James Zern [Sat, 9 May 2015 03:17:20 +0000 (20:17 -0700)]
build_intra_predictors*: reduce above_data size
currently this needs to be 2x (NEED_ABOVERIGHT) the size of the largest
block (32) + 1 (for above_left). reduce the buffer size from 128 + 16
(alignment) to 64 + 16.
hkuang [Fri, 8 May 2015 22:42:50 +0000 (15:42 -0700)]
Fix clang ioc warning due to NULL mi pointer.
The warning only happens in VP9 encoder's first pass due to src_mi
is not set up yet. But it will not fail the encoder as left_mi and
above_mi are not used in the first_pass and they will be set up again
in the second pass.
paulwilkins [Fri, 1 May 2015 15:04:52 +0000 (16:04 +0100)]
Skip the last frame update for some frame repeats.
Where a frame appears to be a repeat of an earlier
frame or frame buffer, but the first pass code
does not anticipate this (usually because it is matching
the GF or ARF buffer not the last frame buffer), do not
update the last frame buffer.
This helps ensure that the content of the last frame buffer
is kept "different" where possible, and not updated to
match the GF or ARF. This is particularly helpful in some
animated sequences where there are groups of repeating
frames. Here it has quite a big impact. However, in most
of our standard test clips it has little or no impact.
James Zern [Sat, 2 May 2015 20:24:16 +0000 (13:24 -0700)]
replace DECLARE_ALIGNED_ARRAY w/DECLARE_ALIGNED
this macro was used inconsistently and only differs in behavior from
DECLARE_ALIGNED when an alignment attribute is unavailable. this macro
is used with calls to assembly, while generic c-code doesn't rely on it,
so in a c-only build without an alignment attribute the code will
function as expected.
Yunqing Wang [Fri, 1 May 2015 20:20:50 +0000 (13:20 -0700)]
Add intra mode early termination in non-rd mode
Added the intra mode early termination in order to
speed up the mode search in non-rd case since we
started to include more intra modes in the search
list. Borg tests(rtc set) showed a 0.048% PSNR gain
and 0.061 SSIM gain. No speed change.
Marco [Thu, 30 Apr 2015 18:39:02 +0000 (11:39 -0700)]
Allow for H and V intra modes for non-rd mode.
For non-rd mode (speed >=5): use mask based on prediction block size, and
(for non-screen content mode) allow for checking horiz and vert intra modes
for blocks sizes < 16x16.
Avg psnr/ssim metrics go up by about ~0.2%.
Only allowing H/V intra on block sizes below 16x16 for now, to keep
encoding time increase very small, and also when allowing H/V on 16x16 blocks,
metrics went down on a few clips which need to be further examined.
paulwilkins [Fri, 1 May 2015 12:45:43 +0000 (13:45 +0100)]
Image size restriction to rd auto partition search.
Impose a limit on the rd auto partition search based on
the image format. Smaller formats require that the search
includes includes a smaller minimum block size.
This change is intended to mitigate the visual impact of
ringing in some problem clips, for smaller image formats.
paulwilkins [Fri, 1 May 2015 10:31:13 +0000 (11:31 +0100)]
Remove CONSTRAIN_NEIGHBORING_MIN_MAX.
Remove one of the auto partition size cases.
This case can behaves badly in some types of animated content
and was only used for the rd encode path. A subsequent patch
will add additional checks to help further improve visual quality.
Previously limit on max interval set to 0.5 seconds.
Though this helped some low frame rate material it
appears to be a bit too aggressive for some 24 and 25 fps
content. This patch relaxes the limit to 0.75 seconds.
The patch also adds a new minimum interval variable
to replace the current hard wired value. This allows us
to impose a limit on the maximum number of primary
arfs per second for high frame rate (e.g. 50 & 60fps)
content. This is to address concerns regarding playback
performance on some platforms if there is a high base
frame rate and very frequent arfs.
James Zern [Sat, 25 Apr 2015 05:33:17 +0000 (22:33 -0700)]
vp9_idct_intrin_*: RECON_AND_STORE: remove dest offset
offsetting by a variable stride prevents instruction reordering,
resulting in poor assembly.
additionally reroll 16x16/32x32 loops to reduce register spill with this
new format
Yunqing Wang [Thu, 30 Apr 2015 18:03:06 +0000 (11:03 -0700)]
Reduce intra_cost_penalty for BLOCK_8X8
This patch reduced the BLOCK_8X8's intra_cost_penalty, which
allows 8x8 blocks to conduct intra mode search. Borg test
result(rtc set): 0.077% PSNR gain, 0.228% SSIM gain. No speed
changes.
The rotation computation using 2X of cos(pi/16) has a potential to
overflow 32 bit, this commit disable the function to allow further
investigation and optimization.
Yunqing Wang [Mon, 27 Apr 2015 19:40:00 +0000 (12:40 -0700)]
Adjust the vbp early termination threshold slightly
Calculated cpi->vbp_threshold_sad from this frame's dequant value.
The encoding quality and speed didn't change much. Borg test
result: PSNR: -0.002%, SSIM: -0.003%.
The version is currently producing different result from c version
for some input. Disable the use of it for now to allow time for
investigation the source of mismatch.
Yunqing Wang [Fri, 24 Apr 2015 19:12:17 +0000 (12:12 -0700)]
Improve golden frame refreshing in non-rd mode
The default golden frame interval was doubled. After encoding a
frame, the background motion was measured. If the motion was high,
the current frame was set as the golden frame. Currently, the
changes were applied only while aq-mode 3 was on.
Borg tests(rtc set) showed a 0.226% PSNR gain and 0.312% SSIM gain.
No speed changes.
For color sampling format other than 420, valid partion size in Y may
not work for UV plane. This commit adds validation of UV partition
size before select the partition choice.
This fixes a crash for real time encoding of 422 input.
The existing test was triggering a lot of false positives on some types
of animated material with very plain backgrounds. These were triggering
code designed to catch key frames in letter box format clips.
This patch tightens up the criteria and imposes a minimum requirement
on the % blocks coded intra in the first pass and the ratio between the
% coded intra and the modified inter % after discounting neutral (flat)
blocks that are coded equally well either way.
On a particular problem animation clip this change eliminated a large
number of false positives including some cases where the old code
selected kf several times in a row. Marginal false negatives are less
damaging typically to compression and in the problem clip there are now
a couple of cases where "visual" scene cuts are ignored because of well
correlated content across the scene cut.
Replaced some magic numbers related to this with #defines and added
explanatory comments.
projects / libvpx / log