Yaowu Xu [Fri, 20 Jan 2012 23:30:31 +0000 (15:30 -0800)]
Reworked context conversion between 8x8 and 4x4
The commit rationized and simplified the entropy context conversion
betwen MB using 8x8 transform and MB using 4x4 transform. The old version
had a number of weirdness in how 4x4 transform MB's context is used for
8x8 blocks other than the first 8x8 within a MB.
Test showed the change has a gain ~.1% for avg psnr, glb psnr and ssim on
the limited HD set.
Yaowu Xu [Thu, 16 Feb 2012 01:35:48 +0000 (17:35 -0800)]
optmized rounding for transforms
the changes are still temporary, the final transforms, especially
inverse ones should take in account both accuracy, complexity, and
sign-bias, which should be decided at a later time.
Yaowu Xu [Wed, 15 Feb 2012 23:01:35 +0000 (15:01 -0800)]
re-scaled 2nd order haar transform
During the work of extend_qrange, we have rolled a factor of 2 from
quantization/dequatnization into 2nd order walsh-hadamard transform.
This commit does the same for the 2nd order haar transform. so they
can share the same quantizaiton process as the 2nd order WHT.
Yaowu Xu [Tue, 14 Feb 2012 15:40:26 +0000 (07:40 -0800)]
moved scaling from dequantization to inverse transform for T8x8
Previously, the scaling related to extended quantize range happens in
dequantization stage, which implies the coefficients form forward
transform are in different scale(4x) from dequantization coefficients
This worked fine when there was not distortion computation done based
on 8x8 transform, but it completely wracked the distortion estimation
based on transform coefficients and dequantized transform coefficients
introduced in commit f64725a00 for macroblocks using 8x8 transform.
This commit fixed the issue by moving the scaling into the stage of
inverse 8x8 transform.
TODO: Test&Verify the transform/quantization pipeline accuracy.
Results with the new filter coefficients compared with the previous
versions are here:
http://www.corp.google.com/~debargha/vp8_results/enhinterp.html (derf)
http://www.corp.google.com/~debargha/vp8_results/enhinterp_hd.html (HD)
Overall, the derf set improves by 0.94% with 8-tap filters while
the HD set improves by 0.58%.
Ronald S. Bultje [Thu, 16 Feb 2012 00:38:04 +0000 (16:38 -0800)]
Remove dual prediction frame re-encoding loop.
I'm basically not convinced that the concept works at all, let alone
that this is the right place to do it. I think if we want something
like this at all, I should integrate it with the main encoding loop
and re-encode checks in onyx_if.c, and show that it has a significant
benefit (which right now, it doesn't; removing this re-encode check
actually increases all metrics by ~0.15%).
Yaowu Xu [Wed, 15 Feb 2012 01:52:18 +0000 (17:52 -0800)]
moved segment based LPF level selection under CONFIG_FEATUREUPDATES
This commit moved segment based loop filter level selection into
the experiment of CONFIG_FEATUREUPDATES. As previous commit noted,
the segment based loop filter selection helps the compression by
~0.1% on cif set, the ongoing experiment CONFIG_FEATUREUPDATES
made encoding updates of the segment based LPF level more efficient,
hence, another .04% gain on cif set. The commit also fixed an issue
previously where encoder/decoder may use different loop filter level
for one of the segments.
Yaowu Xu [Mon, 13 Feb 2012 22:34:19 +0000 (14:34 -0800)]
added 8x8 based Rate estimation for dualpred case
This commmit added logic for MB using dual-pred to compute rate
estimation based on correct transform size. The section of code
was previously located under #if CONFIG_DUALPRED, that was made
to be working with T8x8 experiment at the same time.
Jim Bankoski [Mon, 13 Feb 2012 16:21:24 +0000 (08:21 -0800)]
vp8 - config_featureupdates
Added a bit to signify that the feature changed since
the last time we sent it, or not so that we don't need
to send all the databits for every feature change.
Yaowu Xu [Mon, 6 Feb 2012 23:10:13 +0000 (15:10 -0800)]
Changed how coefficient probability table is updated
Added a frame level flag to indicate if coef probabilities are updated
at all for the frame.
During the experimental work with 8x8 transform, it is discovered that
even in the case of no probability is ever update, cost of transmitting
"no update" for each of probabilities can run up to become a significant
overhead cost. A single bit to indicate no-update for all coef probs
is therefore helpful, which is also demonstrated by the test results:
2. On HD set:
http://www.corp.google.com/~yaowu/no_crawl/t8x8/HD_t8x8_updprob.html
(avg psnr: .02% glb psnr: .01% SSIM: .02%)
It should be noted that the gain on HD is smaller because the average bit
rate is much higher in contrast to the overhead bit cost.
Yaowu Xu [Fri, 10 Feb 2012 19:49:22 +0000 (11:49 -0800)]
fixed an issue related to 2nd order size due to merge artifacts.
For 8x8 transformed macroblock, the 2nd order transform is a 2x2 haar
transform, here there is only 4 coefficients total. A previous merge
changed these to 64, causing crashes when encoding with 8x8 transform
enabled. (i.e. when input video image size > 640x360 ) This commit
reverts them back to 4 and fixes the crashes.
Paul Wilkins [Fri, 10 Feb 2012 16:02:10 +0000 (16:02 +0000)]
Removal of threading code.
For the experimental branch we are trying to slim the codebase
down removing features such as threading for now which complicate
the process of development and testing.
Ronald S. Bultje [Fri, 10 Feb 2012 00:12:23 +0000 (16:12 -0800)]
Improved coding using 8x8 transform
In summary, this commit encompasses a series of changes in attempt to
improve the 8x8 transform based coding to help overall compression
quality, please refer to the detailed commit history below for what
are the rationale underly the series of changes:
a. A frame level flag to indicate if 8x8 transform is used at all.
b. 8x8 transform is not used for key frames and small image size.
c. On inter coded frame, macroblocks using modes B_PRED, SPLIT_MV
and I8X8_PRED are forced to using 4x4 transform based coding, the
rest uses 8x8 transform based coding.
d. Encoder and decoder has the same assumption on the relationship
between prediction modes and transform size, therefore no signaling
is encoded in bitstream.
e. Mode decision process now calculate the rate and distortion scores
using their respective transforms.
Overall test results:
1. HD set
http://www.corp.google.com/~yaowu/no_crawl/t8x8/HD_t8x8_20120206.html
(avg psnr: 3.09% glb psnr: 3.22%, ssim: 3.90%)
2. Cif set:
http://www.corp.google.com/~yaowu/no_crawl/t8x8/cif_t8x8_20120206.html
(avg psnr: -0.03%, glb psnr: -0.02%, ssim: -0.04%)
It should be noted here, as 8x8 transform coding itself is disabled
for cif size clips, the 0.03% loss is purely from the 1 bit/frame
flag overhead on if 8x8 transform is used or not for the frame.
---patch history for future reference---
Patch 1:
this commit tries to select transform size based on macroblock
prediction mode. If the size of a prediction mode is 16x16, then
the macroblock is forced to use 8x8 transform. If the prediction
mode is B_PRED, SPLITMV or I8X8_PRED, then the macroblock is forced
to use 4x4 transform. Tests on the following HD clips showed mixed
results: (all hd clips only used first 100 frames in the test)
while the results are mixed and overall negative, it is interesting to
see 8x8 helped a few of the clips.
Patch 2:
this patch tries to hard-wire selection of transform size based on
prediction modes without using segmentation to signal the transform size.
encoder and decoder both takes the same assumption that all macroblocks
use 8x8 transform except when prediciton mode is B_PRED, I8X8_PRED or
SPLITMV. Test results are as follows:
Interestingly, by removing the overhead or coding the segmentation, the
results on this limited HD set have turn positive on average.
Patch 3:
this patch disabled the usage of 8x8 transform on key frames, and kept the
logic from patch 2 for inter frames only. test results on HD set turned
decidedly positive with 8x8 transform enabled on inter frame with 16x16
prediction modes: (avg psnr: .81% glb psnr: .82 ssim: .55%)
http://www.corp.google.com/~yaowu/no_crawl/t8x8/hdintermode8x8_0125.html
results on cif set still negative overall
Patch 4:
continued from last patch, but now in mode decision process, the rate and
distortion estimates are computed based on 8x8 transform results for MBs
with modes associated with 8x8 transform. This patch also fixed a problem
related to segment based eob coding when 8x8 transform is used. The patch
significantly improved the results on HD clips:
http://www.corp.google.com/~yaowu/no_crawl/t8x8/hd8x8RDintermode.html
(avg psnr: 2.70% glb psnr: 2.76% ssim: 3.34%)
results on cif also improved, though they are still negative compared to
baseline that uses 4x4 transform only:
http://www.corp.google.com/~yaowu/no_crawl/t8x8/cif8x8RDintermode.html
(avg psnr: -.78% glb psnr: -.86% ssim: -.19%)
Patch 5:
This patch does 3 things:
a. a bunch of decoder bug fixes, encodings and decodings were verified
to have matched recon buffer on a number of encodes on cif size mobile and
hd version of _pedestrian.
b. the patch further improved the rate distortion calculation of MBS that
use 8x8 transform. This provided some further gain on compression.
c. the patch also got the experimental work SEG_LVL_EOB to work with 8x8
transformed macroblock, test results indicates it improves the cif set
but hurt the HD set slightly.
Tests results on HD clips:
http://www.corp.google.com/~yaowu/no_crawl/t8x8/HD_t8x8_20120201.html
(avg psnr: 3.19% glb psnr: 3.30% ssim: 3.93%)
Test results on cif clips:
http://www.corp.google.com/~yaowu/no_crawl/t8x8/cif_t8x8_20120201.html
(avg psnr: -.47% glb psnr: -.51% ssim: +.28%)
Patch 6:
Added a frame level flag to indicate if 8x8 transform is allowed at all.
temporarily the decision is based on frame size, can be optimized later
one. This get the cif results to basically unchanged, with one bit per
frame overhead on both cif and hd clips.
Patch 8:
Rebase and Merge to head by PGW.
Fixed some suspect 4s that look like hey should be 64s in regard
to segmented EOB. Perhaps #defines would be bette.
Bulit and tested without T8x8 enabled and produces unchanged
output.
Patch 9:
Corrected misalligned code/decode of "txfm_mode" bit.
Limited testing for correct encode and decode with
T8x8 configured on derf clips.
Paul Wilkins [Thu, 9 Feb 2012 15:44:18 +0000 (15:44 +0000)]
Dual pred flag
Further changes to make experiments with the context
used for coding the dual pred flag easier.
Current best performing method tested on derf is a two
element context based on reference frame. I also tried
various combinations of mode and reference frame as
shown in commented out case using up to 6 contexts.
Derf +0.26 overall psnr +0.15% ssim vs original method.
We should only change the dual prediction mode if we actually entered
the recode branch. Else, it may potentially undo beneficial changes
to the dual prediction mode in the first encode iteration.
Some conditions were conditional under a threshold, whereas they should
always execute. Also, some conditions were testing an array instead of
the values within it.
Paul Wilkins [Fri, 3 Feb 2012 17:08:37 +0000 (17:08 +0000)]
Move update of ref frame probabilities in encode loop.
The existing code updated the reference frame probabilities before
the test to evaluate the impact of using updated probabilities
in vp8_estimate_entropy_savings().
The estimate of cost and savings is still basic and does not reflect
the new prediction code but this would require per MB costings
and the benefit is probably marginal, as this is really just used for
rate estimation in the loop.
Paul Wilkins [Fri, 3 Feb 2012 15:10:44 +0000 (15:10 +0000)]
Merged NEWNEAR experiment
This commit merges the NEWNEAR experiment such that it
is effectively always on.
The fact that there were changes in the threading code again
highlights the need to strip out such features during the
bitstream development phase as trying to maintain this code
(especially as it is not being tested) slows the development cycle.
Paul Wilkins [Thu, 2 Feb 2012 17:30:27 +0000 (17:30 +0000)]
Coding the hybrid dual prediction signal.
Initial modifications to make limited use of common prediction
functions.
The only functional change thus far is that updates to the probabilities are
no longer "damped". This was a testing convenience but in fact seems to
help by a little over 0.1% over the derf set.
Paul Wilkins [Fri, 3 Feb 2012 13:46:18 +0000 (13:46 +0000)]
Modified prediction behavior for reference frame.
Trial of a modified prediction function that ranks each possible
reference frame based on a combination of local usage and
frame level probability. The code is a bit cleaner and simpler.
In direct comparison with old unpredicted method with segment level
coding turned off for mode,ref & EOB the prediction gives a gain on derf
of around 0.4%. There is some further gain from bug fixes over earlier code.
With segment coding on the prediction method is slightly -ve on some very
easy clips (at low rates) due to slightly higher overheads, but better on harder
clips. Overall neutral on derf in direct comparison on latest code base, but
compared to earlier code without bug fixes about +0.7% overall psnr
+0.3% SSIM.
Adrian Grange [Thu, 2 Feb 2012 17:04:40 +0000 (09:04 -0800)]
Added encoding in Superblock Order
As a precursor to encoding 32x32 blocks this cl adds the
ability to encode the frame superblock (=32x32 block) at
a time. Within a SB the 4 indiviual MBs are encoded in
raster-order (NW,NE,SW,SE).
This functionality is added as an experiment which can be
enabled by ispecifying --enable-superblocks in the
command line specified to configure (CONFIG_SUPERBLOCKS
macro in the code).
To make this work I had to disable the two intra
prediction modes that use data from the top-right of the
MB.
On the tests that I have run the results produce
almost exactly the same PSNRs & SSIMs with a very
slightly higher average data rate (and slightly higher
data rate than just disabling the two intra modes in
the original code).
NOTE: This will also break the multi-threaded code.
This replaces the abandoned change:
Iebebe0d1a50ce8c15c79862c537b765a2f67e162
Yaowu Xu [Wed, 1 Feb 2012 22:49:10 +0000 (14:49 -0800)]
Fixes a decoder bug
The bug was introduced by the commit that added I8X8 intra prediction
mode for inter frames, the decoder was not update to accept the additional
probability update from encoder. This causes the decoder typicall to crash
when encoder sends intra mode probability update.
Paul Wilkins [Sat, 28 Jan 2012 10:07:08 +0000 (10:07 +0000)]
Added common prediction modules.
This function adds the common prediction modules, some data structures
and a config option but does not use them.
It also corrects a bug in clearing down the MODE_INFO border and introduces
a new element that indicates if an entry corresponds to an "in image" macro block
or is part of the border.
Deb Mukherjee [Mon, 30 Jan 2012 17:27:44 +0000 (09:27 -0800)]
Refining the 8-tap interpolation filters.
Fixes a rounding issue with the 8-tap filters, which allows us to
use sharper filters that before with a little better performance.
Results on derf (0verall gain 0.76):
http://www.corp.google.com/~debargha/vp8_results/enhinterp.html
Results on a 720P set (Overall gain 0.61):
http://www.corp.google.com/~debargha/vp8_results/enhinterp_hd.html
Yaowu Xu [Thu, 26 Jan 2012 22:36:20 +0000 (14:36 -0800)]
fixed an issue with 8x8 token cost in trellisquant
changed the token cost for 8x8 transformed macroblock used in trellisquant
from those derived from 4x4 transform coefficient distribution to those
derived from 8x8 transform coefficient distribution. Test results show
this fix help 8x8 transform based compression consistently on cif and hd
sets:
Note: To test the effect of this change, 8x8 transform was forced to be used
only on 16x16 predicted macroblocks on inter frames, the effect would be
bigger had all macroblocks been forcd to use 8x8 transform.
Deb Mukherjee [Mon, 16 Jan 2012 19:21:12 +0000 (11:21 -0800)]
Adds support for enhanced interpolation for subpel motion
using an 8-tap filter.
The results with 3 different 8-tap filters on the derf set are in:
http://www.corp.google.com/~debargha/vp8_results/enhinterp.html
The one that gives the most gain achieves an overall gain of about
0.6%. The results for a set of 12 hd (720p) videos are in:
http://www.corp.google.com/~debargha/vp8_results/enhinterp_hd.html
with max gain of 0.55% with the same filter. The best filter apparently
achieves the best trade-off between pass band ripple and stop band
attenuation.
Jim Bankoski [Mon, 23 Jan 2012 17:55:23 +0000 (09:55 -0800)]
vpn common -> implicit segmentation
This introduces base functions for introducing implicit segmentation.
The code that actually stores the results to the segment map isn't
here yet. This just prints out the segmentation map results
if you call it.
Uses connected component labeling technique on mbmi info so that only
if 2 mbs are horizontally or vertically touching do they get the same
segment.
vp8next - plumbing for rotation
code to produce taps for rotation ( tapify. py ), code
for predicting using rotation ( predict_rotated.c ) , code
for finding the best rotation find_rotation.c.
didn't checkin code that uses this in the codec. still work
in progress.
Yaowu Xu [Wed, 30 Nov 2011 16:05:45 +0000 (08:05 -0800)]
changed loop filter for MBs using 8x8 transform
This commit added a set of loop filter functions for macroblocks
using 8x8 transform. First we turned off the regular loop filtering
on 4x4 block boundaries that do not exist in macroblocks using 8x8
transform. Second, we change to use the same loop filter(mask and
7 tap filter) that used for macroblock edge filtering.
Yaowu Xu [Fri, 20 Jan 2012 00:37:32 +0000 (16:37 -0800)]
Added code to prevent I8X8_PRED mode for MBs using 8x8 transform
This fixed a conflict introduced by the change of adding 8x8 intra
prediction modes. The 8x8 intra prediction mode code assumed the
use of 4x4 transform, and causes encoder crashes when the codec is
configured with --enable-t8x8.
Yaowu Xu [Fri, 13 Jan 2012 22:37:17 +0000 (14:37 -0800)]
new loop filter functions for macroblock boundaries
The commit adds a new set of loop filter for macroblock edge filtering.
The new loop filter has a mask to detect so-called "flat" regions. The
detection checks 5 pixels of each side of an edge. If the all pixels
have value with +/-1 from the edge pixel on the same side, the region
is treated as a "flat" region. For such case, a 7 tap filter is used
to change 3 pixel values on each side. The 7 taps are:
[1, 1, 1, 2, 1, 1, 1]/8
The furthest away pixels used as input are +/-5 away from edge. For
non-flat region, we fall back to old filtering. It should be noted
here that the thresholds and filter taps may require more optimization
for best possible results.
Tests on a set of hd clips showed consistent gains:
http://www.corp.google.com/~yaowu/no_crawl/mblpf_hd.html
(avg psnr: .83% glb psnr: .77% ssim: .82%)
Tests on derf set also showed consistent gains:
http://www.corp.google.com/~yaowu/no_crawl/mblpf_derf.html
(avg psnr: .24% glb psnr: .22% ssim: .48%)
Paul Wilkins [Wed, 11 Jan 2012 14:05:57 +0000 (14:05 +0000)]
Rate control on static scenes plus Y2dc delta Q fix.
A problem can arise on static clips with force key frames where
attempts to avoid popping lead to a progressive reduction in key
frame Q that ultimately may lead to unexpected overspend against
the rate target.
The changes in this patch help to insure that in such clips the
quality of the key frames across the clip is more uniform (rather
than starting bad and getting better - especially at low target rates).
This patch also includes a fix that removes a delta on the Y2DC
when the baseline q index < 4 as this is no longer needed.
There is also a fix to try and prevent repeat single step Q adjustment in
the recode loop leading to lots of recodes, especially where the use
of forced skips as part of segmentation has made the impact of Q on
the number of bits generated much smaller.
Patch 2: Amend "last_boosted_qindex" calculation for arf overlay frames.
Yaowu Xu [Thu, 5 Jan 2012 19:22:06 +0000 (11:22 -0800)]
Added an emms to prevent invalid stats output
In certain hardware configuration, where mmx code is enabled and
other simd (sse2/sse3) disabled, lacking of this emms caused invalid
internal stats outputs.
Yaowu Xu [Wed, 21 Dec 2011 20:05:10 +0000 (12:05 -0800)]
changed mode_context update strategy
Previously, the mode context is always udpated based on stats of current
frame, when there is no count, 50% is used for both left and right branch.
However, it is observed that with such strategy, a small count or no count
at all can skew the probability distribution significantly. This commmit
changed the mode_context update strategy to prevent small counts from
skewing the probability distributions.
Tests on derf set showed a small gain: .06% in psnr and .09% in ssim
Paul Wilkins [Fri, 16 Dec 2011 10:23:01 +0000 (10:23 +0000)]
Extend to 256 Q steps.
This commit extends the number of Q steps to 256 from 128.
The q_trans[] array has been altered to distribute available Q index values
(using the current 64 steps available as input parameters) evenly across the
available range. This is coupled with the fact that each Q step where possible
now equates to a fixed % change in the quantizer. This may want refinement
later especially in terms of the granularity at the high quality end but is a
reasonable starting point.
Paul Wilkins [Mon, 12 Dec 2011 18:27:25 +0000 (18:27 +0000)]
Further QIndex realted Fixes:
Added code to support 256 index steps instead of 128 but disabled for now.
Replace hard wired table vp8cx_base_skip_false_prob[128]
Observed Qindex problem with setting minimum loop filter value.
(Experiment code using real Q in place but for now just returning 0. This has a big
beneficial effect on some clips, particularly waterfall which shows 5% ssim gain)
Adrian Grange [Fri, 16 Dec 2011 22:00:36 +0000 (14:00 -0800)]
Reset segment_id to 0 when segmentation is disabled
Whilst the encoder explicitly set the segment_id to 0
when segmentation is diabled, the decoder would allow
the segment_id to persist from the previous frame.
This fix attempts to make the decoder behave the same
as the encoder by explicitly setting the segment_id to
0 in this case.
Adrian Grange [Thu, 15 Dec 2011 18:00:46 +0000 (10:00 -0800)]
Fixed bug to use mode_info_stride rather than mb_cols
Both encoder & decoder were using mb_cols to
offset from one row of MODE_INFO structures to the next
when they should have been using mode_info_stride.
Fixing this in both encoder and decoder gives around
a 3KB size saving and 0.025dB PSNR improvement on the one
720P clip I tried.
(Also removed "index" which was being updated but not used)
Yaowu Xu [Thu, 8 Dec 2011 19:43:09 +0000 (11:43 -0800)]
Enabled adaptive UV intra coding for inter frames
Previously, Y-adaptive UV intra coding only enabled on key frames in
UVINTRA experiment. This commit enabled the same coding for inter
frames, so the encoding of UV intra modes are consistent cross all
frame types. Tests on derf set showed a very small overall gain around
.04%:
Yaowu Xu [Thu, 8 Dec 2011 19:43:09 +0000 (11:43 -0800)]
Revised coding using adaptive mode context to depend on frame type
A previous commit 76feb965 made the vp8_mode_context adaptive on a frame
frame basis, this commit further made the coding context adaptive to two
frame types separately. Tests on derf set showed a further small gain on
all metrics: avg psnr 0.10%, glb psnr: 0.11%, ssim: 0.08%
Yaowu Xu [Thu, 8 Dec 2011 04:08:31 +0000 (20:08 -0800)]
fixed a crash caused invalid Q choice
The commit fixed a problem by capping cpi->active_best_quality to be
smaller than cpi->worst_quality. Also fixed a few line of code that
was misplaced.
Yaowu Xu [Wed, 7 Dec 2011 21:55:58 +0000 (13:55 -0800)]
Removed #if CONFIG_MULCONTEXT
This commit removed the macro CONFIG_MULCONTEXT, which was used to
indicate the experiment code for using separate context for altref
and normal frames. This commit made the change fully merged in.
Yaowu Xu [Tue, 6 Dec 2011 20:03:42 +0000 (12:03 -0800)]
made vp8_mode_context adaptive
vp8_mode_contexts[] is an entropy table used to code inter mode
choices. It was a fixed constant table. This commit made the entropy
context adaptive. Tests on derf set showed very good consistent gains
on all metrics: avg psnr .47%, overall psnr .46% and ssim .40%.
Yaowu Xu [Tue, 6 Dec 2011 15:37:39 +0000 (07:37 -0800)]
fixed a crash when MODE_STATS is enabled
The MODE_STATS macro was used to #ifdef around code for mode entropy
stats collection, this commit fixed a crash when MODE_STATS is on.
The commit also changed a number of array definitions to use defined
macros instead of hard-coded numbers.