GitLab

mode_context[ref_frame] is calculated in find_mv_refs_idx(), but is
set to 0 in setup_ref_mv_list. Therefore, the calculation in
find_mv_refs_idx() is not needed.

Change-Id: I65ca06a2000278ad21c2eaa81eb12c48a7c1fcb8

This change is not normative and produces the same results as before.
TPL_MV_REF data structure is about 5x smaller.
Observed overall decoder run time reduction is about 4%.
No observed change in encoder run time.

Change-Id: Id68a492bac3bf28f48b7ceeedf85cd29981238ee

Change-Id: I591badb15e714e23f2c8a2c2df8514fdd75fc880

Avoid potential out of boundary buffer access issue.

Change-Id: Ic3766cd5bc2c1c8b4e1cbf38310fe6689aa38b59

When the reference frame size differs from the current coding
frame, skip the use of its motion information for motion field
estimation.

BUG=aomedia:1194

Change-Id: I1a32395367228f4ac86a21099b6b7b2c25e60270

To calculate wmmat[2] to wmmat[5], we don't need the absolute
coordinates which include the global offset components. Hence,
while finding samples, we only need to gather the relative coordinates
that are the neighbor block center point's coordinates relative to the
left-top pixel of current block. This patch won't change bitstream.

Change-Id: Ifbe7769effe6253b27cf97ea9c942a3a29eb7f50

The original sample selection process involves finding best 8 sorted
samples according to motion vector difference(MVD) between neighbor
block and current block, and then trimming samples. To reduce the
complexity, use the current block width/height as the MVD threshold,
and trim the samples without sorting.

This gives slightly less gain than the original method.
AWCY result:
         PSNR   PSNR HVS   SSIM
Average  -0.07   -0.13     -0.12
Borg test result:
             avg_psnr ovr_psnr ssim
cam_lowres:  -0.112   -0.112  -0.180
lowres:      -0.068   -0.073  -0.125

Change-Id: Ic2f79a170441d5bcb04ea87dddf490ef7fbba8bc

Change-Id: I69916bb6390dd9275341d8cd3fae2d8961e1cae3

Sub8x8 blocks will not check the extended region in motion field.
For regular block sizes, limit the extended region to be 3 points,
down from 9 points check.

Change-Id: I70f2631aa726ad01ee6bb83fffdf71ef82505888

When either component of a motion vector is above 4096, drop this
motion vector from the motion vector reference frame for later
motion field projection use. The coding performance change is close
to 0 for lowres and midres. This ensures that the motion vector
and reference frame tuple can be efficiently stored within 32 bits.

Change-Id: I9ae60a5caab2d3f49200abb5415532d82986839f

Part2 of ext-warped-motion experiment was removed.

Change-Id: I1b7aebad1fd7cb60d92c3ce959d5101228addae2

Reduce the reference motion vector search region over the spatial
neighbor blocks for sub8x8 block sizes, in order to reduce the
worst case context model parsing latency.

Change-Id: I77a2a25483836cc02cf1784c93566fa7cff40fc8

Removing the drl dependency on the candidate list length appears
to incur more than 0.3% compression performance loss. Hence remove
this option from opt-ref-mv to allow better latency vs compression
performance trade off.

Change-Id: I6edaeb2d437996082b7bdd6cda7351426c5584b9

This is a bitstream simplification.
It will reduce motion vector context model generating latency.

Change-Id: I98a496f5d72402ff51a478d5387a0653fa306dc1

For ext-skip, if a partition block uses the new skip mode,
NEAREST_NEARESTMV is the prediction mode. Other predictions were once
tried out but did not show noticeable coding gains.

Change-Id: I76c0aa0c68fe71f95cc4a09ef0de533e4ceb2f60

This experiment has been abandonned for AV1.

Change-Id: Ic706b3533ecf502ab316505ba0a9a8eeeb296612

All compound predicted frames now have the choice to either turn on
or off the use of the skip mode. The current encoder design only
turns on the use of skip mode when the temporal distances of the two
reference frames to the current frame are only different by 1.

This patch also includes a fix on the calculating of the temporal
distance between the two reference frames to the current frame for
skip mode.

Performance wise, this patch does not have noticeable impact as the
encoder stays with the same choice with the frame-level skip mode
flag.

Change-Id: I34f370940b3b25d2ab429b8721344133ae6288ad

For both low delay and high delay scenarios, the reference pair in
skip mode are specified as the closest fwd ref, together with the
closest bwd ref if there is any bwd ref, otherwise with the two
closest fwd refs.

Skip mode by default uses COMPOUND_AVERAGE. When all the reference
frames are on the same side, temporal-distance weighted compound is
considered, and a compound index is signaled to indicate whether
distance-weighted compound or compound-average is usd.

Whether to use distance-weighted compound for skip mode is still
under experimenting, hence a flag is temporarily added:
SKIP_MODE_WITH_JNT_COMP.

Following experimental results are obtained over 30 frames, using the
setup of --disable-ext-partition --disable-ext-partition-types
--disable-txmg --enable-jnt-comp --enable-mfmv --enable-ext-skip:

(1) High Latency:
For Google test sets (lowres/midres), the BDRate coding gain is ~0.2%;
For AWCY, the coding gain is ~0.1%.
(2) Low Latency:
No gain has been observed over Google sets and ~0.1% gain is obtained
only when temporal-distance weighted prediction is used.

Change-Id: I8c433357adebed0126ebfdd5c4d51aa71e64be57

Make the force reset condition under the opt-ref-mv flag.

Change-Id: I2c5f317e064d9517abee98937f3982c8e50dcfa5

The get_y_mode function, is superfluous, not used consistently, and requires a
useless block_idx parameter than gets pass around a lot inside the
codebase for no apparent reason.

The block parameter is misleading, as it could cause people to think all these
functions actually use this value.

Change-Id: I7ae0a8d1282c009b9114c83771cce10f5c2ee397

Support the motion field projection for both interpolation and
extrapolation situations.

Change-Id: I74e2d3c8956e62a2731f342f0d7a60dbc6b91ed1

Make the motion field projection process independent of the
reference frame type syntax.

Change-Id: Ie5d024d76ddb990fe623c83b3f4b4b6fadc72d2c

The motion vector storage location in the reference frame is set
to be dependent on the relative frame position, instead of the
reference frame index.

Change-Id: Id8f7d1e6ba32a5ed9e182194178fd2a70094a309

Change-Id: Idac0e6f4928e18c115c1055ca43f5b817ca51518

Also, make them const when appropriate.

Change-Id: I96d544e2cc9a0bce4d52fd33e44a4eaa40edda3c

This experiment has been abandonned for AV1.

Change-Id: Ief159fa7cf41dafa1e71b140f585cbc5a7d0a67d

Re-design the dynamic motion vector index coding system to remove
the context parsing dependency on the reconstructed motion vectors.

Change-Id: I01dd6eda239a0bed32d8dc98f0f10f18249a76d4

BUG=aomedia:1098

Change-Id: If7086930520d740d2610adeb699b68cdafab3ed3

This commit resolves the inter mode context model dependency on
the reconstructed motion vectors.

Change-Id: I3fd885dba6c10be8b1dcd072c1a5b3925ef4d1f5

Multiple above rows and left columns are checked while generating reference
MV candidate list for the current block. But, for warped ref_mv, we only
generate warped reference MV for current block if one neighbor block is
WARPED_CAUSAL mode and is located in the immediate above row or left column.

Change-Id: Ia9e9c2b7f97b61e0e4d2eeffd8d91e9e5f93d1a0

A new block mode, referred to as skip_mode is added. If a block is
coded as skip_mode, it will be inter-coded, with its references and
motion vectors derived from its neighboring blocks with zero-residue.
Otherwise, the block can be coded in the current intra or inter mode.

The computational load on skip_mode evaluation at the encoder should
be kept minimum. No transform size / type evaluations are needed.

Change-Id: I5aef0159c7d5ecd64258510835903375d6c536d6

If a block's motion_mode is WARPED_CAUSAL and its mode is NEARESTMV, search
its immediate above and left neighbors to get the set of neighbor blocks
using WARPED_CAUSAL motion mode, pick the one with largest block size, and
use that neighbor's warped motion parameters directly for the current block.
If none of the neighbors uses WARPED_CAUSAL motion mode, we estimate the
current block's warped motion parameters.

Before this patch, for every block, we estimate its warped motion parameters.
With this patch, we reduce the number of blocks doing parameter estimation.
Here are results by testing on clips with camera motions.
                    WARPED_CAUSAL blocks   blocks reusing parameters
station2_240p(30f):     3857                    1678
netflix_arieal(30f):     692                     223

No noticable changes in coding gain. Borg test result showed a PSNR
change of +0.006% on cam_lowres set, and -0.014% on lowres set.

Change-Id: If12387ad0ca8a1996ea4c3f1bedcb269ebf78c6c

This does not affect the coding performance or hw throughput. It
prepares for sw decoder speed up.

Change-Id: I5965ba2269e8d843f562b63ccaade7df3429f4f5

Change-Id: Icd0a8c4a9cf47cb93e8329ed0bdc0f79787baaaf

The numerator in the mv projection can be negative, e.g. cur_to_bwd
or cur_to_alt2, since either bwdref or altref2 can be a forward
predictive reference, whereas the denominator always stays positive.
The rounding inside mv projection hence should use signed operation.

Change-Id: I42a105835754a002dd31fcfa7c845e4c105ec54f

When generating ref mv list, only consider neighbor blocks that use intrabc.

Change-Id: I69f838a38478ce118f936dbf431048d3d42d4f34

The control of using reference frame motion vector is a separate
factor from the existence of previous frame motion vectors. This
commit decouples these two, such that the encoder can control the
use of reference motion vector. When it is used, one can further
identify if the previous frame exists or not, then to decide if
need to force use_prev_frame_mvs to be zero.

This solves the issue where the previous frame mvs is set to be
0 and it accidentally shuts off the access to all other existing
referece frames mvs in the mfmv system. It brings back the coding
performance gains to normal.

Change-Id: I2531f73e55582a9bb5b3e0ff47e361a199ec8082

This name better reflects the functionality of the mode.

Change-Id: Idc2372e7c7b226f70f63fa4f272a88cb4d857089

The experiment of mfmv has already extensively leveraged the use of
the motion vectors from the reference frames, hence there is no need
to further add the motion vectors from the previous frame to serve
as the motion vector reference candidates, let alone such motion
vectors have not taken into account of the use of frame offsets.

Experimental results collected so far have shown this patch has zero
coding performance impact.

Change-Id: I57a2f8b60797bd5e8f0a95206689b6e443b28ed1

Change-Id: Ie9152c5a631cd53ac69cc09469682505a88b045d