The 2-D inverse transform X = M1*Z*Transposed_M2 was calculated
in 2 steps from left to right:
1. Vertical transform: Y = M1*Z
2. Horizontal transform: X= Y*Transposed_M2
In SIMD, a transpose is needed in vertical transform.
Here, switched the calculation order to do it from right to left.
In this way, we could eliminate that transpose by writing the
intermediate results out to their transposed positions.
Change-Id: I34dfe5eb01292f6e363712420d99475e2e81e12c
Various fixups to resolve issues when building vp9-preview under the more stringent
checks placed on the experimental branch.
Change-Id: I21749de83552e1e75c799003f849e6a0f1a35b07
Adds an experiment to derive the previous context of a coefficient
not just from the previous coefficient in the scan order but from a
combination of several neighboring coefficients previously encountered
in scan order. A precomputed table of neighbors for each location
for each scan type and block size is used. Currently 5 neighbors are
used.
Results are about 0.2% positive using a strategy where the max coef
magnitude from the 5 neigbors is used to derive the context.
Change-Id: Ie708b54d8e1898af742846ce2d1e2b0d89fd4ad5
For coefficients, use int16_t (instead of short); for pixel values in
16-bit intermediates, use uint16_t (instead of unsigned short); for all
others, use uint8_t (instead of unsigned char).
Change-Id: I3619cd9abf106c3742eccc2e2f5e89a62774f7da
MSVC 2012 (_MSC_VER=1600) introduced the definition, this commit
prevents the redefinition of the macro
Change-Id: I7de92e7e9e865a342f2bcc4b071f8d3c9b2a508c
Modifies the scanning pattern and uses a floating point 16x16
dct implementation for now to handle scaling better.
Also experiments are in progress with 2/6 and 9/7 wavelets.
Results have improved to within ~0.25% of 32x32 dct for std-hd
and about 0.03% for derf. This difference can probably be bridged by
re-optimizing the entropy stats for these transforms. Currently
the stats used are common between 32x32 dct and dwt/dct.
Experiments are in progress with various scan pattern - wavelet
combinations.
Ideally the subbands should be tokenized separately, and an
experiment will be condcuted next on that.
Change-Id: Ia9cbfc2d63cb7a47e562b2cd9341caf962bcc110
Gives 0.5-0.6% improvement on derf and stdhd, and 1.1% on hd. The
old tables basically derive from times that we had only 4x4 or
only 4x4 and 8x8 DCTs.
Note that some values are filled with 128, because e.g. ADST ever
only occurs as Y-with-DC, as does 32x32; 16x16 ever only occurs
as Y-with-DC or as UV (as complement of 32x32 Y); and 8x8 Y2 ever
only has 4 coefficients max. If preferred, I can add values of
other tables in their place (e.g. use 4x4 2nd order high-frequency
probabilities for 8x8 2nd order), so that they make at least some
sense if we ever implement a larger 2nd order transform for the
8x8 DCT (etc.), please let me know
Change-Id: I917db356f2aff8865f528eb873c56ef43aa5ce22
As suggested by Yaowu, we can use eob to reduce the complexity
of the vp9_ihtllm_c function. For the 1080p test clip used, the decoder
performance improved by 17%.
Change-Id: I32486f2f06f9b8f60467d2a574209aa3a3daa435
Add a function clip_pixel() to clip a pixel value to the [0,255] range
of allowed values, and use this where-ever appropriate (e.g. prediction,
reconstruction). Likewise, consistently use the recently added function
clip_prob(), which calculates a binary probability in the [1,255] range.
If possible, try to use get_prob() or its sister get_binary_prob() to
calculate binary probabilities, for consistency.
Since in some places, this means that binary probability calculations
are changed (we use {255,256}*count0/(total) in a range of places,
and all of these are now changed to use 256*count0+(total>>1)/total),
this changes the encoding result, so this patch warrants some extensive
testing.
Change-Id: Ibeeff8d886496839b8e0c0ace9ccc552351f7628
Some further changes and refactoring of mv
reference code and selection of center point for
searches. Mainly relates to not passing so many
different local copies of things around.
Some place holder comments.
Change-Id: I309f10ffe9a9cde7663e7eae19eb594371c8d055
This commit changed the ENTROPY_CONTEXT conversion between MBs that
have different transform sizes.
In additioin, this commit also did a number of cleanup/bug fix:
1. removed duplicate function vp9_fix_contexts() and changed to use
vp8_reset_mb_token_contexts() for both encoder and decoder
2. fixed a bug in stuff_mb_16x16 where wrong context was used for
the UV.
3. changed reset all context to 0 if a MB is skipped to simplify the
logic.
Change-Id: I7bc57a5fb6dbf1f85eac1543daaeb3a61633275c
Use these, instead of the 4/5-dimensional arrays, to hold statistics,
counts, accumulations and probabilities for coefficient tokens. This
commit also re-allows ENTROPY_STATS to compile.
Change-Id: If441ffac936f52a3af91d8f2922ea8a0ceabdaa5
This adds Debargha's DCT/DWT hybrid and a regular 32x32 DCT, and adds
code all over the place to wrap that in the bitstream/encoder/decoder/RD.
Some implementation notes (these probably need careful review):
- token range is extended by 1 bit, since the value range out of this
transform is [-16384,16383].
- the coefficients coming out of the FDCT are manually scaled back by
1 bit, or else they won't fit in int16_t (they are 17 bits). Because
of this, the RD error scoring does not right-shift the MSE score by
two (unlike for 4x4/8x8/16x16).
- to compensate for this loss in precision, the quantizer is halved
also. This is currently a little hacky.
- FDCT and IDCT is double-only right now. Needs a fixed-point impl.
- There are no default probabilities for the 32x32 transform yet; I'm
simply using the 16x16 luma ones. A future commit will add newly
generated probabilities for all transforms.
- No ADST version. I don't think we'll add one for this level; if an
ADST is desired, transform-size selection can scale back to 16x16
or lower, and use an ADST at that level.
Additional notes specific to Debargha's DWT/DCT hybrid:
- coefficient scale is different for the top/left 16x16 (DCT-over-DWT)
block than for the rest (DWT pixel differences) of the block. Therefore,
RD error scoring isn't easily scalable between coefficient and pixel
domain. Thus, unfortunately, we need to compute the RD distortion in
the pixel domain until we figure out how to scale these appropriately.
Change-Id: I00386f20f35d7fabb19aba94c8162f8aee64ef2b
Only declare the functions in vpx_scale RTCD and include the relevant
header.
Remove unused files and functions in vpx_scale to avoid wasting time
renaming. vpx_scale/win32/scaleopt.c contains functions which have not
been called in a long time but are potentially optimized.
The 'vp8' functions have not been renamed yet. That is for after the
cleanup.
Change-Id: I2c325a101d60fa9d27e7dfcd5b52a864b4a1e09c
This patch reduces the cpu cost of the MV ref
search by only allowing insert for candidates
that would be in the current top 4.
This could alter the outcome and slightly favors
near candidates which are tested first but also
limits the worst case loop count to 4 and means in
many cases it will drop out and not happen.
Change-Id: Idd795a825f9fd681f30f4fcd550c34c38939e113
Only declare the functions in vpx_scale RTCD and include the relevant
header.
Remove unused files and functions in vpx_scale to avoid wasting time
renaming. vpx_scale/win32/scaleopt.c contains functions which have not
been called in a long time but are potentially optimized.
The 'vp8' functions have not been renamed yet. That is for after the
cleanup.
Change-Id: I2c325a101d60fa9d27e7dfcd5b52a864b4a1e09c
Adds support for compound inter-intra prediction with superblocks.
Also, fixes a bug that disabled intra modes for superblocks.
Change-Id: I4d711317e1bc19df8c2f32dc645429f7fff31036
Allows switchbale filters to be used without mismatch when the
superblock experiment is on.
Also removes a spurious clamping code in decodemv.c which causes
rare encode/decode mismatches.
Change-Id: I809d9ee0b2859552b613500b539a615515b863ae
This patch allows use of 8x8 and 4x4 ADST correctly for Intra
16x16 modes and Intra 8x8 modes when the block size selected
is smaller than the prediction mode. Also includes some cleanups
and refactoring.
Rebase.
Change-Id: Ie3257bdf07bdb9c6e9476915e3a80183c8fa005a
This change included:
1. Aligned reads in vp9_mbloop_filter_vertical_edge function.
Since we actually read 16 bytes, we can align the reads to read
starting at (s - 8) instead of (s - 5).
2. Combined u, v loop filters.
3. Added 8x16 transpose.
This gave 2% decoder performance gain (tulip clip).
Change-Id: Ib14c2f1645c4a3436df17fe2f24789506bf0bb58
This commit removed a couple of redundant data structures in frame
coding contextsm, mode_context and mode_context_a, and changed to
use vp9_mode_contexts only. The switch of the context for different
frame type now relies on the switch of frame coding context between
lfc and lfc_a. This commit also removed a number of memcpy among
these redundant data structure.
Change-Id: I42e8174bd60f466b0860afc44c1263896471b0f3
Not all segment feature data elements are full-range powers of two, so
there are values that can be encoded that are invalid. Add a new function
to clamp values to the maximum allowed.
Change-Id: Ie47cb80ef2d54292e6b8db9f699c57214a915bc4
Support for gyp which doesn't support multiple objects in the same
static library having the same basename.
Change-Id: Ib947eefbaf68f8b177a796d23f875ccdfa6bc9dc
Vp9_sad3x16_sse2() is heavily called in decoder, in which the
unaligned reads consume lots of cpu cycles. When CONFIG_SUBPELREFMV
is off, the unaligned offset is 1. In this situation,
we can adjust the src_ptr to be 4-byte aligned, and then do the
aligned reads. This reduced the reading time significantly. Tests
on 1080p clip showed over 2% decoder performance gain with
CONFIG_SUBPELREFM off.
Change-Id: I953afe3ac5406107933ef49d0b695eafba9a6507
Experiments with a larger set of contexts and some
clean up to replace magic numbers regarding the
number of contexts.
The starting values and rate of backwards adaption
are still suspect and based on a small set of tests.
Added forwards adjustment of probabilities.
The net result of adding the new context and forward
update is small compared to the old context from the
legacy find_near function. (down a little on derf but
up by a similar amount for HD)
HOWEVER.... with the new context and forward update
the impact of disabling the reverse update (which may be
necessary in some use cases to facilitate parallel decoding)
is hugely reduced.
For the old context without forward update, the impact of
turning off reverse update (Experiment was with SB off) was
Derf - 0.9, Yt -1.89, ythd -2.75 and sthd -8.35. The impact was
mainly at low data rates.
With the new context and forward update enabled the impact
for all the test sets was no more than 0.5-1% (again most at
the low end).
Change-Id: Ic751b414c8ce7f7f3ebc6f19a741d774d2b4b556
added additional motion vectors at close neighborhood of a superblock
to the list of candiate motion vectors, and removed a couple that are
further away.
The change helped std-hd set about .8% (all metrics) and smaller gain
for derf set.
Change-Id: Iaa69b98614db43420ed3fd4738d0ca5587b90045
A patch on compound inter-intra prediction.
In compound inter-intra prediction, a new predictor for
16x16 inter coded MBs are obtained by combining a single
inter predictor with a 16x16 intra predictor, in a manner
that the weight varies with distance from the top/left
boundary. The current search strategy is to combine the best
inter mode with the best intra mode obtained independently.
Results so far:
derf +0.31%
yt +0.32%
std-hd +0.35%
hd +0.42%
It is conceivable that the results would improve somewhat
with a more thorough search strategy where all intra modes
are searched given the best mv, or even a joint search for
the best mv and the best intra mode.
Change-Id: I7951f1ed0d6eb31ca32ac24d120f1585bcd8d79b
Modify the decoder to return the ending position of the bool decoder and
use that as the starting position for the next frame.
The constant-space algorithm for parsing the appended frame lengths is
O(n^2), which is a potential DoS concern if n is unbounded. Revisit
the appended lengths for use as partition lengths when multipartition
support is added.
In addition, this allows decoding of raw streams outside of a container
without additional framing information, though it's insufficient to
be able to remux said stream into a container.
Change-Id: I71e801a9c3e37abe559a56a597635b0cbae1934b
Update decode_coefs() to break when c >= eob, since it's possible that
c starts the loop from 1 and eob is 0. The loop won't terminate in that
case.
Add new get_eob() function to consistently clamp the eob based on the
segment level EOB and the block size. It's possible to code a segment
level EOB that's greater than the block size, and that leads to an
out of bounds access.
Change-Id: I859563b30414615cf1b30dcc2aef8a1de358c42d
This is in line with other cases where we disable ADST if prediction
size and transform size don't match. Before this patch, the RD loop
will use ADST for superblocks, but frame encoding/decoding won't.
Change-Id: I700368c632eb72b5e089c22ef25649d99d7697d0
There are now more than 16 possible modes so 5
bits required for segment mode feature.
Note that it is likely that the mode feature and how it is
coded will change but for now the 4 bits was a bug.
Change-Id: I63348ae3a9cc31566a656c2dc78f09f5e1a9dcc9
Similar to 16x16 dequant and idct, based on the value of eobs, the
8x8 dequant and idct calculation was simplified to improve decorder
performance.
Combined vp9_dequant_idct_add_8x8 and vp9_dequant_dc_idct_add_8x8
to eliminate duplicate code.
Change-Id: Ia58e50ab27f7012b7379c495837c9c0b5ba9cf7f
This change is a fix / extension of the newbestrefmv
experiment. As such it is presented without IFDEF.
The change creates a new context for coding inter modes
in vp9_find_mv_refs(). This replaces the context that
was previously calculated in vp9_find_near_mvs().
The new context is unoptimized and not necessarily
any better at this stage (results pending), but eliminates
the need for a legacy call to vp9_find_near_mvs().
Based on numbers from Scott, this could help decode
speed by several %.
In a later patch I will add support for forward update of
context (assuming this helps) and refine the context as
necessary.
Change-Id: I1cd991b82c8df86cc02237a34185e6d67510698a
Experiment to test speed trade off of reducing the
extent of the ref mv search.
Reducing the maximum number of tested candidates to 9 had
minimal net effect on quality in any of the tests sets.
Reduction to 7 has a small negative impact (worst was STD-HD
at about -0.2%).
This change is in response to the apparently high number of
decode cycles reported in regard to mv-ref selection.
Change-Id: I0e92e92e324337689358495a1ec9ccdeb23dc774
Preliminary patch on a new 4x4 intra mode B_CONTEXT_PRED where the
dominant direction from the context is used to encode. Various decoder
changes are needed to support decoding of B_CONTEXT_PRED in conjunction
with hybrid transforms since the scan order and tokenization depends on
the actual direction of prediction obtained from the context. Currently
the traditional directional modes are used in conjunction with the
B_CONTEXT_PRED, which also seems to provide the best results.
The gains are small - in the 0.1% range.
Change-Id: I5a7ea80b5218f42a9c0dfb42d3f79a68c7f0cdc2
As suggested by Yaowu, simplified 16x16 dequant and idct. In decoder,
after detoken step, we know the number of non-zero dct coefficients
(eobs) in a macroblock. Idct calculation can be skipped or simplified
based on eobs, which improves the decoder performance.
Change-Id: I9ffa1cb134bcb5a7d64fcf90c81871a96d1b4018
Creates a merge between the master and experimental branches. Fixes a
number of conflicts in the build system to allow *either* VP8 or VP9
to be built. Specifically either:
$ configure --disable-vp9 $ configure --disable-vp8
--disable-unit-tests
VP9 still exports its symbols and files as VP8, so that will be
resolved in the next commit.
Unit tests are broken in VP9, but this isn't a new issue. They are
fixed upstream on origin/experimental as of this writing, but rebasing
this merge proved difficult, so will tackle that in a second merge
commit.
Change-Id: I2b7d852c18efd58d1ebc621b8041fe0260442c21
In the variance calculations the difference is summed and later squared.
When the sum exceeds sqrt(2^31) the value is treated as a negative when
it is shifted which gives incorrect results.
To fix this we force the multiplication to be unsigned.
The alternative fix is to shift sum down by 4 before multiplying.
However that will reduce precision.
For 16x16 blocks the maximum sum is 65280 and sqrt(2^31) is 46340 (and
change).
This change is based on:
1698234 Missed some variance casts
fea3556 Fix variance overflow
Change-Id: I2c61856cca9db54b9b81de83b4505ea81a050a0f
s/([vV][pP])8/$19/
additionally dct.h was removed; declare the _c functions that are used
in the tests. the TODO for conversion to parameterized tests still
remains.
Change-Id: I73db9425a57075bbb78a92693ba6b320578981cd
there are still a couple type of warning left, which are related to
double constants assigned to float type. As those would be addressed
by the conversion of transforms into integer version. This commit
has left those un-dealt with.
Change-Id: I48fd9b489c0c27ad6b543f4177423419f929f2bb
The block sizes for decoding tokens are up to 16x16, which means
eobs is within [0, 256]. Using (signed) char is not enough. Changed
eobs data type to unsigned short to fix the problem.
Change-Id: I88a7d3098e1f1604c336d6adb88ffec971fb03a6
