This patch allows the encoder to skip the search for partition
SPLIT, HORZ, VERT after the search for partition NONE is done
in RD optimization. It uses the first pass block-wise statistics
to make the decision. If all 16x16 blocks in the current partition
have zero motions and small residues from the frist pass statistics,
and it has small difference variance, further partition search is
skipped.
For speed 2 setting, experiments on general youtube clips show that
the speedup varies from 1% - 10%, 5% on average. On the performance
side in PSNR, derf 0.004%, yt -0.059%, hd -0.106%, stdhd 0.032%.
For hard stdhd clips:
park_joy_1080p, 502952 ms -> 503307 ms (-0.07%)
pedestrian_area_1080p, 227049 ms -> 220531 ms (+3%)
This feature is under the compilation flag CONFIG_FP_MB_STATS and
it is off in current setting.
Change-Id: I554537e9242178263b65ebe14a04f9c221b58bae
Remove the variable that indicates the relative block index. This
is explicitly covered by the use of pc_tree.
Change-Id: Ib13142582fff926c85e375bde656aa050add8350
This commit enables a chessboard pattern constrained partition
search for 720p and above resolutions. The scheme applies stricter
partition search to alternative blocks based on its above/left
neighboring blocks' partition range, as well as that of the
collocated blocks in the previous frame. It is currently turned
on at 16x16 block size level. The chessboard pattern is flipped
per coding frame.
The speed 3 runtime is reduced:
park_joy_1080p, 652832 ms -> 607738 ms (7% speed-up)
pedestrian_area_1080p, 215998 ms -> 200589 ms (8% speed-up)
The compression performance is changed:
hd -0.223%
stdhd -0.295%
Change-Id: I2d4d123ae89f7171562f618febb4d81789575b19
vp9_variance16x16(), and vp9_get16x16var().
On a Nexus 7, vpxenc (in realtime mode, speed -12)
reported a performance improvement of ~16.7%.
Change-Id: Ib163aa99f56e680194aabe00dacdd7f0899a4ecb
This commit replace the repetitive retrieve of max and min allowed
partition from speed_feature with local variables max_size and
min_size.
Change-Id: Ib06f11f16615e4876e4dd5fb6a968c6bf5f7b216
The get_chessboard_index() used to call the entire VP9_COMMON
struct pointer to retrieve the chessboard pattern index. This cl
makes it call the frame index directly.
Change-Id: I3cad9d209ea2e77a358085a04fe1ff0ddec5ba03
Remove the redundant index computation when store the first
pass block-wise statistics. Currently, a single byte is
allocated for a 16x16 blocks, and all the frame statistics
saved during the first pass will be kept in memory for use
in the second pass. For a 1920x1080 300-frame clip, it will
take about 2.3 MB memory. This feature is off in current
setting.
Change-Id: I135a95b348ec093d54c6a07e1e8237626909e3bd
Remove all the redundant dct functions (dct4x4, dct8x8)
in avx2 except dct32x32 those functions were copied originally from dct_sse2
Change-Id: I742576fbf5175f3ac09f2076976a9247b259323e
The partition search for 4x4 blocks takes unnecessary steps to
reconstruct pixels and an extra partition type update. This commit
removes such operations. No visible compression/speed difference.
Thanks to Yue (yuec@) for finding this issue.
Change-Id: I3f83824aa3fd3717d63be0b280fa57258939a70a
The assignment of the variable mode_excluded in
vp9_rd_pick_inter_mode_sub8x8 takes redundant conditional jump.
This commit removes it.
Change-Id: Ie195fbe6e54ec2ade7093d562c456a2e93143704
1. Remove last reference flag for first frame upper layers in one pass mode.
2. Disable refresh golden frame flag for key frames.
Change-Id: I44ac1bd2c795169e4fbfdd078ea79a1d33a204d6
The value of mode_excluded has been properly set in
vp9_rd_pick_inter_mode_sb(). It is redundant to send it in
handle_inter_mode() and re-set the value again.
Change-Id: I408d4731f2f42e0bcf3ae62e85757717bb410471
This commit extends the chessboard pattern prediction filter search.
If the above and left blocks have the same prediction filter type,
the encoder will skip the prediction filter type search and use the
reference one.
The overall chessboard pattern prediction filter type search reduces
speed 3 runtime for hard clips. Experiments on park joy at 1080p
and 15000 kbps show that the runtime goes from 723265 ms to 65832 ms,
i.e., about 10% speed-up. Compression performance wise, it affects
the coding quality by
Change-Id: I880975497c7ad166532e9eea9bf46684d77ff327
derf: -0.326%
yt: -0.257%
hd: -0.241%
stdhd: -0.417%
This commit enables a chessboard pattern prediction filter type
search scheme for rate-distortion optimization speed-up. For the
inferred motion vector modes, the encoder can re-use its above/left
neighbor blocks' prediction filter type and skip a full test on
all possible filter types. Such operation is turned on/off
alternatively in a chessboard manner.
It is turned on in speed 3. For test clip pedestrian 1080p, the
runtime is reduced from 231500 ms -> 221700 ms. The compression
performance is changed:
derf: -0.147%
yt: -0.134%
hd: -0.079%
stdhd: -0.220%
Change-Id: I1912f278e7576c2dc632688e3ad7a257410c605a
When OUTPUT_YUV_DENOISED is enabled the encoder outputs the uncompressed,
denoised video to a separate file. Moved the point at which the file is
written to in order to avoid an extra blank frame at the beginning of the video.
Change-Id: I805f6a912b18b3d9cae59b13c5b8108279439ce3
This should be a local variable. Move the definition from
vp9_rd_pick_inter_mode_sb to handle_inter_mode.
Change-Id: I14f4168bb1c896ed04e8f6d4cd89fbf4c9839944
For sequences of resolution below 720p, the encoder will check
intra prediction modes and inter prediction modes from LAST_FRAME.
This commit turns on adaptive prediction filter scheme for sub8x8
blocks, where inter prediction modes are enabled. For the test
sequence bus at CIF, the speed 2 runtime goes down from 17879 ms
to 16783 ms, i.e., 6% speed up. The compression performance of
derf set is down by -0.128%.
Change-Id: I01d5321a5ceab4e0666ac5be56c52d896c7a8d45
The commit moved a call to vp9_clear_system_state() to a correct
location, i.e. prior function calls using floating point numbers.
This was to fix a mismatch mmx code and sse2 version, where a
floating point number used in adjust_frame_rate(cpi) gets NAN due
to mmx registers being in wrong state.
Change-Id: I40e0a6de98812000ccee6a729badb630604fd7e6
For gcc, when libvpx config option debug is disabled, added the
flag -DNDEBUG to disable the assertions in libvpx for some speedup.
Change-Id: Ifcb7b9e8ef5cbe5d07a24407b53b9a2923f596ee
Separates HBD profile int two profiles (2 and 3) consistent with the
highbitdepth branch. This patch is ported from the original highbitdepth
branch patch: https://gerrit.chromium.org/gerrit/#/c/70460/
Two of the invalid file tests needed to be updated.
Change-Id: I6a4acd2f7a60b1fb4cbcc8e0dad4eab4248431e3
The bug sets the wrong pointer to the first pass mb stats
if the encoder does the re-coding in the second pass.
Change-Id: I8a11f45dd7dceb38de814adec24cecccae370d00
This patch is the first step toward simplifying the
frame buffer handling.
The final goal is to have a common frame buffer handling
framework for both encoder and decoder that incorporates
the existing ability to use externally allocated memory.
Change-Id: I2c378a4f54a39908915f46c4260e17a080db7ff1
This commit changed the hard-coded DEFAULT_INTERP_FILTER to a speed
feature with the same default value: SWITCHABLE.
Change-Id: I7f54f40f1bd3f5277841d04b85db7a84e47313f1
and vp9_sad16x16_neon()
On a Nexus 7, vpxenc (in realtime mode, speed -6)
reported a performance improvement of ~17%.
Change-Id: I91e070cde2973451083d3f3d63b49b7886de9a85
2 pass only change to calculation of rd mult based on Q.
Make a small adjustment based on frame type and also
replace adjustment based on iifactor with an one based
on the ambient GF/ARF boost level.
Also fix multi arf bug / issue.
Overall these change give an slight improvement in ssim
but hurt psnr a little.
Change-Id: I5e1751e3ff5390a26f543d7855059e6fbcce105e
We target this speed to achieve similar encoding speed and better
compression than vp8 rt mode with cpu-used at -12.
Change-Id: Ic1bb4371c81a17ea80e83459c1cbf4c09a3498e8
In vp8, statistics are collected about the different modes as they are searched.
This process is more complicated due to the variable block size. Fields were
added to the PICM_MODE_CONTEXT struct to hold this information for each point in
the search. The information is then taken from the appropriate part of the tree
during denoising.
Change-Id: I89261ab77ad637821287ae157dfdf694702b8e77
All changes are for spatial svc only.
1. Enable encoding hidden frames in each layer and use alt reference idex to reference the hidden frame in each layer
2. Use golden reference idx for spatial reference
3. For those layers that don't have hidden frames (caused by lack of frame buffers), reference a hidden frame in lower layers
4. Add "auto-alt-refs" in svc options
Change-Id: Idf27d1fd2fb5f3ffd9e86d2119235e3dad36c178
This commit fixes a potential out-of-boundary memory access due to
the use of reuse_inter_pred_sby in the non-RD coding flow. It
resolves the corresponding asan error.
Change-Id: Iff605f5921230966990013541cd855d698810922
This commit fixes a mismatched use case of block size in non-RD
intra prediction check. The residual SSE and variance should be
calculated per transform block size, instead of operating block
size, which caused chrome valgrind warning on conditional jump
based on uninitialized value (webm issue 823). This commit
resolves this issue.
Change-Id: I595c06599c7e0fd0e4a08736519ba68fc14bc79a
Refactoring to remove some duplication of probability
tables between tokenization and detokenization.
Change-Id: I2fc6a6497f9c0410021a9b41f828bc58a864e466
Use a weaker filter for second level arf frames.
Average gain across all sets and metrics ~0.3%
Remove code for arnr_type which is no longer
supported in VP9 which always uses a centered blur.
Re-factor and some cleanup.
Change-Id: Ieb4b8940e99e4e02b3fcc9fca6f2d4109e6ed639
Currently, vp9_diamond_search_sadx4() is only called when sse3 is
enabled, which is improper since sse2 optimization of sdx4df
functions are available. Changed to always use
vp9_diamond_search_sadx4().
Change-Id: I4b95d6b7a3c6c645783c373f0ba8d645ece24717
There are sse2 optimization of sdx4df functions. Instead of calling
vp9_refining_search_sadx4 only when sse3 is enabled, call it always.
Change-Id: I24f93818f7d4209d1425039e0eb099ff9ff08fe9
Use FAST_HEX in speed 5 and 6, which covers more points than
FAST_DIAMOND and improves motion search quality.
At speed 6, RTC set borg tests showed slight quality gain (psnr
gain: 0.143%, ssim gain: 0.226%). No noticeable encoding speed
change.
Change-Id: Ifa62875d9a52ee382ec494f271382bb77d8c67bf
This commit combined the full pel and sub pel motion search into a
single function to avoid code duplication. The commit does not change
encoder outputs.
Change-Id: Ibe18342c4f64073bef20f9cf6c6ca0a20d01bf0d
This commit enables a new quantization process for 32x32 2D-DCT
transform coefficient blocks. It improves the compression
performance of speed 5 by 1.4%. The overall compression gains of
speed 5 due to the new quantization scheme is 4.7%. It also includes
the SSSE3 implementation of the 32x32 quantization process.
Change-Id: I0855b124fd6462418683f783f5bcb44255c9993b
In the previous version, only certain buffers in the macroblockd were saved and
the restored. In this version, all of the buffers are saved and restored. The
code was then rolled into a loop for readability.
Also contains a tiny fix for when the -DOUTPUT_YUV_DENOISED flag is used.
Change-Id: Id925ef8b3fa122ae88acfa1d9a1e4df45df83518
* Replace max_step_search_steps with constant MAX_MVSEARCH_STEPS
* Fold (reduce_first_step_size + speed > 5) into reduce_first_step_size
replacing uses of reduce_first_step_size that don't add the speed
check with zero.
Change-Id: Iae46395dbf3eaca138bf4d18b838a9e364b5a198
vp9_rdopt is for making rd optimal mode decisions. vp9_rd is for all
other rd related routines. Anything used outside of making an rd optimal
decision belongs in rd.
Change-Id: I772a3073f7588bdf139f551fb9810b6864d8e64b
Moved the threshold adjustment before reference flag checking,
which could set the threshold to INT_MAX for disabled reference
frame, and cause overflow if the adjustment is done after that.
Change-Id: I85e94f8726d5e3ae93f65965aa978721dddc9957
Renamed updating_running_avg() to filter(). Extended function with the rest of
the filter procedure. Made all of the empirically-determined constants used in
VP8 into functions so they can be tweaked more easily.
Change-Id: I41730c8c92370c76885950a43742347477ca4e7e
As in VP8.
Currently, this parameter is set with the VP8E_SET_NOISE_SENSITIVITY flag.
The flag was not renamed so that we don't break the interface for webrtc. This
should probably be changed at some point in the future.
Change-Id: Ic73fcb0dde9d1d019e9d042050b617333ac65472
Add test code to turn multi-arf on and off depending
on group length and zero motion.
Changes to active max group length for mult-arf.
Fund second arf only from normal frame bits.
Change-Id: I920287fac1c886428c15a39f731a25d07c2b796c
This commit added a speed feature to control the step_param used in
full pixel motion search. The intention is to reduced the search
steps for high speed real time coding.
Change-Id: I21d2f0105c2b647783a6688615da7fcf2b6d670b
Adapt the use of segmentation in AQ mode 2 based on
the ambient kf/arf/gf Q.
Disable segmentation where the rate per SB is very
low and overheads are likely to outweigh the benefits.
This patch reduces the -ve average metrics impact
of AQ mode 2 while allowing stronger 3 segment AQ
in some cases. Average improvement ~0.5-1.0%.
Change-Id: I5892dfcc7507c5cc6444531cc7fe17554cf8d0c7
This commit re-designs the quantization process for transform
coefficient blocks of size 4x4 to 16x16. It improves compression
performance for speed 7 by 3.85%. The SSSE3 version for the
new quantization process is included.
The average runtime of the 8x8 block quantization is reduced
from 285 cycles -> 255 cycles, i.e., over 10% faster.
Change-Id: I61278aa02efc70599b962d3314671db5b0446a50
Add a conditional compile flag for this feature. Also add a
switch to enable the encoder to use these statistics in the
second pass. Currently, the switch is turned off.
Change-Id: Ia1c858c35ec90e36f19f5cffe156b97ddaa04922
The current threshold is knid of low, and in many cases NEWMV
mode is checked but not picked as the best mode. This patch
added a speed feature to increase NEWMV threshold, so that
less partition mode checking goes to check NEWMV. This feature
is enabled for speed 6 and 7.
Rtc set borg tests showed:
1. Speed 6, overall psnr: -0.088%, ssim: -1.339%;
Average speedup on rtc set is 11.1%.
2. Speed 7, overall psnr: -0.505%, ssim: -2.320%
Average speedup on rtc set is 12.9%.
Change-Id: I953b849eeb6e0d5a1f13eacba30c14204472c5be
Encoder still uses SWITCHABLE as default via DEFAULT_INTERP_FILTER,
but does not override the default if it is not SWITCHABLE.
Change-Id: I3c0f6653bd228381a623a026c66599b0a87d01d5
When the frame is intra coded only, the encoder takes the RD
coding flow. Hence the function set_mode_info is not practically
in use. This commit removes it and the associated conditional
branches.
Change-Id: I1e42659ceb55b771ba712d1cdecacb446aa6460d
For real time speed 7, once encode breakout is on(i.e. encoding
setting --static-thresh=1), a proper encode breakout threshold
is set to speed up the encoder.
Set --static-thresh=1, RTC set borg test showed a slight overall
psnr loss of 0.162%, but ssim gain of 0.287%. The average speedup
on RTC set is 6%, and for some clips, the speedup can be 10+%.
Change-Id: Id522d9ce779ff7c699936d13d0c47083de4afb85
Before encoding a frame, calculate and store each 16x16 block's
variance of source difference between last and current frame.
Find partitioning threshold T for the frame from its variance
histogram, and then use T to make partition decisions.
Comparing with fixed 16x16 partitioning, rtc set test showed an
overall psnr gain of 3.242%, and ssim gain of 3.751%. The best
psnr gain is 8.653%.
The overall encoding speed didn't change much. It got faster for
some clips(for example, 12% speedup for vidyo1), and a little
slower for others.
Also, a minor modification was made in datarate unit test.
Change-Id: Ie290743aa3814e83607b93831b667a2a49d0932c
Bug introduced in I930dced169c9d53f8044d2754a04332138347409. If
svc.number_temporal_layers == 1 and svc.number_spatial_layers == 1, the system
attempt to do spatial SVC. It no longer does that.
Change-Id: Ie6b130a72b1eea40c547c9a64447e40695f811c5
For the primary arf in a group, if multiple arfs
are enabled and we were using arfs in the previous
group, then allow the second arf from the previous
group to be used as an additional reference.
Change-Id: Iaf41706a52f54ef21548026851cd77100d6aebda
This commit enables an adaptive transform size selection method
for speed -6. It uses largest transform size when the sse is more
than 4 times of variance, i.e., most energy is compacted in the
DC coefficient. Otherwise, use the default TX_8X8. It improves
the compression efficiency for rtc set of speed -6 by 0.8%, no
speed change observed.
Change-Id: Ie6ed1e728ff7bf88ebe940a60811361cdd19969c
This patch allows the encoder to skip the partition search for the
frame if it is an inter frame and only zero motion vectors have
been detected in the first pass. The partition size is directly
assigned according to the difference variance.
Borg tests show overall little performance changes in term of PSNR
(derf -0.027%, yt 0.152%, hd 0.078%, stdhd 0%). The worst case of
PSNR loss is -0.514% from yt. The best PSNR gain is 4.293% from yt.
The second pass encoding speedup for slideshow clips is 15%-40%.
Change-Id: I881f347d286553ee5594a9ea09ba1a61ac684045
This commit enables a fast reference motion vector search scheme.
It checks the nearest top and left neighboring blocks to decide the
most probable predicted motion vector. If it finds the two have
the same motion vectors, it then skip finding exterior range for
the second most probable motion vector, and correspondingly skips
the check for NEARMV.
The runtime of speed -5 goes down
pedestrian at 1080p 29377 ms -> 27783 ms
vidyo at 720p 11830 ms -> 10990 ms
i.e., 6%-8% speed-up.
For rtc set, the compression performance
goes down by about -1.3% for both speed -5 and -6.
Change-Id: I2a7794fa99734f739f8b30519ad4dfd511ab91a5
Bug introduced during multiple iterations on: I3831*
gf_group->arf_update_idx[] cannot currently be used
to select the arf buffer index if buffer flipping on overlays
is enabled (still currently the case when multi arf OFF).
Change-Id: I4ce9ea08f1dd03ac3ad8b3e27375a91ee1d964dc
This commit fixes the potential issue in the non-RD mode decision
flow that only checks part of the block to estimate the cost. It
was due to the use of fixed transform size, in replacing the
largest transform block size. This commit enables per transform
block cost estimation of the intra prediction mode in the non-RD
mode decision.
Change-Id: I14ff92065e193e3e731c2bbf7ec89db676f1e132
Cosmetic patch only in response to comments on
previous patches suggesting a couple of name changes
for consistency and clarity.
Change-Id: Ida3a359b0d5755345660d304a7697a3a3686b2a3
This commit replaces a few use cases of cpi->common with preset
variable cm, to avoid unnecessary pointer fetch in the non-RD
coding mode.
Change-Id: I4038f1c1a47373b8fd7bc5d69af61346103702f6
In real-time speed 6, no partition search is done. The inter
prediction results got from picking mode can be reused in the
following encoding process. A speed feature reuse_inter_pred_sby
is added to only enable the resue in speed 6.
This patch doesn't change encoding result. RTC set tests showed
that the encoding speed gain is 2% - 5%.
Change-Id: I3884780f64ef95dd8be10562926542528713b92c
There is a normative scaling range of (x1/2, x16)
for VP9. This patch fixes the maximum downscaling
tests that are applied in the convolve function.
The code used a maximum downscaling limit of x1/5
for historic reasons related to the scalable
coding work. Since the downsampling in this
application is non-normative it will revert to
using a separate non-normative scaler.
Change-Id: Ide80ed712cee82fe5cb3c55076ac428295a6019f
Add indirection to the section of buffer indices.
This is to help simplify things in the future if we
have other codec features that switch indices.
Limit the max GF interval for static sections to fit
the gf_group structures.
Change-Id: I38310daaf23fd906004c0e8ee3e99e15570f84cb
Fix some bugs relating to the use of buffers
in the overlay frames.
Fix bug where a mid sequence overlay was
propagating large partition and transform sizes into
the subsequent frame because of :-
sf->last_partitioning_redo_frequency > 1 and
sf->tx_size_search_method == USE_LARGESTALL
Change-Id: Ibf9ef39a5a5150f8cbdd2c9275abb0316c67873a
This patch implements a mechanism for inserting a second
arf at the mid position of arf groups.
It is currently disabled by default using the flag multi_arf_enabled.
Results are currently down somewhat in initial testing if
multi-arf is enabled. Most of the loss is attributable to the
fact that code to preserve the previous golden frame
(in the arf buffer) in cases where we are coding an overlay
frame, is currently disabled in the multi-arf case.
Change-Id: I1d777318ca09f147db2e8c86d7315fe86168c865
