bit, and returning from the function early, appears to make
> no difference to decoder speed.
> 
> moving 'v += 1 << size++' to go outside the inner loop makes it more clear
> what the algorithm is doing. i was hoping for a speed improvement, but see no change.
> 
> -- Peter
> (A907 E02F A6E5 0CD2 34CD 20D2 6760 79C5 AC40 DD6B)
> 
> 
> _______________________________________________
> ffmpeg-devel mailing list
> ffmpeg-devel at ffmpeg.org
> https://ffmpeg.org/mailman/listinfo/ffmpeg-devel
> 
> To unsubscribe, visit link above, or email
> ffmpeg-devel-request at ffmpeg.org with subject "unsubscribe".
>