GitLab

In addition to PARTITION_SPLIT, fully enable early-exit for HORZ
and VERT partition types. This gives further reduction of encoding time
at speed 0, now x3 and x2 for Qindex 252 and 220, resp.

It turns out that the broken directional intra prediction has been
the curprit of bdrate change when enabling full early-exit with bottomup
partition search.

Vladimir Kazakov authored Feb 12, 2019 and Josh Holmer committed Feb 12, 2019

https://rust-lang.github.io/rust-clippy/master/index.html#deprecated_cfg_attr

Working around an outage of https://www.nasm.us/

We use the AC quantizer as the source quantizer since its quantizer
 tables have unique entries, while the DC tables do not.

Pick the nearest DC entry since an exact match may be unavailable,
 and target the midpoint of the chosen entries.

We are limited by the resolution of the tables, but by picking
 lambda based on the desired quantizer (which can have much higher
 precision), this should be reasonably close to a truely flat
 quantizer.

This should improve PSNR and will likely harm the other metrics.

This only implements the "constant quantizer" strategy, and uses
 the same per-frame-type QP index adjustments as the current code.
However, it is one step towards the true quantizer-based
 adjustments required by the rate control model.

This should result in no change in encoded quality.

Instead of calculating it on the fly from the quantizer wherever it
 happens to be needed.
This will allow us to choose lambda with more precision than a
 single quantizer step size later on, for more precise rate
 targeting.

This results in no change in quality.

https://rust-lang.github.io/rust-clippy/master/index.html#range_plus_one

Vladimir Kazakov authored Feb 09, 2019 and Thomas Daede committed Feb 09, 2019

Only those that don't cause warnings and errors.

Automated by running `cargo fix --edition-idioms`

Nathan E. Egge authored Feb 08, 2019 and Nathan Egge committed Feb 08, 2019

This commit does not require the caller of Frame::new() to know to pad
 the width and height to multiples of 8x8 so that the internal buffers
 are allocated properly.
When filling the allocated buffers, callers should use the
 PlaneConfig.stride variable which is set by Frame::new().

The 'needs_rec' parameter for one of the two rdo_mode_decision() calls by
topdown parition search has been passed with wrong value.
It should be false, since re-encoding step follows once rdo_mode_decision()
decided the mode.
In fact, we don't need 'needs_rec' parameter in rdo_mode_decision(), so remove it.

This patch will cause loss of coding efficiency (0.1250% at speed 1)
with a small speedup.

However, if there is more than one intra prediction blocks in a
partition, i.e. tx_size < partition size, rdo_mode_decision() will need
generate reconstrtucted pixels for intra prediction modes. Sine tx
partition is not provided by rav1e yet, this condition can be
implemented in the future.

Thomas Daede authored Feb 08, 2019 and Thomas Daede committed Feb 08, 2019

Fixes #954

Enable tx domain distortion at speed >= 4 instead of >= 1

This way, the default speed (3) always uses pixel domain distortion
which is easier to understand while we figure out what our quantizers
should look like, how to set lambda, what distortion metrics to use,
etc.

Note that this only affects --tune Psnr since --tune Psychovisual always
uses (weighted) pixel domain distortion.

At speed 3 with --tune Psnr the metrics change are:

   PSNR | PSNR Cb | PSNR Cr | PSNR HVS |    SSIM | MS SSIM | CIEDE 2000
-0.1422 | -0.4571 | -0.2197 |  -0.1806 | -0.1451 | -0.1617 |    -0.5266

And the encoding speed seems to be only very slightly reduced.

* Mention rectangular blocks
* Mention directional modes