InspireFaceInspireFace1.2.4.d3
Home
Get started
Get and build the SDK
Examples
  • English
  • 简体中文
GitHub
Home
Get started
Get and build the SDK
Examples
  • English
  • 简体中文
GitHub
  • Introduction
  • Get started
  • Features
  • Guides

    • Architecture and lifetime
    • Model packs
    • Image inputs and coordinates
    • Sessions and tracking
    • Face analysis
    • Recognition and FeatureHub
    • Facial landmarks
    • Liveness detection
    • Face capture
    • More API recipes
  • Language and platform

    • C API
    • C++
    • Python
    • Android
    • Apple
    • iOS
    • macOS
    • HarmonyOS
  • Get and build the SDK

    • Overview and downloads
    • Source and common options
    • Linux
    • macOS
    • Android
    • iOS
    • HarmonyOS
    • NVIDIA TensorRT
    • Rockchip NPU
    • Python packaging
  • Hardware deployment

    • ARM
    • NVIDIA TensorRT
    • Rockchip NPU
    • Python on Rockchip
  • InspireCV
  • Complete examples
  • API coverage
  • Performance
  • Image processing benchmarks
  • Troubleshooting

Rockchip NPU builds

Build for the board’s CPU architecture, C library and NPU generation. Prepare the SDK source and the cross toolchain supplied for the board’s Linux root filesystem. These builds run on a Linux host; run the resulting library on the board.

Match the board and toolchain

TargetCompiler prefixModel pack
RV1109 / RV1126arm-linux-gnueabihfGundam_RV1109
RV1103 / RV1106arm-rockchip830-linux-uclibcgnueabihfGundam_RV1106
RK3566 / RK3568aarch64-linux-gnuGundam_RK356X
RK3588aarch64-linux-gnuGundam_RK3588

ARM_CROSS_COMPILE_TOOLCHAIN points to the directory containing bin/. The compiler’s sysroot, glibc/uClibc and C++ runtime must match the board image. A generic ARM64 CPU build does not enable RKNN inference.

RV1109 and RV1126

export ARM_CROSS_COMPILE_TOOLCHAIN=/opt/rv1109-toolchain
"$ARM_CROSS_COMPILE_TOOLCHAIN/bin/arm-linux-gnueabihf-gcc" --version
VERSION=1.2.4 bash command/build_cross_rv1109rv1126_armhf.sh

The script selects ISF_RK_DEVICE_TYPE=RV1109RV1126 with RKNN enabled. The SDK directory is build/inspireface-linux-armv7-rv1109rv1126-armhf-1.2.4/InspireFace.

RV1106 with uClibc

export ARM_CROSS_COMPILE_TOOLCHAIN=/opt/rv1106-toolchain
"$ARM_CROSS_COMPILE_TOOLCHAIN/bin/arm-rockchip830-linux-uclibcgnueabihf-gcc" --version
VERSION=1.2.4 bash command/build_cross_rv1106_armhf_uclibc.sh

This route selects RKNPU2, ISF_RK_DEVICE_TYPE=RV1106, ISF_RK_COMPILER_TYPE=armhf-uclibc and RGA. The script also downloads its MNN 2.3.0 source into .rknpu2_cache on the first build. The SDK is under build/inspireface-linux-armv7-rv1106-armhf-uclibc-1.2.4/InspireFace.

Use the board’s uClibc toolchain for the application too. A glibc executable or Python wheel cannot be made compatible by renaming its file.

RK356x and RK3588

export ARM_CROSS_COMPILE_TOOLCHAIN=/opt/rk-aarch64-toolchain
"$ARM_CROSS_COMPILE_TOOLCHAIN/bin/aarch64-linux-gnu-gcc" --version
VERSION=1.2.4 bash command/build_cross_rk356x_rk3588_aarch64.sh

RK356x and RK3588 share this ARM64 RKNPU2 build. The script uses ISF_RK_DEVICE_TYPE=RK356X, ISF_RK_COMPILER_TYPE=aarch64 and enables RGA. Select the model pack for the actual SoC. The SDK is under build/inspireface-linux-aarch64-rk356x-rk3588-1.2.4/InspireFace.

Inspect the Linux output

Each Linux script builds a shared library with samples, tests and benchmarks disabled, then keeps the installed SDK and removes intermediate compilation files. Inside the SDK directory, use include/ with lib/libInspireFace.so from the same build.

file build/inspireface-linux-aarch64-rk356x-rk3588-1.2.4/InspireFace/lib/libInspireFace.so
"$ARM_CROSS_COMPILE_TOOLCHAIN/bin/aarch64-linux-gnu-readelf" -d \
  build/inspireface-linux-aarch64-rk356x-rk3588-1.2.4/InspireFace/lib/libInspireFace.so

This inspection example uses the RK356x/RK3588 build and toolchain. Check the listed shared-library dependencies against the board’s filesystem, including RKNN/RGA libraries and their driver pairing. Run ldd on the target board if its environment provides it; the host’s loader cannot validate a foreign-architecture library by loading it.

Android on RK356x and RK3588

For a Rockchip Android device, use the NDK script instead of the Linux toolchain. Set ANDROID_NDK to an installed NDK directory:

export ANDROID_NDK=/path/to/android-ndk
VERSION=1.2.4 bash command/build_android_rk356x_rk3588.sh

The script builds arm64-v8a and armeabi-v7a, both with Android API 24, c++_static, RKNN enabled and RGA disabled. The reorganized output contains:

build/inspireface-android-rk356x-rk3588-1.2.4/
  lib/arm64-v8a/libInspireFace.so
  lib/armeabi-v7a/libInspireFace.so
  version.txt

Use the Android packaging instructions to pair Java classes and JNI libraries. Include any required vendor runtime libraries for each ABI, and validate on the Rockchip device with its matching NPU model and driver. This script’s final layout does not retain the installed headers; use the same source revision’s public headers when building your own JNI adapter.

Run an application or package Python

The Rockchip deployment guide includes an inline C example, application cross-compilation, device file layout and RGA selection. For Python, first follow Python on Rockchip to test the wrapper with the board library, then use Python packaging to distribute that library in a wheel. Models and board runtime libraries remain separate deployment inputs.

Edit this page
Last Updated:: 9/28/26, 3:20 PM
Contributors: Jingyu
Prev
NVIDIA TensorRT
Next
Python packaging