mirror of
https://github.com/shumaiOne/shumai.git
synced 2026-11-22 11:27:14 +00:00
* feat(transcode): decode and scale on the GPU for VA-API transcodes When the selected encoder is h264_vaapi, decode with -hwaccel vaapi and scale with scale_vaapi so frames stay in GPU memory, instead of decoding and scaling on the CPU and hwupload-ing before the encoder. On a Haswell iGPU this makes 4K proxy/HLS renditions 3-5x faster. Fallback order: GPU decode + VA-API encode -> software decode + VA-API encode (previous behavior) -> libx264. Watermark overlays, HDR sources and rotated sources skip GPU decode: ffmpeg cannot auto-insert its rotation filter on hardware frames, so rotated input would otherwise come out sideways instead of failing. SHUMAI_VAAPI_HW_DECODE=false disables it. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * docs(transcode): document VA-API GPU decoding and SHUMAI_VAAPI_HW_DECODE Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * refactor(transcode): generic hardware decode config, stall watchdog, reuse probed rotation Follow-ups from comparing with Immich and Jellyfin: - Replace the VA-API-specific helpers with a per-encoder HW_DECODE_CONFIGS table (input args, output args, GPU scale filter) and build the MP4 and HLS input/filter graph in one shared method so the paths cannot drift. Only h264_vaapi has an entry; other backends can be added once tested on real hardware. The opt-out is now SHUMAI_HW_DECODE=false. - Kill a GPU-decode attempt with SIGKILL when its output stops growing for SHUMAI_HW_DECODE_STALL_TIMEOUT (default 120s) and fall back, since some VA-API driver failures hang ffmpeg instead of exiting (immich#31933). - Add -noautoscale on the GPU-decode path so a mid-stream decoder reinit (e.g. 1080 -> 1088) cannot auto-insert a software scaler into the GPU graph (jellyfin#17133). - Pass the rotation already probed into MediaMetadata through the workflow and activities as sourceRotation; ffprobe only runs when it is missing. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>