VPI - Vision Programming Interface

4.0 Release

Release Notes v4.0.5

VPI-4.0.5

This VPI version supports

  • Jetson AGX Thor DevKit
  • Jetson AGX Thor T5000 and T4000
  • IGX Thor DevKit + Blackwell RTX6000 dGPU
  • IGX Thor T7000 and T5000
  • IGX Thor Safety DevKit

Linux x86_64 with NVIDIA dGPUs

  • CUDA 12: Maxwell or newer (sm_52 or newer)
  • CUDA 13: Volta or newer (sm_70 or newer)
  • tested with Ubuntu 24.04

Key Highlights

Major Features

  • Multistream Stereo Sample - High-performance concurrent stream processing
  • VPI_DEBUG Environment Variable - Runtime debugging without recompilation

Critical Bug Fixes

  • ORB PVA Output Array Size - Fixed values in array size reporting
  • Sample 18 Python PVA - Fixed visualization for PVA backend
  • FAST on Hopper/Blackwell dGPUs - Fixed nondeterministic negative scores
  • OptFlow Dense (v4.0.1) - Fixed pyramid size rounding error
  • Sample 19 DCF Tracker (v4.0.3) - Fixed target tracking list management

Performance Improvements

  • Gaussian Pyramid PVA - Multiple optimizations (unroll, vectorization, agen fixes)
  • Multistream Stereo - Optimized sync logic for 4ms per frame

Compatibility

  • API: Fully compatible with VPI 4.0.0
  • ABI: Maintained across all 4.0.x releases
  • Platform: All VPI 4.0.x supported platforms
  • Requirements: BSP 38.2.1-PRC2+, CUDA 13.0

VPI_DEBUG Environment Variable

# Enable exception backtraces
VPI_DEBUG=exception ./your_application
# Shows thread names with VPIStream handles
# Backtraces at exception throw points

Known Issues Resolved

  • Dense Optical Flow: Pyramid levels > 3 is now working.
  • Sample 18: ORB Python PVA backend now produces FAST keypoints
  • FAST/ORB algo on Hopper/Blackwell dGPUs works
  • Sample 19: DCF tracker works for all nvpmodel settings
  • Fisheye sample work with default opencv version

Known Issues

  • StereoDisparity: The following block linear formats are not working: VPI_IMAGE_FORMAT_NV12_ER_BL, VPI_IMAGE_FORMAT_NV12_ER_BL16, VPI_IMAGE_FORMAT_NV12_BL, VPI_IMAGE_FORMAT_NV12_BL16. Workaround is to convert those image formats.
  • Dense Optical Flow: There is no discernible difference in quality from low, medium to high.
  • Background Subtractor: Initial frames may produce unstable background model.
  • PyTorch: 16-vpi_pytorch sample fails due to the absence of a compatible PyTorch module for CUDA Toolkit 13.0.
  • ORB: pva backend does only produce FAST keypoints.
  • Jetson Thor - VIC, Interpolation Limitation: VPI_INTERP_NEAREST not available.