platform/upstream/libvpx.git
11 years agoMerge changes Ife0d8147,I7d469716,Ic9a5615f into experimental
John Koleszar [Fri, 8 Feb 2013 21:19:47 +0000 (13:19 -0800)]
Merge changes Ife0d8147,I7d469716,Ic9a5615f into experimental

* changes:
  Restore SSSE3 subpixel filters in new convolve framework
  Convert subpixel filters to use convolve framework
  Add 8-tap generic convolver

11 years agoRestore SSSE3 subpixel filters in new convolve framework
John Koleszar [Fri, 8 Feb 2013 01:00:37 +0000 (17:00 -0800)]
Restore SSSE3 subpixel filters in new convolve framework

This commit adds the 8 tap SSSE3 subpixel filters back into the code
underneath the convolve API. The C code is still called for 4x4
blocks, as well as compound prediction modes. This restores the
encode performance to be within about 8% of the baseline.

Change-Id: Ife0d81477075ae33c05b53c65003951efdc8b09c

11 years agomove dct/idct constants to a header file
Yaowu Xu [Thu, 7 Feb 2013 19:51:23 +0000 (11:51 -0800)]
move dct/idct constants to a header file

also removed some un-unsed functions.

Change-Id: Ie363bcc8d94441d054137d2ef7c4fe59f56027e5

11 years agoButterfly ADST based hybrid transform
Jingning Han [Tue, 5 Feb 2013 20:37:13 +0000 (12:37 -0800)]
Butterfly ADST based hybrid transform

Refactor the 8x8 inverse hybrid transform. It is now consistent
with the new inverse DCT. Overall performance loss (due to the
use of this variant ADST, and the rounding errors in the butterfly
implementation) for std-hd is -0.02.

Fixed BUILD warning.

Devise a variant of the original ADST, which allows butterfly
computation structure. This new transform has kernel of the
form: sin((2k+1)*(2n+1) / (4N)). One of its butterfly structures
using floating-point multiplications was reported in Z. Wang,
"Fast algorithms for the discrete W transform and for the discrete
Fourier transform", IEEE Trans. on ASSP, 1984.

This patch includes the butterfly implementation of the inverse
ADST/DCT hybrid transform of dimension 8x8.

Change-Id: I3533cb715f749343a80b9087ce34b3e776d1581d

11 years agoAdded skip switches for SB32 and SB64
Paul Wilkins [Tue, 5 Feb 2013 12:54:09 +0000 (12:54 +0000)]
Added skip switches for SB32 and SB64

Added switches and code to skip/breakout from
doing SB32 and SB64 tests based on whether
the 16x16 MB tests used split modes. Also to
optionally skip 64x64 if 16x16 was chosen over
32x32.

Impact varies depending on clip from a few %
up to almost 50% on encode speed. Only the
split mode breakout is currently enabled.

Change-Id: Ib5836140b064b350ffa3057778ed2cadcc495cf8

11 years agoUse fdct8x4 instead of fdct4x4 where the block size allows it.
Ronald S. Bultje [Wed, 6 Feb 2013 22:13:05 +0000 (14:13 -0800)]
Use fdct8x4 instead of fdct4x4 where the block size allows it.

This allows for faster SIMD implementations in the future (currently
there is no speed impact).

Change-Id: I732647e9148b5dcb44e6bc8728138f0141218329

11 years agoUse configure checks for various inline keywords.
Ronald S. Bultje [Wed, 6 Feb 2013 20:45:28 +0000 (12:45 -0800)]
Use configure checks for various inline keywords.

Change-Id: I8508f1a3d3430f998bb9295f849e88e626a52a24

11 years agoAdd sse2 versions of sub_pixel_variance{32x32,64x64}.
Ronald S. Bultje [Wed, 6 Feb 2013 19:20:59 +0000 (11:20 -0800)]
Add sse2 versions of sub_pixel_variance{32x32,64x64}.

7.5% faster overall encoding.

Change-Id: Ie9bb7f9fdf93659eda106404cb342525df1ba02f

11 years agoMerge "Reindent segmentation code." into experimental
Ronald S. Bultje [Wed, 6 Feb 2013 18:07:30 +0000 (10:07 -0800)]
Merge "Reindent segmentation code." into experimental

11 years agoReindent segmentation code.
Ronald S. Bultje [Wed, 6 Feb 2013 17:08:30 +0000 (09:08 -0800)]
Reindent segmentation code.

Indentation was off by 2 spaces for this particular block.

Change-Id: I1e587b7ad3eff77ade5521252d20c7bb2daa0f6d

11 years agoEliminate tautology
John Koleszar [Wed, 6 Feb 2013 16:22:48 +0000 (08:22 -0800)]
Eliminate tautology

      Unreachable code
  that does nothing anyway
      removed forever.

Change-Id: I14105d2dd9dbc9d558f36464055e350dbeb45488

11 years agoMerge "Change definition of NearestMV." into experimental
Paul Wilkins [Wed, 6 Feb 2013 12:06:31 +0000 (04:06 -0800)]
Merge "Change definition of NearestMV." into experimental

11 years agoFix mismatch after merge of the tiling patch.
Ronald S. Bultje [Wed, 6 Feb 2013 01:15:04 +0000 (17:15 -0800)]
Fix mismatch after merge of the tiling patch.

Change-Id: I8ecc178b4d4069e721c7fec6d7631c00e4a3e5d5

11 years ago[WIP] Add column-based tiling.
Ronald S. Bultje [Fri, 1 Feb 2013 17:35:28 +0000 (09:35 -0800)]
[WIP] Add column-based tiling.

This patch adds column-based tiling. The idea is to make each tile
independently decodable (after reading the common frame header) and
also independendly encodable (minus within-frame cost adjustments in
the RD loop) to speed-up hardware & software en/decoders if they used
multi-threading. Column-based tiling has the added advantage (over
other tiling methods) that it minimizes realtime use-case latency,
since all threads can start encoding data as soon as the first SB-row
worth of data is available to the encoder.

There is some test code that does random tile ordering in the decoder,
to confirm that each tile is indeed independently decodable from other
tiles in the same frame. At tile edges, all contexts assume default
values (i.e. 0, 0 motion vector, no coefficients, DC intra4x4 mode),
and motion vector search and ordering do not cross tiles in the same
frame.
t log

Tile independence is not maintained between frames ATM, i.e. tile 0 of
frame 1 is free to use motion vectors that point into any tile of frame
0. We support 1 (i.e. no tiling), 2 or 4 column-tiles.

The loopfilter crosses tile boundaries. I discussed this briefly with Aki
and he says that's OK. An in-loop loopfilter would need to do some sync
between tile threads, but that shouldn't be a big issue.

Resuls: with tiling disabled, we go up slightly because of improved edge
use in the intra4x4 prediction. With 2 tiles, we lose about ~1% on derf,
~0.35% on HD and ~0.55% on STD/HD. With 4 tiles, we lose another ~1.5%
on derf ~0.77% on HD and ~0.85% on STD/HD. Most of this loss is
concentrated in the low-bitrate end of clips, and most of it is because
of the loss of edges at tile boundaries and the resulting loss of intra
predictors.

TODO:
- more tiles (perhaps allow row-based tiling also, and max. 8 tiles)?
- maybe optionally (for EC purposes), motion vectors themselves
  should not cross tile edges, or we should emulate such borders as
  if they were off-frame, to limit error propagation to within one
  tile only. This doesn't have to be the default behaviour but could
  be an optional bitstream flag.

Change-Id: I5951c3a0742a767b20bc9fb5af685d9892c2c96f

11 years agoMerge "Add SSE3 versions for sad{32x32,64x64}x4d functions." into experimental
Ronald S. Bultje [Tue, 5 Feb 2013 23:40:46 +0000 (15:40 -0800)]
Merge "Add SSE3 versions for sad{32x32,64x64}x4d functions." into experimental

11 years agoMerge "fix a build issue with MSVC on windows" into experimental
Yaowu Xu [Tue, 5 Feb 2013 23:26:47 +0000 (15:26 -0800)]
Merge "fix a build issue with MSVC on windows" into experimental

11 years agoMerge "rewrite 4x4 idct and fdct" into experimental
Yaowu Xu [Tue, 5 Feb 2013 23:26:36 +0000 (15:26 -0800)]
Merge "rewrite 4x4 idct and fdct" into experimental

11 years agoAdd SSE3 versions for sad{32x32,64x64}x4d functions.
Ronald S. Bultje [Tue, 5 Feb 2013 23:21:47 +0000 (15:21 -0800)]
Add SSE3 versions for sad{32x32,64x64}x4d functions.

Overall encoding about 15% faster.

Change-Id: I176a775c704317509e32eee83739721804120ff2

11 years agoConvert subpixel filters to use convolve framework
John Koleszar [Tue, 29 Jan 2013 00:59:03 +0000 (16:59 -0800)]
Convert subpixel filters to use convolve framework

Update the code to call the new convolution functions to do subpixel
prediction rather than the existing functions. Remove the old C and
assembly code, since it is unused. This causes a 50% performance
reduction on the decoder, but that will be resolved when the asm for
the new functions is available.

There is no consensus for whether 6-tap or 2-tap predictors will be
supported in the final codec, so these filters are implemented in
terms of the 8-tap code, so that quality testing of these modes
can continue. Implementing the lower complexity algorithms is a
simple exercise, should it be necessary.

This code produces slightly better results in the EIGHTTAP_SMOOTH
case, since the filter is now applied in only one direction when
the subpel motion is only in one direction. Like the previous code,
the filtering is skipped entirely on full-pel MVs. This combination
seems to give the best quality gains, but this may be indicative of a
bug in the encoder's filter selection, since the encoder could
achieve the result of skipping the filtering on full-pel by selecting
one of the other filters. This should be revisited.

Quality gains on derf positive on almost all clips. The only clip
that seemed to be hurt at all datarates was football
(-0.115% PSNR average, -0.587% min). Overall averages 0.375% PSNR,
0.347% SSIM.

Change-Id: I7d469716091b1d89b4b08adde5863999319d69ff

11 years agoAdd 8-tap generic convolver
John Koleszar [Fri, 25 Jan 2013 17:47:09 +0000 (09:47 -0800)]
Add 8-tap generic convolver

This commit introduces a new convolution function which will be used to
replace the existing subpixel interpolation functions. It is much the
same as the existing functions, but allows for changing the filter
kernel on a per-pixel basis, and doesn't bake in knowledge of the
filter to be applied or the size of the resulting block into the
function name.

Replacing the existing subpel filters will come in a later commit.

Change-Id: Ic9a5615f2f456cb77f96741856fc650d6d78bb91

11 years agofix a build issue with MSVC on windows
Yaowu Xu [Mon, 4 Feb 2013 23:22:55 +0000 (15:22 -0800)]
fix a build issue with MSVC on windows

for idct 16x16 unit test

Change-Id: I51da9405c3a4d7bb3f4cdf062aaccaa90b33dca4

11 years agorewrite 4x4 idct and fdct
Yaowu Xu [Mon, 4 Feb 2013 23:22:32 +0000 (15:22 -0800)]
rewrite 4x4 idct and fdct

This commit changes the 4x4 iDCT to use same algorithm & constants as
other iDCTs. The 4x4 fDCT is also changed to be based on the new iDCT.

Change-Id: Ib1a902693228af903862e1f5a08078c36f2089b0

11 years agoChange definition of NearestMV.
Paul Wilkins [Tue, 5 Feb 2013 10:28:26 +0000 (10:28 +0000)]
Change definition of NearestMV.

This commit makes the NearestMV match the chosen
best reference MV. It can be a 0,0 or non zero vector
which means the the compound nearest mv mode can
combine a 0,0 and a non zero vector.

Change-Id: I2213d09996ae2916e53e6458d7d110350dcffd7a

11 years agoMerge "Added vp9_short_idct1_32x32_c" into experimental
Scott LaVarnway [Tue, 5 Feb 2013 16:56:05 +0000 (08:56 -0800)]
Merge "Added vp9_short_idct1_32x32_c" into experimental

11 years agoMerge "Re-factor code for rd thresholds." into experimental
Paul Wilkins [Tue, 5 Feb 2013 10:12:45 +0000 (02:12 -0800)]
Merge "Re-factor code for rd thresholds." into experimental

11 years agoAdded vp9_short_idct1_32x32_c
Scott LaVarnway [Tue, 5 Feb 2013 00:49:17 +0000 (16:49 -0800)]
Added vp9_short_idct1_32x32_c

and called this function in vp9_dequant_idct_add_32x32_c when
eob == 1.  For the test clip used, the decoder performance improved
by 21+%.  Based on Yaowu's 16 point idct work.

Change-Id: Ib579a90fed531d45777980e04bf0c9b23c093c43

11 years agoRe-factor code for rd thresholds.
Paul Wilkins [Mon, 4 Feb 2013 17:43:02 +0000 (17:43 +0000)]
Re-factor code for rd thresholds.

Separate out code to set the main encode speed
related rd thresholds. Some values changed from
the initial defaults for various new modes.

Quality test results pending but even the addition
of some further non-zero defaults helps encode speed
somewhat in limited testing on derf clips.

Adjustment of thresholds for quality / speed tradeoff
to follow.

Change-Id: I117ee473157e151a1b93193d5f393449328de20d

11 years agoAdded INT16_MIN and INT16_MAX for MSVC builds
Yaowu Xu [Mon, 4 Feb 2013 18:16:59 +0000 (10:16 -0800)]
Added INT16_MIN and INT16_MAX for MSVC builds

These macros were not defined in earlier version of MSVC

Change-Id: I8270a3abb7c6e9ead1931a653d7e41f877a1017b

11 years agoenable 16x16 iDCT unit test
Yaowu Xu [Fri, 1 Feb 2013 19:02:42 +0000 (11:02 -0800)]
enable 16x16 iDCT unit test

test for forward transform will be enabled later after re-do forward
transform

Change-Id: Ie7c7cf88baf7ecbebbe52fe027e1c3b33d3b9d49

11 years agore-write 8 point idct
Yaowu Xu [Fri, 1 Feb 2013 23:34:28 +0000 (15:34 -0800)]
re-write 8 point idct

to be consistent with idct16 and idct32.

Change-Id: Ie89dbd32b65c33274b7fecb4b41160fcf1962204

11 years agoa couple of minor fixes
Yaowu Xu [Sat, 2 Feb 2013 00:45:54 +0000 (16:45 -0800)]
a couple of minor fixes

fixed a function prototypes to prevent compiler warnings;
removed a function not in use;
un-capitialize "Refstride" to ref_stride

Change-Id: Ib4472b6084f357d96328c6a06e795b6813a9edba

11 years agoMerge "Changes 16 point idct" into experimental
Yaowu Xu [Fri, 1 Feb 2013 16:22:20 +0000 (08:22 -0800)]
Merge "Changes 16 point idct" into experimental

11 years agoMerge "fix a small bug in 16 point forward dct" into experimental
Yaowu Xu [Fri, 1 Feb 2013 13:57:41 +0000 (05:57 -0800)]
Merge "fix a small bug in 16 point forward dct" into experimental

11 years agoChanges 16 point idct
Yaowu Xu [Fri, 1 Feb 2013 00:16:28 +0000 (16:16 -0800)]
Changes 16 point idct

This commit changes the inverse 16 point dct to use the same algorithm
as the one for 32 point idct. In fact, now 16 point dct uses the exact
version of the souce code for even portion of the 32 point idct.

Tests showed current implementation has significant better accuracy
than the previous version. With this implementation and the minor bug
fix on forward 16 point dct, encoding tests showed about 0.2% better
compression of CIF set, test results on std-hd setting pending.

Change-Id: I68224b60c816ba03434e9f08bee147c7e344fb63

11 years agofix a small bug in 16 point forward dct
Yaowu Xu [Thu, 31 Jan 2013 23:39:41 +0000 (15:39 -0800)]
fix a small bug in 16 point forward dct

The commit fixes a minor error in 16 point fdct where in a rotation can
produce result of -1 instead of 0.

Change-Id: I45aac4a52bcd06225c6d04e643547a13e1c1aade

11 years agoMerge "A fix point implementation of 32x32 idct" into experimental
Yaowu Xu [Thu, 31 Jan 2013 18:48:01 +0000 (10:48 -0800)]
Merge "A fix point implementation of 32x32 idct" into experimental

11 years agoA fix point implementation of 32x32 idct
Yaowu Xu [Wed, 30 Jan 2013 21:01:49 +0000 (13:01 -0800)]
A fix point implementation of 32x32 idct

This commit changes the 32x32 idct to use integer only. The algorithm
was taken directly from "A Fast Computational Algorithm for the
Discrete Cosine Tranform" by W. Chen, et al., which was published in
IEEE Transaction on Communication Vol. Com.-25 No. 9, 1977. The signal
flow graph in the original paper is for a 32 point forward dct, the
current implementation of inverse DCT was done by follow the graph in
reversed direction.

With this implementation, the 32 point inverse dct contains a 16 point
inverse dct in its even portion, similarly the 16 point idct further
contains 8 point and 4 point inverse dcts.

As of patch 4, encoding tests showed there is no compression loss when
compared against the floating point baseline. Numbers even showed very
small postives. (cif: .01%, std-hd: .05%).

Change-Id: I2d2d17a424b0b04b42422ef33ec53f5802b0f378

11 years agoMerge "Adding a frame parallel decoding mode" into experimental
Deb Mukherjee [Wed, 30 Jan 2013 20:03:45 +0000 (12:03 -0800)]
Merge "Adding a frame parallel decoding mode" into experimental

11 years agoMerge "don't code the branch for the predicted seg_id if that flag is false." into...
Ronald S. Bultje [Wed, 30 Jan 2013 18:02:51 +0000 (10:02 -0800)]
Merge "don't code the branch for the predicted seg_id if that flag is false." into experimental

11 years agodon't code the branch for the predicted seg_id if that flag is false.
Ronald S. Bultje [Wed, 30 Jan 2013 17:30:46 +0000 (09:30 -0800)]
don't code the branch for the predicted seg_id if that flag is false.

Change-Id: Icb6e21dc0c2d9918faa33c8bf70943660df7ad88

11 years agoMerge "Default superblock skip flag to 32x32 for skip-blocks." into experimental
Ronald S. Bultje [Wed, 30 Jan 2013 17:12:17 +0000 (09:12 -0800)]
Merge "Default superblock skip flag to 32x32 for skip-blocks." into experimental

11 years agoMerge "Reset skip flag in superblock RD loop." into experimental
Ronald S. Bultje [Wed, 30 Jan 2013 17:12:02 +0000 (09:12 -0800)]
Merge "Reset skip flag in superblock RD loop." into experimental

11 years agoMerge "Further improvement on compound inter-intra expt" into experimental
Deb Mukherjee [Wed, 30 Jan 2013 16:38:17 +0000 (08:38 -0800)]
Merge "Further improvement on compound inter-intra expt" into experimental

11 years agoDefault superblock skip flag to 32x32 for skip-blocks.
Ronald S. Bultje [Wed, 30 Jan 2013 05:46:31 +0000 (21:46 -0800)]
Default superblock skip flag to 32x32 for skip-blocks.

This is identical to the later decisions made in encode_superblock().
This commit doesn't actually change anything, but makes the mbmi state
more consistent between the RD loop and the final encode result.

Change-Id: I9e735afb7c5a52e5b61728cb88c67ef9b9bf59be

11 years agoReset skip flag in superblock RD loop.
Ronald S. Bultje [Wed, 30 Jan 2013 05:42:46 +0000 (21:42 -0800)]
Reset skip flag in superblock RD loop.

This is the superblock equivalent of commit 290b83a.

Change-Id: Ib3945dd9e992fa9ec1fdea5a11e17a3cc0e37637

11 years agoWrite only visible area (for better comparison with rec.yuv).
Ronald S. Bultje [Wed, 30 Jan 2013 00:58:52 +0000 (16:58 -0800)]
Write only visible area (for better comparison with rec.yuv).

Change-Id: I32bf4ee532a15af78619cbcd8a193224029fab50

11 years agoMerge "Fix block pointer corruption in intra8x8 prediction with 4x4 transform." into...
Ronald S. Bultje [Tue, 29 Jan 2013 20:49:42 +0000 (12:49 -0800)]
Merge "Fix block pointer corruption in intra8x8 prediction with 4x4 transform." into experimental

11 years agoMerge "Fix overread/write reported by valgrind if (mb_cols) & 3 != 0." into experimental
Ronald S. Bultje [Tue, 29 Jan 2013 20:49:22 +0000 (12:49 -0800)]
Merge "Fix overread/write reported by valgrind if (mb_cols) & 3 != 0." into experimental

11 years agoMerge "Remove eob_max_offset markers." into experimental
Paul Wilkins [Tue, 29 Jan 2013 17:29:45 +0000 (09:29 -0800)]
Merge "Remove eob_max_offset markers." into experimental

11 years agoMerge "Segment Skip Flag" into experimental
Paul Wilkins [Tue, 29 Jan 2013 17:29:26 +0000 (09:29 -0800)]
Merge "Segment Skip Flag" into experimental

11 years agoMerge "Added eob == 0 check to vp9_dequant_idct_add_32x32_c" into experimental
Scott LaVarnway [Tue, 29 Jan 2013 17:19:58 +0000 (09:19 -0800)]
Merge "Added eob == 0 check to vp9_dequant_idct_add_32x32_c" into experimental

11 years agoFix block pointer corruption in intra8x8 prediction with 4x4 transform.
Ronald S. Bultje [Sat, 26 Jan 2013 00:08:19 +0000 (16:08 -0800)]
Fix block pointer corruption in intra8x8 prediction with 4x4 transform.

The RD loop would change the pointer after the first mode (DC) was tested,
leading to corrupt block objects being provided for the others. This
would essentially render the i8x8 predictor useless.

Change-Id: I16c5906ca64fb34878ac32ce59af8974e4582bb8

11 years agoRemove eob_max_offset markers.
Paul Wilkins [Tue, 29 Jan 2013 13:33:17 +0000 (13:33 +0000)]
Remove eob_max_offset markers.

Remove eob_max_offset markers and replace
with the generic skip_block flag to indicate
to the quantizer that all coeffs to be set to 0
and eob position set to 0;

Change-Id: Id477e8f8d4ec1a5562758904071013c24b76bfd7

11 years agoFurther improvement on compound inter-intra expt
Deb Mukherjee [Thu, 24 Jan 2013 23:22:57 +0000 (15:22 -0800)]
Further improvement on compound inter-intra expt

Adds a special combination mode specific to intra prediciton
mode D45.

Current results with the compound inter/intra experiment:
derf: 0.2%
yt: 0.55%
std-hd: 0.75%
hd: 0.74%

Change-Id: I8976bdf3b9b0b66ab8c5c628bbc62c14fc72ca86

11 years agoSegment Skip Flag
Paul Wilkins [Mon, 28 Jan 2013 15:22:53 +0000 (15:22 +0000)]
Segment Skip Flag

First step in simplifying the segment mode and
segment EOB flags into a simpler segment skip
flag that implies 0,0 mv and EOB at position 0.

Change-Id: Ib750cac31a7a02dc21082580498efd9f7d8d72a5

11 years agoMerge "Simplify Zero bin and zero bin run code." into experimental
Paul Wilkins [Mon, 28 Jan 2013 16:35:36 +0000 (08:35 -0800)]
Merge "Simplify Zero bin and zero bin run code." into experimental

11 years agoSimplify Zero bin and zero bin run code.
Paul Wilkins [Thu, 24 Jan 2013 18:28:33 +0000 (18:28 +0000)]
Simplify Zero bin and zero bin run code.

Simplification to eliminate a number of very large data
data structures. All zero run, zbin boosts for different
transform sizes are now limited to a maximum run length
of 15 before they max out the boost.

Some further work still needs be done to refactor, rationalize
and optimize the multiple quantizer functions.

The simplification coupled with tweaks to the 16 element array
now used for all transform sizes, has minimal effect on quality.

Change-Id: I6f3948b8ca0418b60d4db9030ff19026a34ed423

11 years agoTweak DropFramesWithoutRecovery params
John Koleszar [Sat, 26 Jan 2013 01:48:18 +0000 (17:48 -0800)]
Tweak DropFramesWithoutRecovery params

The previous parameters did not pass with VP8.

Change-Id: I1505171fbd713d21029541828ebdbe4c7dea79b8

11 years agoFix overread/write reported by valgrind if (mb_cols) & 3 != 0.
Ronald S. Bultje [Sat, 26 Jan 2013 00:03:15 +0000 (16:03 -0800)]
Fix overread/write reported by valgrind if (mb_cols) & 3 != 0.

We'd backup and restore all cols for a 64x64 SB, but the array wouldn't
be big enough to hold all that data.

Change-Id: Ic68ea721bf07e0b2f3937bd16b0b734bcc743ce1

11 years agoAdding a frame parallel decoding mode
Deb Mukherjee [Fri, 25 Jan 2013 19:30:28 +0000 (11:30 -0800)]
Adding a frame parallel decoding mode

Adds a flag to disable features that would inhibit frame parallel
decoding. This includes backward adaptation and MV sorting based
on search in ref frame buffer.

Also includes some minor clean-ups.

Change-Id: I434846717a47b7bcb244b37ea670c5cdf776f14d

11 years agoMerge "Remove "update_context" variable from VP9_COMP context." into experimental
Ronald S. Bultje [Fri, 25 Jan 2013 17:43:42 +0000 (09:43 -0800)]
Merge "Remove "update_context" variable from VP9_COMP context." into experimental

11 years agoAdded eob == 0 check to vp9_dequant_idct_add_32x32_c
Scott LaVarnway [Fri, 25 Jan 2013 01:09:56 +0000 (17:09 -0800)]
Added eob == 0 check to vp9_dequant_idct_add_32x32_c

Added a quick eob == 0 check.  Once the integer version of the dct32x32 is
complete, we can check for other eob cases.

For the 1080p clip used, the decoder performance improved by 4%.

Change-Id: I9390b6ed3c8be0c0c0a0c44c578d9a031d6e026e

11 years agoRemove "update_context" variable from VP9_COMP context.
Ronald S. Bultje [Fri, 25 Jan 2013 00:28:53 +0000 (16:28 -0800)]
Remove "update_context" variable from VP9_COMP context.

The variable is always zero.

Change-Id: Id5cdbecad543bca465a5b1d471badaec7e112c8d

11 years agoMvref speedup
Paul Wilkins [Thu, 24 Jan 2013 12:52:13 +0000 (12:52 +0000)]
Mvref speedup

Quality / decode speed trade off changes.
Simpler insert method without sort. Quality impact small.

Change-Id: Id0c0941bc508d985405abd06a13ffe7489170b62

11 years agoMerge "MSVS compiler error" into experimental
Paul Wilkins [Thu, 24 Jan 2013 17:21:49 +0000 (09:21 -0800)]
Merge "MSVS compiler error" into experimental

11 years agoMSVS compiler error
Paul Wilkins [Thu, 24 Jan 2013 11:42:38 +0000 (11:42 +0000)]
MSVS compiler error

Visual Studio reports and error on the line
  static const int bsize2 = bsize >> 1;

"error C2099: initializer is not a constant"

Also warnings on signed/unsigned mismatch

Change-Id: I76cb06f736bd9c90107a26bfb3c130ec3f6c33de

11 years agoMerge "Intrinsic version of loopfilter now matches C code" into experimental
Scott LaVarnway [Thu, 24 Jan 2013 16:45:22 +0000 (08:45 -0800)]
Merge "Intrinsic version of loopfilter now matches C code" into experimental

11 years agoAdds an error-resilient mode with test
Deb Mukherjee [Tue, 15 Jan 2013 14:43:35 +0000 (06:43 -0800)]
Adds an error-resilient mode with test

Adds an error-resilient mode where frames can be continued
to be decoded even when there are errors (due to network losses)
on a prior frame. Specifically, backward updates are turned off
and probabilities of various symbols are reset to defaults at
the beginning of each frame. Further, the last frame's mvs are
not used for the mv reference list, and the sorting of the
initial list based on search on previous frames is turned off
as well.

Also adds a test where an arbitrary set of frames are skipped
from decoding to simulate errors. The test verifies (1) that if
the error frames are droppable - i.e. frame buffer updates have
been turned off - there are no mismatch errors for the remaining
frames after the error frames; and (2) if the error-frames are non
droppable, there are not only no decoding errors but the mismatch
PSNR between the decoder's version of the post-error frames and the
encoder's version is at least 20 dB.

Change-Id: Ie6e2bcd436b1e8643270356d3a930e8989ff52a5

11 years agoMerge "Support multiple codecs in test infrastructure" into experimental
John Koleszar [Wed, 23 Jan 2013 22:52:22 +0000 (14:52 -0800)]
Merge "Support multiple codecs in test infrastructure" into experimental

11 years agoSupport multiple codecs in test infrastructure
John Koleszar [Fri, 18 Jan 2013 19:51:12 +0000 (11:51 -0800)]
Support multiple codecs in test infrastructure

This commit starts to convert the tests to a system where the codec
to be used is provided by a factory object. Currently no tests are
instantiated for VP9 since they all fail for various reasons, but it
was verified that they're called and the correct codec is
instantiated.

Change-Id: Ia7506df2ca3a7651218ba3ca560634f08c9fbdeb

11 years agoMerge "Modifies the comp inter-intra expt" into experimental
Deb Mukherjee [Wed, 23 Jan 2013 17:43:07 +0000 (09:43 -0800)]
Merge "Modifies the comp inter-intra expt" into experimental

11 years agoIntrinsic version of loopfilter now matches C code
Scott LaVarnway [Wed, 23 Jan 2013 17:31:40 +0000 (09:31 -0800)]
Intrinsic version of loopfilter now matches C code

Updated the instrinsic code to match Yaowu's latest loopfilter change.
(I584393906c4f5f948a581d6590959522572743bb)

The decoder performance improved by ~30% for the test clip used.

Change-Id: I026cfc75d5bcb7d8d58be6f0440ac9e126ef39d2

11 years agoMerge changes Ia82cef79,I7324a75a,I7b66daad,I73344451,I91dc210f,I5945b5ce into experi...
John Koleszar [Tue, 22 Jan 2013 16:31:55 +0000 (08:31 -0800)]
Merge changes Ia82cef79,I7324a75a,I7b66daad,I73344451,I91dc210f,I5945b5ce into experimental

* changes:
  Use alt-ref frame context for keyframes
  Preserve the previous golden frame on golden updates
  Generalize and increase frame coding contexts
  Start to anonymize reference frames
  Update encoder to use fb_idx_ref_cnt
  Remove buffer-to-buffer copy logic

11 years agoUse alt-ref frame context for keyframes
John Koleszar [Fri, 18 Jan 2013 22:35:20 +0000 (14:35 -0800)]
Use alt-ref frame context for keyframes

This matches the behavior prior to generalizing the frame context
selection, and intuitively makes sense in that the first forward ref
is immediately after the keyframe, so it's quality is improved a bit
by using the keyframe's entropy context rather than the default.

Change-Id: Ia82cef79382b9d8cfafdc44ba0533d4dc3e44053

11 years agoa minor change to a portion of loop filtering
Yaowu Xu [Thu, 17 Jan 2013 21:26:24 +0000 (13:26 -0800)]
a minor change to a portion of loop filtering

The loop filtering used for MB edge or internal edge of a MB using 8x8
tranform was reading 5 pixel each side and writting 3 pixel each side.
With suggestion from Aki and Scott on hardware&software performance,
this commit changed to read 4 pixel each side and write 3 pixel each
side.

Change-Id: I584393906c4f5f948a581d6590959522572743bb

11 years agoPreserve the previous golden frame on golden updates
John Koleszar [Wed, 16 Jan 2013 20:19:42 +0000 (12:19 -0800)]
Preserve the previous golden frame on golden updates

This commit restores the quality lost when the buffer-to-buffer copy
logic was removed. Note that this is specific to the current use of
golden frames and will need rework when RTC functionality is added.

Change-Id: I7324a75acd96eafd9e0f9b8633d782e390d5dc21

11 years agoGeneralize and increase frame coding contexts
John Koleszar [Tue, 15 Jan 2013 23:57:11 +0000 (15:57 -0800)]
Generalize and increase frame coding contexts

Previously there were two frame coding contexts tracked, one for normal
frames and one for alt-ref frames. Generalize this by signalling the
context to use in the bitstream, rather than tieing it to the alt ref
refresh bit. Also increase the number of contexts available to 4, which
may be useful for temporal scalability.

Change-Id: I7b66daaddd55c535c20cd16713541fab182b1662

11 years agoStart to anonymize reference frames
John Koleszar [Tue, 15 Jan 2013 21:49:44 +0000 (13:49 -0800)]
Start to anonymize reference frames

Remove lst_fb_idx, gld_fb_idx, alt_fb_idx, refresh_last_frame,
refresh_golden_frame, refresh_alt_ref_frame from common. Gold/Alt are
encode side conventions. From the decoder's perspective, we want to be
dealing with numbered references.

Updates to active_ref 2 signal mode context switches, vestigial from
refresh_alt_ref_frame. This needs some clean up to make sense with
increased numbers of reference frames, as well as reimplementing the
swapping of alt/golden which was previously done using the
buffer-to-buffer copy mechanism removed in an earlier commit.

Change-Id: I7334445158b7666f9295d2a2dd22aa03f4485f58

11 years agoUpdate encoder to use fb_idx_ref_cnt
John Koleszar [Mon, 14 Jan 2013 23:52:20 +0000 (15:52 -0800)]
Update encoder to use fb_idx_ref_cnt

Do reference counting the same way on the encoder as the decoder does,
rather than maintaining the 'flags' member of YV12_BUFFER_CONFIG.

Change-Id: I91dc210ffca081acaf9d5c09a06e7461b3c3139c

11 years agoRemove buffer-to-buffer copy logic
John Koleszar [Mon, 14 Jan 2013 22:44:39 +0000 (14:44 -0800)]
Remove buffer-to-buffer copy logic

This is the first in a series of commits to add additional reference
frames to the codec. Each frame will be able to update any of the
available references, but copying between references is not
supported.

Change-Id: I5945b5ce6cc3582c495102b4e7eed4f08c44d5a1

11 years agofix a number issues that cause failures
Yaowu Xu [Tue, 15 Jan 2013 02:03:34 +0000 (18:03 -0800)]
fix a number issues that cause failures

During master jenkins verification proces

Change-Id: I3722b8753eaf39f99b45979ce407a8ea0bea0b89

11 years agoModifies the comp inter-intra expt
Deb Mukherjee [Tue, 15 Jan 2013 01:32:26 +0000 (17:32 -0800)]
Modifies the comp inter-intra expt

Uses a single 1D table to implement the weighting of the predictors
for the compound inter-intra experiment.

Change-Id: I204ffbe4f9fc79d5d43b6c724ad253d800461012

11 years agoUse INT64_MAX instead of LLONG_MAX
John Koleszar [Mon, 14 Jan 2013 19:49:30 +0000 (11:49 -0800)]
Use INT64_MAX instead of LLONG_MAX

These variables have the type int64_t, not long long. long long could
be a larger type than 64 bits. Emulate INT64_MAX for older versions of
MSVC, and remove the unreferenced vpx_ports/vpxtypes.h

Change-Id: Ideaca71838fcd3849d816d5ab17aa347c97d03b0

11 years agoRemove compound intra-intra experiment.
Ronald S. Bultje [Mon, 14 Jan 2013 22:37:53 +0000 (14:37 -0800)]
Remove compound intra-intra experiment.

This experiment gives little gains and adds relatively much code
complexity (and it hinders other experiments), so let's get rid of
it.

Change-Id: Id25e79a137a1b8a01138aa27a1fa0ba4a2df274a

11 years agoMerge experiment "subpelrefmv"
Yaowu Xu [Mon, 14 Jan 2013 22:39:54 +0000 (14:39 -0800)]
Merge experiment "subpelrefmv"

Change-Id: Iac7f3d108863552b850c92c727e00c95571c9e96

11 years agoMerge experiment "widerlpf"
Yaowu Xu [Mon, 14 Jan 2013 20:11:06 +0000 (12:11 -0800)]
Merge experiment "widerlpf"

Change-Id: I0c94475075e66e13cfe4c20fab7db6474441ae86

11 years agoMerge "changed UV plane loop filtering for TX_8X8" into experimental
Yaowu Xu [Mon, 14 Jan 2013 22:47:31 +0000 (14:47 -0800)]
Merge "changed UV plane loop filtering for TX_8X8" into experimental

11 years agoMerge "change to evaluate reference mvs using above only" into experimental
Yaowu Xu [Mon, 14 Jan 2013 22:40:38 +0000 (14:40 -0800)]
Merge "change to evaluate reference mvs using above only" into experimental

11 years agochanged UV plane loop filtering for TX_8X8
Yaowu Xu [Mon, 14 Jan 2013 17:28:35 +0000 (09:28 -0800)]
changed UV plane loop filtering for TX_8X8

In commit 9a1d73d, loop filtering was added for UV 4x4 boundaries
when TX_8X8 is used by a MB. This commit further refined the decision
to be based on the actual transform used for the UV planes. When
UV planes use 4x4 transform, i.e. when prediction mode used is either
I8X8_PRED or SPLITMV, UV planes are filtered on 4x4 boundaries, and no
filtering is applied on 4x4 block boundaries when UV planes use 8X8
transform.

Change-Id: Ibb404face0a1d129b4b4abaf67c55d82e8df8bec

11 years agoMerge "Fix compiler warnings" into experimental
Paul Wilkins [Mon, 14 Jan 2013 22:20:57 +0000 (14:20 -0800)]
Merge "Fix compiler warnings" into experimental

11 years agoMerge "Merge prediction filter" into experimental
Adrian Grange [Mon, 14 Jan 2013 22:18:21 +0000 (14:18 -0800)]
Merge "Merge prediction filter" into experimental

11 years agochange to evaluate reference mvs using above only
Yaowu Xu [Thu, 10 Jan 2013 23:38:29 +0000 (15:38 -0800)]
change to evaluate reference mvs using above only

Change-Id: Ibcc342efac0a9be7a21d9b2c09984d9e16bbb225

11 years agoFix compiler warnings
Yaowu Xu [Mon, 14 Jan 2013 21:56:52 +0000 (13:56 -0800)]
Fix compiler warnings

The warnings caused verify failure with gerrit for several  commits

Change-Id: I030df8638bd69b8783a3ac58e720ff9f0bfd546c

11 years agoMerge prediction filter
Adrian Grange [Mon, 14 Jan 2013 20:57:07 +0000 (12:57 -0800)]
Merge prediction filter

Removed the experimental flag from around the prediction filter.

Change-Id: Ic1dd2db8fe8ac17ed5129f83094d4c5cdd5527d2

11 years agoReset x->skip for each iteration in the RD loop.
Ronald S. Bultje [Mon, 14 Jan 2013 20:43:12 +0000 (12:43 -0800)]
Reset x->skip for each iteration in the RD loop.

This prevents ill-defined behaviour, such as setting x->skip for a mode
that is excluded because of frame-level flags (e.g. filter selection,
compound prediction selection), then not breaking out of the RD loop
because the mode is not allowed, but keeping the flag on. Whatever mode
is iterated through next in the RD loop will then carry this flag, and
all sort of bad stuff happens, such as x->skip being set on intra pred
modes.

Change-Id: I5bec46b36e38292174acb1c564b3caf00a9b4b9a

11 years agoFix unused variable warnings
John Koleszar [Mon, 14 Jan 2013 20:10:11 +0000 (12:10 -0800)]
Fix unused variable warnings

Previous commit does not build cleanly on Jenkins with the DWT/DCT
hybrid experiment enabled (--enable-dwtdcthybrid).

Change-Id: Ia67e8f59d17ef2d5200ec6b90dfe6711ed6835a5

11 years agoFurther enhancements/fixes on dct/dwt hybrid txfm
Deb Mukherjee [Wed, 9 Jan 2013 14:26:54 +0000 (06:26 -0800)]
Further enhancements/fixes on dct/dwt hybrid txfm

Fixes some scaling issues. Adds an option to only compute the
dct on the low-low subband for 32x32 and 64x64 blocks using
only a single 16x16 dct after 1 and 2 wavelet decomposition
levels respectively. Also adds an option to use a 8x8 dct
as building block.

Currenlty with the 2/6 filter and with a single 16x16 dct on
the low low band, the reuslts compared to full 32x32 dct is
as follows:
derf: -0.15%
yt: -0.29%
std-hd: -0.18%
hd: -0.6%
These are my current recommended settings, since the 2/6 filter
is very simple.

Results with 8x8 dct are about 0.3% worse.

Change-Id: I00100cdc96e32deced591985785ef0d06f325e44

11 years agoMerge "WIP: Added sse2 version of vp9_mb_lpf_horizontal_edge_w" into experimental
Jim Bankoski [Sat, 12 Jan 2013 01:15:41 +0000 (17:15 -0800)]
Merge "WIP: Added sse2 version of vp9_mb_lpf_horizontal_edge_w" into experimental

11 years agoWIP: Added sse2 version of vp9_mb_lpf_horizontal_edge_w
Scott LaVarnway [Sat, 12 Jan 2013 01:11:04 +0000 (17:11 -0800)]
WIP: Added sse2 version of vp9_mb_lpf_horizontal_edge_w

and vp9_mb_lpf_vertical_edge_w_sse2.  This was quickly done so we can
run some tests over the weekend.  Future commits will optimize/refactor these
functions further.

The decoder performance improved by ~17% for the clip used.

Change-Id: I612687cd5a7670ee840a0cbc3c68dc2b84d4af76

11 years agoMerge "Upstream changes from Chromium Android Clang build." into experimental
Jim Bankoski [Sat, 12 Jan 2013 01:06:26 +0000 (17:06 -0800)]
Merge "Upstream changes from Chromium Android Clang build." into experimental