Hui Su [Thu, 14 Jun 2012 02:03:31 +0000 (19:03 -0700)]
Add lossless compression mode.
This commit adds lossless compression capability to the experimental
branch. The lossless experiment can be enabled using --enable-lossless
in configure. When the experiment is enabled, the encoder will use
lossless compression mode by command line option --lossless, and the
decoder automatically recognizes a losslessly encoded clip and decodes
accordingly.
To achieve the lossless coding, this commit has changed the following:
1. To encode at lossless mode, encoder forces the use of unit
quantizer, i.e, Q 0, where effective quantization is 1. Encoder also
disables the usage of 8x8 transform and allows only 4x4 transform;
2. At Q 0, the first order 4x4 DCT/IDCT have been switched over
to a pair of forward and inverse Walsh-Hadamard Transform
(http://goo.gl/EIsfy), with proper scaling applied to match the range
of the original 4x4 DCT/IDCT pair;
3. At Q 0, the second order remains to use the previous
walsh-hadamard transform pair. However, to maintain the reversibility
in second order transform at Q 0, scaling down is applied to first
order DC coefficients prior to forward transform, and scaling up is
applied to the second order output prior to quantization. Symmetric
upscaling and downscaling are added around inverse second order
transform;
4. At lossless mode, encoder also disables a number of minor
features to ensure no loss is introduced, these features includes:
a. Trellis quantization optimization
b. Loop filtering
c. Aggressive zero-binning, rounding and zero-bin boosting
d. Mode based zero-bin boosting
Lossless coding test was performed on all clips within the derf set,
to verify that the commit has achieved lossless compression for all
clips. The average compression ratio is around 2.57 to 1.
(http://goo.gl/dEShs)
Adrian Grange [Mon, 25 Jun 2012 23:23:58 +0000 (16:23 -0700)]
Added Prediction Filter to Mode Selection
Added the ability to optionally filter the prediction data
when inter modes are selected (excludes SPLITMV, for now).
The mode selection loop considers both the filtered and
non-filtered prediction data when choosing mode. The filter
can be turned on/off at the frame-level, or signaled for
each MB.
Deb Mukherjee [Tue, 5 Jun 2012 22:25:07 +0000 (15:25 -0700)]
Adaptive entropy coding of coefficients, modes, mv.
This patch incorporates adaptive entropy coding of coefficient tokens,
and mode/mv information based on distributions encountered in a frame.
Specifically, there is an initial forward update to the probabilities
in the bitstream as before for coding the symbols in the frame, however
at the end of decoding each frame, the forward update to the
probabilities is reverted and instead the probabilities are updated
towards the actual distributions encountered within the frame.
The amount of update is weighted by the number of hits within each
context.
Results on derf/hd/std-hd are all up by 1.6%.
On derf, the most of the gains come from coefficients, however for the
hd and std-hd sets, the most of the gains come from the mode/mv
information updates.
Yaowu Xu [Thu, 24 May 2012 19:02:12 +0000 (12:02 -0700)]
changed to use a fixed number for update probabity
This commit is to remove two arrays, which contain the probabilities
of how likely each probability in coef_probs table is updated. The
commit changed to use a fixed number "252".
Surprisedly, the overall impact on quality is close to zero, which
basically says the two big static arrays are not helpful at all.
derf: -0.016%, -0.020%
std-hd: 0.000%, -0.013%
yt: -0.022%, +0.007%
yt-hd: -0.038%, +0.034%
Yaowu Xu [Tue, 8 May 2012 19:38:39 +0000 (12:38 -0700)]
Added the ability to accumulate coef stats across encodings
This commit added the ability to accumulate the coef stats across
different encodings using an intermediate binary stats files. The
accumulation happens only the binary stats file exists in current
directory. The encoder needs to be built with "ENTROPY_STATS" to
allow the output. The commit also fixed a few formating issues in
output stats file.
Paul Wilkins [Wed, 30 May 2012 11:54:34 +0000 (12:54 +0100)]
Merge of further two pass rc clean up and adjustments.
Changes to calculation of sr_coded_error to include 0,0 case.
Experimental use of sr_coded_error in calculating correction factor
for estimating the allowable Q range.
Reinstated some code needed for calculating section_intra_rating.
Add flash detection in calculation of KF boost
Increased tolerance in testing candidate key frames (needed with
longer motion search as this tends to slightly increase inter %.
Zbin changes for 8x8.
Other minor adjustments, refactoring and bug fixes.
Reinstated some motion break out clauses in boost loop
as their removal hurt a few 50fps clips badly in the std set.
It may be possible to remove them again later if a better way
can be found of preventing overly long gf intervals.
Paul Wilkins [Wed, 30 May 2012 10:29:49 +0000 (11:29 +0100)]
Rate control fix.
This fix addresses some problems with very complex clips like
handling of flashes on clips like crew (which was made worse
by an earlier patch (derf and std-hd)).
Yaowu Xu [Wed, 16 May 2012 19:07:46 +0000 (12:07 -0700)]
changed the way that default probs for 8x8 is set.
The commit changed how baseline 8x8 coefficient probabilities are
initialized, to be consistent with the initialization of baseline
4x4 coefficient probabilities.
The commit does not have any effect on compression.
Ronald S. Bultje [Tue, 15 May 2012 00:39:42 +0000 (17:39 -0700)]
Rewrite reference frame costing in the RD loop.
I now see I didn't write a very long description, so let's do it
here then. We took a pretty big quality hit (0.1-0.2%) from my
recent fix of the inversion of arguments to vp8_cost_bit() in the
RD reference frame costing. I looked into it and basically the
costing prevented us from switching reference frames. This is of
course silly, since each frame codes its own prob_intra_coded, so
using last frame cost indications as a limiting factor can never
be right.
Here, I've rewritten that code to estimate costings based partially
on statistics from progress on current frame encoding. Overall,
this gives us a ~0.2%-0.3% improvement over what we had previously
before my argument-inversion-fix, and thus about ~0.4% over current
git (on derf-set), and a little more (0.5-1.0%) on HD/STD-HD/YT.
Yaowu Xu [Wed, 9 May 2012 16:31:14 +0000 (09:31 -0700)]
Changed to use integer 8x8 dct
The commit added an integer version of 8x8 forward DCT, based on the
orginal forward DCT from VP6. The constants, roundings, and shifts
were adjusted to improve the accuracy. The latest patch has a very
similar accuracy in term of round trip error against the floating
point version.
It should be noted here that the purpose of the patch is to help
encoding speed and facilitate all other experiments. There will be
futher review in combination with inverse DCT before finalization.
configure with "--enable--int_8x8fdct" to use the integer version
Paul Wilkins [Tue, 15 May 2012 13:58:13 +0000 (14:58 +0100)]
Firstpass.c refactoring
Removed unused function.
Added tentative code to take error score of an older frame
into account when calculating Q range. However, for now
it is disabled pending merging other changes and testing.
Paul Wilkins [Mon, 14 May 2012 14:13:26 +0000 (15:13 +0100)]
Two pass refactoring continued.
Remove testing of whether we estimate that it will be possible
to code an arf at a lower Q than the ambient Q. This adds quite
a bit of extra code and complexity for marginal gain.
Factored out some code relating to ARNR selection to a separate
function as this is likely to be changed / simplified soon.
Paul Wilkins [Mon, 14 May 2012 13:54:54 +0000 (14:54 +0100)]
Two pass rc refactoring.
Removed odd *100 >> 4 factor from boost calculations. Not all the
calculations exactly match what was there before so there may be
some minor impact on results.
Some other minor tidying up in regard to coding conventions.
The specific values of factors and thresholds will likely change as
part of subsequent patches.
Ronald S. Bultje [Sat, 12 May 2012 00:48:20 +0000 (17:48 -0700)]
Don't use compound prediction for golden frames based on alt-ref frames.
These frames can force reference frame (arf), mode (zeromv) and skip,
which means that if we use compound prediction (i.e. arf+last), we
might use a blend of a perfect (arf) and an imperfect (last) predictor,
leading to semi-garbage display and thus a huge drop in SSIM/PSNR (up
to 10dB for some frames I analyzed).
Paul Wilkins [Fri, 11 May 2012 17:07:33 +0000 (18:07 +0100)]
First pass overhaul preparatory change.
This is the first patch in a series of changes to the first
pass code. (Broken down for ease of testing/merging/review).
This patch introduces a new stats element "sr_coded_error".
This is the coded error recorded vs the second reference
frame (which is updated such that it lags by at least one frame).
No use is made of the new structure in this change so this patch
should have no material effect.
Removed some ifdefs and deprecated code (#if NEW_BOOST).
Removed twopass.gf_decay_rate (not used any more)
Yaowu Xu [Thu, 10 May 2012 00:19:16 +0000 (17:19 -0700)]
Reversible WHT pair
This commit changed the forward and the inverse 4x4 Walsh Hadamard
transform to a new pair, where the inverse transform can pefectly
reconstuct the input to forward transform. It also does so without
changing the input and output value range. Even more, it does not
change the complexity of the transforms.
While it was not expected to improve the results of our current test,
it does improve std-hd set by 0.2% on all metrics. No change on derf.
Yaowu Xu [Wed, 2 May 2012 20:53:15 +0000 (13:53 -0700)]
a number of fixes to entropy stats collection
1. block types
There are only three types of blocks for 8x8 transformed MBs, i.e. Y
block with DC does not exist for 8x8 transformed MBs as all MB using
8x8 transform have 2nd order haar transform. This commit introduced
a new macro BLOCK_TYPES_8X8 to reflect such fact.
2. context counters
This commit also fixed the mixed of context_counters between 4x4 and
8x8 transformed MBs. The mixed use of the counters leads me to think
the existing the context probabilities were not properly generated
from 8x8 transformed MBs.
3. redundant collecting in recoding
The commit also corrected the code that accumulates entropy stats by
making sure stats only collected for final packing, not during the
recode loop
Deb Mukherjee [Wed, 21 Mar 2012 22:22:21 +0000 (15:22 -0700)]
Expanding the coefficient encoding contexts
This patch expands the set of prev contexts used for video coding
from 3 to 4.
There is a small improvement of the order of 0.08% for derf and
0.15% on the HD set. The tests were rerun after the various merges
last week. There are two columns in each test - the first are the
results with the mbskip change, and the second with expanded contexts
added on top of that.
Yaowu Xu [Wed, 2 May 2012 02:54:19 +0000 (19:54 -0700)]
slight adjustment to coef band definition
This commit adjusted slightly the 4x4 coefficents band definition to
better classify coefficients with similar distributions and usages.
It helps derf set about .1%, it is alos slightly positive for std-hd
set, where 4x4 blocks are used less frequently.
Adrian Grange [Thu, 26 Apr 2012 17:37:52 +0000 (10:37 -0700)]
Removed MV costing from ARNR filtering
The ARNR filter uses a motion compensated temporal filter,
but the motion estimation implementation accounts for the
cost of the mv in its decision making process. The ARNR
filter uses a dummy cost table initialized to 0 as a way
to ignore the mv costs (which are irrelevant to the filter).
This CL modifies the ARNR filter implementation to so that
the mv costing is ignored without the requirement for
dummy tables.
This commit merge the QI mode experiment. As the experiment affects
the encoding of intra coding modes on key frame only, the overall
effect of the experiment on encoding tests is insignificant.
Deb Mukherjee [Thu, 12 Apr 2012 16:24:03 +0000 (09:24 -0700)]
Differential encoding of probability updates
Adds differential encoding of prob updates using a subexponential
code centered around the previous probability value.
Also searches for the most cost-effective update, and breaks
up the coefficient updates into smaller groups.
This commit changed to enable the usage 8x8 transform for all frame
type, all resolution and all quantizer range. This has an overall
benefit .2% to .3% in term of compression, but more importantly,
the difficult clips benefits much more, up to 2% to 3% on clips
like football, harbour and so on.
We observed some weird humps on very high end on a couple of youtube
clips, but have determined the underly cause was the aggressive zbin
having an effect of lowering rate with lower quality, which have
an impact on slide show clips around 60DB.
The commit does not change the association between prediction mode
and transform size.
Deb Mukherjee [Wed, 11 Apr 2012 21:32:20 +0000 (14:32 -0700)]
Turning off interpolation filter selection
Turning off the interpolation filter selection based on edge
proportion. This heuristics has not been working as well as
expected and I have started a more rigorous investigation into
this. We can turn this off for now since it is unnecessarily
slowing things down.
This commit tries to address an issue related to the oddity shown on
HD _mobcal clip, where some rather ugly blocks shown in the second
frame at low-mid bit rates if the third frame is not made a key frame
by he encoder. The fixes include: 1) made calls to sad_16x16 to be
consistent with function prototype. 2) remove the error bias to intra
and golden in mbgraph search. 3) changed the error accumulation on
inter_segment encoding to avoid potential out-of-range. 1) has no
effect on encoding results.
Encoding test show that the overall effect of the commit helps about
.2%(HD) to .3%(cif)
The commit removed a limit on key frame detection, which caused a big
drop in all metric measurements for standard HD clip such as _mobcal.
This single change helps two standard HD clips by a huge amount, which
help the overall std-hd set by 2.4% (glb psnr), 0.9% (avg_psnr), 2.1%
(vpxssim).
In the result page:
http://pafr9.prod.google.com:26163/?/cns/rc-d/home/on2-prod/sunkaras/borg-test/yaowu
2012_04_02_1649_yaowu_bugfix_std-hd
2012_04_03_1452_yaowu_hump_std-hd
represent the encoding test results and std-hd set prior and after this
commit respectively.
This commit has made macro_block_yrd_8x8 and macro_block_yrd_8x8 to
take same parameters. It also removed a few unnecessary shifts that
has the potential to create out-of-range distortion values.
Paul Wilkins [Tue, 27 Mar 2012 12:01:34 +0000 (13:01 +0100)]
T8x8 zbin and rate control changes.
Some adjustments to zbin for t8x8.
Changes to rules for sizing forced key frames.
Some extra stats output in tmp.stt.
Approximate gain on YT-hd set 0.5%
There are still issues in sizing key frames and gf/arf frames
when the image is largely static. These in part relate to
problems with cost estimates in the recode loop.
Adrian Grange [Fri, 6 Apr 2012 23:38:34 +0000 (16:38 -0700)]
Superblock encoding order
This is the first patch to add superblock (32x32) coding
order capabilities. It does not yet do any mode selection
at the SB level, that will follow in a further patch.
This patch encodes rows of SBs rather than
MBs, each SB contains 2x2 MBs.
Two intra prediction modes have been disabled since they
require reconstructed data for the above-right MB which
may not have been encoded yet (e.g. for the bottom right
MB in each SB).
Results on the one test clip I have tried (720p GIPS clip)
suggest that it is somewhere around 0.2dB worse than the
baseline version, so there may be bugs.
It has been tested with no experiments enabled and with
the following 3 experiments enabled:
--enable-enhanced_interp
--enable-high_precision_mv
--enable-sixteenth_subpel_uv
in each case the decode buffer matches the recon buffer
(using "cmp" to compare the dumped/decoded frames).
Note: Testing these experiments individually created
errors.
Some problems were found with other experiments but it
is unclear what state these experiments are in:
--enable-comp_intra_pred
--enable-newentropy
--enable-uvintra
This code has not been extensively tested yet, so there
is every likelihood that further bugs remain. I also
intend to do some code cleanup & refactoring in tandem
with the next patch that adds the 32x32 modes.
Deb Mukherjee [Wed, 28 Mar 2012 22:19:45 +0000 (15:19 -0700)]
Bug fix in probability update savings computation
Found this bug while tracking down some anomalies in my experiments.
Since vp8_cost_one and vp8_cost_zero return unsigned int, the
bit shift by 8 will be incorrect if the value is negative.
I am cautiously optimistic that this fix will make the prob
updates more correct and somewhat improve results across the board.
But the update probabilities will need to be retuned I think.
Patch 2: Adding more of the same fixes using a macro.
Yaowu Xu [Mon, 19 Mar 2012 18:03:27 +0000 (11:03 -0700)]
enable 8x8 transform for MBs in intra frames
When ac_yquant>171, a key frame is enabled to use 8x8 transform. In
such case, MBs with DC_PRED or TM_PRED are selected to use T8x8. This
change helped the full STD-HD set by ~.1% or so, which is reasonable
considering how often key frame occurs in these encodings.
projects / libvpx / log