5.2. Release notes - 09_02_00¶
5.2.1. Features¶
Note
New features vs previous release are marked in bold
5.2.1.1. Hardware support¶
J722S HS support
Fusion1 RevC daughter card with UB960 FPDLink deserializer
IMX390 2MP CSI2 image sensor + UB953 FPDLink serializer
AR0233 2.5MP CSI2 image sensor + UB953 FPDLink serializer
AR0820 8MP CSI2 image sensor + UB953 FPDLink serializer
5.2.1.2. Software features¶
Important
Below is a summary of key features. Refer individual SDK Components release notes for more details
5.2.1.2.1. Edge AI TIOVX Apps¶
TI OpenVx based application for demonstrating DL inference with multiple input and output sources
Single input Single inference dataflow
Single input Multi-network inference dataflow
Multi input Multi-network dataflow
Supports all models (Classification, Detection, Segmentation) from Edge AI Model Zoo
Agnostic to DL runtime, runs purely using TIDL OpenVx Node
Optimal dataflows with RTOS based camera and display OpenVx nodes
Optimal dataflows with V4L2 based decoder and encoder accelerators with zero-buffer-copy overheads
ARM NEON based SIMD optimized pre/post processing kernels
Refer to Edge AI TIOVX Apps Reference Guides for further details
5.2.1.2.2. Vision Apps (ADAS, Vision, DL demos)¶
OpenVX based demos for ADAS, Vision, deep learning applications
FreeRTOS on C7x, R5F and Linux/QNX on A53
Integrates all major PSDK RTOS and PSDK LINUX SW components like TIDL, MMALIB, MCU+ SDKs, OpenVX, C7x algorithms, imaging/sensors, IPC, Linux, FreeRTOS, SPL/uboot.
Integrates all major HW components like CSI2 camera, HDMI display, UART, I2C, ethernet, SD card
Deep Learning demos
Image Classification demo
File based
Camera based
Debug support for the ability to debug intermediate layer information from TIDL node
ADAS/Vision demos
Dense optical flow (DOF) demo
Stereo disparity engine demo
Single and multi camera to display (CSI2RX + VISS + LDC + MSC + Display) demo
C7x algorithm offload demo
OpenVX target nodes
image post-processing and visualization nodes for DOF, Stereo demos
Profiling, logging and other utility APIs
All included MCU tasks are running on MCU2_0 (VPAC HWA)
5.2.1.2.3. MCU-PLUS-SDK¶
RTOS device drivers running on R5F, C7x
Refer to MCU-PLUS-SDK release notes LINK for more details.
5.2.1.2.4. TI Deep learning Product (TIDL) and MMALIB¶
TensorFlow Lite Runtime with Delegate API support for heterogeneous execution on ARM+C7x_MMA
ONNX Runtime with Execution Provider support for heterogeneous execution on ARM+C7x_MMA
NEO-AI-DLR support - TVM model compilation for heterogeneous execution on ARM+C7x_MMA
TIDL deep learning inference engine library running on C7x/MMA
Supports Caffe, ONNX, TFlite network exchange formats
Supported layers: Convolution, Deconvolution, Pooling, ReLU, Elt-wise, Inner product, Soft-max, and many more
TIDL network import tool and graph visualizer tool
Post Training Quantization options for 8-bit, 16-bit and mixed precision inference
Support for batch processing
Support for auto C7x code generation with TVM
Refer to edgeai-tidl-tools for further details
5.2.1.2.5. TI OpenVX (TIOVX)¶
OpenVX v1.1 compliant implementation with graph pipelining, user data object, bidirectional parameters and safe casts extensions support
OpenVX API on A53 running Linux/QNX
OpenVX target kernels on C7x, R5F, A53
OpenVX v1.1 C6x optimized kernels recompiled for C7x
OpenVX nodes for
TIDL (C7x-MMA)
VISS (R5F)
LDC (R5F)
MSC (multi-scalar) (R5F)
Pyramid (using MSC) (R5F)
Dense Optical Flow (R5F)
Stereo Disparity Engine (R5F)
CSI2-RX camera (R5F)
CSI2-TX (R5F)
Display (HDMI) (R5F)
OpenVX tutorials for getting started
PyTIOVX tools for target kernel code generation
Performance Analyzer tool for run time logging of OpenVX graphs
VISS node supports heterogeneous cameras
5.2.1.2.6. Imaging¶
IMX390 2MP sensor driver (30/60fps) with IQ tuning
AR0233 2.5MP sensor driver with IQ tuning
AR0820 8MP sensor driver with IQ tuning
TI Auto-exposure and auto white-balance algorithms
Support for UB960 broadcast mode
Sensor framework supports heterogeneous cameras
5.2.1.2.7. TI Autonomous Driving Algorithms (TIADALG)¶
Image pre-processing APIs for color plane conversion and separation YUV to RGB for DL applications (C6x optimized. recompiled for C7x)
Camera Pose estimation API using solve pnp technique (C6x optimized. recompiled for C7x)
Visual localization algorithm (C6x optimized. recompiled for C7x)
Structure from Motion algorithm
5.2.1.2.8. FreeRTOS¶
FreeRTOS for R5F and C7x
5.2.1.2.9. Code Gen Tools (CGT)¶
TI Compiler, assembler, linker for R5F, C7x
TI LLVM (CLANG) Compiler, assembler, linker for R5F
GCC compiler for A53 (needs to be downloaded separately)
5.2.2. Device Support and Validation Information¶
SoC |
Build Host (OS) |
Run Target (OS) |
Test Plaform |
---|---|---|---|
J722S |
x86_64 (Linux Ubuntu 22.04) |
R5F, C7x running FreeRTOS A53 running Linux or QNX |
J722S EVM with daughter cards |
PC emulation mode (1) |
x86_64 (Linux Ubuntu 22.04) |
x86_64 (Linux Ubuntu 22.04) |
x86_64 (Linux Ubuntu 22.04) |
(1) Requires download of addon package from mySecure Software
5.2.3. Upgrade and Compatibility¶
In this section, we only highlight a few significant changes in this SDK. For full list of upgrade and compatibility topics, please refer to individual components release notes SDK Components.
5.2.4. Known Issues¶
Refer to individual SDK Components release notes for known issues in each component