the same code runs on macOS and iOS，but get different results

[LokiThorn]

I used the c++ api of onnxruntime to write a simple demo based on cpu, but it can run normally on the macos platform, but a segment error occurs in the call of run on ios. Does anyone know what the problem is?

macos: 2.8 GHz quad-core Intel Core i7
ios:iphone 11, ios15.0.2
demo:
Ort::Env env(ORT_LOGGING_LEVEL_WARNING, "ort_run");
Ort::SessionOptions session_options;
// session_options.SetGraphOptimizationLevel(GraphOptimizationLevel::ORT_ENABLE_ALL);
Ort::Session session(env, model_path, session_options);

// Prepare to enter data
size_t input_nums = session.GetInputCount();
size_t output_nums = session.GetOutputCount();
std::cout << "input_nums = " << input_nums << std::endl;
std::cout << "output_nums = " << output_nums << std::endl;
std::vector<const char*> input_names;
std::vector<Ort::Value> input_tensors;
std::vector<const char*> output_names;
std::vector<Ort::Value> output_tensors;

Ort::AllocatorWithDefaultOptions allocator;
std::vector<std::string> input_strs;
for (size_t i = 0; i < input_nums; ++i) {
    auto input_name = session.GetInputNameAllocated(i, allocator);
    std::cout << "input_name : " << input_name << std::endl;
    auto input_type_info = session.GetInputTypeInfo(i);
    auto data_type = input_type_info.GetTensorTypeAndShapeInfo().GetElementType();

    std::cout << "input_type_info : " << data_type << std::endl;
    Ort::Value input_tensor(nullptr);
    auto tensor_info = input_type_info.GetTensorTypeAndShapeInfo();
    std::vector<int64_t> input_shape(tensor_info.GetShape());
    std::cout << "input_shape : { ";
    for (auto& dim : input_shape) {
        std::cout << dim << "  ";
    }
    std::cout << " }" << std::endl;

    // The value of the batch dimension must >= 1
    if (input_shape[0] < 1) {
        input_shape[0] = 1;
    }
    size_t input_size = 1;
    for (auto& element : input_shape) {
        input_size *= element;
    }
    if (data_type == ONNX_TENSOR_ELEMENT_DATA_TYPE_FLOAT) {
        std::vector<float> input_data(input_size);
        for (int j = 0; j < input_size; ++j) {
            input_data[j] = (float)j / (float)input_size;
        }
        Ort::MemoryInfo memory_info =
            Ort::MemoryInfo::CreateCpu(OrtDeviceAllocator, OrtMemTypeCPU);
        input_tensor = Ort::Value::CreateTensor<float>(
            memory_info, input_data.data(), input_data.size(), input_shape.data(),
            input_shape.size());
    } 
    std::string str = std::string(input_name.get());
    input_strs.emplace_back(str);
    input_tensors.emplace_back(std::move(input_tensor));
}
for (auto& name : input_strs) {
    input_names.emplace_back(name.c_str());
}
std::vector<std::string> output_strs;
for (size_t i = 0; i < output_nums; ++i) {
    auto output_name = session.GetOutputNameAllocated(i, allocator);
    std::cout << "output_name : " << output_name << std::endl;
    std::string str = std::string(output_name.get());
    output_strs.emplace_back(str);
}
for (auto& name : output_strs) {
    output_names.emplace_back(name.c_str());
}
// inference model and get the output tensor
_ort_output =
    session.Run(Ort::RunOptions(nullptr), input_names.data(), input_tensors.data(),
                input_tensors.size(), output_names.data(), output_names.size());

[LokiThorn]

[LokiThorn]

mobilenetv2-7.onnx is used.