popstas / talks-reducer / 31034849737

79%

Build Type

push

github

Committed by popstas

Commit Message

feat(ffmpeg): Use VideoToolbox for HEVC

macOS builds only ever looked for CUDA/NVENC, so check_cuda_available() always returned False and every export ran on the CPU encoders. Add check_videotoolbox_available() and route the HEVC plan through hevc_videotoolbox with a libx265 fallback.

Measured on a 29-minute 1080p30 recording through the full pipeline: HEVC drops from 387s to 108s at a comparable output size (58.8MB -> 61.3MB). H.264 deliberately stays on libx264 -- Apple's media engine caps near 290 fps at 1080p while libx264 -preset veryfast reaches roughly twice that across the CPU cores, so at matched size the hardware encoder finished in 101s against 60s. AV1 keeps the software encoder because Apple ships none.

-q:v 68 was calibrated inside the pipeline rather than on a bare clip, since the -g 900 -force_key_frames arguments shift the rate control substantially.

encoder_supports_option() probes and caches per-encoder help so -spatial_aq is only emitted where it exists: it arrived in FFmpeg 7.1 and the bundled static-ffmpeg is 7.0, where an unknown option rejects the whole command line.

ProcessingResult.used_cuda becomes used_gpu and gains gpu_backend, and both reset when a hardware encode falls back to the CPU so the summary stops crediting the GPU for CPU work.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

Coverage Stats

57 of 59 new or added lines in 4 files covered. (96.61%)

83 existing lines in 3 files now uncovered.

12276 of 15518 relevant lines covered (79.11%)

0.79 hits per line