Rust utilities for YUV format handling and conversion.

Fast and simple YUV approximation conversion in pure Rust. At most the same as libyuv does. Performance will be equal to libyuv or slightly higher on platforms where SIMD is implemented. Otherwise, equal or slower.

Mostly implemented AVX-512, AVX2, SSE, NEON, WASM. Waiting for SVE to be available in nightly.

x86 targets with SSE and AVX uses runtime dispatch to detect available cpu features.

Supports:

• [x] YCbCr ( aka YUV )
• [x] YCgCo
• [x] YCgCo-Re/YCgCo-Ro
• [x] YUY2
• [x] Identity ( GBR )
• [x] Sharp YUV

All the methods support RGB, BGR, BGRA and RGBA

Scaling

If need scaling support then pic-scale is available.

SIMD

Runtime dispatch is used for use if available sse4.1, avx2, avx512bw, avx512vbmi, avx512vnni, avxvnni, rdm for NEON.

For the sse4.1, avx2, rdm there is no action needed, however for AVX-512 activated feature nightly_avx512 along nightly rust channel is required.

Always consider to compile with nightly_avx512 when compiling for x86, it adds noticeable gain on supported cpus.

Wasm simd128 as target feature should be enabled for implemented SIMD wasm paths support,

Rayon

Some paths have multi-threading support. However, YUV conversion usually makes more sense in single-threaded mode, thus use this feature with care.

Adding to project

cargo add yuv

RGB to YCbCr

let mut planar_image =
    YuvPlanarImageMut::<u8>::alloc(width as u32, height as u32, YuvChromaSubsampling::Yuv420);
rgb_to_yuv422(
    &mut planar_image,
    &src_bytes,
    rgba_stride as u32,
    YuvRange::Limited,
    YuvStandardMatrix::Bt601,
)
.unwrap();

YCbCr to RGB

yuv420_to_rgb(
    &yuv_planar_image,
    &mut rgba,
    rgba_stride as u32,
    YuvRange::Limited,
    YuvStandardMatrix::Bt601,
)
.unwrap();

HDR Formats

The following planar formats are used for HDR content, with either 10-bit or 12-bit color depth:

Supported:

• I010, I012, P010, P012, P016
  Half width, half height
• I210, I212, P210, P212, P216
  Half width, full height
• I410, I412, P410, P412
  Full width, full height

Format Breakdown:

• I: Represents the color space (as defined above), and these formats use 3 planes: Y (luminance), U (chrominance), and V (chrominance).
• ICgC: Represents the color space (as defined above), and these formats use 3 planes: Y (luminance), Cg (chrominance), and Co (chrominance).
• P: A biplanar format (similar to NV12), but with 16-bit precision, where the valid bits are stored in the high bits. This format has:
  • A Y plane
  • A UV plane (with U and V interleaved)
• GB: Represents the GBR planar format, which uses 3 separate planes for each of the red, green, and blue color components. This format has:
  • G plane (green channel)
  • B plane (blue channel)
  • R plane (red channel)

Subsampling:

• 0: Represents 4:2:0 chroma subsampling.
• 2: Represents 4:2:2 chroma subsampling.
• 4: Represents 4:4:4 chroma subsampling.

Bit Depth:

• 10,12,14,16: Indicates the bits per channel.

Benchmarks

Tests performed on the image 1997x1331

YUV 8 bit-depth conversion

aarch64 tested on Mac Pro M3.

AVX2 tests performed on m4.large with Ubuntu 24.04 droplet. AVX2 Win test performed on Windows 11 Intel Core i9-14900HX.

cargo bench --bench yuv8 --manifest-path ./app/Cargo.toml --features fast_mode,professional_mode,rdm

AVX-512 tests performed on AWS c5.large with Ubuntu 24.04 instance with command

cargo +nightly bench --bench yuv8 --manifest-path ./app/Cargo.toml --features nightly_avx512,fast_mode,professional_mode

Encoding 8-bit

Decoding 8-bit

YUV 16 bit-depth conversion

cargo bench --bench yuv16 --manifest-path ./app/Cargo.toml --features rdm

AVX-512 tests performed on AWS c5.large with Ubuntu 24.04 instance with command

cargo +nightly bench --bench yuv16 --manifest-path ./app/Cargo.toml --features nightly_avx512

Encoding 10-bit

10-bit encoding is not implemented in libyuv

Decoding 10-bit

Geometry and mirroring

cargo bench --bench geometry --manifest-path ./app/Cargo.toml

This project is licensed under either of

• BSD-3-Clause License (see LICENSE)
• Apache License, Version 2.0 (see LICENSE)

at your option.