GitLab

Change-Id: I01cecc829e2d57517427a1de6387e91ba3c64312

The SSE4.1 and AVX2 implementations of the self-guided filter have been updated
to match the updated FAST_SGR C implementation in restoration.c.

The self-guided filter speed tests have been altered to compare the speeds of
the SIMD and C implementations of the relevant functions.

Speed Tests (code compiled with CLANG)
===========

For LowBD:
- The SSE4.1 implementation is ~220% faster (~69% less time) than the C code
- The AVX2 implementation is ~314% faster (~76% less time) than the C code

For HighBD:
- The SSE4.1 implementation is ~240% faster (~71% less time) than the C code
- The AVX2 implementation is ~343% faster (~77% less time) than the C code

Change-Id: Ic2734bb89ccd3f66667c68647e5f677a5a496233

Change-Id: I8dcaa6882d47a097498c8f8af515b1185df4fdf3

In preparation for supporting q_adapt_probs.

Change-Id: I4a39b81b0d2c4ceb1586ae411a1216c6c20d896d

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

Change-Id: I91f18c498c694829b933bb73812ad94d66962994

Added an AVX2 version of the Wiener filter, along with associated tests. Speed
tests have been added for all implementations of the Wiener filter.

Speed Test results
==================

GCC
---

Low bit-depth filter:
- SSE2 vs C: SSE2 takes ~92% less time
- AVX2 vs C: AVX2 takes ~96% less time
- SSE2 vs AVX2: AVX2 takes ~43% less time (~74% faster)

High bit-depth filter:
- SSSE3 vs C: SSSE3 takes ~92% less time
- AVX2  vs C: AVX2  takes ~96% less time
- SSSE3 vs AVX2: AVX2 takes ~46% less time (~84% faster)

CLANG
-----

Low bit-depth filter:
- SSE2 vs C: SSE2 takes ~84% less time
- AVX2 vs C: AVX2 takes ~88% less time
- SSE2 vs AVX2: AVX2 takes ~27% less time (~36% faster)

High bit-depth filter:
- SSSE3 vs C: SSSE3 takes ~85% less time
- AVX2  vs C: AVX2  takes ~89% less time
- SSS3  vs AVX2: AVX2 takes ~24% less time (~31% faster)

Change-Id: Ide22d7c09c0be61483e9682caf17a39438e4a208

Changes the CONFIG_FAST_SGR=1 strategy to not use any
subsampling for the r=1 filter, but for the r=2 filter
sub-sample vertically but combine only by filtering
horizontally in the last stage for odd rows.

Coding efficiency loss sems quite minimal.

Change-Id: I5644ac400b387c37a2d278db7f6ad3ac0a6b5e93

Change-Id: I6138519456b2ad3ffc8bced803ddc4418b246e74

Some parameter tuning included.

lowres (q, 30 frames, speed 1):
-1.243% av PSNR, -2.337% ov PSNR, +0.577% SSIM

lowres (vbr, 30 frames, speed 1):
-0.327% av PSNR, -1.007% ov PSNR, +0.182% SSIM

A few videos become a lot worse in SSIM, which needs to be
investigated. But PSNR-wise the patch seems pretty good.

Change-Id: I17c8d812c96ee49ddae7d3959a459aa3ffcea208

Since aom_comp_mask_upsampled_pred just call aom_upsampled_pred
and aom_comp_mask_pred, no need to separate c version from simd
version any more.

Change-Id: I1ff8bcae87d501c68a80708fd2dc6b74c6952f88

BUG=aomedia:1306

Change-Id: I5a8bdbd472213ded2de706c5b044a1bf24823670

The current aq mode encoder setting would alter the segment_id
between the rate-distortion optimization and the block encoding
stages. Disable the corresponding consistency check in this case.

BUG=aomedia:1251

Change-Id: Ic910a23fd64a9b4554567d3c8c9a9ae5f6062c7b

until av1_idct*_new are optimized this function sits high on the decode
perf

Change-Id: Ic55c9a92b9926fc09eaee211a45fde00333b7c15

Change-Id: I1d7f33546053615a334b67b75147bd5e027a545b

Change-Id: Ia34909cc6edc20f17a777e0b7bff97a62e0ac0c2

This reverts Change-Id: Ie11dd055255d200954b704b8c2ad8ca3dff7bf5c

BUG=aomedia:1305

Change-Id: I6894928dcadc99a79417034a7096a215693a46f2

Change-Id: I1e6a8a74d0ca1e6aa01d2da12bd9b19c8307154e

Change-Id: Ia448b44ca734fe111422de9afdad97ac48e78b66

When cb_partition_scan is true, only DCT_DCT is considered. Therefore
there's no need to prune transform types; and if DCT_DCT is pruned, we
end up with no transform type to use.

Change-Id: I1d65fe94e72de66fde18e271a598f9e67ade9cfb

Change-Id: Ibe4f7bb61837b6bae6717f0c683fa23f78de5b80

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

Add square block partition to serve as the first pass partition
search.

Change-Id: Ib637bba205d2cd0f6b0a5e2e91b270e22dce5580

BUG=aomedia:1274

Change-Id: Ib1d814db4ef1bcb075444e4da855fd840e945a7d

Same as https://aomedia-review.googlesource.com/c/aom/+/42901
Adopt the same code refactoring to aom_comp_mask_pred_c
(Should be bitwise identical).

Change-Id: Ieea71d370f5df48d216f40515842ad62499432c8

SSE2 version already extended to support
13 TAPs

Change-Id: I58e04527b297256b6ca63b12097d9345196a12bd

This reverts commit fbeee067.

The assembly which was failing has been fixed.

BUG=aomedia:102

Change-Id: Ide75630b38603a2553f6e231085994251c77b26c

Change-Id: If0b1d2fe31569104f2d8eef3cfd42cab30162c7e

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

- Add explicit cast of bool to int to silence a test warning.
- Add explicit cast of size_t to int for same in dump_obu.

Change-Id: I90846eb5c88880d921f20cb66b116ab7d2799af5

BUG=aomedia:717

Change-Id: Ib06a12039cb72665c1ee534cc2246ac3d23f878d

cmake: -DCONFIG_SCALABILITY=1

Change-Id: Ifa908f809bcf904bdf0ed87b351e1ef3accc2b3f

Clear linker error when building with gcc 6:
relocation R_X86_64_32 against `.rodata' can not be used when making a
shared object; recompile with -fPIC

BUG=aomedia:102

Change-Id: I6c06de1e9dac1c044a4b07125abcaba0943a29b6

3~5% encoding speedup for speed 0; no quality loss.

Change-Id: I0e31755f45253e5e99d8d9eed0d7a6fe6050f49f

(1) Explicitly reset RD stats for each partition.

Earlier,
PARTITION_SPLIT was the only one resetting the RD_STATS in 'sum_rdc'.

But this was working because:
- PARTITION_SPLIT was tried before VERT, HORZ, VERT_4 and HORZ_4; and
- RD cost calculations in VERT, HORZ, VERT_4 and HORZ_4 partitions
implicitly discarded existing value in sum_rdc

However, that was very fragile; explicitly resetting the stats every
time is much safer.

(2) Using a separate variable 'temp_best_rd_cost' was fragile as someone
may forget to update the same. So, we use best_rdc.rdcost directly.

BUG=aomedia:1246

Change-Id: Icd75f25c34bb0f1806e691784648bcffce2417e6

AVX2 implementation of av1_convolve_x_sr, av1_convolve_y_sr and
av1_convolve_2d_sr have been added.

Improvements have been made to av1_convolve_x_avx2, av1_convolve_y_avx2
and av1_convolve_2d_avx2.

Change-Id: I62a699dd9dcf42de94dd72cc2d43affc0dc31404

BUG=aomedia:1296

Change-Id: If9f944b58f23cdb71f919bd391f6b37e27b271f1

Serialize the adst4 operations
Update stage range accordingly
Change the cos_bit precision accordingly.
Correct 4x8/8x4 inv_start_range

BUG=aomedia:1271

Change-Id: I10bc91585a61d790decdc24cb91659102e043620

As per the linked bug report, the distance-weighted compound
prediction has two separate round operations, first by 3
bits (inside the various convolve functions), then by 10 bits
(after the convolution functions).

We can improve on this by right shifting by 3 bits inside the
convolve functions - this is equivalent to doing a single round
by 13 bits at the end.

Note: In the encoder, when doing joint_motion_search(), we do
things a bit differently: So that we can try modifying the two
"sides" of the prediction independently, we predict each side as
if it were a single prediction (including rounding), then blend
these single predictions together.

This is already an approximation to the "real" prediction, even
in the non-jnt-comp case. So we leave that code path as-is.

BUG=aomedia:1289

Change-Id: I9ad1fbcb3e12db2b5fc3c82b407f0fd9e6b39750