GitLab

Includes unit tests for conformance and speed.

SSSE3/CFLPredictHBDTest:
4x4: C time = 1436 us, SIMD time = 358 us (~4x)
8x8: C time = 4821 us, SIMD time = 598 us (~8.1x)
16x16: C time = 18528 us, SIMD time = 1793 us (~10x)
32x32: C time = 72998 us, SIMD time = 6400 us (~11x)

AVX2/CFLPredictHBDTest:
4x4: C time = 1436 us, SIMD time = 398 us (~3.6x)
8x8: C time = 4924 us, SIMD time = 644 us (~7.6x)
16x16: C time = 18624 us, SIMD time = 1617 us (~12x)
32x32: C time = 73509 us, SIMD time = 3635 us (~20x)

Change-Id: Icbcfefbf165facdbd77c9b3861af2bbf464254a0

Includes unit tests for conformance and speed.

SSSE3/CFLPredictTest:
4x4: C time = 2063 us, SIMD time = 313 us (~6.6x)
8x8: C time = 6656 us, SIMD time = 493 us (~14x)
16x16: C time = 24970 us, SIMD time = 1327 us (~19x)
32x32: C time = 59020 us, SIMD time = 5178 us (~11x)

AVX2/CFLPredictTest:
4x4: C time = 2052 us, SIMD time = 333 us (~6.2x)
8x8: C time = 6712 us, SIMD time = 513 us (~13x)
16x16: C time = 25292 us, SIMD time = 1023 us (~25x)
32x32: C time = 58994 us, SIMD time = 2828 us (~21x)

Change-Id: I08690a548be981ff10e184de468b9e0e691ee812

Includes unit tests for conformance and speed.

SSSE2/SubsampleSpeedTest:
4x4: C time = 868 us, SIMD time = 200 us (~4.3x)
8x8: C time = 3054 us, SIMD time = 293 us (~10x)
16x16: C time = 11887 us, SIMD time = 760 us (~16x)

AVX2/SubsampleSpeedTest:
4x4: C time = 784 us, SIMD time = 205 us (~3.8x)
8x8: C time = 2774 us, SIMD time = 307 us (~9x)
16x16: C time = 10978 us, SIMD time = 489 us (~22x)

Change-Id: I7d5958097542599d57d1a9f9a0a1b809c6a345b0

By default, the DC_PRED is not cached (this includes
decoding). During cfl_rd_pick_alpha(), DC_PRED caching
is enabled, the DC_PRED is cached after the first time it
is computed (for each plane) and then it is reused when
testing all the other scaling parameters.

Change-Id: Ie8ba0bb0427c4d9be8de5b44e6330e8a78b9c7d9

Removes unused BLOCK_2X2, BLOCK_2X4 and BLOCK_4X2 from the
BLOCK_SIZE enum.

Change-Id: I964d99718026c51a1eaf30d4a1fc83cc52f94083

Based on the HW Subgroup call of December 4th 2017, we limit luma partition to
32X32.

Regression on Subset 1
  PSNR | PSNR Cb | PSNR Cr | PSNR HVS |    SSIM | MS SSIM | CIEDE 2000
0.0881 |  1.3504 |  1.2936 |   0.0572 |  0.0182 |  0.0227 |     0.5204

https://two.arewecompressedyet.com/?job=CfL-PartU%402017-12-12T15%3A39%3A36.794Z&job=CfL-Max32x32%402017-12-12T16%3A10%3A09.989Z

Change-Id: I7e3cfd68097c0bc24b1426348b5fd574c4f638a0

To avoid a cascade of encodes when performing CfL RDO,
we compute DC_PRED on the partition unit. To do so, we
change the tx_size of CfL to match the size of the
partition unit (i.e. CfL partitions only contain 1
transform block).

This change requires disabling CfL when a chroma
partition-unit-sized DC_PRED is unavailable
(i.e. 4:1, 1:4 partitions and chroma partitions > 32X32).

Results on Subset1 (compared to disabling 4:1 and 1:4 partU):
   PSNR | PSNR Cb | PSNR Cr | PSNR HVS |    SSIM | MS SSIM | CIEDE 2000
-0.1243 | -1.9286 | -2.0140 |  -0.1514 | -0.1512 | -0.1947 |    -0.8066

https://two.arewecompressedyet.com/?job=master%402017-12-12T14%3A53%3A01.451Z&job=CfL-PartU%402017-12-12T15%3A39%3A36.794Z

Change-Id: I2a4adde79c10089130775b8e0df5f9c198855cad

Moving CfL to using partition unit DC_PRED requires 4:1 and 1:4 DC_PRED,
which are not currently implemented. A simple solution is to disable CfL
for 4:1 and 1:4 partitions.

CfL is also disabled for luma intra partitions < 4x4. This is inherent
to luma intra prediction partition sizes. We add an assert to enforce
this.

Resulting in the following regression for Subset1
   PSNR | PSNR Cb | PSNR Cr | PSNR HVS |   SSIM | MS SSIM | CIEDE 2000
-0.0093 |  0.1803 |  0.1519 |  -0.0180 | 0.0256 |  0.0226 |     0.0352

https://two.arewecompressedyet.com/?job=CfL%402017-11-30T19%3A05%3A05.639Z&job=CfL-Disable-4to1%402017-11-30T19%3A04%3A00.761Z

Change-Id: Ie2c8b4d9cb6b6f33a103b540209e1a2fb6df74a7

Results on Subset 1
   PSNR | PSNR Cb | PSNR Cr | PSNR HVS |    SSIM | MS SSIM | CIEDE 2000
-0.0354 | -0.2567 | -0.3941 |   0.0104 | -0.0084 |  0.0120 |    -0.0996

https://arewecompressedyet.com/?job=master%402017-11-03T15%3A57%3A30.643Z&job=cfl-av1-DC_PRED%402017-11-03T16%3A00%3A10.866Z

BUG=aomedia:928

Change-Id: I4e26e8c56d2246ca32b8d86145ef67f6df90d8d1

Pixels are subsampled when they are stored in the CfL prediction
buffer. This allows to avoid having two buffers. No impact on the
bitstream.

Results on subset1 (Compared to parent)

  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

https://arewecompressedyet.com/?job=cfl-avg-on-fly%402017-09-23T21%3A38%3A04.440Z&job=cfl-Sub-On-Store%402017-09-24T15%3A01%3A41.161Z

Change-Id: If051e34d6d7078d77609542305a2afebef5cb177

Instead of storing the transform block average, it is immediately
subtracted from the subsampled pixels. This change does not alter the
bitstream and it reduces CfL complexity.

Change-Id: Ia5038b336abf1ec01e295b235734318906d3bae6

Result from luma subsampling is left-shifted by 3. This avoids having to
do it during averaging, in alpha search and when building the
prediction. This change does not alter the bitstream.

Results on Subset1
  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

https://arewecompressedyet.com/?job=cfl-baseline%402017-09-06T17%3A41%3A38.041Z&job=cfl-SubsampleQ3%402017-09-06T17%3A42%3A01.252Z

Change-Id: I6e89eac6496f7c36e46364c9223fbcbca6759032

Like for intra block in intra frames, an extra call to
txfm_rd_in_plane is added to the RDO of intra blocks in inter frames.
This extra call is performed using the best parameters found during
RDO and the reconstructed luma pixel are stored.

Results on objective-1-fast (compared to CfL on Intra frames only)

   PSNR | PSNR Cb | PSNR Cr | PSNR HVS |    SSIM | MS SSIM | CIEDE 2000
-0.2497 | -3.5526 | -3.5048 |  -0.2456 | -0.2392 | -0.2508 |    -1.4811

https://arewecompressedyet.com/?job=cfl-no-inter%402017-09-13&job=cfl-inter%402017-09-13T14%3A13%3A13.918Z

Change-Id: I70ea2c01859b6c55d7c3eb9680d492c0bfc2aad4

The cfl_init function is moved out of cfl.h simplifying the includes and
removing the need for forward declarations.

Change-Id: I47312b25410b718a830b001391e386647005d57e

Instead of forward-declaring AV1_COMMON and MACROBLOCKD,
move the dependent struct and function prototype closer
to where they are used and after these types are defined.

Change-Id: I75f005b46ef322a6fcbc01377b8dded1637c5f73

When Chroma from Luma is combined with chroma_sub8x8, the prediction
used for sub8x8 blocks originates from multiple luma blocks. Extra
asserts are added to validate that the prediction buffer contains all
the required information.

Change-Id: I305c46ce9b8292697e1d5b181d123461026da11c

Since the scaled luma can be negative, ROUND_POWER_OF_TWO_SIGNED must be used.
This changes the behavior from rounding toward -infinity to rounding towards 0.

Results for Subset1 (compared with 35545dd5 with CfL enabled)
  PSNR | PSNR Cb | PSNR Cr | PSNR HVS |    SSIM | MS SSIM | CIEDE 2000
0.0082 | -0.1061 | -0.0119 |  -0.0126 | -0.0011 | -0.0121 |     0.0094

Change-Id: Ie7258a17a199368339d4794fba6b5916e607c95b

With recent changes, it is now possible to store the storage
flag inside the CFL_CTX. This simplifies the implementation
and will allow reuse in the decoder.

This change does not alter the bitstream.

Change-Id: Ibb8aebdd3d06f8765d40248ece8a038892e87032

Also, move body of update_cfl_costs() to av1_fill_mode_rates().

Results on Subset1 (Compared to 1cfe474b with CFL enabled)

  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

No change in bitstream, for an average encode speed-up of 2.3%.

Change-Id: I3948abcd70cfecad8086edfe4c45552b576ae06f

Expand the range of alpha to [-2, 2] in Q3.
Jointly signal the signs, including zeros.
Use the signs to give context for each quadrant
and half-axis. The (0, 0) point is excluded.
Symmetry in alpha_u == alpha_v yields 6 contexts.

Results on Subset1 (Compared to 9136ab7d with CFL enabled)

   PSNR | PSNR Cb | PSNR Cr | PSNR HVS |    SSIM | MS SSIM | CIEDE 2000
-0.0792 | -0.7535 | -0.7574 |  -0.0639 | -0.0843 | -0.0665 |    -0.3324

Change-Id: I250369692e92a91d9c8d174a203d441217d15063
Signed-off-by: David Michael Barr <b@rr-dav.id.au>

CfL is now an independent mode.

Results on Subset1 (Compared to 4266a7ed with CFL enabled)

   PSNR | PSNR Cb | PSNR Cr | PSNR HVS |    SSIM | MS SSIM | CIEDE 2000
-0.1645 | -0.4017 |  0.2475 |  -0.1851 | -0.2179 | -0.2338 |    -0.2897

Change-Id: I2e86e7ea7bfc12bb1d763e70a136ca992d57a3c5

Alpha's biggest fraction is 1/8, so Q3 does not change the bitstream.

Results on Subset1 (compared to 503aca74 with CfL enabled)

  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: I1fe5b2ace97179d5f950d7406a4f3d391924f89d

The block level DC_PRED computed by CfL goes down from Q6 to Q0. This
will allow to reuse existing assembly for DC_PRED and also reduce the
requirements on the multilpy required to scale the reconstructed luma
values

Results on Subset1 (compared to f9684d222 with CfL enabled)

   PSNR | PSNR Cb | PSNR Cr | PSNR HVS |    SSIM | MS SSIM | CIEDE 2000
-0.0347 |  0.0229 | -0.1326 |  -0.0420 | -0.0057 | -0.0072 |    -0.0644

Change-Id: I6ba82cc9e04fa4ab7c8ec40a7856deb273881748

Since alpha is Q3, we reduce y_average from Q10 to Q3. As such, the
prediction is reduced from Q13 to Q6. Chroma dc_pred is reduced from Q7
to Q6 in order to match with the prediction.

Results on Subset1 (compared to 209de2e5b with CfL enabled)

  PSNR | PSNR Cb | PSNR Cr | PSNR HVS |   SSIM | MS SSIM | CIEDE 2000
0.0010 |  0.0176 | -0.0538 |  -0.0043 | 0.0027 | -0.0097 |    -0.0018

Change-Id: Ib7dd3968a764e0380ddc0ad2333ebacf1e9699cd

The dc_pred values stored in the CfL context are in Q8.7 (Worst case
division will be of 1/128).

Results on Subset1 (compared to f9684d222 with CfL enabled)

  PSNR | PSNR Cb | PSNR Cr | PSNR HVS |   SSIM | MS SSIM | CIEDE 2000
0.0118 | -0.0181 | -0.0109 |   0.0086 | 0.0086 |  0.0196 |     0.0018

Change-Id: I0701e04fb76f03eff12ed01fd5fda675fbb15e32

This change does not impact the bitstream as no loss is incured by using
a fixed point value for the transform size average.

For low bit depth, the transform size average is stored using Q8.10
fixed point format. Worst case, smallest fraction is 1/1024.

Results on Subset1 (Compared to 366b74 with CfL)

  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: Ia5b046b92a0e4c40e413b16af3394bdc0a8c8cd9

When computing alpha, multiple averages are computed, one for each
transform block. The CfL prediction now uses the transform block average
instead of partition block average.

This allows the decoder to build the CfL prediction by using only the
collocated reconstructed luma values for the current transform size and
not the entire partition.

Results on Subset 1 (Compared to 0e81b97c with CfL)

  PSNR | PSNR Cb | PSNR Cr | PSNR HVS |   SSIM | MS SSIM | CIEDE 2000
0.0180 |  0.2627 |  0.2274 |   0.0233 | 0.0301 |  0.0312 |     0.1506

A small regression is expected, this change was made to simplify
hardware implementations.

Change-Id: Ib2ce2a3053b85300c5c62ef0e3270af489568a38

Adjust row and col offset for sub8x8 blocks to allow the CfL prediction
to use all available reconstructed luma pixels.

Results on Subset 1 (Compared to b03c2f44 with CfL)

   PSNR | PSNR Cb | PSNR Cr | PSNR HVS |    SSIM | MS SSIM | CIEDE 2000
-0.1355 | -0.8517 | -0.4481 |  -0.0579 | -0.0237 | -0.0203 |    -0.2765

Change-Id: Ia91f0a078f0ff4f28bb2d272b096f579e0d04dac

The function cfl_compute_parameters is added and contains the logic
related to building the CfL context parameters. As such, many cfl
functions can now be encapsulated inside of cfl.c and not exposed to the
rest of AV1.

This also allows for supplemental asserts that validate that the CfL
context is properly built.

Results on Subset1 (compared to 9c6f8547 with CfL)

  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: I6d14a426416b3af5491bdc145db7281b5e988cae

Results on Subset 1 (Compared to a0f8c145 with CfL)

  PSNR | PSNR Cb | PSNR Cr | PSNR HVS |   SSIM | MS SSIM | CIEDE 2000
0.0677 | -0.3359 | -0.2115 |   0.0529 | 0.0735 |  0.0495 |    -0.0907

Change-Id: Ib61ff862e8cfbdf0c693a4eba5f2712a6e9ab819
Signed-off-by: David Michael Barr <b@rr-dav.id.au>

In 84a44dbe, there was a transcription error in the table.
Point { 5, 3 } was duplicated and { 5, 0 } was missing.

While we are here, updated the order and CDF from subset3.

Results on subset1 (compared to cab68ae6 with CfL)
   PSNR | PSNR Cb | PSNR Cr | PSNR HVS |    SSIM | MS SSIM | CIEDE 2000
 0.0416 | -0.2732 | -0.3848 |   0.0385 | 0.0309 |  0.0230 |    -0.0178

Change-Id: I3244245792c5ab99b4149ae5f8a2439d4214ed69
Signed-off-by: David Michael Barr <b@rr-dav.id.au>

Extract the compution of the luma reconstructed average out of cfl_load
and into cfl_compute_average. The reconstructed luma average is stored
in the CFL_CONTEXT to avoid computing it for each transform block and
for each plane.

Results on subset1 (compared to 803bea26 with CfL)
   PSNR | PSNR Cb | PSNR Cr | PSNR HVS |    SSIM | MS SSIM | CIEDE 2000
-0.0474 | -0.1486 | -0.2931 |  -0.0358 | -0.0397 | -0.0127 |    -0.1162

Change-Id: I9e34af0fe5961ce8dbe70cb80aea2a16221d0d92

plane_bsize is now computed properly. This also includes support for the
special case of blocks < 4X4

Results on subset1 (compared to 8e689e4b with CfL)
   PSNR | PSNR Cb | PSNR Cr | PSNR HVS |    SSIM | MS SSIM | CIEDE 2000
-0.0218 | -0.2328 | -0.2555 |  -0.0230 | -0.0379 | -0.0723 |    -0.1205

Change-Id: I6ec87d818d8df6a40ecf3bb1b86954e59c952930

This change does not impact the bitstream
  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: I6e131e91bad5efa345ed2542ae970eb6122eff51

Separate the codes into a table of distinct values
and an index into that table.
Pull the SSE calculation of the RDO loop and avoid
repeating for the same alpha values.

Change-Id: I8c4bd7eab6f8000e6aca9687d9190abc3e270c37
Signed-off-by: David Michael Barr <b@rr-dav.id.au>

Encapsulates the logic to update the rate of each CfL codeword.
The if statements are removed from the loop and the arrays are
stored in CFL_CTX instead of being declared every time.

Change-Id: I0cb208b14e6c6a888210dd33c5e8fe8d74dd87f4

Move cfl_idx_to_alpha in the header to facilitate inlining.
Remove the forward MB_MODE_INFO forward declaration

Change-Id: Id33fb0228d88b6285252843e2345a0d3ae875cd2

The prediction block level DC_PRED is stored and computed as double
instead of int.

Change-Id: I22766c102a7b62d4b5e7621438185808cc0ea8f4

Since the size used with cfl_load can either be based on the transform
block size and the prediction block size, width and height are used as
parameters instead of TX_SIZE.

This resolves a problem where cfl_compute_alpha_ind was reading
uninitialized memory.

Change-Id: I187dbdd5b2e8bd85e82bb77eb74859bee2cd3f1e

Writes and reads alpha to and from the bitstream.

A special case is needed on the encoder side to handle prediction block
skips. Since whether or not a prediction block is skipped during CfL, a
rollback is required if the block was skipped and the alpha index was
not zero. The advantage of this is that no signaling is required when
the prediction block is skipped as it is assumed tha the alpha index is
zero.

A encode facade is added to the intra prediction facade as CfL requires
special encoder side operations.

Change-Id: Ic3b11d0fdbd51389d862112eb09d8785127a6b06