baracuda-nvjpeg

Safe Rust wrappers for NVIDIA nvJPEG — GPU-accelerated JPEG decode and encode. Comprehensive coverage of single + batched paths and the modern three-phase hybrid pipeline.

Coverage

• Single-image decode: nvjpegDecode with planar / interleaved output, JpegStream parser, DecodeParams (output format, ROI, CMYK handling).

• Batched decode: simple batched + the three-phase hybrid pipeline:

  1. Host parse (CPU-side bitstream parsing).
  2. Host transfer (D2H of phase-1 outputs + H2D of decode tables).
  3. Device decode (the actual GPU kernels).

  Decoupling the phases lets the caller overlap CPU parsing of batch N+1 with GPU decode of batch N.

• Buffer pools: BufferPinned for host pinned memory, BufferDevice for GPU memory; both poolable.

• DecodeParams: pick output format (RGB, BGR, RGBI, BGRI, YUV420, YUV422, YUV444), ROI cropping, CMYK input handling.

• Encoder: quality, chroma subsampling (Css420 / Css422 / Css444), optimized Huffman tables, encode from packed image or planar YUV, retrieve bitstream.

use baracuda_nvjpeg::{Handle, JpegState};
use baracuda_driver::{Context, Device};

# fn demo() -> Result<(), Box<dyn std::error::Error>> {
let ctx = Context::new(&Device::get(0)?)?;
let handle = Handle::new(&ctx)?;
let mut state = JpegState::new(&handle)?;

let bitstream: Vec<u8> = std::fs::read("photo.jpg")?;
let info = handle.get_image_info(&bitstream)?;
// ... decode + retrieve into device buffers ...
# Ok(()) }

Pairs with baracuda-nvjpeg-sys for the raw FFI surface.

Part of the baracuda workspace.

License

Dual MIT / Apache-2.0.