View Single Post
Old 13th December 2018, 19:50   #1314  |  Link
Beelzebubu
Registered User
 
Join Date: Feb 2003
Location: New York, NY (USA)
Posts: 109
Quote:
Originally Posted by benwaggoner View Post
So, decoding will be limited by serial decoding of entropy decoding? Do non-reference frames still update and thus serialize the entropy state? If decoding the "bbbb" in an IbbbbBbbbbP" sequence is serialized, that'll really impact decoder parallelization. but if all the non-ref b frames inherit the CABAC state of the most recently decoded reference frame, than it'll be a lot easier.
Frames with a "similar entropy" reference each other, so a high-level P might use the previous P (which is coded 16 frames back) as its entropy reference, and a non-reference inner B frame (which might not be a reference picture at all for pixel purposes) may actually use the previous inner B-frame (which may well be the one directly before this, or usually 2 and sometimes 3 frames back) as its reference. So this certainly influences how well frame-multithreading scales, not in the worst possible way but not ideal either.

And that's why you see weird things where using 256 instead of 128 threads (I think this is 32/16 frame threads x 8 tile threads) on a 32 core leads to pretty significant speedups (like this).
Beelzebubu is offline   Reply With Quote