GitLab

This experiment has been adopted, we can simplify the code
by dropping the associated preprocessor conditionals.

Change-Id: Ic3438799335c6cd05f170302f49bd47e1f705c24

This experiment has been abandonned for AV1.

Change-Id: I924d172f7ce85b2422bace44684329fcbf71cb5e

The CL makes the context design for single reference frame coding the
same as that for the compound reference frame coding. There are 3
contexts designed for each of the binary symbols for the single
reference frame scenario, and the designed contexts simply rely on the
counts of the references used in the neighboring two blocks.

Once this CL is merged, the coding of the reference frames, regardless
of single prediction or compound prediction, will all follow the same
context design pattern for all the binary symbols. The design logic is
much simpler and the lines of code for each binary symbol context
identification are reduced by 80%.

Further, this CL has obtained a small coding gain for 30 frames with
the default coding tools:

lowres: avg_psnr -0.015%; ovr_psnr -0.021%; ssim -0.002%
midres: avg_psnr -0.108%; ovr_psnr -0.139%; ssim -0.135%

BUG=aomedia:1402
BUG=aomedia:973

Change-Id: Ia72a1d18e85ac3a05308675b60b95f80f2219c46

Reduce the blk_skip array size to the maximum txb numbers in a
coding block. This makes the speed 1 about 5% faster for smaller
resolution clips.

Change-Id: I1812dc4d403e08f0566ead8a14c6046e505bb849

The experiment is fully adopted.

Change-Id: I6cc80a2acf0c93c13b0e36e6f4a2378fe5ce33c3

This patch uses the neighboring ref counts to design the contexts
for the coding of the first reference frame of a reference pair for
the compound prediction. This aligns the context design with that
for the second reference frame of a reference pair for the
compound prediction.

The new designed contexts are much simpler than that in the baseline.
The number of contexts for each binary symbol is reduced from 5 to
3. Further, the logic for each context only depends on the collected
neighboring ref counts, which is straightforward to derive.

The default CDFs for the first reference frame coding have been
updated using aom_entropy_optimizer.

Experimental results demonstrate a small coding gain for Google test
sets of both lowres and midres, with 30 frames coded for the default
coding tool setup:

lowres: avg_psnr -0.077%; ovr_psnr -0.076%; ssim -0.106%
midres: avg_psnr -0.059%; ovr_psnr -0.066%; ssim -0.037%

BUG=aomedia:1356

Change-Id: I781abbe4616dc3f3a7213ec663946ff9844eb830

Change-Id: If88516ac3dcd72b528f4f7e27aab181a5137b285

Search for tx size using only DCT and 1D DCT first; then do an
additional round of tx type search with tx sizes already decided.

If enabled on speed 1, quality drop is 0.2% on lowres and midres;
average encoding speed up 20% (15~30% depending on sequences).

Currently enabled for speed >= 2.

Change-Id: I62526a501ab4db54c91d2bc506dadf9fc141bd03

For the coding of compound backward refs, this patch aligns the
allocated memory with the actual number of contexts designed.

Change-Id: Ib387d411b1f154fb3eb4656558078c54016cb9c1

The txk-sel allows each transform block to select its own
transform kernel. Such locality enables one to store the selected
RD cost including tx_type selection per transform block size.
It reduces the needed hash map size to 1 / 16 of what is needed
without txk-sel.

This commit re-works the hash map RD cost fetch for txk-sel. Tested
on red_kayak_480p in speed 1, enabling txk-sel makes the encoding
speed 12% faster than the baseline without txk-sel on. Further
enabling reduced hash map size gains speed 1 another 10%.

Change-Id: I4a5d99d27e2a76b10e76c00a8178f692c95fdf13

Reduce the length of inter_tx_size[] from 1024 to 16.

On a cif test sequence,
encoder memory consumption decreases by 18% (380MB -> 312MB);
decoder memory consumption decreases by 56% (21.4MB -> 9.4MB).

Change-Id: I42928eb9312748f96f4393c8d8040791f38f98b6

This reverts Change-Id: Ie11dd055255d200954b704b8c2ad8ca3dff7bf5c

BUG=aomedia:1305

Change-Id: I6894928dcadc99a79417034a7096a215693a46f2

Obtain the most likely partition range from a first pass square
block base partition search. Use the constrained partition search
region for full rate-distortion optimization search in the second
pass.

Tested on pedestrian 1080p at 2000 kbps, it makes the encoding
speed 40% faster for speed 0 and 30% faster for speed 1. The
average coding performance loss is around 0.15%.

Change-Id: Ifc83d48e6413d1b887e68cd1962084e018a2258f

Use simple rate-distortion search route for the first pass coding
block partition.

Change-Id: Iaaec3e1af83f46f625d3de8361eddd79a2bc6cef

Reduce the length of inter_tx_size[] from 1024 to 16.

On a cif test sequence,
encoder memory consumption decreases by 18% (380MB -> 312MB);
decoder memory consumption decreases by 56% (21.4MB -> 9.4MB).

Change-Id: Ie11dd055255d200954b704b8c2ad8ca3dff7bf5c

So they can be generated at prediction block, and then easily
accessed by transform block.

Change-Id: I376042e8d57e00586d3cf90e237544e705b77e8b

Change-Id: I83ca47c1469d8e383a815058c02c4826c6282873

Make the per transform block kernel selection process unified with
the rate distortion model used in preliminary mode search. This
makes the txk-sel model search space same as baseline.

Change-Id: I82a2d94e88a03c88154582575ced500197f8a409

Change-Id: I69e07309b708caa59c313ae2156828d8c089cf4d

The x0 parameter can now be chosen on the encoder side
independently of the dq offset.

Change-Id: I4a63bae6ae81d5a7f65c53efe1b1ecd713355b01

This is to exclude coding CFL for block sizes where it is normatively
not allowed.

Reuse the alternate CDFs from the --disable-cfl path.

This yields some improvement on subset1:
   PSNR | PSNR Cb | PSNR Cr | PSNR HVS |    SSIM | MS SSIM | CIEDE 2000
-0.0113 | -0.2630 | -0.2138 |   0.0021 | -0.0110 | -0.0157 |    -0.0511

BUG=aomedia:1182

Change-Id: I1e5f23ef7f2debcf8edce788bfa63ec95db4a679

This experiment has been adopted, we can simplify the code
by dropping the associated preprocessor conditionals.

Change-Id: I3e08eec385c40360e3934fa4f66f7c671e860517

At this point, av1_adapt_inter_frame_probs() can be removed.

Change-Id: Ic8f9cc4f6b874aeec84a57fb7dbfacdb33309b51

Change-Id: I7c09b63aadefef3350d713e03b822447c9472509

Change-Id: Ib4b83335ba9c61435f774136bc57638d8b88e963

-Use the CDF values to calculate the cost for ref frame signaling
-Update the cost per superblock
-Fix some errors in the default CDF

Test on midres speed-1 30 frames shows ~0.1% coding gain.

Change-Id: I6f3e5896ef1cae9b515fc8287bc7565d68d05a57

Change-Id: I657e43999f45ec173dd169b511eb96c6dae1f981

Change-Id: I7c7f193dea8e99be31d048b3fe2b198887a52822

1. code the range of eob in a transform block as a single symbol.
2. initialize cdfs directly.
3. initial cdfs are trained by using laplace prob estimates as in base
level coding (CL 38638)

Change-Id: I65c88d6dd0da3381feeec837928ac5bd81be4296

1. Train and init cdfs directly.
2. Use laplace prob estimates in aom_entropy_optimizer to avoid zero
probabilities.

Change-Id: I878fc0a306cbffe3eb51c5b86d5872459b6705c5

Part2 of ext-warped-motion experiment was removed.

Change-Id: I1b7aebad1fd7cb60d92c3ce959d5101228addae2

Use the number of pixels in a block as context, rather than the bsize
itself. The rectangular blocks therefore share the same context, e.g.
BLOCK_8X16 and BLOCK_16X8.

The number of contexts is reduced from 10 to 7.
Almost no coding performance changes.

Change-Id: Ib3241194580c2b93ad0e953957cdc9188393d055

Performance change is within noise range.

Change-Id: I9ca2ea6dfdc629513836637283ae3f964a1dd497

Clean up: clean up unused entropy related arrays
Bug-fixes: resolve compiling errors when it is enabled with
entropy_stats and ext_skip

Change-Id: I2893e1b5c09e847fb4677f6e62709464208b76f8

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

Account for the txb_optimize effect on level map context modeling
in the transform block rate distortion optimization search.
Properly update such information in the txb hash rd cost.

BUG=aomedia:1118

Change-Id: I16cf80c57fa37929f94291a7500d84f9e0778400

Fix the block context model used in the hash rd for lv-map.

Change-Id: Id590413bf8282b820ceca8b61132094df93be892

Instead of applying different offsets for each of quadri-split, horizontal split,
vertical split, store the start addresses of src and dst 8x8 blocks in each of
YUV palnes then pass it to distortion compute function for dist-8x8, dist_8x8_yuv().

Also, don't save decoded pixels for 8x8 block in additinal buffer but directly
access the xd->dst buffer. Thus removed the decoded-8x8[8x8] buffer in struct macroblock.

However, this patch does NOT fix the case where dist-8x8's assert fails with
CfL enabled.

Fix part of issue (i.e. 4:4:4 chroma iput) in
BUG=aomedia:1094

Change-Id: I6399d0eef3e17cacc9ec6467480a7959c85d419b

BUG=aomedia:1093

Change-Id: I0871c67d9a2ddc6aa9d3084f645305f6b74a999f

Remove COMPOUND_AVERAGE from compound_type_cdfs since it is now grouped
to compound_idx. However, COMPOUND_AVERAGE is still used elsewhere.

Change-Id: Ie0d460aabf9252e80eb4130cfef9aaf0efc3969d