3. NPU

RK3568 has a NPU(Neural Process Unit) that Neural network acceleration engine with processing performance up to 1 TOPS. Using this NPU module needs to download RKNN SDK which provides programming interfaces for RK3566/RK3568 chip platforms with NPU. This SDK can help users deploy RKNN models exported by RKNN-Toolkit2 and accelerate the implementation of AI applications

The contents of what you will get:

└── RK_NPU_SDK_1.3.0
    ├── rknn-toolkit2-1.3.0        // toolkit2, for X86_64 PC
    │   ├── ...
    │   └── rknn_toolkit_lite2     // lite version of toolkit2, for arm64 platform
    ├── rknpu2_1.3.0.tar.gz        // this is the RKNN SDK, remember to unzip it
    └── Rockchip_Quick_Start_RKNN_SDK_V1.3.0_CN.pdf

3.1. RKNN Model

RKNN is the model type used by the Rockchip NPU platform. It is a model file ending with the suffix .rknn . RKNN SDK provides a complete model transformation Python tool for users to convert their self-developed algorithm model into RKNN model

The RKNN model can run directly on the RK3568 platform. There are demos under rknpu2_1.3.0/examples. Refer to the README.md to compile Android or Linux Demo (Need cross-compile environment). You can also just download compiled demo.

  • First prepare the runtime environment for ITX-3568Q :

Android

adb root && adb remount
adb push rknpu2_1.3.0/runtime/RK356X/Android/librknn_api/arm64-v8a/* /vendor/lib64
adb push rknpu2_1.3.0/runtime/RK356X/Android/librknn_api/arm64-v8a/* /vendor/lib

Linux

adb push rknpu2_1.3.0/runtime/RK356X/Linux/librknn_api/aarch64/* /usr/lib

  • Push demo to ITX-3568Q and run:

:/ # cd /data/rknn_ssd_demo_Android/    (use rknn_ssd_demo_Linux in Linux System)
:/data/rknn_ssd_demo_Android # chmod 777 rknn_ssd_demo
:/data/rknn_ssd_demo_Android # export LD_LIBRARY_PATH=./lib
:/data/rknn_ssd_demo_Android # ./rknn_ssd_demo model/RK356X/ssd_inception_v2.rknn model/road.bmp (In linux it's bus.jpg)
Loading model ...
rknn_init ...
model input num: 1, output num: 2
input tensors:
  index=0, name=Preprocessor/sub:0, n_dims=4, dims=[1, 300, 300, 3], n_elems=270000, size=270000, fmt=NHWC, type=UINT8, qnt_type=AFFINE, zp=0, scale=0.007812
output tensors:
  index=0, name=concat:0, n_dims=4, dims=[1, 1917, 1, 4], n_elems=7668, size=30672, fmt=NHWC, type=FP32, qnt_type=AFFINE, zp=53, scale=0.089455
  index=1, name=concat_1:0, n_dims=4, dims=[1, 1917, 91, 1], n_elems=174447, size=697788, fmt=NHWC, type=FP32, qnt_type=AFFINE, zp=53, scale=0.143593
rknn_run
loadLabelName
ssd - loadLabelName ./model/coco_labels_list.txt
loadBoxPriors
person @ (13 125 59 212) 0.984696
person @ (110 119 152 197) 0.969119
bicycle @ (171 165 278 234) 0.969119
person @ (206 113 256 216) 0.964519
car @ (146 133 216 170) 0.959264
person @ (49 133 58 156) 0.606060
person @ (83 134 92 158) 0.606060
person @ (96 135 106 162) 0.464163

3.2. RKNN-Toolkit-lite2

The demo mentioned above is using C/C++ program to deploy models. This requires developers to be familiar with RKNN API to customize.

Toolkit-lite2 with python can simplify the deployment and operation of the model, making it easy for developers to get started quickly, and it is highly recommended.

Toolkit-lite2 can only be used to deploy models on arm64 platform, NOT to transfer models. For Non-RKNN models, use the PC tools mentioned later

3.2.1. Environment Dependence

  • System dependency: Toolkit-lite2 needs to work on Debian 10/11 (aarch64)，This tool can only be installed on RK3568 Debian, Ubuntu is not supported yet

  • Python version: 3.7/3.9

  • Python library:

numpy
ruamel.yaml
psutils

3.2.2. Toolkit-lite2 installation

# 1）Install Python3.7/3.9 and pip3
sudo apt-get install python3 python3-dev python3-pip gcc
# 2）Install dependent libraries
sudo apt-get install -y python3-opencv
sudo apt-get install -y python3-numpy
# PS: Toolkit-lite2 itself does not rely on opencv-python, but demo does. If you don't need image processing, you can choose not to install it.
# 3）Install Toolkit-lite2
# Debian10 ARM64 with python3.7
pip3 install rknn_toolkit_lite2-1.2.0-cp37-cp37m-linux_aarch64.whl
# Debian11 ARM64 with python3.9
pip3 install rknn_toolkit_lite2-1.2.0-cp39-cp39m-linux_aarch64.whl

3.2.3. Running Demo

First prepare the runtime environment, put libraries in RK3568 platform. Methods mentioned above, not repeat here.

Place the demo inference_with_lite which under rknn-toolkit2-1.3.0/rknn_toolkit_lite2/examples directory into ITX-3568Q, then run it:

root@firefly:~# cd inference_with_lite/
root@firefly:~/inference_with_lite# python3 test.py
--> Load RKNN model
done
--> Init runtime environment
I RKNN: [03:46:35.193] RKNN Driver Information: version: 0.4.2
I RKNN: [03:46:35.193] RKNN Runtime Information: librknnrt version: 1.2.0 (9db21b35d@2022-01-14T15:16:23)
I RKNN: [03:46:35.194] RKNN Model Information: version: 1, toolkit version: 1.2.0(compiler version: 1.1.2b17 (2d31041c6@2022-01-10T17:56:44)), target: RKNPU lite, target platform: rk3566, framework name: PyTorch, framework layout: NCHW
done
--> Running model
resnet18
-----TOP 5-----
[812]: 0.9996383190155029
[404]: 0.00028062614728696644
[657]: 1.6321087969117798e-05
[833 895]: 1.015903580992017e-05
[833 895]: 1.015903580992017e-05

done

3.3. Non-RKNN Model

For other models like Caffe, TensorFlow, etc, to run on RK3568 platform, conversions are needed. Use RKNN-Toolkit2 to convert other model into RKNN model.

3.4. RKNN-Toolkit2

3.4.1. Introduction of Tool

RKNN-Toolkit2 is a development kit that provides users with model conversion, inference and performance evaluation on PC and Rockchip NPU platforms. Users can easily complete the following functions through the Python interface provided by the tool:

  • Model conversion: support to convert Caffe / TensorFlow / TensorFlow Lite / ONNX / Darknet / PyTorch model to RKNN model, support RKNN model import/export, which can be used on Rockchip NPU platform later

  • Quantization: support to convert float model to quantization model, currently support quantized methods including asymmetric quantization(asymmetric_quantized-8, asymmetric_quantized-16). and support hybrid quantization. Asymmetric_quantized-16 not supported yet

  • Model inference: Able to simulate Rockchip NPU to run RKNN model on PC and get the inference result. This tool can also distribute the RKNN model to the specified NPU device to run, and get the inference results

  • Performance evaluation: distribute the RKNN model to the specified NPU device to run, and evaluate the model performance in the actual device

  • Memory evaluation: Evaluate memory consumption at runtime of the model. When using this function, the RKNN model must be distributed to the NPU device to run, and then call the relevant interface to obtain memory information

  • Quantitative error analysis: This function will give the Euclidean or cosine distance of each layer of inference results before and after the model is quantized. This can be used to analyze how quantitative error occurs, and provide ideas for improving the accuracy of quantitative models

3.4.2. Environment Dependence

  • The system needs: Ubuntu 18.04 (x64) or later. The Toolkit can only be installed on PC, and Windows, MacOS or Debian not supported yet

  • Python version: 3.6/3.8

  • Python rely on：

#Python3.6
cat doc/requirements_cp36-1.3.0.txt
numpy==1.16.6
onnx==1.7.0
onnxoptimizer==0.1.0
onnxruntime==1.6.0
tensorflow==1.14.0
tensorboard==1.14.0
protobuf==3.12.0
torch==1.6.0
torchvision==0.7.0
psutil==5.6.2
ruamel.yaml==0.15.81
scipy==1.2.1
tqdm==4.27.0
requests==2.21.0
opencv-python==4.4.0.46
PuLP==2.4
scikit_image==0.17.2
# if install bfloat16 failed, please install numpy manually first. "pip install numpy==1.16.6"
bfloat16==1.1
flatbuffers==1.12

#Python3.8
cat doc/requirements_cp38-1.3.0.txt
numpy==1.17.3
onnx==1.7.0
onnxoptimizer==0.1.0
onnxruntime==1.6.0
tensorflow==2.2.0
tensorboard==2.2.2
protobuf==3.12.0
torch==1.6.0
torchvision==0.7.0
psutil==5.6.2
ruamel.yaml==0.15.81
scipy==1.4.1
tqdm==4.27.0
requests==2.21.0
opencv-python==4.4.0.46
PuLP==2.4
scikit_image==0.17.2
# if install bfloat16 failed, please install numpy manually first. "pip install numpy==1.17.3"
bfloat16==1.1

3.4.3. RKNN-Toolkit2 installation

It is recommended to use virtualenv to manage the python environment because there may be multiple versions of the python environment in the system at the same time. Let’s start with Python 3.6 as an example:

# 1）Install virtualenv、Python3.6 and pip3
sudo apt install virtualenv \
sudo apt-get install python3 python3-dev python3-pip
# 2）Install dependent libraries
sudo apt-get install libxslt1-dev zlib1g zlib1g-dev libglib2.0-0 libsm6 \
libgl1-mesa-glx libprotobuf-dev gcc
# 3）Use virtualenv and install Python dependency, such as requirements_cp36-1.3.0.txt 
virtualenv -p /usr/bin/python3 venv
source venv/bin/activate
pip3 install -r doc/requirements_cp36-*.txt
# 4）Install RKNN-Toolkit2，such as rknn_toolkit2-1.3.0_11912b58-cp36-cp36m-linux_x86_64.whl
sudo pip3 install packages/rknn_toolkit2*cp36*.whl
# 5）Check if RKNN-Toolkit2 is installed successfully or not，press key ctrl+d to exit
(venv) firefly@T-chip:~/rknn-toolkit2$ python3
>>> from rknn.api import RKNN
>>>

The installation is successful if the import of RKNN module doesn’t fail. One of the failures is as follows:

>>> from rknn.api import RKNN
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
ImportError: No module named 'rknn'

3.4.4. Model Conversion Demo

Toolkit Demos are under rknn-toolkit2-1.3.0/examples. Here we run a model conversion demo for example, this demo shows the process of converting tflite to RKNN, exporting model, inferencing, deploying on NPU and fetching results. For detailed implementation of the model conversion, please refer to the source code in Demo and the documents at the end of this page.

3.4.4.1. Simulate the running example on PC

RKNN-Toolkit2 has a built-in simulator, simply run a demo to deploy on NPU simulator.

(venv) firefly@T-chip:~/rknn-toolkit2-1.3.0$ cd examples/tflite/mobilenet_v1
(venv) firefly@T-chip:~/rknn-toolkit2-1.3.0/examples/tflite/mobilenet_v1$ ls
dataset.txt  dog_224x224.jpg  mobilenet_v1_1.0_224.tflite  test.py
(venv) firefly@T-chip:~/rknn-toolkit2-1.3.0/examples/tflite/mobilenet_v1$ python3 test.py
W __init__: rknn-toolkit2 version: 1.3.0-11912b58
--> Config model
W config: 'target_platform' is None, use rk3566 as default, Please set according to the actual platform!
done
--> Loading model
INFO: Initialized TensorFlow Lite runtime.
done
--> Building model
Analysing : 100%|█████████████████████████████████████████████████| 58/58 [00:00<00:00, 1869.33it/s]
Quantizating : 100%|████████████████████████████████████████████████| 58/58 [00:00<00:00, 68.07it/s]
W build: The default input dtype of 'input' is changed from 'float32' to 'int8' in rknn model for performance!
                      Please take care of this change when deploy rknn model with Runtime API!
done
--> Export rknn model
done
--> Init runtime environment
Analysing : 100%|█████████████████████████████████████████████████| 60/60 [00:00<00:00, 1434.93it/s]
Preparing : 100%|██████████████████████████████████████████████████| 60/60 [00:00<00:00, 373.17it/s]
W init_runtime: target is None, use simulator!
done
--> Running model
mobilenet_v1
-----TOP 5-----
[156]: 0.9345703125
[155]: 0.0570068359375
[205]: 0.00429534912109375
[284]: 0.003116607666015625
[285]: 0.00017178058624267578

done

3.4.4.2. Run on ITX-3568Q NPU connected to the PC

RKNN Toolkit2 runs on the PC and connects to the ITX-3568Q through the PC’s USB. RKNN Toolkit2 transfers the RKNN model to the NPU device of ITX-3568Q to run, and then obtains the inference results, performance information, etc. from the ITX-3568Q

  • First prepare ITX-3568Q environment: update librknnrt.so and run rknn_server

Android

adb root && adb remount
adb push rknpu2_1.3.0/runtime/RK356X/Android/rknn_server/arm64-v8a/vendor/bin/rknn_server /vendor/bin
adb push rknpu2_1.3.0/runtime/RK356X/Android/librknn_api/arm64-v8a/librknnrt.so /vendor/lib64
adb push rknpu2_1.3.0/runtime/RK356X/Android/librknn_api/arm64-v8a/librknnrt.so /vendor/lib
adb shell reboot

# Android System will automatically run rknn_server, chekc it by using command "ps -ef | grep rknn_server"

Linux

adb push rknpu2_1.3.0/runtime/RK356X/Linux/librknn_api/aarch64/* /usr/lib

# The system generally comes with rknn_server, use "systemctl status rknn_server" to check if the service is running
# If service doesn't exist or is not runnung, manually add and run rknn_server
adb push rknpu2_1.3.0/runtime/RK356X/Linux/rknn_server/aarch64/usr/bin/* /usr/bin/

# We can use "systemctl status rknn_server" to check whether the rknn_server is running.
# Without running rknn_server, run it on the serial terminal
chmod +x /usr/bin/rknn_server
/usr/bin/rknn_server

  • Then modify the demo file rknn-toolkit2-1.3.0/examples/tflite/mobilenet_v1/test.py on PC, add the target platform in it.

diff --git a/rknn-toolkit2-1.3.0/examples/tflite/mobilenet_v1/test.py b/examples/tflite/mobilenet_v1/test.py
index 0507edb..fd2e070 100755
--- a/examples/tflite/mobilenet_v1/test.py
+++ b/examples/tflite/mobilenet_v1/test.py
@@ -24,11 +24,11 @@ def show_outputs(outputs):
 if __name__ == '__main__':
 
     # Create RKNN object
-    rknn = RKNN(verbose=True)
+    rknn = RKNN()
 
     # Pre-process config
     print('--> Config model')
-    rknn.config(mean_values=[128, 128, 128], std_values=[128, 128, 128])
+    rknn.config(mean_values=[128, 128, 128], std_values=[128, 128, 128], target_platform='rk3568')
     print('done')
 
     # Load model
@@ -62,7 +62,7 @@ if __name__ == '__main__':
 
     # Init runtime environment
     print('--> Init runtime environment')
-    ret = rknn.init_runtime()
+    ret = rknn.init_runtime(target='rk3568')
     if ret != 0:
         print('Init runtime environment failed!')
         exit(ret)

  • Run test.py on host PC

(venv) firefly@T-chip:~/rknn-toolkit2-1.3.0/examples/tflite/mobilenet_v1$ python3 test.py 
W __init__: rknn-toolkit2 version: 1.3.0-11912b58
--> Config model
done
--> Loading model
INFO: Initialized TensorFlow Lite runtime.
done
--> Building model
Analysing : 100%|█████████████████████████████████████████████████| 58/58 [00:00<00:00, 1730.77it/s]
Quantizating : 100%|███████████████████████████████████████████████| 58/58 [00:00<00:00, 366.86it/s]
W build: The default input dtype of 'input' is changed from 'float32' to 'int8' in rknn model for performance!
                      Please take care of this change when deploy rknn model with Runtime API!
done
--> Export rknn model
done
--> Init runtime environment
I NPUTransfer: Starting NPU Transfer Client, Transfer version 2.1.0 (b5861e7@2020-11-23T11:50:36)
D RKNNAPI: ==============================================
D RKNNAPI: RKNN VERSION:
D RKNNAPI:   API: 1.3.0 (121b661 build: 2022-04-29 11:07:20)
D RKNNAPI:   DRV: rknn_server: 1.3.0 (121b661 build: 2022-04-29 11:11:34)
D RKNNAPI:   DRV: rknnrt: 1.3.0 (9b36d4d74@2022-05-04T20:16:47)
D RKNNAPI: ==============================================
done
--> Running model
mobilenet_v1
-----TOP 5-----
[156]: 0.93505859375
[155]: 0.057037353515625
[205]: 0.0038814544677734375
[284]: 0.0031185150146484375
[285]: 0.00017189979553222656

done

3.4.5. Other Toolkit Demo

Other Toolkit demos can be found under rknn-toolkit2-1.3.0/examples/, such as quantization, accuracy analysis demos. For detailed implementation, please refer to the source code in Demo and the detailed development documents.

3.5. Detailed Development Documents

Please refer to pdf files under docs directory in RKNN and Toolkit SDK for development.