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

This opens the path to finer control by the user over the encoding
options, and centralizes the speed preset logic in one location.

Closes #290

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

Straightforward addition of 14 more 2:1 and 4:12 rectangular
sized intra predictions, following current trait approach.

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

I started with all tolerances at zero then increased when seeing a failure.
With these tolerances I have now run the test 150 times without failure.

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

Tests are simple roundtrip of transforms, for all currently supported
sizes and types. Tolerance of +-1 is accepted.

Josh Holmer authored 6 years ago and Luca Barbato committed 6 years ago

Adds a new `--verbose` flag to display frame-by-frame info similar to
the old behavior.

It seems to make no difference in benchmark reliability for me.

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

Followup to #710

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

rav1e is no longer dependent on libaom unless running the comparative bench suite.

Inverse 64-point transform is added.
Forward 64-point transform computes all 4096 coefficients such
that distortion from zeroing high-frequency coefficients can
be easily computed

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

It is more performant to pass small types which implement `Copy` by
value, because they can be passed directly through registers. Clippy is
conservative with this lint and optimizes for 32-bit architectures. The
setting can be overridden in `clippy.toml` if we choose to further
optimize for 64-bit architectures.

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

`add` takes a `usize` parameter, so this method can be used to avoid an
unnecessary cast to `isize`.

* Modify frame management to allow for deeper hierarchies

Half the slots are used for top-level frames and the other half
for other frames, where each level in the hierarchy has a single slot
assigned to it

* Reuse probabilities from previous frames only when coding top level

* Increase order hint bits and remove outdated assert

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

`write` may not process the entire buffer, so if using it, it is
recommended to handle the return value (number of bytes written), or
strange runtime errors may occur. Since we are clearing the buffers
immediately after, it seems the desired behavior is to write the entire
buffer, for which `write_all` will guarantee correct behavior.

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

It is more idiomatic in Rust to omit the `return` keyword when possible.
This is a purely stylistic lint and has no effect on performance.

It is currently used for the extended tests.

Fixes #651 and Completes #630.

They are already called by the API.

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

And restrict it to u64 instead of usize.

They take slightly more than 1 minute to execute,
do not twice the normal tests while at it.

Reduces encoder runtime by about 5% at default speed setting

The inputs are generated with dav1d layout.

Address #686

Also, add simple reference tests for u8 prediction.

* Generate downsampled versions of input frames and store them with reference frames

* Add hierarchical ME, where initial ME is done using 4x subsampled 64x64 blocks. MVs are refined using 2x subsampled 32x32 blocks. In the refinement step, search is done around MV found for colocated 64x64 block, and also around MVs found for neighboring 64x64 blocks

Use transform domain distortion during RDO-based mode decision.
- Turn on tx-domain distortion when speed >= 1.
-  For CfL mode, use pixel domain distortion since reconstructed luma pixels are required for CfL mode of chroma channels.

For default speed = 3, there is a regression of 0.35% PSNR bd-rate increase on AWCY.

  PSNR | PSNR Cb | PSNR Cr | PSNR HVS |   SSIM  | MS SSIM | CIEDE 2000
0.3440 |  0.4580   |  0.3700   |   0.2730     | 0.2714 |  0.2616    |     0.5526

Frame reordering should be applied when low_latency is false