The project is still in active development; expect breaking changes.

Comparison to Video-Processing-Framework

Methodologies

VPF is written fully in C++ and uses pybind to expose Python interfaces. PNC is written fully in Python and uses ctypes to access Nvidia C interfaces. Our codes tends to be more concise, less duplicative and easier to read and write. It also allows us better interpolation with other python libraries.

Performance

Preliminary tests shows little to no difference in terms of performance, because the heavy lifting is done on the GPU anyway. Both library can saturate GPU decoder. PNC uses more CPU than VPF as expected from Python vs. C++, but still negligible (less than 10% of Ryzen 3100 single core for 8K*4K HEVC)

Resource Management

• In VPF Surface given to user are not owned by the user. It will be overwritten by new frames which is counter-intuitive; Picture are not exposed to user at all - they are always mapped (post-processed and copied) to Surface so the picture can be ready for new frames. The latter is inefficient when only a subset of Pictures are needed (e.g. screenshots).
• The above is because VPF allocates the bare minimum of resources needed for most decoding tasks. PNC allows the user to specify the amount of resources to be allocated for advanced applications. Users own the resources and decide when and whether to deal with them.
• Managing resources is not painful: similar to pycuda, we shift the burden of managing host/device resources to the Python garbage collector. Resources (such as Picture and Surface) are automatically freed when the user drops the reference.

Please open issues if you have ideas for improvement.

• [X] Decoding
• Color Conversion
  • Source Format:
    • [X] NV12
    • [X] P016
    • [X] YUV444
    • [X] YUV444_16Bit
  • Target Format:
    • [X] RGB24
    • [X] RGB48
    • [X] RGB444P
    • [X] RGB444P16
  • Color Ranges
    • [X] MPEG (limited range)
    • [X] JPEG (full range)
  • Color Space
    • [X] BT. 601
    • [X] BT. 709
    • [X] BT. 2020
    • [ ] …You tell me
• [X] support for built-in cropping and scaling provided by NV decoder
• [X] make the library thread-safe and thread-friendly
• [ ] Encoder
• [X] interpolation with deep learning frameworks (PyTorch)
  • [ ] interpolation between different CUDA stream wrappers across framesworks

• Many thanks to @rarzumanyan for all the helps and explanations!
• The blog posts (https://mymusing.co/tag/color/) from myMusing is very helpful.