AminGeng / yolov8-demo

System

• Ubuntu 16.04.7 LTS
• gcc version 5.4.0 20160609 (Ubuntu 5.4.0-6ubuntu1~16.04.12)

MNN

• clone MNN Rep date: 2024/06/14
• compile command

  git clone https://github.com/alibaba/MNN.git
  cd MNN
  mkdir build && cd build
  cmake -DMNN_BUILD_OPENCV=ON -DMNN_IMGCODECS=ON -DMNN_BUILD_SHARED_LIBS=OFF ..
  make -j4

• copy lib and files by sh copy_MNN.sh after compiled

Program

yolov8_demo

in test/yolov8 folder.

yolov8_demo.h, yolov8_demo.cpp:

• original codes copy from https://github.com/wangzhaode/mnn-yolo/blob/main/cpp/yolov8_demo.cpp
• I convert model load and infer process into class
• yolov8n.mnn download url: https://github.com/wangzhaode/mnn-yolov8/releases/download/v1.0/yolov8n.mnn

multi-thread test program

in test folder.

mult_thread_test.cpp

compile

cd test
mkdir build && cd build
cmake ..
make -j4

Issue

run ./multi_thread_test 1 no problem. but n>1 like ./multi_thread_test 12 error!

This error does not always occur, but the likelihood increases with a higher number of threads. I have tested it on Ubuntu 16 (GCC 5.4), Ubuntu 18 (GCC 7.5), and Ubuntu 22.04 (WSL with GCC 11.4), and the issue persists across all versions.”

Fix Method

1. using MNN::CV in multithreaded scenarios, locking is required

2. using MNN::Module api, multithreaded scenarios, reference: https://mnn-docs.readthedocs.io/en/latest/inference/module.html#id6