Architecture 2.0#
Note: The Architecture 2.0 approach is being deprecated in favor of aligning Deep Learning Streamer with GStreamer analytics library.
Deep Learning Streamer Pipeline Framework release 2022.2 introduces new architecture 2.0, aiming for the following goals:
Memory interop on CPU and GPU. Provide independent sub-component (library) for zero-copy buffer sharing and memory mapping between various frameworks and memory handles on CPU and GPU. The zero-copy optimization technique is especially and critically important for devices connected with host via PCI, including dGPU. This sub-component also defines abstract C++ interfaces Source, Transform and Sink for C++ elements implemented in next sub-component. See Memory Interop and C++ abstract interfaces for details.
“non-GStreamer” backend support via C++/Python programming. Expose Deep Learning Streamer functional blocks (elements) as both GStreamer elements and C++/Python interfaces accessible by any C++/Python application (not using GStreamer). These out-of-box elements support typical media analytics operations such as post-processing and visualization of inference results. C++/Python applications (including samples in Direct Programming section) can use mix of out-of-box elements, custom/additional elements, memory interop library and pipeline management logic on top to build complete performance-optimized media analytics pipeline. See C++ elements and Elements list and Samples 2.0 for details.
GStreamer bin elements. Reimplement most gva* elements as GStreamer bin elements which internally build sub-pipeline of other (low-level) GStreamer elements. Most low-level elements are C++ elements (#2 above) automatically wrapped and registered as GStreamer elements. Each low-level element performs only single operation (no multi-stage processing inside element and no intermediate memory allocation) with all pipelining done on either bin-elements or application level. See pages GStreamer Elements and GStreamer Bin Elements for details.
The following diagram visualizes Architecture 2.0 sub-components ①, ② and ③ inside block Deep Learning Streamer Pipeline Framework
Note: Architecture 2.0 currently available in preview mode as transition not completed yet. There is no strict backward compatibility in preview mode (some changes/renaming may occur). It could be used for evaluation and prototyping, it’s not recommended for production usage yet.