Refactor per-transform copy & paste into a common function
update_coef_probs_common() and read_coef_probs_common(). The dry-run and
bit-writing loops in the encoder are still obvious candidates to be made
common, but they start to diverge a bit in the next commit, so are left
as-is for now.
Change-Id: I896bd3f4a073a6296eab7e92463fef79d8c6c08c
There is a macro DEFAULT_INTERP_FILTER defined in encoder/onyx_if.c that
is set as EIGHTTAP for now - so SWITCHABLE is not really used. Ideally,
this should be SWITCHABLE but that would make the encoder quite a bit slower.
We will change the default filter to SWITCHABLE once we find a faster way to
search for switchable filters.
Change-Id: Iee91832cdc07e6e14108d9b543130fdd12fc9874
The vp8_post_proc_down_and_across_mb_row_sse2() needs space for an
even number of macroblocks, as they are read two at a time for the
chroma planes. Round up the width during the allocation of
pp_limits_buffer to support this.
Change-Id: Ibfc10c7be290d961ab23ac3dde12a7bb96c12af0
* changes:
Fix another typo in 4x4-transform-for-i8x8-intra-pred coeff contexts.
8x8 transform support in splitmv.
Use SPLITMV_PARTITIONING instead of a plain integer type.
For splitmv, where partitioning is 8x16, 16x8 or 8x8, this patch
uses the 8x8 transform (instead of the 4x4) if txfm_mode is
ALLOW_8X8 or ALLOW_16X16. For TX_MODE_SELECT, splitmv can indicate
which of the 2 transform sizes (4x4 or 8x8) it wants to use.
Gains (with hybridtx4x4/8x8/16x16 and tx_select experiments
enabled) on derf: +0.9%, HD: +0.4%, STD/HD: +0.8% (SSIM or overall
PSNR, both metrics show similar improvements).
Change-Id: Ide954b000b415548ed92a7ac78e24f36e60fcf06
This can be used to distinguish between 16x8, 8x16, 8x8 and 4x4
partitioning modes when choosing splitmv as a MB mode.
Change-Id: Idc8b59772e1a80ccc4ad44d63c5c2ec3fc061a3c
It currently counts the probability that the branch is true, but it
should count the probability that the branch is false.
Change-Id: I963825da2e7a7ed3a613eb23ffd085e427dc36e5
Allows B_VL_PRED & B_LD_PRED modes to be used for all blocks
within a MB in B_PRED mode. These modes were temporarily
disabled with super-block coding.
Change-Id: I973b9bdb82c3da5f12d7cc963162a28805f25303
First sse2 version of vp8_mbloop_filter_vertical_edge(). For now,
intrinsics are being used until the bitstream is finalized. This function
will be revisited later for further performance improvements.
For the test clip used, a 34+% decoder performance improvement
was seen. This will vary depending on material.
Change-Id: I455b438bc8d8af76cf7533ac42eda5f689b21f7c
There were several different methods for calculating bitstream
probabilities in use. Consolodate these into a pair of functions,
get_prob() and get_binary_prob().
Change-Id: I5534f517f74027fee16d89c9baefaafea8156b2f
Pass the bool coder to be used explicitly. This avoids cases where two
different bool coders can be addressed from the same function. Also be
more consistent with bool coder variable naming, start to standardize
on 'bc'.
Change-Id: I1c95e2fdbe24ebe8c0f84924daa1728e3b054a31
Separates the logic on transform type selection previously spread out
over a number of files into a separate function. Currently the tx_type
field in b_mode_info is not used, but still left in there to eventually
use for signaling the transform type in the bitstream.
Also, now for tx_type = DCT_DCT, the regular integer DCT is used, as
opposed to the floating point DCT used in conjuction with hybrid
transform.
Results change somewhat due to the transform change, but are within
reasonable limits. The hd/std-hd sets are slightly up, while derf/yt
are slightly down.
Change-Id: I5776840c2239ca2da31ca6cfd7fd1148dc5f9e0f
First sse2 version of vp8_mbloop_filter_horizontal_edge(). For now,
intrinsics are being used until the bitstream is finalized. This function
will be revisited later for further performance improvements.
For the test clip used, a 31+% decoder performance improvement
was seen. This will vary depending on material.
Change-Id: I03ed3a7182478bdd1f094644ff3e0442625600e7
This commit moves a bit of data that ended up packed with the
modes/mv/residual partition during the change to interleaved encoding
into partition 0 where it belongs.
Change-Id: Ic711a378c58d9d6a17254384f492c213a15bad92
Packs the bitstream with each mb's residual following its mode/mv
information.
TODO: There are still a few fields that should be packed into partition
0 but are included in partition 1, due to them being serialized from
write_kfmodes/pack_inter_mode_mvs, which execute after the first
partition is finalized. These need to be separated out into a separate
function, similar to mb_mode_mv_init() in decodemv.c.
Change-Id: I43a46c363601ab36954d07ebe498760e1e2e3af4
Rather than decoding all modes/mvs separately, decode them per MB. This
forces the mode which was already used form the CONFIG_NEWBESTREFMV and
CONFIG_SUPERBLOCKS experiments, and is a precursor to changing to
interleaved encoding.
Change-Id: If19ee74ac8a987846d1cd0cf2b2e02a82f1a43ad
The commit changed to avoid using pixels from extended border in
in evaluating and select best reference motion vector.
Change-Id: I39b758889373e42ed2889d59744388e5b9c1a20a
It is essentially a duplicate of mode for RD-only purposes. Removing it
saves us 4 bytes per B_MODE_INFO, or ~0.5MB for a 1080p video encode.
Change-Id: I0a54db5f51658b3946d7efb1ca6e8cfbda0cdf88
The variable is essentially a duplicate of mode for RD-only purposes.
Removing it gives identical results, and saves 4 bytes per macroblock
(i.e. 32.5kB for a 1080p HD video encode).
Change-Id: I22d5058fdb80ab0b69862caee825e9d86bb148b3
This way a caller doesn't need to implement the logic for which (and how
many) tokens to write out to stuff one macroblock worth of EOBs. Make
the actual function implementations static, since they are now only used
in tokenize.c; also do some minor stylistic changes so it follows the
style guide a little more closely; use PLANE_TYPE where appropriate,
remove old (stale) frame_type function arguments; hardcode plane type
where only a single one is possible (2nd order DC or U/V EOB stuffing);
support stuffing 8x8/4x4 transform EOBs with no 2nd order DC.
Change-Id: Ia448e251d19a4e3182eddeb9edd034bd7dc16fa3
Change the macros PLANE_TYPE_{Y_NO_DC,Y2,UV,Y_WITH_DC} to a typed enum,
and use this typed enum consistently across all places where relevant.
In places where the type is implied (e.g. in functions that only handle
second order planes or chroma planes), remove it as a function argument
and instead hardcode the proper enum in the code directly.
Change-Id: I93652b4a36aa43163d49c732b0bf5c4442738c47
Also merge the three occurrences of 4x4 chroma block writing into a
single function, and call that function instead of duplicating the
4x4 chroma tokenization code in 3 places.
Change-Id: I7913538d1029f709b0e3ae49fff1148d3be9eeb9
Merge code blocks for different transform sizes; use MACROBLOCKD as a
temp variable where that leads to smaller overall source code; remove
duplicate code under #if CONFIG_HYBRIDTRANSFORM/#else blocks. Some style
changes to make it follow the style guide a little better.
Change-Id: I1870a06dae298243db46e14c6729c96c66196525
Also make some minor stylistic changes to bring the code closer to
the style guide. Remove distinction between inter and intra transform
functions, since both do exactly the same thing except for the check
against SPLITMV for the second-order transform. Remove some commented
out debug code. Remove 8x8/16x16 transform code in encode_inter16x16y(),
since the first-pass only uses 4x4 anyway.
Change-Id: Ife54816ff759825b9141f95dc2ba43c253c14dba
Also make some minor stylistic changes to bring the code closer to
the style guide. Remove checks against i8x8/bpred in the mb-codepath,
since these do individual block reconstruction and thus don't go through
this codepath.
Change-Id: I4dfcf8f78746f4647a206475acf731837aa4fd47
This includes trellis optimization, forward/inverse transform,
quantization, tokenization and stuffing functions.
Change-Id: Ibd34132e1bf0cd667671a57b3f25b3d361b9bf8a
Entropy coding takes care of this anyway, and this causes changes to
the txfm size assigned to skip blocks, which can affect the loopfilter
output, thus causing encoder/decoding mismatches.
Change-Id: I591a8d8a4758a507986b751a9f83e6d76e406998
The update_mb_segmentation_map flag was being signalled earlier than
other data dependent on that flag. Consolidate this data so it's
parsed within the same if-scope as the flag is originally parsed in.
Change-Id: I10e90b4f511856445ef75a85a44ff441e1e5e672
If the threshold(limits) <= 0, skipped filtering and copied the
frame directly. Also, fixed memory allocation checking.
Change-Id: If3d79d5b2bcb71b9777e6eb5cba1384585131e22
Use the common update_skip_probs() function rather than duplicating its
logic in write_kf_modes().
Change-Id: I2890a28f6907cb79ffe0fb21d20f0ef98b85cdd9
If a reference frame is forced because of low dissimilarity, then
shut off the search of intra modes. This change has mixed results. On
one clip (QVGA), it hurt quality by ~1.5% with negligible speed impact.
On another (VGA) it had negligible affect on quality, but a ~0.2% speed
impact.
Change-Id: Ic8b07648979d732f489de5f094957e140f84d2eb
Rather than overloading the parent_ref_frame value to shut off the
search in some cases, add a new validity flag. This cleans up some
of the duplicated mr_encoder_id && mr_low_res_mv_avail checks as
well, for readability.
Change-Id: Iddad93a27066c3d85ff2f25a361ac113b288ab7b
Results: derf (vanilla or +hybridtx) +0.2% and (+hybrid16x16
or +tx16x16) +0.7%-0.8%; HD (vanilla or +hybridtx) +0.1-0.2%
and (+hybrid16x16 or +tx16x16) +1.4%, STD/HD (vanilla or +hybridtx)
about even, and (+hybrid16x16 or +tx16x16) +0.8-1.0%.
Change-Id: I03899e2f7a64e725a863f32e55366035ba77aa62
1. Algorithm modification:
Instead of having same filter threshold for a whole frame, now we
allow the thresholds to be adjusted for each macroblock. In current
implementation, to avoid excessive blur on background as reported
in issue480(http://code.google.com/p/webm/issues/detail?id=480), we
reduce the thresholds for skipped macroblocks.
2. SSE2 optimization:
As started in issue479(http://code.google.com/p/webm/issues/detail?id=479),
the filter calculation was adjusted for better performance. The c
code was also modified accordingly. This made the deblock filter
2x faster, and the decoder was 1.2x faster overall.
Next, the demacroblock filter will be modified similarly.
Change-Id: I05e54c3f580ccd427487d085096b3174f2ab7e86
In some situations, believed to be an interaction between temporal
scalability and dropped frames, the references available to an
encoder may not be the same references available to its parent.
Previously, the code tried to force the reference frame chosen by
the parent to be used on this frame, even if it was disabled. This
was preventing the pick mode loop from running even once, which led
to a crash.
Attempts to reproduce this bug locally were unsuccessful, so it is
still undetermined what the underlying cause of this issue is. In
the specific case that was failing, the application did not set
any flags which influenced the reference selection on that frame.
ref_frame_flags indicated that the golden frame was disabled,
believed to be because the last frame updated the last and golden
frames, so golden was shut off by default. It's not clear why this
wouldn't have also been true in the lower res encoder, ie, why the
lower res encoder decided to use and/or was allowed to use the
golden frame. We weren't able to debug into the non-crashing
lower res encoder as the crash couldn't be reproduced locally.
Change-Id: Ifb265253d26963ac2afde0e20cf6792788be6af7
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.
Change-Id: I7c9a72bd79315246ed880578dec51e2b7c178442
If a parent mb is available but is intra coded, then parent_ref_mv is
invalid. Check that the parent is inter coded before trying to access
the parent_ref_mv. Previously the parent_ref_mv was being read from
an uninitialized stack allocation, causing potential OOB reads and
other undefined behavior.
Change-Id: I0c93cd412a19c3a184bcf6decaa145b3a036a6c0
Protect the call to {Initialize,Delete}CriticalSection() with an
Interlocked{Inc,Dec}rement() pair, rather than the previous static
initialization. This should play better with AppVerifier, and fix issue
http://code.google.com/p/webm/issues/detail?id=467
Change-Id: I06eadbfac1b3b4414adb8eac862ef9bd14bbe4ad
The codec as it stood placed a keyframe one frame after a
real cut scene - and ignored datarate and other considerations.
TODO: Its possible that we should detect a keyframe and recode
the frame ( in certain circumstances) to improve quality.
Change-Id: Ia1fd6d90103f4da4d21ca5ab62897d22e0b888a8
Reset the cyclie refresh mode index in alloc_compressor_data().
This is needed to handle both cases of internal and
external spatial resizing.
Change-Id: I2697e12d45135eae2e8f0d45161811f24722312a
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.
Change-Id: Ic7c0c62dacc68ef551054fdb575be8b8507d32a8
On an internal spatial resize, this mode index was not reset to 0,
and therefore could exceed dimensions of seg_map or cyclic_refresh_map.
Change-Id: I6fe85dbd2765eb0207a9d9f71fda8d8b8c34f075
Fixes some build issues for people building for win32 who have a
pthreads emulation layer installed.
Change-Id: I0e0003fa01f65020f6ced35d961dcb1130db37a8
This should avoid problems with blocks gettings high quality
improvement despite having recently moved:
Change-Id: Ic0af0de2d6577807fa3c553f47b55d547ef36359
Set the seg map to 0 for key frame.
In previous commit on cyclic refresh, the seg map for key frame
was not reset, and instead used the seg map from last frame.
Change-Id: I848eb2face420dfcd2f7daca6f070b9127ca938b
-Increase the amount of mbs to be refreshed.
-Replace the delta qp with a fixed and reduced delta.
-Change to the mb update loop to try to always update same amount of mbs.
Change-Id: I93ac88002fd8dc677d2337f77998ff93f64e4ff9
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.
Change-Id: I9afb3880906fd0a1368a374041fc08efcf060c54
This change is necessary for the frame-based multithreading
implementation.
Since the postproc occurs in this call, vpxdec was modified to time around
vpx_codec_get_frame()
Change-Id: I389acf78b6003cd35e41becc16c893f7d3028523
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.
Change-Id: Ifad3b528d0b4b6e3c22518af789d76eff23c1520
Initialize the top line at the beginning of each frame and
the left column at the beginning of each row.
Change-Id: I5412f7ea49ffc490215cf65a62715a6c5e3a5a29
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:
derf: +0.279%
std-hd: +0.617%
hd: +1.299%
yt: +0.822%
With threshold 8 integer pels are:
derf: +0.295%
std-hd: +0.623%
hd: +1.365%
yt: +0.847%
Patch: rebased
Patch: rebase fixes
Change-Id: I4ed14600df3c457944e6541ed407cb6e91fe428b
Multiple decoders were getting allocated per frame.
If the decoder crashed we exitted with out freeing
them and the next time in we'd allocate over.
This fix removes the allocation and just has 8
boolcoders in the pbi structure
Change-Id: I638b5bda23b622b43b7992aec21dd7cf6f6278da
Some cleanups that will make it easier to maintain the code
and incorporate upcoming changes on entropy coding for the
hybrid transforms.
Change-Id: I44bdba368f7b8bf203161d7a6d3b1fc2c9e21a8f
