Enable CDEF and restoration filters for 4:2:2

[barrbrain]

As far as I can tell, there are no outstanding issues with this combination.
However, I will attempt to revive the fuzzer to answer this with confidence.
Closes #1149.

[barrbrain]

I breathed life into the rav1e fuzzer, and encountered an attempt to subtract with overflow at src/lrf.rs:1416 with w = 16, h = 262 and tiles=2.

w = 16
h = 262
speed = 10
q = 255
limit = 1
min_keyint = 3
max_keyint = 3
low_latency = false
bitrate = 0
--
uv_unit_size = 32
uv_unit_log2 = 5
uv_sb_h_log2 = 5
uv_sb_v_log2 = 6
fi.sb_width = 1
fi.sb_height = 5
stripe_uv_decimate = 0
uv_cols = 1
uv_rows = 8

I was able to reproduce from the rav1e CLI:

ffmpeg -f lavfi -i "testsrc=duration=1:size=16x262:rate=30" -pix_fmt yuv422p 422.y4m
cargo run -- -o /dev/null 422.y4m --tiles 2

[barrbrain]

The cause may be at src/lrf.rs:1368:

uv_unit_size =
  uv_unit_size.min(tile_aligned_uv_h_unit_size.min(tile_aligned_uv_v_unit_size));

[lu-zero]

It sounds possibly related to #2212

[xiphmont]

It is, but I'm still figuring out in detail how much overlap there is.

…

On Wed, Apr 8, 2020 at 12:42 PM Luca Barbato ***@***.***> wrote: It sounds possibly related to #2212 <#2212> — You are receiving this because your review was requested. Reply to this email directly, view it on GitHub <#2224 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ACWMHECEAJ22HREKZZPX7T3RLSSQHANCNFSM4MDTIFMQ> .

[rzumer]

Also see #970.

[coveralls]

Coverage increased (+0.005%) to 82.595% when pulling 54d236b on barrbrain:enable-restoration-filters-422 into 4e922fe on xiph:master.

[barrbrain]

The newest commit resolves the desync in #1149. The crash in #2212 occurs independent of enabling CDEF. Further analysis is needed for the decoding error when enabling LRF for 4:2:2.
See Cdef_Uv_Dir in AV1 spec 7.15.1 - CDEF block process.

AWCY results on derf-720p-422 at the default speed show a 9.9% decrease in encoding speed:

PSNR	PSNR Cb	PSNR Cr	PSNR HVS	SSIM	MS SSIM	CIEDE 2000
-2.5572	0.4582	-1.0744	-2.0057	-0.1799	-1.0506	-2.9736

[barrbrain]

Added back enabling of LRF for 4:2:2 now that there is no desync locally or in AWCY.
Added as a separate commit so that we may document the impact.

AWCY results on derf-720p-422 at the default speed show a 28% decrease in encoding speed:

PSNR	PSNR Cb	PSNR Cr	PSNR HVS	SSIM	MS SSIM	CIEDE 2000
-1.4957	-1.2369	-0.5441	-1.0947	-0.9361	-0.7724	-2.1054

[barrbrain]

The total effect of enabling both filters for 4:2:2 is substantial:

PSNR	PSNR Cb	PSNR Cr	PSNR HVS	SSIM	MS SSIM	CIEDE 2000
-4.0057	-0.8299	-1.6084	-3.0774	-1.1085	-1.8073	-5.0115

[lu-zero]

Looks fine to me, the cdef logic might or might not live in a stand alone function since it got a little longer than I'd like.

[barrbrain]

Until #2212 is resolved, I will minimise changes to LRF, CDEF or tiling parts of the code to avoid confusing anyone looking at that bug. I agree that they could be more readable with some work.