Encoder crashes with target-cpu=native on Ryzen R7 2700X

[quietvoid]

Hi,

Build from master, 1622e66
I've been building rav1e with this command in WSL:

RUSTFLAGS="-C target-cpu=native" cargo build --release --target=x86_64-pc-windows-gnu

The produced binary crashes with this command line:

.\rav1e.exe tos_8b.y4m --keyint 24 --threads 16 --tile-rows 8 --tile-cols 8 --bitrate 4000 --range limited --first-pass out.stats --output out.ivf

And error

index out of bounds: the len is 134912 but the index is 134944
in rav1e/src/frame/plane.rs, line 295

However building without target-cpu and the p20191215 binary works.
For a 10 bit input, rav1e also stops before the encoding starts, without error or panic.

The selected CPU level is AVX2, for a Ryzen R7 2700X.

[rzumer]

As of a few days ago I had no problem encoding in native mode, so unless one of the most recent two commits broke it, a sample to reproduce it would be helpful.

[quietvoid]

Both 8 bit and 10 bit samples can be found here: https://mega.nz/#F!dFUWkCKK!vB73_UAit1v2v_KjI-qH2Q
They are from Tears of Steel, downsampled to 1950x816 and split in 24 frames scenes.

[lu-zero]

Tested on linux with your 8bit sample 1622e66 and rustc 1.39.0 (4560ea788 2019-11-04)

$ ./rav1e-p20191215 ~/Samples/sample_8b.y4m --keyint 24 --threads 16 --tile-rows 8 --tile-cols 8 --bitrate 4000 --range limited --first-pass out.stats --output out.ivf
 INFO  rav1e > Using y4m decoder: 1950x816p @ 24/1 fps, 4:2:0, 8-bit
 INFO  rav1e > Encoding settings: keyint_min=12 keyint_max=24 quantizer=255 bitrate=4000000 min_quantizer=0 low_latency=false tune=Psychovisual rdo_lookahead_frames=40 min_block_size=8x8 multiref=true fast_deblock=false reduced_tx_set=true tx_domain_distortion=false tx_domain_rate=false encode_bottomup=false rdo_tx_decision=false prediction_modes=Complex-KFs include_near_mvs=false no_scene_detection=false diamond_me=true cdef=true quantizer_rdo=false use_satd_subpel=true non_square_partition=false enable_timing_info=false
 INFO  rav1e::api::config > CPU Feature Level: AVX2
 INFO  rav1e::api::internal > Using 56 tiles (8x7)
 INFO  rav1e::stats         > encoded 49 frames, 1.924 fps, 3138.62 Kb/s
 INFO  rav1e::stats         > ----------
 INFO  rav1e::stats         > Key frame:             3 | avg QP:  66.00 | avg size:   80119 B
 INFO  rav1e::stats         > Inter frame:          46 | avg QP: 125.00 | avg size:   12187 B
 INFO  rav1e::stats         > Intra only frame:      0 | avg QP:   0.00 | avg size:       0 B
 INFO  rav1e::stats         > Switching frame:       0 | avg QP:   0.00 | avg size:       0 B

On windows using mingw64:

$ ./target/release/rav1e.exe ../sample_8b.y4m --keyint 24 --threads 16 --tile-rows 8 --tile-cols 8 --bitrate 4000 --range limited --first-pass out.stats --output out.ivf
 INFO  rav1e > Using y4m decoder: 1950x816p @ 24/1 fps, 4:2:0, 8-bit
 INFO  rav1e > Encoding settings: keyint_min=12 keyint_max=24 quantizer=255 bitrate=4000000 min_quantizer=0 low_latency=false tune=Psychovisual rdo_lookahead_frames=40 min_block_size=8x8 multiref=true fast_deblock=false reduced_tx_set=true tx_domain_distortion=false tx_domain_rate=false encode_bottomup=false rdo_tx_decision=false prediction_modes=Complex-KFs include_near_mvs=false no_scene_detection=false diamond_me=true cdef=true quantizer_rdo=false use_satd_subpel=true non_square_partition=false enable_timing_info=false
 INFO  rav1e::api::config > CPU Feature Level: AVX2
 INFO  rav1e::api::internal > Using 56 tiles (8x7)
 INFO  rav1e::stats         > encoded 49 frames, 0.906 fps, 3138.62 Kb/s
 INFO  rav1e::stats         > ----------
 INFO  rav1e::stats         > Key frame:             3 | avg QP:  66.00 | avg size:   80119 B
 INFO  rav1e::stats         > Inter frame:          46 | avg QP: 125.00 | avg size:   12187 B
 INFO  rav1e::stats         > Intra only frame:      0 | avg QP:   0.00 | avg size:       0 B
 INFO  rav1e::stats         > Switching frame:       0 | avg QP:   0.00 | avg size:       0 B

$ rustc --version
rustc 1.39.0 (4560ea788 2019-11-04)

[lu-zero]

Can you please tell me the compiler version and if using

RUSTFLAGS="-C target-cpu=+avx2"

Is still causing problems?

[quietvoid]

It seems I can't build some dependencies with RUSTFLAGS="-C target-cpu=+avx2" on WSL with --target=x86_64-pc-windows-gnu.

LLVM ERROR: 64-bit code requested on a subtarget that doesn't support it!
error: could not compile lazy_static.
'+avx2' is not a recognized processor for this target (ignoring processor)
...
LLVM ERROR: 64-bit code requested on a subtarget that doesn't support it!
error: could not compile regex-syntax.

Building for Linux with RUSTFLAGS="-C target-cpu=native" cargo build --release works when running straight from WSL.

So there's somehow an issue with targeting x86_64-pc-windows-gnu
rustc 1.39.0 (4560ea788 2019-11-04)

[lu-zero]

I typoed target-cpu for target-feature= sorry.

[lu-zero]

Passing `RUSTFLAGS="-C target-feature=+avx2" I cannot still reproduce the problem here.

[easyfab]

For info,
Building rav1e under windows 10 ( msys2) for my 2700x with:
RUSTFLAGS="-C target-cpu=native" cargo build -j16 --release
No problem here with your samples and command lines

[quietvoid]

Indeed compiling with RUSTFLAGS="-C target-feature=+avx2" cargo build --release --target=x86_64-pc-windows-gnu, rav1e does not crash.
So this might just be a cross compiling issue with rustc, and target-cpu=native when using WSL.

[lu-zero]

I guess we should suggest to not use the -C target-cpu=native and explicitly enable certain features.

[shssoichiro]

Specifically with WSL, anyway. I have no problems using target-cpu=native on Linux with a 2700X.

[quietvoid]

Since it's so specific, this can be closed?

[lu-zero]

Probably it could be reported upstream, but I guess there isn't much we can do on our side :)

[shssoichiro]

Hey, I believe this is already fixed upstream: rust-lang/rust#63959 although it won't be released until Rust 1.41