Raphaël Zumer authored 6 years ago and Thomas Daede committed 6 years ago

Run with --features=comparative_bench
Applies to intra prediction tests and disables native function inlining
Also refactor benchmarks and reduce code duplication

Also move reference frame buffer into frame invariants to
avoid having to pass an additional parameter to each function
requiring access to reference frame buffer.

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

This is required because the old ones are only available as
sse2 intrinsics.

This disables 4x4 and 8x8 transform assembly.

Monty Montgomery authored 6 years ago and Thomas Daede committed 6 years ago

This change is to allow multiple and throwaway writers to be used in
predictive and out-of-order coding.  A Writer must be passed in (and
checkpointed) separately/explicitly.

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

This affects all intra prediction modes in rav1e except DC_PRED.

* Enable writing/using all possible CDEF setups

No RDO as yet-- this patch implements setting up, passing and
utilizing full cdef information

* Do not write CDEF frame or block data when cdef is disabled

This also requires passing Sequence structure deeper into RDO loop than before.

* Ignore aom.rs in all directories

* Add encode-decode test for various frame dimensions

* Remove redundant .gitignore

* Add testing subsection in README.md

* Reduce runtime of dimensions encode-decode test

* Comment out failing dimensions in decoding test

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

It enables the switch for INTRA_EDGE_FILTER at frame level,
which is turned off by default in rav1e.

The commit be197cf5 has caused
the corrupted decoded video frames with speed=0 or 1, since its landing on
June 26.

The quick fix is to disable skip RDO loop, removing its outermost loop,
and set skip = false.

This bug has been there around a month and believed to be the culprit of
breakage of ext_partition_types PR by TD-Linux as well.

Better fix will come, which is also coincidentally stated earlier in
https://github.com/xiph/rav1e/issues/373.

No need to clutter up lib.rs with self-contained CDEF work.
Split the CDEF functions into their own file.  No functional change.

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

* Define reference frame buffer array

A reference frame buffer with 8 slots is defined. The buffer is
updated after encoding each frame according to the value of
refresh_frame_flags. The value of refresh_frame_flags is also
set and properly encoded.

* Fix test function

Also move update of frame buffer to a separate function

* Use named constants instead of magic numbers

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

* Remove 'EXT_' in'EXT_TX' for the names

Since everything that has been called extended is now a standard.

* Rename TxSetType TxSet and etc

- Rename 'ext_tx_' as 'tx_'.
- Rename 'write_tx_type_lv_map' as 'write_tx_type'.

* Rename '*_set_type' as '*_set'

* Rename '*_ext_tx' as '*_tx'

Because it is not 'extended' anymore but a standard.

Luca Barbato authored 6 years ago and Thomas Daede committed 6 years ago

* Make more structures Debug

Makes easier to println!-debug.

* Fix the cmake flags passed to build the decoding library

* Make sure to not call partially implemented ith

At least the sse4.1 variant is incomplete and triggers an assert.

* Do not use SIMD for TxType::IDTX ith

The implementation of it is incomplete.

* Provide an encode-decode test

Use `cargo test --features=decode_test -- --ignored` to run it.

* Enable decode_test in travis

Monty Montgomery authored 6 years ago and Thomas Daede committed 6 years ago

Correction to end-of-frame bounds checking on skip flags (can't get
skip flag for blocks of the endge of the frame)

* cdef: WIP

* cdef: now calling cdef_find_dir_c (but only on the first block)

* cdef: call cdef_find_dir on every block (only safe for size % 64 == 0)

* cdef: apply CDEF to luma with pri_strength 7 and sec_strength 0.

* cdef: only apply to full 64X64 filter blocks for now

* cdef: signal cdef strength

* cdef: comment out printing

* cdef: first attempt at applying CDEF to chroma planes

* Corrections to first CDEF attempts

This is still not quite first cuts, but it is basically functional.
It still hard-codes specific CDEF strengths.
There is still an occasional block mismatch.

* Implement CDEF paramters, remove hardwired settings

Add CDEF settings to the FrameInvariants, rather than using hardwires
strength values duplicated in multiple locations.

* Remove last CDEF dependency on libaom

* Remove unrelated change from CDEF work

remove -r rec_file.y4m from rav1e command line

Guillaume Martres authored 6 years ago and Thomas Daede committed 6 years ago

It turns out that there's already some code to emit the
skip_mode_present flag, it's just dead code because skip_mode_allowed is
currently always false, added an assertion to revisit this code when it
becomes live because it's incomplete.