yes

nnPerf: A Real-time On-device Tool Profiling DNN Inference on Mobile Platforms

Overview | Key Features | Installation instructions | Visualization tool
Support for LLM | Citation | Get Help | Credits | GitHub | PaperPDF | FAQ

Our paper "nnPerf" won the Best Paper Award Runner-up of ACM SenSys 2023 !

Overview

This webpage contains instructions to use our nnPerf. nnPerf is a real-time on-device profiler designed to collect and analyze the DNN model runtime inference latency on mobile platforms. nnPerf demystifies the hidden layers and metrics used for pursuing DNN optimizations and adaptations at the granularity of operators and kernels, ensuring every facet contributing to a DNN model's runtime efficiency is easily accessible to mobile developers via well-defined APIs.

With nnPerf, the mobile developers can easily identify the bottleneck in model run-time efficiency and optimize the model architecture to meet system-level objectives (SLO).

The figure below compares nnPerf to existing DNN model analyzers designed for mobile platforms.

The effectiveness of nnPerf
The effectiveness of nnPerf

To cite this tool, the best reference is the SenSys 2023 paper.

Key Features

[News 20240305] We will update the features for measuring the frequency, utilization, and temperature of multi-core processors later.


1. Plug-and-play design principles

• Follows a self-contained approach with no need for extra libraries or complex installations.

2. Real-time on-device profiling

• Monitor DNN inference delays directly on the device without external dependencies like adb.

3. Support measuring fine-grained information at the GPU kernel level

• Allows deep inspection of GPU kernels for detailed insights into DNN model optimization.

For more design details and features, please refer to our Sensys 2023 paper.

Installation instructions

Set up nnPerf in just a few steps. You can download the latest version of nnPerf here.

Quick start with apk

1. Use adb to connect to smartphones or mobile platforms (Android basic system)

2. Install the nnPerf_v1.0.apk

adb install -t .\nnPerf_v1.0.apk

Build android project

1. Install Android Studio 3.6.3 (Runtime version: 1.8.0_212-release-1586-b04 amd64).

2. Import Project

File -> Open -> Current file directory

3. Android Studio Setting

Android Gradle Plugin Version:   3.1.3
Gradle Version:                  4.4
NDK Version:                     21.0.6113669
JDK Verison:                     1.8.0_211
Complile Sdk Version:            27
Build Tools Version:             27.0.3

4. Run to profile

Output path: /data/data/com.example.android.nnPerf/

5. Model support list (Support for adding other .tflite models)

mobilenetV3-Large-Float
mobilenetV3-Small-Float
EfficientNet-b0-Float
mobilenetV1-Quant
mobilenetV2-Float
mobilenetV1-Float
Squeezenet-Float
Densenet-Float
MNasNet-1.0
MobileBert
SSDMobileV2
Esrgan

Online timeline visualization tool

We are developing an online timeline visualization tool for nnPerf, we will release it later.

Channel state information for four 1x1 links
Example of visualization tool

Features of our visualization tools:

1. Easily upload test data files and resize the interface using the scroll wheel.

2. Utilize the "Sort" button to organize data within the file based on three distinctcategories.

3. Retrieve and query previously uploaded files conveniently through the History feature.

4. Selectively hide specific filters through the intuitive legend in the upper right corner of the interface.

Support for LLM

We also support profiling LLM, such as GPT, BERT.

latency (ns)
Inference Time (s)
Channel state information for four 1x1 links
Example of GPT-2
Channel state information for four 1x1 links
Example of MobileBERT

Citation

If you find nnPerf useful in your research, please consider citing:

    @inproceedings{nnPerf,
        author = {Chu, Haolin and Zheng, Xiaolong and Liu, Liang and Ma, Huadong},
        title = {nnPerf: Demystifying DNN Runtime Inference Latency on Mobile Platforms},
        year = {2023},
        publisher = {Association for Computing Machinery},
        address = {New York, NY, USA},
        url = {https://dl.acm.org/doi/10.1145/3625687.3625797},
        doi = {10.1145/3625687.3625797},
        booktitle = {Proceedings of the 21st ACM Conference on Embedded Networked Sensor Systems},
        pages = {125–137},
    }

Get Help

If you have any functional requirements or usage feedback for nnPerf, please send it to this email address: buptwins#163.com (replace # with @).

让您拥有移动端侧神经网络推理的“智子”视角@SenSys’23

Credits

Authors: Haolin Chu | Xiaolong Zheng | Liang Liu | Huadong Ma
Maintainers: Haolin Chu | Haiteng Xin | Boyan Lu