Skip to main content

CLIP

This document describes how to run CLIP on the NPU.

info

Refer to Model Zoo Download for the example.

The CLIP example converts the image model and the text model separately, then computes image-text similarity on the board.

CLIP example directory structure:

$ tree ./
./
├── CMakeLists.txt
├── images_convert_model
├── text_convert_model
│   └── python
├── model
│   ├── clip-images_uint8_a733.nb
│   ├── clip-text_int16_a733.nb
│   ├── demo.png
│   ├── demo.txt
│   └── merges.txt
├── main.cpp
└── README.md

Model Conversion​

Enter the container development environment first. See Create and Start Container in the Model Zoo download page.

info

Select the Docker image that matches the NPU:

  • A733: ubuntu-npu:v2.0.10.2
  • T527: ubuntu-npu:v1.8.13

Download the floating-point ONNX models from the Allwinner netdisk:

Put clip-images.onnx in images_convert_model and clip-text.onnx in text_convert_model.

X86 Linux PC
docker exec -it model-zoo /bin/bash

Convert clip-images​

X86 Linux PC
cd /workspace/examples/clip/images_convert_model/
./convert_model_env.sh
./pegasus_import.sh clip-images
./pegasus_quantize.sh clip-images uint8 10
X86 Linux PC
./pegasus_export_ovx_nbg.sh clip-images uint8 a733

Convert clip-text​

X86 Linux PC
cd /workspace/examples/clip/text_convert_model/
./convert_model_env.sh
./pegasus_import.sh clip-text
./pegasus_quantize.sh clip-text int16 10
X86 Linux PC
./pegasus_export_ovx_nbg.sh clip-text int16 a733

The exported models are stored in the ../model directory.

Build the Example​

Then compile the example. Exit the container first, then run the commands below.

Configure the third-party library and the cross-compilation toolchain first.

info

Skip this step if you have already configured them in another example.

X86 Linux PC
cd ../../../3rdparty/opencv/
unzip opencv-4.9.0-aarch64-linux-sunxi-glibc.zip
cd ../../0-toolchains/

Download the toolchain from this link, put it in 0-toolchains/, then run:

X86 Linux PC
tar -xvf gcc-arm-10.2-2020.11-x86_64-aarch64-none-linux-gnu.tar.xz
cd ../examples/clip/
X86 Linux PC
../build_linux.sh -t a733 -s debian11

Model Deployment​

After compilation, the example will be installed in the install directory. You can use scp to transfer it to the board.

Configure NPU Driver​

info

You can skip this step if you have already configured NPU driver in other examples.

Transfer the driver library to the board's lib directory via scp.

  • A733 corresponds to the common/npuruntime/lib_linux_aarch64/A733 directory
  • T527 corresponds to the common/npuruntime/lib_linux_aarch64/T527 directory

Then execute the following command to export to environment variables.

Radxa SBC
echo 'export LD_LIBRARY_PATH=$HOME/lib:$LD_LIBRARY_PATH' >> ~/.bashrc

Run Example​

After configuring the driver, you can run the example.

tip

For T527 platform, you need to first enable NPU by referring to the A5E's "Enable NPU on Board" documentation, then use the following command to grant the current user permission to use /dev/vipcore.

Radxa SBC
sudo chmod 777 /dev/vipcore
Radxa SBC
cd clip_demo_linux_a733/
Radxa SBC
chmod +x ./clip_demo_a733
./clip_demo_a733 -ib model/clip-images_uint8_a733.nb -tb model/clip-text_int16_a733.nb -i model/demo.png -t model/demo.txt

The running result is as follows:

$ ./clip_demo_a733 -ib model/clip-images_uint8_a733.nb -tb model/clip-text_int16_a733.nb -i model/demo.png -t model/demo.txt
text_model_file=model/clip-text_int16_a733.nb, images_model_file=model/clip-images_uint8_a733.nb, input_images_file=model/demo.png, input_text_file=model/demo.txt, loop_count=1, malloc_mbyte=10
VIPLite driver software version 2.0.3.2-AW-2024-08-30
nbg name=model/clip-images_uint8_a733.nb, size: 63747488.
create network 0: 34340 us.
nbg name=model/clip-text_int16_a733.nb, size: 115039816.
create network 1: 60155 us.
prepare network: 2633 us.
prepare network: 1587 us.
images: model/demo.png
text  : a photo of a motorcycle
score : 0.998
clip-text total infer time: 35368 us.
clip-images total infer time: 13695 us.
destroy npu finished.
~NpuUint.

This performance data only calculates the time consumption of model inference. Unless otherwise specified, it does not include the time consumption of pre-processing and post-processing.

SoCNPUModelInput ResolutionNetwork Creation TimeNetwork Preparation TimeSingle Frame Inference TimePost-processing TimeTotal TimeFrame Rate
Allwinner A733Vivante VIP9000clip-images224×22434.3 ms2.6 ms13.7 ms50.6 ms73.0 FPS
Allwinner A733Vivante VIP9000clip-text1×2060.2 ms1.6 ms35.4 ms97.2 ms28.3 FPS

    You need to be logged into GitHub to post a comment. If you are already logged in, please ignore this message.

    Radxa-docs © 2026 by Radxa Computer (Shenzhen) Co.,Ltd. is licensed under CC BY 4.0