cuda保存设备图片到本地

    cv::Mat corrected_img(height_, width_, CV_32FC3);
    CUDA_CHECK(cudaMemcpyAsync(
            corrected_img.data,
            d_corrected,
            3 * width_ * height_ * sizeof(float),
            cudaMemcpyDeviceToHost,
            stream_
    ));
    CUDA_CHECK(cudaStreamSynchronize(stream_));
    cv::Mat corrected_image_8u;
    corrected_img.convertTo(corrected_image_8u, CV_8UC3);
    cv::imwrite("corrected_image.png", corrected_image_8u);