[FFmpeg-devel] LIBMPEG2_BITSTREAM_READER vs. golomb.h

[Måns Rullgård]

Michael Niedermayer <michaelni at gmx.at> writes:

> On Wed, Jul 16, 2008 at 12:06:30AM +0100, M?ns Rullg?rd wrote:
>> Michael Niedermayer <michaelni at gmx.at> writes:
>> 
>> > On Mon, Jul 14, 2008 at 12:02:47AM +0100, M?ns Rullg?rd wrote:
>> >> I just spent some time bug-hunting, only to discover that the
>> >> exp-golomb functions in golomb.h do not work with
>> >> LIBMPEG2_BITSTREAM_READER.  This is because they rely on more than the
>> >> promised 17 bits to be available after an UPDATE_CACHE() call.
>> >> 
>> >> To avoid future mishaps, I was thinking of adding a preprocessor check
>> >> to golomb.h (I'm too lazy to try to make it work). 
>> >
>> > try:
>> > @@ -64,10 +63,12 @@
>> >
>> >          return ff_ue_golomb_vlc_code[buf];
>> >      }else{
>> > -        log= 2*av_log2(buf) - 31;
>> > -        buf>>= log;
>> > +        log= 31 - av_log2(buf);
>> > +        LAST_SKIP_BITS(re, gb, log);
>> > +        UPDATE_CACHE(re, gb);
>> > +        buf= SHOW_UBITS(re, gb, log+1);
>> >          buf--;
>> > -        LAST_SKIP_BITS(re, gb, 32 - log);
>> > +        LAST_SKIP_BITS(re, gb, log+1);
>> >          CLOSE_READER(re, gb);
>> >
>> >          return buf;
>> > @@ -149,10 +150,11 @@
>> >
>> >          return ff_se_golomb_vlc_code[buf];
>> >      }else{
>> > -        log= 2*av_log2(buf) - 31;
>> > -        buf>>= log;
>> > -
>> > -        LAST_SKIP_BITS(re, gb, 32 - log);
>> > +        log= 31 - av_log2(buf);
>> > +        LAST_SKIP_BITS(re, gb, log);
>> > +        UPDATE_CACHE(re, gb);
>> > +        buf= SHOW_UBITS(re, gb, log+1);
>> > +        LAST_SKIP_BITS(re, gb, log+1);
>> >          CLOSE_READER(re, gb);
>> >
>> >          if(buf&1) buf= -(buf>>1);
>> 
>> This seems to work with the samples I've tried it on.
>
> But it likely will slow the code down thus i think we cannot
> add it by default. #ifdef would be ugly. That leaves
> waiting for someone to move this code into a noinline function and
> do some benchmarks to ensure the noinline as well as the new code
> causes no speedloss.

As it happens, enabling unaligned memory access on ARM (see another
thread), makes ALT_BITSTREAM_READER as fast asLIBMPEG2_BITSTREAM_READER.

-- 
M?ns Rullg?rd
mans at mansr.com