You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Preserving hardware memory during hw decoding, exporting/importing via dlpack (#2155)
Add DLPack support for zero-copy GPU frame exchange between PyAV and frameworks like PyTorch/CuPy, avoiding GPU->CPU->GPU round-trips when using cuvid decode + GPU processing + nvenc encode.
- Implement __dlpack__ and __dlpack_device__ on VideoPlane
- Add VideoFrame.from_dlpack for importing DLPack tensors (nv12/p010le/p016le)
- Add CudaContext for managing CUDA device/frames contexts
- Add HWAccel.is_hw_owned to preserve frames on GPU memory
- Switch frame buffer management from av_image_alloc to av_frame_get_buffer
- Handle hw_frames_ctx in planes, reformatter, and to_ndarray paths
0 commit comments