What is rocPyDecode?

AMD GPUs contain one or more media engines (VCNs) that provide fully accelerated, hardware-based video decoding. Hardware decoders consume lower power than CPU-based decoders. Dedicated hardware decoders offload decoding tasks from the CPU, boosting overall decoding throughput. With proper power management, decoding on hardware decoders can lower the overall system power consumption and improve decoding performance.

Using the rocPyDecode Python API, you can decode compressed video streams while keeping the resulting YUV frames in video memory. With decoded frames in video memory, user can run video post-processing using ROCm HIP, thereby avoiding unnecessary data copies via PCIe bus. You can post-process video frames using scaling or color conversion and augmentation kernels (on a GPU or host) in a format for GPU/CPU-accelerated inferencing and training.

rocPyDecode python library uses the rocDecode C++ SDK library in the backend to do single or multiple instance video decoding based on the number of available VCNs in AMD GPU device. In addition, rocPyDecode can seemlessly interop with other machine learning frameworks like PyTorch to facilitate machine learning application on decoded surfaces. It also allows to use FFMpeg demultiplexer to seek and demux packetized media file.

rocDecode Python Binding, rocPyDecode, is a tool that allows users to access rocDecode APIs in Python. It is currently distributed as python wheel packages which can simply be installed using pip. It works by connecting Python and C/C++ libraries, enabling function calling and data passing between the two languages. The rocpydecode.so library is a wrapper that facilitates the use of rocDecode APIs that are written primarily in C/C++ language within Python.