XpressReal T3 NPU SDK

This SDK packages headers, prebuilt Vivante VIP9000 runtime libraries, and a small demo app to exercise the XpressReal T3 single-board computer's NPU.

Repository layout

• acuity-root-dir/: Vivante OpenVX headers plus prebuilt .so artifacts for arm64 targets (lib/arm64/1619b).
• yolov4tiny/: Minimal YOLOv4-tiny demo showing model loading, pre/post-processing, and inference against a sample image.

Prerequisites

• arm64 build environment with make and gcc/g++ (apt install build-essential)
• OpenCV 4 development packages (install with apt install libopencv-dev)

Build and run the YOLOv4-tiny demo

# Change NPU device node permission for user access
sudo chmod 0666 /dev/galcore

cd yolov4tiny/
make -f makefile.linux

# Run the sample against the provided dog image
export LD_LIBRARY_PATH=../acuity-root-dir/lib/arm64/1619b
./bin_r/yolov4tinyuint8 ./model/yolov4tiny.nb ./input/dog.jpg

The build pulls headers from acuity-root-dir/include and links against the prebuilt Vivante/OpenVX libraries in acuity-root-dir/lib/arm64/1619b. Adjust LD_LIBRARY_PATH if you relocate the libraries.