Implements scan order to band map with arrays in both the encoder
and decoder to remove conditional statements.
Encoding seems to be about 1% faster at speed 0, tested on football.
Decoding seems to be about 0.5-1% faster on a set of 25 videos.
Change-Id: Idb233ca0b9e0efd790e30880642e8717e1c5c8dd
This commit enables the dual buffer rate-distortion optimization
and encoding scheme. It stacks the original transform coefficients,
quantized levels, and reconstructed coefficients, in the rate-
distortion optimization search process, hence eliminates the need
to re-run residual generation, forward transform, and quantization
in the encoding stage.
Change-Id: I011bfad3a59a380a869ee552e91dae0394ec492e
Allocate memory space of dual buffer sets that store the coeff, qcoeff,
dqcoeff, and eobs. Connect the pointers of macroblock_plane and
macroblockd_plane to the actual buffer in use accordingly.
Change-Id: I2f0b5f482ca879fae39095013eaf8901db20a5a4
The pointer was asigned only once with vp9_regular_quantize_b_4x4, calling
this function directly now. Also removing unused declarations:
prototype_quantize_block
prototype_quantize_block_pair
prototype_quantize_mb
vp9_regular_quantize_b_4x4_pair
vp9_regular_quantize_b_8x8
Change-Id: I14325bc2f082336820671eafbc06126651b79f73
Renames for consistency with other constants:
NUM_FRAME_TYPES -> FRAME_TYPES
NUM_PARTITION_CONTEXTS -> PARTITION_CONTEXTS
Change-Id: I3db30acb2868eb0a424237c831087b2e264ec47f
This commit makes the buffer allocation of zcoeff_blk array in
pick_mode_context block size aware. It calculates the number of
4x4 blocks in the partition and assigns the memory space accordingly.
This process (and the uninitialization) is done once for each encoding
pass. It allows memory copy of smaller buffer when possible.
For football at 600kbps, the runtimes improve by about 1%:
speed 1, 45961ms -> 45472ms
speed 2, 23863ms -> 23598ms
Change-Id: Id2ca24906fa89f46fa5fe742ec4b8efc2a61f877
This should be similar to what x264 does with --aq-mode 1.
It works well with clips like parkjoy and touhou
(http://x264.nl/developers/Dark_Shikari/LosslessTouhou.mkv).
At low bitrates, the segmentation signaling overhead may negate the
benefits of this feature.
(PGW) Default changed to feature OFF to allow provisional merge.
Change-Id: I938abf9bb487e1d4ad3b0264ea03d9826275c70b
Use the zcoeff_blk buffer of PICK_MODE_CONTEXT to store the indexes
of all-zero-coeff block of the current best mode. Remove the temporary
buffer best_zcoeff_blk defined in the rate-distortion optimization
loop. This improves the speed performance by about 0.5% in all speed
settings.
Change-Id: Ie3e15988ddfa581eafa2e19a8228d3fe4a46095c
This commit moves token_cache buffer into macroblock struct, instead
of defining as a local variable in cost_coeffs. This avoids repeatedly
re-allocating memory space in the rate-distortion optimization loop.
The runtime at speed 0 reduces:
bus 2000kbps, 161692ms to 159951ms
football 600kbps, 229505ms to 225821ms
Change-Id: If7da6b0b6d8c5138a16271a33c4548fba33d8840
This commit re-designs the per transformed block rate-distortion
costs tracking buffers. It removes redundant buffer usage, makes
the needed context memory allocation per VP9_COMP instance and
reuses the same buffer sets inside the rate-distortion optimization
search loop, thereby avoiding repeatedly requiring memory space.
It reduces speed 0 runtime:
bus at 2000 kbps from 166763ms to 158967ms,
football at 600 kbps from 246614ms to 234257ms.
Both about 5% speed-up. Local tests suggest about 2% to 5% speed-up
for speed 1 and 2 settings. This does not change compression
performance.
Change-Id: I363514c5276b5cf9a38c7251088ffc6ab7f9a4c3
Use b_mode_info to store the inter prediction mode of sub8x8 block,
in replacement of the use of partition_info. Remove redundant buffer
update for partition_info. For bus_cif at 2000 kbps, this seem to make
speed 0 about 1% faster.
Change-Id: Id1b3be45e75a24fb4b42335ac480c23e440978f6
This commit enables forcing all coefficients zero per transformed
block, when its rate-distortion cost is lower than regular coeff
quantization.
The overall performance improvement (including its parent patch on
calculating rd cost per transformed block) at speed 1:
derf: 0.298%
yt: 0.452%
hd: 0.741%
stdhd: 0.006%
Change-Id: I66005fe0fd7af192c3eba32e02fd6d77952accb5
This commit enables adaptive constraint on motion search range for
smaller partitions, given the motion vectors of collocated larger
partition as a candidate initial search point.
It makes speed 0 runtime of bus at CIF and 2000 kbps goes from
167s down to 162s (3% speed-up), at 0.01dB performance gains. In
the settings of speed 1, this makes the runtime goes from 33687 ms
to 32142 ms (4.5% speed-up), at 0.03dB performance gains.
Compression performance wise, it gains at speed 1:
derf 0.118%
yt 0.237%
hd 0.203%
stdhd 0.438%
Change-Id: Ic8b34c67810d9504a9579bef2825d3fa54b69454
The low precision 32x32 fdct has all the intermediate steps within
16-bit depth, hence allowing faster SSE2 implementation, at the
expense of larger round-trip error. It was used in the rate-distortion
optimization search loop only.
Using the low precision version, in replace of the high precision one,
affects the compression performance by about 0.7% (derf, stdhd) at
speed 0. For speed 1, it makes derf set down by only 0.017%.
Change-Id: I4e7d18fac5bea5317b91c8e7dabae143bc6b5c8b
Adds a speed feature to skip all intra modes other than
DC_PRED if the source variance is small. This feature is
made part of speed 1 and up.
Results on derf300: psnr -0.07%, speedup about 1-2%
Also uses the source variance to fine-tune the early
termination criteria when FLAG_EARLY_TERMINATE is on.
This feature is made part of speed 2 and up.
Results on derf300: psnr -0.52%, speedup about 5-7%
Change-Id: I59e38aa836557cfa5405ae706fc64815cbfe4232
This allows us to increment the position at the band-level only as
we go from one band to the next; more importantly, that allows us to
use an add instead of multiply instruction, and omit the instruction
altogether if the band doesn't change from one coef to the next, thus
being slightly faster (probably more noticeable on systems where a
multiply is expensive, like arm).
Change-Id: I4343fe35b9f9a47fa00b217bdcbf5f91ff96c381
This patch modifies the auto_mv_step_size speed feature to
use a combination of the maximum magnitude mv from the last
inter frame, and the maximum magnitude mv for the two reference
mvs with the same reference. For arf frames, the max mav step
for the resolution is used.
The bounds therefore are slightly tighter. The feature is made
a speed 1 feature.
Rebased.
Results (when this feature is turned on over speed 0):
derfraw300: -0.046% psnr, about 5+% speedup
(tested on football: goes from 4m30.760s to 4m17.410s).
Change-Id: If492797a61b0b4b3e58c0b8f86afb880165fc9f6
Current partition checking starts from small sizes, and then goes up
to large sizes. This experiment uses the small partitions' motion
estimation result, which is already available, to speed up the
large partition's motion estimation. We can decide to skip some
patition checkings if they are unlikely choices. We could use the
motion vector(MV) result as current partition's prediction MV, limit
the search range and reference frame.
Current result at speed 1:
psnr loss: 1.19% for stdhd, 0.287% for derf.
speed gain: 14% for sunflower(hd), 11% for akiyo.
Further improvement will be done later.
Change-Id: I5abfd070e9cace2e91e2a0247d1325df313887ab
This speed feature allows the encoder to largely remove the spatial
dependency between blocks inside a 64x64 superblock, thereby removing
the need to repeatedly encode superblocks per partition type in the
rate-distortion optimization loop.
A major challenge lies in the intra modes tested in the rate-distortion
optimization loop. The subsequent blocks do not have access to the
reconstructed boundary pixels without the intermediate coding steps.
This was resolved by using the original pixels for intra prediction
in the rd loop, followed by an appropriately designed distortion
modeling on the quantization parameters. Experiments also suggested
that the performance impact is more discernible at lower bit-rate/psnr
settings. Hence a quantizer dependent threshold is applied to deactivate
skip of block coding.
For bus_cif at 2000 kbps,
speed 0: runtime 269854ms -> 237774ms (12% speed-up) at 0.05dB
performance loss.
speed 1: runtime 65312ms -> 61536ms, (7% speed-up) at 0.04dB
performance loss.
This operation is currently turned on in settings of speed 1.
Change-Id: Ib689741dfff8dd38365d8c1b92860a3e176f56ec
Changes cost_mv_ref() into doing a LUT into pre-calculated cost
arrays instead. Encode time of first 50 frames of bus (speed 0)
@ 1500kbps goes from 2min11.6 to 2min10.9, i.e. 0.5% faster overall.
Change-Id: If186e92c34c201b29cbbc058785a15c9c09e433a
Overall, on all test sets, this gains about +0.2% on all metrics.
City is a clip where this really hurts (-1.0% on all metrics), I'm
not quite sure why yet. Maybe interesting to look into in the future.
Change-Id: I6f0eecb20e72f0194633270d30bf00d76d9eae78
Total encoding time for first 50 frames of bus (speed 0) @ 1500kbps
goes 2min34.8 to 2min14.4, i.e. a 10.4% overall speedup. The code is
x86-64 only, it needs some minor modifications to be 32bit compatible,
because it uses 15 xmm registers, whereas 32bit only has 8.
Change-Id: I2df53770c2e850813ffa713e1a91b45b0082b904
Cycle timings for first 3 frames of bus (speed 0) at 1500kbps:
4x4: 298 -> 234 cycles
8x8: 1227 -> 878 cycles
16x16: 23426 -> 18134 cycles
32x32: 4906 -> 3664 cycles
Total encode time of first 50 frames of bus @ 1500kbps (speed 0) goes
from 3min0.7 to 2min51.6 seconds, i.e. 5.3% faster.
Change-Id: I68a0e1b530b0563b84a67342cca4b45146077e95
This commit replaces zrun_zbin_boost, a method of biasing non-zero
coefficients following runs of zero-coefficients to be rounded towards
zero, with an explicit skip-block choice in the RD loop.
The logic is basically that if individual coefficients should be rounded
towards zero (from a RD point of view), the trellis/optimize loop should
take care of it. If whole blocks should be zero (from a RD point of
view), a single RD check is much more efficient than a complete
serialization of the quantization loop.
Quality change: derf +0.5% psnr, +1.6% ssim; yt +0.6% psnr, +1.1% ssim.
SIMD for quantize will follow in a separate patch. Results for other
test sets pending.
Change-Id: Ife5fa641163ac5150ac428011e87188f1937c1f4
This commit makes use of dual fdct32x32 versions for rate-distortion
optimization loop and encoding process, respectively. The one for
rd loop requires only 16 bits precision for intermediate steps.
The original fdct32x32 that allows higher intermediate precision (18
bits) was retained for the encoding process only.
This allows speed-up for fdct32x32 in the rd loop. No performance
loss observed.
Change-Id: I3237770e39a8f87ed17ae5513c87228533397cc3
Migrates costing changes/fixes from the rebalance expt to the head
without the expt on.
Rebased.
Change-Id: I51677d62f77ed08aca8d21a4c9a13103eb8de93f
Results:
derfraw300: +0.126%
This patch changes the coefficient tree to move the EOB to below
the ZERO node in order to save number of bool decodes.
The advantages of moving EOB one step down as opposed to two steps down
in the other parallel patch are: 1. The coef modeling based on
the One-node becomes independent of the tree structure above it, and
2. Fewer conext/counter increases are needed.
The drawback is that the potential savings in bool decodes will be
less, but assuming that 0s are much more predominant than 1's the
potential savings is still likely to be substantial.
Results on derf300: -0.237%
Change-Id: Ie784be13dc98291306b338e8228703a4c2ea2242
Removed one 4x4 prediction step that was unnessary in the rd loop.
Removed a unused modecosts estimate from encoder side.
Change-Id: I65221a52719d6876492996955ef04142d2752d86
1. remove prediction mode conversion
2. unified bmode, same for key and non-key frame
3. set I4X4_PRED count for pdf to 0, as I4X4_PRED is no longer
coded ever. It is determined by ref_frame and block partition
Change-Id: If5b282957c24339b241acdb9f2afef85658fe47d
Also do per-partition motion vector referencing in <sb8x8 partitions,
and adjust mvref finding for sub8x8 partitions.
Change-Id: Id3ed1ed4d2a8910d11d327db6cc63b8eb79f941f
Move 4x4/4x8/8x4 partition coding out of experimental list.
This commit fixed the unit test failure issues. It also resolved
the merge conflicts between 4x4 block level partition and iterative
motion search for comp_inter_inter.
Change-Id: I898671f0631f5ddc4f5cc68d4c62ead7de9c5a58
This commit allows the rate-distortion optimization recursion
at encoder to go down to 4x4 block size. It deprecates the use
of I4X4_PRED and SPLITMV syntax elements from bit-stream
writing/reading. Will remove the unused probability models in
the next patch.
The partition type search and bit-stream are now capable of
supporting the rectangular partition of 8x8 block, i.e., 8x4
and 4x8. Need to revise the rate-distortion parts to get these
two partition tested in the rd loop.
Change-Id: I0dfe3b90a1507ad6138db10cc58e6e237a06a9d6
This commit enables the search for the optimal superblock
partition types in the recursion form. The intention is to
make the optimization process more concise and ready to
support partition down to 4x4 block size next.
Change-Id: Iae279a67df3a7cc372553c84c775bc4d2f3e4336
This allows removing a large number of transform size specific functions,
as well as supporting 444/alpha by routing all code through the
subsampling-aware path.
Change-Id: Ieb085cebe9f37f24fc24de179898b22abfda08a4
Work-in-progress, not yet ready for review. TODO items:
- bitstream writing (encoder) and reading (decoder)
- decoder reconstruction
Change-Id: I5afb7284e7e0480847b47cd0097cb469433c9081
All members can be referenced from their per-plane counterparts, and
removes assumptions about 24 blocks per macroblock.
Change-Id: I593fb0715e74cd84b48facd1c9b18c3ae1185d4b
This commit enables selecting probability models for recursive block
partition information syntax, depending on its above/left partition
information, as well as the current block size. These conditional
probability models are reasonably stationary and consistent across
frames, hence the backward adaptive approach is used to maintain and
update the contextual models.
It achieves coding performance gains (on top of enabling rectangular
block sizes):
derf: 0.242%
yt: 0.391%
hd: 0.376%
stdhd: 0.645%
Change-Id: Ie513d9673337f0d27abd65fb566b711d0844ec2e
Quantizers can vary per plane, but not per block. Move these values to
the per-plane part of MACROBLOCK.
Change-Id: I320a55e38b7b28b29aec751a4aca5ccd0c9b9326
This commit moves the coeff storage from the MACROBLOCK struct to its
per-plane part. The next commit will remove the coeff member from the
BLOCK structure so that it is consistently accessed per-plane.
Also refactors vp9_sb_block_error_c and vp9_sb_uv_block_error_c to be
variable subsampling aware.
Change-Id: I18c30f87f27c3a012119b6c1970d5fa499804455
First in a series of commits making certain MACROBLOCK members
addressable per-plane. This commit also refactors the block subtraction
functions vp9_subtract_b, vp9_subtract_sby_c, etc to be
loops-over-planes and variable subsampling aware.
Change-Id: I371d092b914ae0a495dfd852ea1a3d2467be6ec3
This version of speed 1 only disables modes at higher resolution that
had distortions >2x the best mode we found...
The hope is that this could be a replacement for speed 0 ...
Change-Id: I7421f1016b8958314469da84c4dccddf25390720
Adds RD integration for 32x16, 16x32, 64x32 and 32x64 rectangular blocks.
Derf almost +0.6%, HD a little over +1.0%, STDHD +1.3%.
Change-Id: Id651fdb6a655fdbb5c47009757e63317acfb88a5
Merge sb32x32 and sb64x64 functions; allow for rectangular sizes. Code
gives identical encoder results before and after. There are a few
macros for rectangular block sizes under the sbsegment experiment; this
experiment is not yet functional and should not yet be used.
Change-Id: I71f93b5d2a1596e99a6f01f29c3f0a456694d728
These are mostly just for experimental purposes. I saw small gains (in
the 0.1% range) when playing with this on derf.
Change-Id: Ib21eed477bbb46bddcd73b21c5c708a5b46abedc
This patch revamps the entropy coding of coefficients to code first
a non-zero count per coded block and correspondingly remove the EOB
token from the token set.
STATUS:
Main encode/decode code achieving encode/decode sync - done.
Forward and backward probability updates to the nzcs - done.
Rd costing updates for nzcs - done.
Note: The dynamic progrmaming apporach used in trellis quantization
is not exactly compatible with nzcs. A suboptimal approach has been
used instead where branch costs are updated to account for changes
in the nzcs.
TODO:
Training the default probs/counts for nzcs
Change-Id: I951bc1e22f47885077a7453a09b0493daa77883d
Split macroblock and superblock tokenization and detokenization
functions and coefficient-related data structs so that the bitstream
layout and related code of superblock coefficients looks less like it's
a hack to fit macroblocks in superblocks.
In addition, unify chroma transform size selection from luma transform
size (i.e. always use the same size, as long as it fits the predictor);
in practice, this means 32x32 and 64x64 superblocks using the 16x16 luma
transform will now use the 16x16 (instead of the 8x8) chroma transform,
and 64x64 superblocks using the 32x32 luma transform will now use the
32x32 (instead of the 16x16) chroma transform.
Lastly, add a trellis optimize function for 32x32 transform blocks.
HD gains about 0.3%, STDHD about 0.15% and derf about 0.1%. There's
a few negative points here and there that I might want to analyze
a little closer.
Change-Id: Ibad7c3ddfe1acfc52771dfc27c03e9783e054430
1. Added a bit in frame header to to indicate if a frame is encoded
in lossless mode, so decoder does not make the decision based on Q0
2. Minor changes to make sure that lossy coding works same as when
the lossless experiment is not enabled.
3. Renamed function pointers for transforms to be consistent, using
prefix fwd_txm and inv_txm for forward and inverse respectively
To encode in lossless mode, using "--lossless=1 --min-q=0 --max-q=0"
with vpxenc.
Change-Id: Ifae53b26d2ffbe378d707e29d96817b8a5e6c068
Remove eob_max_offset markers and replace
with the generic skip_block flag to indicate
to the quantizer that all coeffs to be set to 0
and eob position set to 0;
Change-Id: Id477e8f8d4ec1a5562758904071013c24b76bfd7
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
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
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
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
