struct Av1Block: narrow field types
#1460
Conversation
…MODES` table with a `const fn` w/ a `match`
…f a `match` since it optimizes better
Note that `rav1d_msac_decode_symbol_adapt16` already truncates its return to `< 16`, so the check is also a no-op here.
… handle negatives
There was a problem hiding this comment.
Looks fine, though I'm a bit worried about all of the
unwraps added thanks to usingInRangein various places. Are you sure that adding all those checks doesn't degrade performance? Can you confirm that they're optimized out?
They generally should be. I'll go add the // Elided comments. Ones like Av1BlockInterRefIndex::new(N).unwrap() or Av1BlockInterRefIndex::new(N + rav1d_msac_decode_bool_adapt(...) as i8).unwrap() are elided since it knows N and the range of a bool. Some others using rav1d_msac_decode_symbol_adapt8 or rav1d_msac_decode_symbol_adapt16 are elided because they truncate to 8/16.
I think Av1BlockInterRefIndex::new(frame_hdr.skip_mode.refs[0]).unwrap(), doesn't, but we can also make frame_hdr.skip_mode.refs use an InRange type (didn't want to do too much just in this PR).
I also measured performance, and there's basically no difference (just noise, < 0.1% different).
2 participants
The benchmark script found that the merge commit 8eb6932 (#951) was one of the largest perf regressions (+4.7% on 80 threads, +4.3% on 150 threads). While it touched a bunch of things, one significant part was
f.frame_thread.b, aDisjointMut<Vec<Av1Block>>.Av1Blockis currently 32 bytes, but could be 24 bytes. This PR narrows all of the field types (recursively) as much as possible so we know their more precise range. This will help us then re-arrange the layout (Av1Blockis not passed to asm) and get it down to 24 bytes (this PR itself is perf neutral, although it should remove a bunch of bounds checks). Plus, this PR just makes things more idiomatic. Moreover, I think it may be possible to switch fromDisjointMuttoRelaxedatomics forAv1Block. It should also help makeAv1Block: FromZeroes, which will let me optimize the.resizewithnew_zeroed_slice.