test_ort_one_session_multiple_threads.cc

/*
Program to demonstrate using one session and multiple threads to call Run on that session.
g++ -std=c++17 -o test_ort_one_session_multiple_threads test_ort_one_session_multiple_threads.cc -I onnxruntime-linux-x64-1.15.1/include/ -lonnxruntime -Lonnxruntime-linux-x64-1.15.1/lib/ -lpthread -Wl,-rpath,/home/pranav/onnxruntime-linux-x64-1.15.1/lib/
Author: Github id: pranavsharma)
*/

#include <onnxruntime_cxx_api.h>
#include <vector>
#include <string>
#include <iostream>
#include <thread>
#include <functional>
#include <numeric>

int main()
{
    Ort::Env env;
    const char *model_path = "mul_1.onnx";
    Ort::Session sess{env, model_path, Ort::SessionOptions{nullptr}}; // Note session is created only once for both the threads
    Ort::MemoryInfo mem_info = Ort::MemoryInfo::CreateCpu(OrtAllocatorType::OrtDeviceAllocator, OrtMemType::OrtMemTypeDefault);
    std::vector<int64_t> shape{3, 2};
    int num_inputs = std::accumulate(shape.begin(), shape.end(), 1, std::multiplies<>());
    std::vector<const char *> input_names{"X"};
    std::vector<const char *> output_names{"Y"};

    std::thread t1([&]()
                   {
        auto id = std::this_thread::get_id();
        std::cout << "Inside thread " << id << "\n";
        std::vector<float> input(num_inputs, 3.0);
        auto ort_value = Ort::Value::CreateTensor(mem_info, input.data(), input.size(), shape.data(), shape.size());
        auto output = sess.Run(Ort::RunOptions{nullptr}, input_names.data(), &ort_value, input_names.size(), output_names.data(), output_names.size()); });

    std::thread t2([&]()
                   {
        auto id = std::this_thread::get_id();
        std::cout << "Inside thread " << id << "\n";
        std::vector<float> input(num_inputs, 6.0);
        auto ort_value = Ort::Value::CreateTensor(mem_info, input.data(), input.size(), shape.data(), shape.size());
        auto output = sess.Run(Ort::RunOptions{nullptr}, input_names.data(), &ort_value, input_names.size(), output_names.data(), output_names.size()); });

    t1.join();
    t2.join();
    std::cout << "End of program\n";
    return 0;
}

it will crush when use DMLEP or CUDAEP.