Supported GStreamer Plugins
vvas_xinfer - ML inference (classification, detection, segmentation) with ONNXRT and VART backends
vvas_xabrscaler - HW-accelerated multi-output scaler with mean/scale normalization
vvas_xoverlay - Draw bounding boxes, labels, and shapes on video
vvas_xmetaconvert - Convert inference metadata to overlay format
vvas_xmetaaffixer - Transfer metadata between streams of different resolutions
vvas_xfilter - Generic filter for custom processing kernels
vvas_xmulticrop - Crop multiple regions of interest
vvas_xfunnel - Merge multiple input branches into one synchronized output stream
vvas_xdefunnel - Split one input stream into multiple output branches
Inference Backends
ONNX Runtime with Vitis-AI Execution Provider (for subgraphs containing operators supported by NPU); remaining subgraphs execute on CPU Execution Provider
VART (AMD Vitis™ AI Runtime) with zero-copy support
Supported Data Formats and Configurations
Expanded supported data types and layouts for vvas_xinfer and vvas_xabrscaler plugins. Refer to Custom AI Tensor Formats for the full list of supported formats.
Model formats: combinations of data types INT8, BF16, FP16 and layouts NCHW, NHWC are now supported in HW mode. This eliminates buffer copying and SW/CPU overhead when the model is compiled to accept these data types and layouts.
Mixed-precision inference support.
Support for relocatable models with data parallelism (dp size) and tensor parallelism (tp size) configurations.
Inference Pipeline Features
DMA buffer (DMA-BUF) input support in vvas_xinfer for zero-copy data sharing between hardware accelerators and the inference engine, reducing CPU usage and memory copies.
Auto-detection of quantization parameters from the model.
Batch mode support (VART backend).
Multi-level (cascaded) inference.
Temporal and spatial sharing of AIE with configurable column allocation.
Unified VART-based post-processing library.
ARM NEON-optimized YOLO post-processing (sample) for improved inference throughput.
Post-processing support for ResNet50, ResNet18, YOLOv5n, and YOLOX-M models.
Benchmarking and time measurement support in vvas_xinfer.
Updated compatibility with ONNX Runtime 1.24.3.
Build and Packaging
VVAS is now bundled in the VAI 6.2 Yocto rootfs - no separate installation required.
Post-processing source code for ResNet50, ResNet18 and YOLO family of models included in the VVAS source package for user reference and customization.