JackyChen [Mon, 5 Jan 2015 18:41:12 +0000 (10:41 -0800)]
Adopt weighted averaging in MFQE.
By using weighted averaging in the calculation of the frames to be
displayed, we get an average gain of more than 1 db for key frames
whose base qp are 20 higher than non-key frames.
Jingning Han [Sat, 3 Jan 2015 20:34:44 +0000 (12:34 -0800)]
Fix denoised video output function
This commit fixes the buffer alignment control in denoised video
output function. The encoder is now able to properly store the
denoised input video into provided file when enabled.
Yaowu Xu [Tue, 30 Dec 2014 16:35:40 +0000 (08:35 -0800)]
Properly set size based on actual buffer layout
VP9FrameSizeTestsLarge.OneByOneVideo has been causing a failure in
jenkins libvpx__unit_tests-valgrind_long for "using of uninitialized
memory", the root cause was that the input image for this test was
not initialized with proper size, therefore plan U and V were not
initialized at all.
This commit fixes the size initialization, and resolves the issue.
Jingning Han [Mon, 22 Dec 2014 21:38:34 +0000 (13:38 -0800)]
Enable sub8x8 inter block search for RTC coding mode
This commit enables sub8x8 inter block coding for RTC mode. The
use of sub8x8 blocks can be turned on by allowing
choose_partitioning function to select 4x4/4x8/8x4 block sizes.
Jingning Han [Tue, 23 Dec 2014 17:30:24 +0000 (09:30 -0800)]
Set ref frame scaling factor in RTC inter mode decision
Properly set the corresponding scaling factor of the reference
frame in the non-RD mode decision process. This allows the mode
search process to account for the scaled reference frame when
selecting coding mode.
Yunqing Wang [Tue, 16 Dec 2014 19:10:20 +0000 (11:10 -0800)]
vp9_ethread: add the multi-threaded encoder unit tests
The unit tests for VP9 multi-threaded encoder are added, which
carry out tests for all three modes(i.e. kTwoPassGood, kOnePassGood
and kRealTime), and speeds ranging from 1 to 8. A 1280x720 test
clip is used, which is encoded into multiple tiles. The number of
threads is num_of_tiles.
Jingning Han [Thu, 18 Dec 2014 19:54:13 +0000 (11:54 -0800)]
Add a guard on intra mode skip control for RTC mode
This commit adds a guard condition to the intra mode test skip
control in RTC coding mode. If all inter modes are skipped, force
the encoder to check intra mode. It avoids situations where the
encoder processes without properly assigning required mode
information.
Jingning Han [Wed, 17 Dec 2014 22:26:00 +0000 (14:26 -0800)]
Rework mode search threshold update for RTC coding mode
In RTC coding mode, the alternate reference frame modes and compound
inter prediction modes are disabled. This commit reworks the
related mode search threshold update process to skip interacting
with these coding modes. It provides about 1.5% speed-up for speed
-6 on average.
vidyo1
16551 b/f, 40.451 dB, 6261 ms -> 16550 b/f, 40.459 dB, 6190 ms
nik720p
33316 b/f, 38.795 dB, 6335 ms -> 33310 b/f, 38.798 dB, 6237 ms
mmmoving
33265 b/f, 41.055 dB, 7176 ms -> 33267 b/f, 41.064 dB, 7084 ms
dark720
33329 b/f, 39.729 dB, 11235 ms -> 33331 b/f, 39.733 dB, 10731 ms
Jingning Han [Wed, 17 Dec 2014 00:02:58 +0000 (16:02 -0800)]
Properly store the tx_size of selected intra mode
Use a temporary variable to store the transform size associated
with the best intra mode and restore the mode_info if the overall
best mode is intra mode.
Yunqing Wang [Fri, 12 Dec 2014 22:34:30 +0000 (14:34 -0800)]
Improve the libvpx encoder test driver
The encoder initialization is called in EncodeFrame(). Therefore,
in the unit tests, the set control is done when video->frame() is 1.
This didn't cause problem since current tests mainly test lag_frame
> 0 case, or no encoding option that needs to allocate memory before
1st frame is used. If use lag_frame = 0 and encoding multiple tiles,
the unit tests crash. The issue is fixed by doing the initialization
before encoding frames.
Jingning Han [Tue, 16 Dec 2014 17:21:22 +0000 (09:21 -0800)]
Remove reset mode_info array per frame
The mode_info array was unnecessarily reset to zero every frame
when error resilient mode turned on, given that the mode info
values per block will be assigned during mode search stage.
This commit removes this reset operation. It reduces the runtime
cost on memset operation to 1/3. The overall speed -6 runtime is
reduced by 2%.
Jingning Han [Tue, 16 Dec 2014 23:46:59 +0000 (15:46 -0800)]
Set second ref frame to be NONE in key frame coding
This commit explicitly set the second reference frame type to be
NONE in key frame coding mode. This fixes a subtle dependency of
reference motion vector used by next inter frame on mode_info
reset before key frame coding.
Paul Wilkins [Thu, 11 Dec 2014 16:22:03 +0000 (16:22 +0000)]
Improve motion detection for low complexity regions.
Where there is very subtle motion, especially when combined
with low spatial complexity, the codec sometimes fails to quickly
pick up the ambient motion field.
Once it has been established though the field propagates well using
Nearest and Near MV.
This patch looks specifically at the case where the Nearest and Near
have not been established as non zero vectors and in this case
discounts the cost of searching for a new vector in the rd code.
This will almost certainly have some implications in terms of encode
speed but it should be possible to mitigate the impact in a subsequent
using first pass stats and the local spatial complexity.
Average results for test sets approximately neutral.
projects / libvpx / log