[lowlevelmetal]

A ground-up rewrite. The goal is the same (stream media as an NDI source from Linux), but the architecture, timing model and feature set are new. See CHANGELOG.md for the full list.

Summary

• One pipeline: demux → decode → process → present, connected by bounded queues and a shared presentation clock. This replaces the three duplicated software/VAAPI/CUDA apps.
• Accurate A/V timing: frames are sent at their timestamps, audio is the master clock while it plays, and NDI timecodes come from media timestamps.
• Hardware decoding: -t auto|vaapi|cuda|… accepts any FFmpeg device type. A stream the device can't decode falls back to software, and GPU frames are downloaded straight into NDI's buffer layout.
• Lower CPU: the default -p auto passes through layouts NDI accepts. At 1080p60 H.264 with the real NDI 6.3 runtime, sender CPU is 33% of a core with software decode and 19% with VAAPI, versus 65% when forcing UYVY. All 720/720 frames were delivered with 0 dropped.
• Graceful overload: late video is dropped, then late frames skip processing, then only keyframes are decoded. Audio is never starved. Live sources tolerate clock drift and mid-GOP joins.
• New features: seamless --loop, --vf filters, any FFmpeg input or capture device, multichannel float audio, clean SIGINT/SIGTERM handling, --stats.
• No NDI SDK needed to build: the runtime is loaded at start-up. The old /opt/ndi SDK path is still searched.
• Compatible CLI: -i, -s and -t software|vaapi|cuda still work.

Fixes compared to 1.x

• No longer crashes on files without audio.
• The NDI frame rate is no longer 0/0.
• The last frames of a file are no longer lost.
• Builds with FFmpeg 8+ (avcodec_close was removed upstream).

Testing

• 85 tests pass locally with GCC 16 and FFmpeg 9.0. The 83 that don't need the NDI SDK also pass under ASan/UBSan, and ThreadSanitizer is clean for project code.
• The whole-pipeline tests stream generated clips into a fake libndi.so. They check frame counts, real-time pacing, A/V alignment, loop continuity, pixel layouts, async frame lifetime, and overload handling.
• With the NDI SDK 6.3.2:
  • ndi_abi checks the project's NDI declarations against the official headers with static_asserts.
  • ndi_end_to_end streams to a real NDI receiver through a private discovery server.
  • Both pass.
• VAAPI decoding was checked on an AMD GPU, with both software and hardware filters.
• An independent code review found 11 issues. All were fixed and checked against the reviewer's reproductions, and regression tests were added.
• This PR's CI covers Ubuntu 24.04 (FFmpeg 6.1) with GCC, Clang, and ASan/UBSan.

🤖 Generated with Claude Code