Build an FFmpeg hardware-accelerated transcoding pipeline choosing between NVENC, VAAPI, and VideoToolbox based on the host platform

Steps

1. Check available hardware acceleration methods with ffmpeg -hwaccels and available hardware encoders with ffmpeg -encoders, filtering for nvenc, vaapi, or videotoolbox
2. On NVIDIA hosts, decode with -hwaccel cuda -hwaccel_output_format cuda and encode with h264_nvenc or hevc_nvenc, using scale_cuda/scale_npp for any resizing so frames stay on the GPU instead of round-tripping to system memory
3. On Linux with VAAPI, set -hwaccel vaapi -vaapi_device /dev/dri/renderD128 -hwaccel_output_format vaapi on decode and encode with h264_vaapi or hevc_vaapi, using scale_vaapi for in-pipeline resizing
4. On macOS, decode with -hwaccel videotoolbox where supported and encode with h264_videotoolbox or hevc_videotoolbox, keeping in mind rate control is coarser than software x264 (no true multi-pass CRF)
5. Avoid mixing hardware-surface frames with CPU-only filters (drawtext, overlay, etc.) unless you explicitly hwdownload/hwupload around them, since that round-trip can dominate pipeline cost
6. Benchmark hardware encoder output quality-per-bit against a software x264/x265 reference at the same bitrate before committing, since fixed-function encoders typically trade some quality for speed