Commit Graph

521 Commits

Author SHA1 Message Date
Yunqing Wang
394020383d Remove an unused first pass statistic
One of the first pass stats "new_mv_count" is no longer used in VP9,
and is removed. This also makes it easy to implement a multi-threaded
first pass. This change doesn't affect the coding performance, which
has been verified by borg tests.

Change-Id: I4c7c7bf9465fda838eb230814ef0c631c068c903
2016-12-07 15:32:25 -08:00
Paul Wilkins
715c65914b Change to KF boost calculation.
This  change is a step in a larger change to the way boost and interval are
determined for ARF and Key frames.

This patch contains some pluming for the general case but focuses on the
key frame boost calculation. This now relies more heavily on the rate at
which the error score increases between the primary and secondary reference
frame. This seems to be less fragile when dealing with different frame sizes.
For example larger image formats tend in the first pass to see a higher
% of intra coded blocks and the use of this number in calculating the frame
decay factor was leading to much lower boost numbers for 4K, for example,
than the same clip coded at 2K.

This change does give overall gains but they are MUCH larger for the 4K Netflix
set. For the 4K Netflix set the average gain is around 3% with some clips > 20%
whereas for the same set at 2K the average gain is 0.5-1%.

In general for small image formats the boost is most often reduced a little whereas
4K clips the boost is increased. There are some -ve cases such as Akiyo at 352x288
where the reduced boost hurts the metrics, especially for SSIM, even while
the set as a whole improves. This is most notable at very low Q and may be the
subject of a future patch.

Some common code for KF and ARF was separated in this patch for the purposes of
tuning but may later be re-merged if appropriate.

Change-Id: Iaa15ac5a58d2be89181100d95cef6a8dc4b12d0d
2016-10-28 15:35:59 +01:00
Paul Wilkins
de859676dd Changes to KF boost calculation.
Remove double counting of decay. Limit maximum KF boost.

Change-Id: I0fb2344d0f78b5e95bb899dfad12b0ca84034b2c
2016-10-26 17:53:29 +01:00
paulwilkins
ccd6a8e2fa Removal of a couple of two pass adjustments.
Removed a couple of adjustments that no longer move the needle
much but complicate the process of tuning.

Change-Id: Ie320f5cf155e6aac14a4757ea9ada2cd59f27590
2016-10-26 17:52:37 +01:00
Yunqing Wang
9afe2cf599 Merge "Fix an issue in vp9_first_pass for non-mulitple of 16 resolutions" 2016-09-30 00:49:06 +00:00
Deepa K G
2745f94deb Fix an issue in vp9_first_pass for non-mulitple of 16 resolutions
This patch sets the 16x16 src_diff to zero and ensures correct calculation
of this_error for block sizes smaller than 16x16.

Change-Id: I7b7c02d267433c9f22c8ac9b8d5df2f499175172
2016-09-29 16:19:23 -07:00
paulwilkins
0421d8e318 Limit max arf boost and scale motion breakout for image size.
Added a cap on the maximum boost for an arf based on interval length.
Fixed bug where by the image size was not accounted for in determining
two of the motion breakout thresholds.

Overall small gains of 0.2-0.4% psnr but on large image format clips with
slow zooms the gain may be as much as 20% or more (e.g. in_to_tree
at 1080P)

Change-Id: Id0a47391203026742daa9c97afac5705fd8c4dfb
2016-09-26 15:38:29 +01:00
paulwilkins
3e9e77008c Casts to remove some warnings.
Added casts to remove warnings:
BUG=webm:1274

In regards to the safety of these casts they are of two types:-

- Normalized bits per (16x16) MB stored in a 32 bit int (This is safe as bits
per MB even with << 9 normalization cant overflow 32 bits. Even raw 12
bits hdr source even would only be  29 bits :- (4+4+12+9) and the encoder
imposes much stricter limits than this on max bit rate.

- Cast as part of variance calculations.  There is an internal cast up to 64 bit
for the Sum X Sum calculation, but after normalization dividing by the number
of points the result will always be <= the SSE value.

Change-Id: I4e700236ed83d6b2b1955e92e84c3b1978b9eaa0
2016-09-01 16:10:12 +01:00
paulwilkins
af3b0de732 Add casting to fix warning.
Frame bits can safely be stored int but group bits
(kf or arf) use 64bit.

Change-Id: I0800f2a28070f8749110a95721c116fc56987885
2016-08-17 11:18:07 +01:00
Paul Wilkins
fe4dd4f43f Merge "Modified ARF group allocation." 2016-08-15 09:42:30 +00:00
paulwilkins
656f4a88cf Modified ARF group allocation.
Small average gains in the range 0.05 - 0.1

Change-Id: I30e85c04be615cc84726427c5057388b20a6ff60
2016-08-10 14:22:01 -07:00
Alex Converse
6554333b59 Refactor mv limits.
Change-Id: Ifebdc9ef37850508eb4b8e572fd0f6026ab04987
2016-08-08 11:54:00 -07:00
clang-format
e0cc52db3f vp9/encoder: apply clang-format
Change-Id: I45d9fb4013f50766b24363a86365e8063e8954c2
2016-08-02 16:47:11 -07:00
Yaowu Xu
039f9e08f0 change HBD pixel value from uint8_t to uint16_t
This fixes a regression in 10/12 bit encoding results.

Change-Id: I438877352a41aae0a864a8d9979afe4aa2061d81
2016-08-02 11:01:39 -07:00
Yaowu Xu
dc5618f3bb Add pointer conversion for HBD buffers
This fixes a crash in HBD build.

Change-Id: I7f688f50227323e69bba65df0d56f4360f01771b
2016-08-01 15:56:43 -07:00
Paul Wilkins
830fa866a5 Merge "Sample points to reduce encode overhead." 2016-07-22 09:27:34 +00:00
Paul Wilkins
063e4a2914 Merge "Noise energy Experiment in first pass." 2016-07-22 09:27:19 +00:00
Yury Gitman
bdfdd7d993 cosmetics: Correct grammar/spelling in comments
Change-Id: I1954f9d6e33abff9081fe7a5cf59d5497768e0df
2016-07-18 12:49:00 -07:00
paulwilkins
3a986eac57 Sample points to reduce encode overhead.
Only noise filter sampled points in first pass to reduce
any first pass speed overhead.

Change-Id: Ic80d4400e59146d1c3332336c4350faf28ff8b17
2016-07-11 11:45:52 +01:00
paulwilkins
2580e7d63e Noise energy Experiment in first pass.
Use a measure of noise energy to adjust Q estimate and
arf filter strength.

Gains 0.3-0.5% on Lowres and |Netflix sets.
Hdres and Midres neutral.

Change-Id: Ic0de552e7b6763e70eeeaa3651619831b423e151
2016-07-07 14:50:21 +01:00
paulwilkins
e25d6252a4 Fix error in get_ul_intra_threshold() for 10/12 bit.
The scaling of the threshold for 10 and 12 bit here appears
to be in the wrong direction. For 10 and 12 bit we expect sse
values to be higher and hence the threshold used should be
scaled up not down.

Change-Id: I2678116652b539aef48100e0f22873edd4f5a786
2016-06-30 13:38:57 +01:00
paulwilkins
f9a3d08f1b Fix error in get_smooth_intra_threshold() for 10/12 bit.
This function seems to scale the threshold for testing an
SSE value in the wrong direction for 10 and 12 bit inputs.

Also for a true SSE the scalings should probably be << 4 and 8

Change-Id: Iba8047b3f70d04aa46d9688a824f3d49c1c58e90
2016-06-30 13:34:11 +01:00
Yury Gitman
67611119b5 cosmetics: Beautify whitespaces and line wrapping
Change-Id: I9afa02cae671bd3527cf344695e53d0cc767f549
2016-06-24 10:18:06 -07:00
paulwilkins
fba94b8cfd Small re-arrangement.
Move initialization of a some new "twopass" values
to the function vp9_init_second_pass() and some other
small  changes.

Remove #if GROUP_ADAPTIVE_MAXQ as this is always
enabled now.

Change-Id: I1dbec2fd7c419779848aa987c4cd7824d4df8456
2016-06-09 11:32:34 +01:00
paulwilkins
45a26dd9c8 Slightly more damped VBR adjustment.
Increase in the damping used in adjusting the active Q range.
This does hurt rate accuracy a little in a few extreme cases
especially if the clip is very short*, but helps metrics.

* Note that the adjustment is applied at the GF/ARF group level based
on what happened in the last group.  Hence for very short clips where
the length of a single group may be a significant % of the clip length
there is still scope for some drift that cannot be accommodated.

In practice most data points in our test sets are now much closer to target
than was previously the case with default settings and in some cases are
better even than they were with the command line undershoot and overshoot
parameter was set very low (e.g. 2%). For example in bridge_close at high rates
the old mechanism was unable to adapt enough to prevent extreme overshoot.

Change-Id: I634f8f0e015b5ee64a9f0ccaa2bcfdbc1d360489
2016-06-03 13:19:51 +01:00
paulwilkins
552fd02cf0 Change to get_twopass_worst_quality()
Change to the calculation of the error divisor used in
get_twopass_worst_quality(). This follows on from other
changes to the rate control that impact the output of this
function.

Change-Id: I414fa9aa1e6a68a64dccea17c3712f44b8a0c10c
2016-06-03 13:18:29 +01:00
paulwilkins
f9865d1701 Removed unused data structure.
Removed unused element from TWOPASS data structure.

Change-Id: I9b662fd8eea727a7978055bc14f7c7328f048a5e
2016-06-03 13:18:09 +01:00
paulwilkins
cd700e1ab9 Adjustment calculation of active worst quality.
Added a factor based on the bit spend in the last arf group vs the
target to adjust the choice of the active worst quality in subsequent
groups.

Helps clips where previously there was a big overshoot or undershoot
to adapt and get closer to the target rate.

Change-Id: I67034b801679b99024409489a2273ea6fe23b8e6
2016-06-03 13:17:21 +01:00
paulwilkins
4328b08521 Remove gf_zeromotion_pct.
The use of this value is preventing rate adjustment on clips
or sections that have very little motion but high noise and
this can give rise to some sections with massive overshoot.

Change-Id: I9a65c7c1148dc5d3a7d8b23e50fc1733f3661621
2016-06-03 12:13:03 +01:00
Paul Wilkins
5fd142e763 Merge "Fixed 8K two pass encoder crash." 2016-05-11 16:25:25 +00:00
paulwilkins
65732c36a8 Fixed 8K two pass encoder crash.
Bug found by Yunqing relating to the correction for size at 8K and
above in get_twopass_worst_quality().

The basis for the correction was changed to the linear size relative to
1080P as a baseline and the adjustment has been clamped to prevent
problems at extreme images sizes.

For 1080P the results on our test sets were neutral but the low res and
mid res sets saw a small gain (0.1%-0.2% average).

I would also expect some gains on 4k and larger content where the
previous correction was overly aggressive.

Change-Id: I30b026b5f4535e9601e3178d738066459d19c8fb
2016-05-11 14:45:50 +01:00
Hui Su
f16cf9fbcc Merge "VP9: enable trellis quantization optimization for intra blocks" 2016-04-26 19:11:49 +00:00
hui su
c8f56d2303 VP9: enable trellis quantization optimization for intra blocks
Coding gain:
lowres  0.18%
midres  0.23%
hdres   0.36%

Change-Id: I044c8afbc481fc55b23d440352941071355b0afb
2016-04-26 10:17:29 -07:00
paulwilkins
015c43f0c1 Resolve a couple of TODOs in firstpass.c
Resolved two TODO items.

Force a minimum value of 1.0 for frame duration as per section duration.

Column inactive zone is currently set to 0 as most of the serious issues
relating to inactive regions relate to letter boxing.

Change-Id: Ifbab3acf2c089d7305620a7ff7ed7c3536cc9235
2016-04-25 12:38:10 +01:00
paulwilkins
4b590058c8 Experiment to adapt for net AQ offset.
In Aq mode 1 the segment and AQ delta for each block is based
on spatial variance. There may be a net imbalance between blocks
that have lower Q than the baseline value and those that have higher Q.

This patch monitors that imbalance and extends the allowed baseline
Q range for the frame to accommodate adjustment of that baseline value
to compensate.

Change-Id: Iae8a48c7c01fe2af94a141e149d03acf467237ca
2016-04-25 12:07:07 +01:00
Paul Wilkins
2e0841931c Merge "Adjustment to prediction decay." 2016-04-18 18:47:13 +00:00
Debargha Mukherjee
ff72cca8bb Merge "Step towards making the 2-pass cq mode perceptual" 2016-04-12 19:42:06 +00:00
Paul Wilkins
1c187c4be0 Adjustment to prediction decay.
Adjustment to stop excessive prediction decay triggered by blocks
or frames with extremely low spatial complexity which rendered the
comparison of intra and inter coded errors meaningless.

This was causing much shorter than expected groups on some 4k
test content.

Change-Id: I3f2c64200ef6dcef4721fc9f2ec09e480056ffc2
2016-04-11 13:25:33 +01:00
Paul Wilkins
f659c7e99e Merge "Adjust motion component of prediction decay." 2016-04-11 10:27:30 +00:00
Paul Wilkins
c5a89b46b9 Merge "Trap very short arf group just before a kf." 2016-04-11 10:27:12 +00:00
Debargha Mukherjee
c47c460f69 Step towards making the 2-pass cq mode perceptual
Uses a metric on fraction of smooth blocks derived from first pass
stats in a frame to adjust down the cq_level modestly in the cq mode.
The current implementation does not add much complexity, and is
fairly light in the adaptation.

Change-Id: Ic484e810d5bd51b7bb6b8945f378c7c3d9d27053
2016-04-09 06:24:18 -07:00
paulwilkins
81d1e7bf28 Adjust motion component of prediction decay.
Adjust the motion decay component to account for image size.
This has very little impact for smaller image sizes.

Average bdrate results for our HD test sets:-

Hdres set:  opsnr +0,92%,  Fast SSIM +1.6%
Netflix hd set:  opsnr + 1.5%, Fast SSIM +3.1%

There are a couple of notable -ve clips such as cyclist and sunflower
which seem to be better with a shorter interval but also a few very big
wins such as Jets >12% psnr 22% Fast SSIM and from the Netflix
Netflix set PierSeaside 9.7% psnr and 18.2% Fast SSIM.

Change-Id: Ie43aaedaa74331ed83d624a13548094ac64fed9e
2016-04-08 10:28:40 +01:00
paulwilkins
107de19698 Trap very short arf group just before a kf.
Trap the case where we end up with a very short arf group just before
a key frame. Such a group often has poor quality and may cause pulsing.

For example if the KF is 17 frames away we are better doing two mid-size
groups of 9 and 8 than a group of 15 followed by a group of 2.

This becomes more and more important when coding with a short forced
kf interval though it may not impact our standard tests much.

Change-Id: I29d83d6637b203eac69be320dd35a7401a4678c1
2016-04-07 14:56:24 +01:00
Scott LaVarnway
ba962a5f37 VP9: Eliminate up_available and left_available
Use above_mi and left_mi instead.

Change-Id: I0b50e232c31d11da30aa2fb6f91a695aaf725e0c
2016-03-30 04:47:39 -07:00
Paul Wilkins
ee35f4dede Fix some HBD scaling bugs.
Fixed a bug for HBD in stats output and scaling of a threshold
within the first pass code.

Change-Id: Icd0463d78cc9ce4661b027b0612c55c0628d212f
2016-03-22 12:28:02 -07:00
Zoe Liu
0f7dd40324 Define a macro to replace hardcoded values
The definition is for the number of frames to check to determine the
recent decay rate, further to determine the next key frame in the
first pass of the encoder.

Change-Id: Ic696d6eb518a86fa296842273cf8767ef0b0e27a
2016-02-04 05:33:22 -08:00
Zoe Liu
b6f1f369af Cleaned code in firstpass related to altref frame
Change-Id: I6dbddf500fd4f64216774685700efea972165470
2016-01-28 16:37:17 -08:00
Zoe Liu
9e612763f0 Cleaned a code in define_gf_group() for firstpass
Change-Id: I0b8eff9e74e3446024482d011ec4fb28b19136d3
2016-01-25 09:38:45 -08:00
Scott LaVarnway
78fadcd063 VP9: Do not set up/left available in vp9_first_pass()
These values are set in set_mi_row_col()

Change-Id: Ie44c9d2dcb79797a8473b5fc29394277a7117547
2016-01-22 11:30:51 -08:00
Scott LaVarnway
5232326716 VP9: Eliminate MB_MODE_INFO
Change-Id: Ifa607dd2bb366ce09fa16dfcad3cc45a2440c185
2016-01-19 16:40:20 -08:00