Paul Wilkins [Fri, 5 Oct 2012 10:16:46 +0000 (11:16 +0100)]
Fix SIMD unsafe use of floating point.
This commit fixes unsafe simd / floating point interactions arising
from the current hybrid and 16x16 transform implementation.
These led to a raft of bugs and issues when the project was
built using VS2008 for Win32 though they did not show up with
the unix builds.
Gerrit makes a meal out of presenting the fix but all I have actually
done is indent the body of each function that uses floating point by
one level and bracket with emms instructions using the function
vp8_clear_system_state(). See below.
function () {
vp8_clear_system_state();
{
... function body
}
vp8_clear_system_state();
}
This is almost certainly over the top in terms of number of emms
instructions but is a temporary measure pending implementation of
integer variants of each function to replace the floating point.
Limited testing suggests that this fixes the problems that arose for
Win32 VS2008 when the hybrid or 16x16 transforms were enabled.
Deb Mukherjee [Mon, 10 Sep 2012 05:42:35 +0000 (22:42 -0700)]
Entropy coding for hybrid transform
Separates the entropy coding context models for 4x4, 8x8 and 16x16
ADST variants.
There is a small improvement for HD (hd/std-hd) by about 0.1-0.2%.
Results on derf/yt are about the same, probably because there is not
enough statistics.
Results may improve somewhat once the initial probability tables are
updated for the hybrid transforms which is coming soon.
Deb Mukherjee [Fri, 21 Sep 2012 21:20:15 +0000 (14:20 -0700)]
Separating 8x8DCT use from hybridtransform8x8
With this change, even if hybridtransform8x8 experiment is off,
8x8 dct is used for the I8x8 mode. However note that the gains
observed with the hybridtransform8x8 experiment will now be less,
since part of the gain is now merged in.
The commit changed to use 3 rows above and 3 cols from left for SAD
scoring for selecting the best reference motion vector. The change
helped std-hd set by >.2% on psnr/ssim metrics.
Deb Mukherjee [Thu, 6 Sep 2012 16:07:42 +0000 (09:07 -0700)]
Adds feature for companded MV encoding
The high-precision (1/8) pel bit is turned off if the reference
MV is larger than a threshold. The motivation for this patch is
the intuition that if motion is likely large (as indicated by
the reference), there is likley to be more motion blur, and as
a result 1/8 pel precision would be wasteful both in rd sense
as well as computationally.
The feature is incorporated as part of the newmventropy experiment.
There is a modest RD improvement with the patch. Overall the
results with the newmventropy experiment with the threshold being
16 integer pels are:
Paul Wilkins [Fri, 7 Sep 2012 13:21:22 +0000 (14:21 +0100)]
Partial merge of NEWBESTREFMV and NEW_MVREF.
This commit merges those parts of the CONFIG_NEW_MVREF
that specifically relate to choosing a better set of candidate
MV references into the NEWBESTREFMV experiment.
CONFIG_NEW_MVREF will then be used for changes relating
to the explicit coding of a cost optimized MV reference in the
bitstream as part of MV coding.
Deb Mukherjee [Thu, 26 Jul 2012 20:42:07 +0000 (13:42 -0700)]
New motion vector entropy coding
Adds a new experiment with redesigned/refactored motion vector entropy
coding. The patch also takes a first step towards separating the
integer and fractional pel components of a MV. However the fractional
pel encoding still depends on the integer pel part and so they are
not fully independent. Further experiments are in progress to see
how much they can be decoupled without affecting performance.
All components including entropy coding/decoding, costing for MV
search, forward updates and backward updates to probability tables,
have been implemented.
Jingning Han [Wed, 29 Aug 2012 18:25:38 +0000 (11:25 -0700)]
hybrid transform of 16x16 dimension
Enable ADST/DCT of dimension 16x16 for I16X16 modes. This change provides
benefits mostly for hd sequences.
Set up the framework for selectable transform dimension.
Also allowing quantization parameter threshold to control the use
of hybrid transform (This is currently disabled by setting threshold
always above the quantization parameter. Adaptive thresholding can
be built upon this, which will further improve the coding performance.)
The coding performance gains (with respect to the codec that has all
other configuration settings turned on) are
Yaowu Xu [Fri, 24 Aug 2012 16:59:49 +0000 (09:59 -0700)]
change to properly account for coef scaling in unit tests
The transform functions in experimental branch absorbed a scaling
factor of 4 to allow quantization steps closer to unit quantizer.
This commit added scaling code in between forward and inverse
transform to properly account for the scaling factor.
Paul Wilkins [Fri, 24 Aug 2012 14:44:01 +0000 (15:44 +0100)]
New Motion Reference Search
Alternative strategy for finding a list of candidate motion
vectors to use as reference values in mv coding and as
nearest and near.
Sort by sad in vp8_find_best_ref_mvs() rather than just
pick the best. Allow 0,0 as a best ref option but not a
nearest or near unless there are no alternatives.
Encode/Decode verified on at least some clips.
Some commented out experimental and stats code still in place.
Gain over existing code averages about 1% on derf (alll metrics)
with improvement on all clips. Other test results pending.
The entropy coding of the mode (nearest/near etc) still
depends upon and requires the old "findnear" code so
this needs looking at and may provide room for further gains.
Ronald S. Bultje [Mon, 20 Aug 2012 21:43:34 +0000 (14:43 -0700)]
Superblock coding.
This commit adds a pick_sb_mode() function which selects the best 32x32
superblock coding mode. Then it selects the best per-MB modes, compares
the two and encodes that in the bitstream.
The bitstream coding is rather simplistic right now. At the SB level,
we code a bit to indicate whether this block uses SB-coding (32x32
prediction) or MB-coding (anything else), and then we follow with the
actual modes. This could and should be modified in the future, but is
omitted from this commit because it will likely involve reorganizing
much more code rather than just adding SB coding, so it's better to let
that be judged on its own merits.
Gains on derf: about even, YT/HD: +0.75%, STD/HD: +1.5%.
Jingning Han [Thu, 16 Aug 2012 20:25:49 +0000 (13:25 -0700)]
Fixed decoder mismatch issue
Resolved the decoder mismatch issue due to quantization parameter
threshold for hybrid transform coding. The macroblock dequantizer
initialization is moved to be performed before coefficient
detokenization, since the (de)tokenization is now dependent on the
macroblock level quantization parameter.
Ronald S. Bultje [Thu, 16 Aug 2012 18:09:31 +0000 (11:09 -0700)]
Set a default mv entropy if no tokens are coded.
This doesn't affect the result, since there are no MVs coded using this
entropy. It does, however, silence valgrind warnings about uninitialized
variables.
Yaowu Xu [Mon, 6 Aug 2012 17:51:20 +0000 (10:51 -0700)]
Changed to use reference mv as nearest mv
The reference motion vector selected by surrounding pixels that has
the best matching score is used as nearest motion vector.
The change has shown consistent gain on all test sets, compression
gains range from .2% to .6%. The variation is largely dependent on
various other experiments on or off.
Paul Wilkins [Fri, 10 Aug 2012 13:12:43 +0000 (14:12 +0100)]
Code cleanup.
Add local variable in several places to reference the MB mode
info structure. Currently this is usually accessed in the code as
x->e_mbd.mode_info_context->mbmi.* or in some places
xd->mode_info_context->mbmi.*
Resolved some uses of x-> for the MACROBLOCKD structure.
Rebased without dependency on motion reference experiment.
Merge "Cleanup and commenting of the sub-pixel interpolation functions. Mostly as a preparation for the upcoming vectorized implementation." into experimental
projects / libvpx / log