Investigate and fix.

For benchmarks, the Rust implementation must be wrapped in another function labeled #[inline(never)], for a fair comparison against the C.

Optionally, this could potentially be made more concise with a macro.

Currently in encode_tile for every superblock, we search for the intra mode with the best RD (based on a 64x64 SSE), for PARTITION_NONE. We should extend this search to one-level of split.

Such as reference list, etc.

If current block is skipped, the encoder should reset coeff contexts of above and left as zeros,
for width and height of the current block size, resp.

Since becced1, frames with show_existing_frame set are blank in the reconstruction stream.

rav1e is great and all, but it doesn't quite roll off the tongue. Let's comment on this issue with name ideas, and then vote on them using 👍.

Tests can be like:

1. Set min partition size as 4x4
2. Set min partition size as 8x8
3. Set min partition size as 16x16
4. Set min partition size as 32x32

yushin@maui:~/workspace/rav1e$ cargo test
Compiling bitstream-io v0.6.3
Compiling bitflags v1.0.1
Compiling bencher v0.1.5
Compiling bitflags v0.4.0
Compiling encode_unicode v0.1.3
Compiling vec_map v0.8.0
Compiling unicode-width v0.1.4
Compiling ansi_term v0.10.2
Compiling y4m v0.1.1
Compiling byteorder v1.2.1
Compiling libc v0.2.36
Compiling strsim v0.7.0
Compiling textwrap v0.9.0
Compiling nix v0.5.1
Compiling atty v0.2.6
Compiling rand v0.4.2
Compiling clap v2.30.0
Compiling rustyline v1.0.0
Compiling rav1e v0.1.0 (file:///home/yushin/workspace/rav1e)
Finished dev [unoptimized + debuginfo] target(s) in 9.73 secs
Running target/debug/deps/rav1e-9b00cc822757fb1d

running 6 tests
test ec::test::booleans ... ok
test ec::test::cdf ... ok
test ec::test::mixed ... ok
test predict::test::pred_max ... ok
test predict::test::pred_same ... ok
test predict::test::pred_matches ... ok

test result: ok. 6 passed; 0 failed; 0 ignored; 0 measured; 0 filtered out

 Running target/debug/deps/rav1e-5895d273f770ce9f

running 0 tests

test result: ok. 0 passed; 0 failed; 0 ignored; 0 measured; 0 filtered out

 Running target/debug/deps/rav1repl-81dcf9202be56248

running 0 tests

test result: ok. 0 passed; 0 failed; 0 ignored; 0 measured; 0 filtered out

Doc-tests rav1e

running 0 tests

test result: ok. 0 passed; 0 failed; 0 ignored; 0 measured; 0 filtered out

yushin@maui:~/workspace/rav1e$ cargo bench
Compiling bencher v0.1.5
Compiling rav1e v0.1.0 (file:///home/yushin/workspace/rav1e)
warning: unused import: rav1e::predict::*;
--> benches/bench.rs:95:5
|
95 | use rav1e::predict::*;
| ^^^^^^^^^^^^^^^^^^
|
= note: #[warn(unused_imports)] on by default

warning: unused variable: mode
--> benches/bench.rs:111:9
|
111 | let mode = PredictionMode::DC_PRED;
| ^^^^
|
= note: #[warn(unused_variables)] on by default
= note: to avoid this warning, consider using _mode instead

warning: foreign function is never used: highbd_dc_left_predictor
--> benches/bench.rs:15:5
|
15 | / fn highbd_dc_left_predictor(dst: *mut u16, stride: libc::ptrdiff_t, bw: libc::c_int,
16 | | bh: libc::c_int, above: *const u16,
17 | | left: *const u16, bd: libc::c_int);
| |______________________________________________________________^
|
= note: #[warn(dead_code)] on by default

warning: foreign function is never used: highbd_dc_top_predictor
--> benches/bench.rs:18:5
|
18 | / fn highbd_dc_top_predictor(dst: *mut u16, stride: libc::ptrdiff_t, bw: libc::c_int,
19 | | bh: libc::c_int, above: *const u16,
20 | | left: *const u16, bd: libc::c_int);
| |______________________________________________________________^

warning: foreign function is never used: highbd_h_predictor
--> benches/bench.rs:21:5
|
21 | / fn highbd_h_predictor(dst: *mut u16, stride: libc::ptrdiff_t, bw: libc::c_int,
22 | | bh: libc::c_int, above: *const u16,
23 | | left: *const u16, bd: libc::c_int);
| |______________________________________________________________^

warning: foreign function is never used: highbd_v_predictor
--> benches/bench.rs:24:5
|
24 | / fn highbd_v_predictor(dst: *mut u16, stride: libc::ptrdiff_t, bw: libc::c_int,
25 | | bh: libc::c_int, above: *const u16,
26 | | left: *const u16, bd: libc::c_int);
| |___________________________________________^

warning: function is never used: pred_h_4x4
--> benches/bench.rs:37:1
|
37 | fn pred_h_4x4(output: &mut [u16], stride: usize, above: &[u16], left: &[u16]) {
| ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^

warning: function is never used: pred_v_4x4
--> benches/bench.rs:44:1
|
44 | fn pred_v_4x4(output: &mut [u16], stride: usize, above: &[u16], left: &[u16]) {
| ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^

Finished release [optimized] target(s) in 3.83 secs
 Running target/release/deps/rav1e-c60027f9b9bd002f

running 6 tests
test ec::test::booleans ... ignored
test ec::test::cdf ... ignored
test ec::test::mixed ... ignored
test predict::test::pred_matches ... ignored
test predict::test::pred_max ... ignored
test predict::test::pred_same ... ignored

test result: ok. 0 passed; 0 failed; 6 ignored; 0 measured; 0 filtered out

 Running target/release/deps/rav1e-ab1af7a55ae80597

running 0 tests

test result: ok. 0 passed; 0 failed; 0 ignored; 0 measured; 0 filtered out

 Running target/release/deps/rav1repl-b79447a2b246d724

running 0 tests

test result: ok. 0 passed; 0 failed; 0 ignored; 0 measured; 0 filtered out

 Running target/release/deps/bench-741fd4a6e67fa1d0

running 4 tests
test aom ... bench: 1,188,882 ns/iter (+/- 38,736)
test native ... bench: 975,183 ns/iter (+/- 59,961)
test native_trait ... bench: 269,857 ns/iter (+/- 10,470)
test write_b_bench ... bench: 123,071 ns/iter (+/- 6,586)

test result: ok. 0 passed; 0 failed; 0 ignored; 4 measured

yushin@maui:~/workspace/rav1e$

Currently they loop through (x,y) pixel coordinates which is slow.

(the code is at the very bottom of src/libr.s)

• a manual cargo clean is needed when the libaom code gets updated
• on some platforms there is no way to avoid the wxwindows dependency
• the encoding-decoding test is not present yet

Reverting #99 would be an improvement.

cargo run --bin rav1e --release
   Compiling rav1e v0.1.0 (file:///Users/.../Workspace/rav1e)
    Finished release [optimized] target(s) in 6.98 secs
     Running `target/release/rav1e`

7s is too long to wait if the code hasn't changed.

Do we want 128x128 from the start? Do this include 4:1 and 1:4 partitions? (use speed levels for this)

Right now, cargo run or cargo run -- --help panics.

This is optional for inter mode coding, i.e. motion compensated blocks.

Want to remove this line, but cannot since it introduces desync, esp. at lower QIndex.

To reproduce, encode akiyo_qcif.y4m. (176x144)

Reconstructed output:

Decoded output:

We need to skip non-visible blocks in write_sb (done), change the dimensions that we write in write_uncompressed_header (done), change how partitions are coded on the edges of the frame (not started), get all sizes of intra-predictors (not needed since we force 4x4 everywhere).

I encoded park_joy_420_720p50.y4m, and then decoded it using the specific git checkout of aom listed in the README. It resulted in this (after being piped to ffmpeg to dump PNGs):

https://themayhaks.com/~gideon/test-av1-png/

IVF file is here (about 88MB):

https://themayhaks.com/~gideon/test.ivf

I'm not sure if this is just a result of me not resizing to 1280x768 before feeding it to rav1e, but it didn't complain and it seems to have converted it to 1280x768 during the encoding process.

That is another lovely built-in feature of Rust. Let's use it asap, and not writing hassle separate unit-tests programs.

Currently a number of experiments are disabled in the decoder. To update the decoder to the final AV1 bitstream, we must enable them. A number of these are quite easy - they simply add a bit to the uncompressed header indicating whether they are used or not, so you must simply add that bit, then remove the flag from the decoder's configuration. Experiments that have simple changes to the header:

--disable-aom_qm
--disable-ext-intra
--disable-ext_partition
--disable-loopfilter_level
--disable-cfl (requires look up table changes)
--disable-kf-ctx
--disable-striped_loop_restoration
--disable-max_tile
--disable-ext-intra-mod
--disable-frame_size
--disable-simplify_tx_mode
--disable-obu (more complex, but still only header changes)

Prerequisite: #74
And, motivation stated in L #75, #75

This requires mostly adding support to the y4m crate:

https://github.com/PistonDevelopers/y4m

Steps:

• Understand LV_MAP_MULTI (multi symbol version of Level Map) algorithm and code in libaom.
• Force partition level skip, enable lv-map, make sure image decodes. (i.e. all changes are self contained in write_coeffs)
• Force lv-map level skip, don't desync.
• Code a +/-1 DC coefficient, don't desync within a superblock.
• Code arbitrary coefficients within a superblock.
• Code all superblocks without desyncing.

LV_Map slides
https://docs.google.com/presentation/d/1nbplj5yuq2OTNQFMbymfASRsfg58GLZjvzWE3MtoZWg/edit#slide=id.g245e1ced9c_1_5

More slides (improvements)
https://docs.google.com/presentation/d/1GcCv543QV10uIehWAaoM9xjJvJkXP-MbLV_Z6GVvIq8/edit#slide=id.p

Paper
https://usercontent.irccloud-cdn.com/file/KUIqFjBi/level_map.pdf

Remove experimental flag of LV_MAP_MULTI
78f7bb98558f513ab69b7d871bf45bf119b1fd5b
https://aomedia.googlesource.com/aom/+/78f7bb98558f513ab69b7d871bf45bf119b1fd5b

Remove experimental flag of LV_MAP
cad5ebce76b50cb413b980a17748941621a747e3
https://aomedia.googlesource.com/aom/+/cad5ebce76b50cb413b980a17748941621a747e3

While the av1 syntax spec, https://aomediacodec.github.io/av1-spec, is for decoder,
we might want to consider following the variable and function names there as much as possible.

For example, instead of those confusing encode_superblock(), encode_b(), encode_sb(), ...etc.,
we can try find better names based on counterparts functions in decoder side,
https://aomediacodec.github.io/av1-spec/#decode-partition-syntax.

• This task includes:

1. Need to add reset_skip_context() if a partition is skipped.

Test can be like:

1. Random skip of partitions.
2. Skip odd rows, skip even rows
3. Skip odd columns, skip even rows

We failed to detect breakage. Requires ci to run nightly.

This task will need sequence_header_read/write(), if not introduced yet.

For now, set a constant CDEF level, and simply call the C CDEF implementation.

rom1v works on #1126

Rename the existing task #127 as "Support Minimum Partition types" then create new task with old name from #127, "Support all partitions types".
As explained in #127, supporting all partition types does not seem essential and not meaningful w/o "ROD based block size decision", so I assume it is not part of MVP.

• Do not support TX_SLELECT at first, but use LARGEST_TX only.

Minimum prerequisite is a tiling support.

Would be happy if we can do this by 2020.

Fuzzing raw input is generally easier, but fuzzing the decisions would be more useful (but probably more involved/intrusive).

Some potentially helpful links:

https://github.com/pingcap/fail-rs
https://rust-fuzz.github.io/book/introduction.html
https://github.com/BurntSushi/quickcheck
https://github.com/altsysrq/proptest

and might as well add this link from @lu-zero for posterity:
https://doc.rust-lang.org/std/num/struct.Wrapping.html

Currently the encoder uses sse to do rate-distortion optimization. Make this function runtime switchable with a function that computes cdef-dist instead.

Found serious problem in our save/restore (i.e. checkpoint/rollback) states, that we are saving "blocks: Vec" of "struct BlockContext", which means all the MI blocks of a frame!

https://docs.travis-ci.com/user/languages/rust/

Speed level decides the minimum partition block size, where recursive quadri-sect splitting stops at.
0(slowest) : BLOCK_4X4
1 : BLOCK_8X8
2 : BLOCK_16X16
3 : BLOCK_32X32
4~10(fastest) : BLOCK_64X64

Encoder option : "--speed=[0..10]" or "-s [0..10]".

Exception: SuperBlock(SB)s on right or bottom frame borders always split down to BLOCK_4X4.

In the future, min partition block size may be controlled by separate encoder option.

Probably by relying on aomdec being on the PATH

https://aomedia-review.googlesource.com/c/aom/+/42861

smells like a decoder bug