GitLab

Change-Id: Ie6b41ac068846da3c87bdcaa94180777edb27d1f

This was missing before. Hopefully it will improve costing.

Change-Id: Id7dbe963fe5e26798b853419804bdd128537a5d9

Cleanup context derivation for base levels by fixing the template size
to 5. No change to bitstream.

Change-Id: I496c74b386b8d8a4fb43c0b69add52a0a798a981

This patch completes the work to add fixed-depth TX domain support to
the quantization and dequantization code.  At present, it is active but
configured to behave identically to current AV1 master as RDO and TX
have not yet been updated to also support this functionality.

subset-1:
monty-rest-of-stack-noshift-s1@2017-11-13T14:37:42.541Z ->
 monty-rest-of-stack-quant-s1@2017-11-13T14:38:43.774Z

  PSNR | PSNR Cb | PSNR Cr | PSNR HVS |   SSIM | MS SSIM | CIEDE 2000
0.0000 |  0.0000 |  0.0000 |   0.0000 | 0.0000 |  0.0000 |     0.0000

objective-1-fast --limit=4:
monty-rest-of-stack-noshift-o1f4@2017-11-13T14:37:16.992Z ->
 monty-rest-of-stack-quant-o1f4@2017-11-13T14:38:28.828Z

  PSNR | PSNR Cb | PSNR Cr | PSNR HVS |   SSIM | MS SSIM | CIEDE 2000
0.0000 |  0.0000 |  0.0000 |   0.0000 | 0.0000 |  0.0000 |     0.0000

Change-Id: I3773a1fc128136c9fea227f4b547576a8aa6efa3

Change-Id: Icca39f1d037a3aca4540e35b70fdfafeae2b094e

Change-Id: I1ce575377c3d971f4488177b82c6e0b32aa5e4be

Change-Id: I3d152b28bb31027156719c1ecd7544df0f9766bb

In TX size search the rate and distortion values are often computed for
residuals that have already been seen before, with exactly the same
entropy context. Such repeated computations occur when the residue
signal is the same for 2 or more partitions for a given prediction mode.
By saving previous RD search results we can achieve some encoder
speed-up at the cost of higher RAM consumption (~20MB).

Test results with all default experiments:
Speedup      15% (measured on 4 lowres 30-frame sequences)
compression   0%

Change-Id: I8d9b8dd483cf862f0864b50ae80466beac978290

Change-Id: I36334c25f904e2a40f29c8ab83da4501f52e2eb3

The probability training tool generates all CDF_PROB_TOP table if
the counts are all zero. Although it means such cases are never
used in actual encoding, in rd cost pre-calculation, we still
sometimes run into such cases and the token cost calcualtion
function cannot handle all-CDF_PROB_TOP cdfs (Will trigger assertion
when debug mode is on).

Change-Id: I259620e210fea0f6e1f7848724fa006fa06da226

Account for the bit-depth in calculating the distortion for lv-map
coefficient level optimization.

Change-Id: I47c60f15bfda5aeba8e32f1ac898015650a057db

Change-Id: I21213ad3a8b9bd018927582ffa7eeab372aa7f9c

Reconstruct dqcoeff inline with transform coefficient level
decoding process. This saves one pass through 0 to eob scan. The
txb coeff decoding speed is up by 5%.

Change-Id: Icd88a9dca4595a180d514ce4b067ecf9dd90a609

This solves an enc/dec mismatch issue due to the fact that
SIMD implementation compares signed int8_t. Hence we need to
limit the base level reference to be under 127, to avoid cmpt
instruction taking it as a negative number.

BUG=aomedia:1045

Change-Id: I2ed616fda52415fbf50451660b9564df1adf77af

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

Replace the multiplication with shifts.

Change-Id: I245efaddea2019d789179569e82e81bb7cb97715

Update the last coefficient position whose level is above
NUM_BASE_LEVELS in the first pass decoding. This increases the
lv-map txb decode speed by 10%.

Change-Id: Ib35409a1d374c1e171ee45ac805a3a318793935e

If CONFIG_FRAME_SUPERRES is disabled, and we don't explicitly signal
a render size for a frame, then the render size would be set to the
crop size of the *previous* frame (or 0 for the first frame in a video).

Return to using the VP9 behaviour in this case (albeit coded slightly
differently to help support superres) - ie, use the crop size of the
current frame if an explicit render size is not provided.

We do this by moving the calls to resize_context_buffers() (which sets
up cm->width and cm->height) before the calls to setup_render_size
(which might use those values). Because resize_context_buffers() does not
read any data, the bitstream format should not be affected, aside from
the intended bugfix.

BUG=aomedia:1044

Change-Id: I4fc71b9aea36cc06e10408715c205d9674f87fc2

Change-Id: Idc72cff4ab9ba1691a7f5c72366a61be9569d370

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

These used to be a combination of some config flags. But as those config
flags are now removed, they were always 1 now.

This simplifies the code a bit.

Change-Id: Ifd3a94b6b786c95c3efc6d646dcf1489cdda7f92

These tests were testing the accuracy of fwd + inverse transform. But,
given that we have an approximation that zeroes out 3/4th of the
coefficients, we cannot put a reasonable upper bound on the error, and
the test itself doesn't make sense.

Change-Id: I0f44e7bcbe56c6be2cabbe3179af844c41836662

BUG=aomedia:1042,aomedia:1043

Change-Id: Ia7e82cbd280132bbeef41c90c38cbb585db8828a

Change-Id: Ifa2cdc2d2230dfa11396ee3e547653180f96b795

Change-Id: I7123717b2d11bca826d650c6e6b6ae137476d541

Change-Id: I9e55a9c9dd546e2e1d5e9c43e3e73fc44c3ba590

Change-Id: I482f95ee0e466d14a893152fdd03d0cfe368a7e1

Speed up the lv-map txb decoding process by 5%.

Change-Id: I5c87666b353a27f7fc67cc55322a30bb417435a3

The stripes are intended to extend down to the full decoded
height of the frame, which is always a multiple of 8 luma pixels,
in order to avoid some nasty edge cases.

This change was partially implemented in previous patches, but
not everywhere was modified, leading to slightly inconsistent code.
This patch finishes making the relevant changes, along with a
slight bit of refactoring.

Change-Id: Ibc8e2f5ace5415815625edbc224557a7c548c38a

Corrected the nz_rate which is used to estimate the cost of changing
the current coeff to EOB (as part of optimize_txb).
(patch suggested by Dake He)

Change-Id: I60d38d86744826e5c3bbc9eaea9ba6c520145f0b

The EOB coefficient cannot be 0 and for that reason it has special base_cdf contexts.
Before this commit there was two contexts (DC and AC). This commit adds two additional
contexts to separate the AC into 3 bands (i<=N/8, i<=N/4, i<=N/2).

Change-Id: If088b20fd891920b7ea7fc988d29bf6d86d93bfc

Do not round the lv_map probabilities but instead
reduce the precision of the entropy coder to 9 bits
for all CDFs, and modify operation so that the EC
interval is never 0.

Change-Id: I1b4e51019613106ae3b68519b20637bc0d7b51be

Allow the codec to time the average transform block coefficient
processing for sw speed check.

Change-Id: Ibdaf15ab5b7f1ea8264604cc00ef45e3ae3114c7

Change-Id: I6ce7aea19e3bdeef24d3fe66ac6eba7b8d585f9a

daala_fwd_txfm and daala_inv_txfm export no symbols when
CONFIG_DAALA_TX is disabled. Don't build them.

Change-Id: I0b991248e4151de51c41fc1006ba3065d1b8d36b

In e72a2091 one lps default probability was accidentally changed from 1 to 128 for
non LV_MAP_MULTI mode. This commit reverts that change and make the change only for
LV_MAP_MULTI mode. Also rather than changing to 128, the probability is changed to 10.

Change-Id: Ia8950379c46c59d40ea388fcd0621bbd78c26ede

This experiment has been abandonned for AV1.

Change-Id: I26d10463bdb592fa1fda022a0da19dd6c37159ca

Change-Id: Ib82549c4fae8bca285062084ed7c9c43b84e1660

In read_intrabc_info() and assign_dv().

BUG=aomedia:1037

Change-Id: Ic430147a9a15024d942bde361be0c4a603f812e4

The prediction can be done in 2x2 or 4x4 processing unit, within
which there is no dependency and the computation can be fully
parallelized.
Also turn < 8x8 filter_intra on, and disable it in > 32x32 txbs.

Change-Id: I4f8a3104019cbb35e88f342d97516f81b19152b0