Marco Paniconi [Mon, 15 Apr 2019 22:59:49 +0000 (15:59 -0700)]
vp9-rtc: tx_size selection for intra mode in nonrd
In nonrd_pickmode for intra modes: add tx_size selection
based on Y prediction signal for the bsize.
The tx selection is done in model_rd, same as inter-modes.
Existing code for intra mode was first setting a tx_size based
only on the bsize, and then in some cases in block_yrd
(during the loop over bsize in units of tx_size) the tx_size
may be set again if model_rd is called in block_yrd.
This CL separates out the tx_size setting (based on Y channel
prediction via model_rd), and then block_yrd is called once
for whole bsize. This allows for better tuning of the tx
selection for intra modes in future change.
Adjust threshold in svc datarate test.
Negligible/neutral change in psnr/ssim metrics
for speed 7 and 8, 1 layer and SVC mode.
Johann [Tue, 23 Apr 2019 16:47:56 +0000 (09:47 -0700)]
remove ARCHITECTURE definition
In the distant past this was used to distinguish between
armv5/6/7 targets when building the assembly files. The
project has not supported armv5/6 for a long time.
For row_mt=1, when mi->skip is set to 1 after parse based on
eobtotal for that partition, dqcoeff and eob need to be restored
as recon_partition doesn't increment these pointers for skip cases
Deepa K G [Tue, 29 Jan 2019 14:28:05 +0000 (19:58 +0530)]
Fix in key frame detection
The frame next to scene cut frame does not usually have
a high second ref useage. Thus the sec ref useage of the
frame next to scene cut frame is tested against a
threshold for scene cut detection.
With this change scene cut detection is improved for
contents where genuine scene cuts were being missed.
Marco Paniconi [Tue, 16 Apr 2019 17:00:53 +0000 (10:00 -0700)]
vp9-rtc: Add speed feature to force SMOOTH filter
Add speed feature for real-time to always force
SMOOTH filter for subpel motion. Can be useful in some
cases for noisy content or high motion at low bitrate.
Also some speedup in avoiding the checking of two filters.
fixes a deadlock with an odd number of threads that go from < number of
tiles to >. the previous calculations were out of sync so going from
e.g., 8 tiles to 2 with 3 threads would result in scheduling only 2
workers, but thread_loop_filter_rows() would expect 3.
Jingning Han [Fri, 12 Apr 2019 22:59:51 +0000 (15:59 -0700)]
Use uniform sampling as initial centers for k-means
The Wiener variance output has been sorted prior to the clustering,
which allows to directly use the uniform sampling as the initial
center points. It avoids empty cluster situations when the samples
are heavily distributed at two far ends and leave the middle empty.
sdeng [Fri, 29 Mar 2019 16:18:56 +0000 (09:18 -0700)]
Add Tune for SSIM
Implementation with some tuning of the paper:
C. Yeo, H. L. Tan, and Y. H. Tan, "On rate distortion optimization using
SSIM," Circuits and Systems for Video Technology, IEEE Transactions on,
vol. 23, no. 7, pp. 1170-1181, 2013.
With switching to clang-7.0.1 we got new warnings. With this change the
warnings are back to 0 for all configurations (excluding warnings in
third_party)
Paul Wilkins [Tue, 2 Apr 2019 15:38:49 +0000 (16:38 +0100)]
Film Mode: Bias based on sub block energy.
Apply a bias for film mode against intra coding
(especially DC_PRED) and compound modes if the sub
blocks of the current block have significantly different
variance in the source.
Frank Galligan [Tue, 26 Mar 2019 06:48:04 +0000 (23:48 -0700)]
Add a test to test rollover of int64 in encoder interface.
The current libvpx encoder interface can potentially rollover an int64_t
value used to calculate the current timestamp. If the timebase was set
to microseconds and first timestamp was 0, then the rollover would
occur in about 10.675 days.
Jingning Han [Wed, 3 Apr 2019 23:46:07 +0000 (16:46 -0700)]
Bypass skip check in tune for sharpness mode
The sharpness mode is enabled for hvc visual quality. Bypass the
skip block check that could potentially force all zero block in
sharpness mode. This resolves the patchy blockiness issue raised
in the 4K SDR HVC encode.
Histogram-based noise estimation algorithm leveraged that low-noise sequences
tend to populate lower-valued histogram bins and high-noise sequences tend to
populate higher-valued histogram bins in a predictable/repeatable manner. The
algorithm compensates for histogram flattening and skewing toward zero as the
scene darkens.
Marco Paniconi [Mon, 1 Apr 2019 21:18:54 +0000 (14:18 -0700)]
vp9-rtc: Move noise estimation to after scene change detection
This allows to use result from scene chage detection to exclude
the current frame from noise estimation analysis if the frame has
scene/ big content change (i.e., high_source_sad flag is set).
The behavior change for noise estimation may be small in practice,
since in the current code, a scene change would have blocks excluded
due to thresh_sum_diff, and the subsequent frames would also be mostly
excluded due to (past) non-zero motion vectors (until the
consec_zeromv > thresh_consec_zeromv is satisfied again).
But its better to completely exclude current frame if its a scene change.
Marco Paniconi [Fri, 29 Mar 2019 22:18:14 +0000 (15:18 -0700)]
vp9-rtc: Speed feature changes for speed 9.
Add threshold multipler for variance partitioning
as speed feature, and increase it by 2x for speed >= 9
for resoln >= VGA. Also only allow simple_interpol
filter when avg_low_motion is below threshold.
Better tradeoff of speed/quality comparing to speed 8.
Paul Wilkins [Tue, 2 Apr 2019 12:00:03 +0000 (12:00 +0000)]
Merge changes Ied3992f4,I83ca669b,I8f1c1a29
* changes:
Further Adjustments to film mode bias.
Add GF group noise weighting in rd_variance_adjustment()
Adjustment to low variance block bias in rd_variance_adjustment()
projects / libvpx / log