This reduces the time for cdef_filter_frame by ~33%.

Change MV mode if NEAREST isn't NEAREST anymore instead of doing repeated RDO optimizations. Improves run time by about 20-25%.

Add assert statement to check that nearest MV is indeed the same
as the MV passed as argument to encode_block_b()

Move MV stack construction out of encoding function such that the
value of the nearest MV can be determined before encoding

* Move motion compensation to separate function

* Motion compensate for every iteration of tx search

Yushin Cho authored 6 years ago and Yushin Cho committed 6 years ago

Also does pulling tx_type_rdo() out of chroma RDO, as requested in #521.

* Move SAD computation into a separate function

* Always inline SAD function for speed

This yields an ~4x speed-up of this function.
Update sameness test with in-range values.
Return early for zero alpha.
Elide bounds checks from SSSE3 implementation
Bounds are enforced in the safe wrapper.

Frank Bossen authored 6 years ago and Thomas Daede committed 6 years ago

Move .take() ealier in the chain as suggested by lu-zero

Frank Bossen authored 6 years ago and Thomas Daede committed 6 years ago

Use single loop for all coefficients

Frank Bossen authored 6 years ago and Thomas Daede committed 6 years ago

Speed up probability update function

Remove array and compute values on the fly instead

Add reference frame structure to capture YUV data and probabilities
of a previously decoded frame.
Symbol counts are reset and not carried from one frame to the next.
Error resilient mode is disabled to support this feature.

Add 1/8th pel motion refinement step but don't enable it by default

Currently this feature yields an average loss of about 1 percent
BD rate

Add storage for block-level loop deblocking filter delta adjustments, as well as infrastructure for writing out these parameters.

Unused as yet.  No funcitonal change to current encoder.

This patch implements the block-level deblocking filter params set by the frame header.

As requested by Yushin in 490

Frank Bossen authored 6 years ago and Yushin Cho committed 6 years ago

Search 1/2 pel square, then 1/4 pel square

Implement motion compensation interpolation filters since chroma
may now require half-sample interpolation

Frank Bossen authored 6 years ago and Yushin Cho committed 6 years ago

Removal suppresses compiler warnings

Motion estimation code should be about 3x faster. Disassembly of
compiled code shows use of SIMD instructions.

* Add the chroma-from-luma predictor and a test
* Add benchmarks for CfL predictor
* Implement CfL signalling
* Plumb from predict_intra to pred_cfl
* Add struct for CfL parameters
* Compute subsampled luma AC and plumb it with CfL alpha
* Plumb CfL parameters up to RDO
* Add CfL to RDO loop, no parameter search

* Add motion vector paramater to inter prediction function

Motion vector is read from added data field in Block structure

* Add reference frame and motion vector fields/parameters

Reference frame and motion vector are added fields to the
RDOOutput structure and added parameters to the block
encoding function

* Fix inter prediction function for nonzero MVs

* Add call to encode new motion vector

* Import default cdfs for MV coding from libaom

* Fix handling of MV precision in MV coding function

* Add coding of DRL mode

* Add motion vector stack

* Add new MV counter to correctly determine context

* Use motion vectors in MV stack for prediction

* Sort MV stack according to weights

Also update weights after near search

* Fix log2() function

* Compute correct context for DRL flag

* Store MVs with 1/8 pel precision and fix MV generation

* Add some basic motion estimation

Motion estimation is full search (+/-16 pel range in 2-pel increments) and
is done independently for each block

* Fix chroma motion compensation for small blocks

Chroma motion compensation sometimes happens on 2x2 basis

This patch adds basic data structures for tracking loop filter state,
as well as being able to encode complete loop filter frame header
syntax.  No functional change.

Loop filter needs to know about both block and tx size.  Right now,
they're implicitly hardwired to the same thing.

This patch makes the hardwiring explicit and adds placeholders to
allow a full loop filter implementation.