> Just for the record, so as to not appear silent here. This is a
> vc2/dirac lowdelay decoder that spends most of its time in Golomb code
> parsing. The commit optimizing Golomb reading only touched some of
> those, for good reason maybe, but not the interleaved versions.
>
> Although the slowdown does appear with the new reader and
> multithreading, single-threaded decoding is faster with the new
> reader, so it's more likely the multithreading or my porting that are
> the cause of this.

I did svq3_get_ue_golomb optimisation and benchmarks for svq3 decoder
with and without the optimisation (dirac allso calls svq3_get_ue_golomb).
I benchmarked with perf stat -r 30, sample:
https://samples.libav.org/V-codecs/SVQ3/FF_VII_AC_High_Res_E3_Trailer.mov
without optimisation: master: 5.61 sec elapsed (+- 0.06%), my_branch:
5.57 (+- 0.09 %) --> 0.7 % speed up
with optimisation: my branch: 5.52 sec --> 1.6 % speed up -- with -threads 1
I also run the bencmark without -threads 1 and the results was similar.

Libav doesn't has vc2/dirac,  I don't know which problem might it
have with threading or anything else and it's not up to me to solve
this.

Best Regards,

Alexandra
_______________________________________________
libav-devel mailing list
[email protected]
https://lists.libav.org/mailman/listinfo/libav-devel

Reply via email to