EdgeFirst Replay v2.3.0

EdgeFirst Replay v2.3.0

Installation

Download Binaries

Platform	Binary
Linux x86_64	edgefirst-replay-2.3.0-linux-x86_64
Linux ARM64	edgefirst-replay-2.3.0-linux-aarch64
Configuration	replay.default

# Download and install (ARM64 example)
wget https://github.com/EdgeFirstAI/replay/releases/download/v2.3.0/edgefirst-replay-2.3.0-linux-aarch64
chmod +x edgefirst-replay-2.3.0-linux-aarch64
sudo mv edgefirst-replay-2.3.0-linux-aarch64 /usr/local/bin/edgefirst-replay

# Install default configuration (optional)
sudo cp replay.default /etc/default/replay

What's Changed

Changed

• Breaking: rt/camera/dma now publishes the decoder-native NV12 buffer,
  matching the live camera wire contract. Replay no longer performs NV12→RGBA
  conversion before publishing this topic. Downstream consumers that were
  reading RGBA from rt/camera/dma should switch to --camera-image-topic
  (sensor_msgs/Image, rgba8).
• Bumped edgefirst-schemas 1.3.1 → 3.4.0 and adopted its zero-copy
  view/builder API (FoxgloveCompressedVideo::from_cdr,
  CompressedImage::from_cdr, Image::builder().encode_into_vec(..),
  DmaBuffer::new(..).into_cdr()).
• Upgraded videostream to 2.5.2. Switched the h264 decoder construction
  from the legacy Decoder::create(H264, fps) to
  Decoder::create_ex(H264, fps, CodecBackend::Auto); the legacy entrypoint
  bypasses the v4l2 device enumeration the 2.5.x backend relies on, while
  the _ex entry uses it. On imx8mp this routes to /dev/video1 vsi_v4l2dec
  with a tighter natural NV12 stride (1920 bytes vs the Hantro path's 2880).
• decode_h264_msg retries on transient decoder backpressure (V4L2 m2m OUTPUT
  queue full): up to 20 attempts, 5 ms sleep between, retrying the same data.
  Mirrors ~/Software/EdgeFirst/camera/src/replay.rs. Without this, the very
  first decode_frame call on a cold pool returns Io("Decoder Error") from
  the C library and is otherwise treated as fatal.
• main is now sync (no #[tokio::main], no tokio::time::sleep, no
  tokio::process::Command). The edgefirst-hal 0.23.1 GL backend calls
  tokio::sync::oneshot::Receiver::blocking_recv during converter init,
  which panics inside a tokio runtime; making main sync sidesteps the
  collision. Service-handler invocations use std::process::Command and
  the inter-frame timing loop uses std::thread::sleep.

Removed

• NXP G2D (g2d-sys), turbojpeg, dma-heap, tokio, and the per-frame
  cdr crate dependency. All image conversion, JPEG decoding, and CMA
  allocation now go through edgefirst-hal 0.23.1.
• src/image.rs (G2D surface helpers, CMA-heap Image, MappedImage).

Added

• --camera-image-topic (env CAMERA_IMAGE_TOPIC) and --camera-image-buffers
  (env CAMERA_IMAGE_BUFFERS, default 4) — opt-in sensor_msgs/Image RGBA
  publisher for consumers that need a non-dma-buf-aware decoded image stream.
  Backed by edgefirst-hal ImageProcessor (auto-selects G2D / OpenGL / CPU).
  Source tensors cached by dma-buf inode; destination tensors pre-allocated
  once and reused.
• JPEG MCAP playback now uses the optimized edgefirst-codec JPEG decoder
  (peek_info + ImageLoad::load_image), writing directly into pre-allocated
  NV12 dma-buf tensors — no host-side intermediate, no memcpy.

---

Full Documentation: https://github.com/EdgeFirstAI/replay#readme