Window编译Opencv CUDA版

编译步骤：

本文写的是注意事项，不是完整流程： 

dnn cuda编译

OpenCV4.4版本 DNN模块使用CUDA加速教程 VS2017 Window10_猛龙不过江的博客-CSDN博客

opencv mat cuda编译

windows下 编译CUDA版的opencv流程记录_sazass的博客-CSDN博客_opencv编译cuda

https://github.com/opencv/opencv

https://github.com/opencv/opencv_contrib

选择tags，

https://github.com/opencv/opencv_contrib/releases/tag/4.5.4

下载source code.zip

cmake gui对着下载后的根目录，如下图：

第一轮编译报错：

1.python版本报错，因为不用python，忽略掉

2.

CMake Warning at cmake/OpenCVDownload.cmake:202 (message):
IPPICV: Download failed: 35;"SSL connect error"

For details please refer to the download log file:

下载失败解决：

进入opencv-4.5.4\3rdparty\ffmpeg目录
将 ffmpeg.cmake 的https://raw.githubusercontent.com
修改为：https://raw.staticdn.net

把代理关掉，下载再失败，自己去url里下载下来，拷贝到目录.cache\ffmpeg

3.

CMake Warning at cmake/OpenCVDownload.cmake:202 (message):
FFMPEG: Download failed: 6;"Couldn't resolve host name"

For details please refer to the download log file:

CMake Warning at cmake/OpenCVDownload.cmake:202 (message):
ADE: Download failed: 28;"Timeout was reached"

For details please refer to the download log file:

4.3等待第一轮编译完成后，勾选：BUILD_opencv_world

 以及，勾选OPENCV_DNN_CUDA，选择解压好的opencv_contrib中modules路径添加进来。

勾选OPENCV_DNN_CUDA，

OPENCV_EXTRA_MODURES_PATH：

 勾选 WITH_CUDA：

在OPENCV_EXTRA_MODULES_PATH 填入opencv_contrib/modules路径：

需要转化为右斜杠，否则会报错，"\r"

D:/project/opencv/opencv_contrib-4.5.4/modules

vs2017 社区版 generaer报错：

opencv threads:threads but the target was not found

vs2017 专业版编译ok。

https://blog.csdn.net/qq_30623591/article/details/82084113

opencv+CUDA9.1+vs2015环境搭建，编译opencv库，调用GPU加速运算_公子_羽的博客-CSDN博客_opencv 编译cuda

windows上编译自己的opencv（含cuda） - 知乎

使用OpenCV操作CUDA 图像_jacke121的专栏-CSDN博客

cuda11变化：

把cub目录提高到根目录下了。

#include <thrust/system/cuda/detail/cub/device/device_radix_sort.cuh>
#include <thrust/system/cuda/detail/cub/iterator/counting_input_iterator.cuh>

使用OpenCV操作CUDA送检图像到TensorRT(SSD模型)_ikevin的专栏-CSDN博客

Cuda::GpuMat imageGpu(cv::Size(640,480),CV_32FC4);

Capture>> frame;
 
cvtColor(frame,imageRGBA,CV_RGB2BGRA);
 
imageRGBA.convertTo(imageRGBA, CV_32FC4,4);



imageGpu.upload(imageRGBA);
...
// Here is the memory data allocated when getting the GPU corresponding to the above upload
// That is, imageNetMean's first parameter parameter
(float4*)imageGpu.ptr<float4>()

cudaError_t cudaPreImageNetMean(float4* input, size_t inputWidth,
                            Size_t inputHeight,float* output,size_t outputWidth, size_t outputHeight, 
                             Const float3& mean_value );

完整代码：GitHub - wkexinw/tensorSSD: Deep-Learn model SSD_300x300 transplante to TensorRT(Nvidia Jetson Tx2)

    for (int i = 0; i < 8; i++) {
        std::string file_name = "D:/work/SDKProject/NeuroDetectionV1/fastertest/testimg/1111/4DA9523CCE2B649D39496741D6E3284E1330013_c.jpg";
        cv::Mat image = cv::imread(file_name);
        //std::cout << "image:" << filelist[i] << std::endl;
        if (image.empty()) {
            std::cout << "读入图片为空，请检查路径！" << std::endl;
            system("pause");
            return -1;
        }

        DWORD start2 = GetTickCount();
        //cv::Mat imageRGB;
        //image.convertTo(imageRGB, CV_32FC3, 3);

        //cv::cuda::GpuMat imageGpu(cv::Size(image.cols, image.rows), CV_32FC3);

        //imageGpu.upload(imageRGB);


        cv::cuda::GpuMat imageGpu(cv::Size(image.cols, image.rows), CV_8UC3);

        cv::cuda::GpuMat imageRGB;
        imageGpu.upload(image);
        imageGpu.convertTo(imageRGB, CV_32FC3, 3);



        DWORD end2 = GetTickCount();
        std::cout << "gpu img time：" << end2 - start2 << "ms " << image.cols << " " << image.rows << std::endl;


        DWORD start3 = GetTickCount();
        std::vector<float> img;


        //width and height of input image
        int height = image.rows;
        int width = image.cols;

        img = convertMat2Vector<float>(image);


        std::vector<float> data(1 * height * width * 3, 0);

        for (int i = 0; i < height * width * 3; i++)
            data[0 * height * width * 3 + i] = img[i];
        DWORD end3 = GetTickCount();
        std::cout << "cpu img time：" << end3 - start3 << "ms " << image.cols << " " << image.rows << std::endl;

    }

segmentation/trt_sample.cpp at d86c237e1191f159d39e115df13d691ab65752c2 · novaHardware/segmentation · GitHub

// preprocessing stage ------------------------------------------------------------------------------------------------
void preprocessImage(const std::string& image_path, float* gpu_input, const nvinfer1::Dims& dims)
{
    // read input image
    cv::Mat frame = cv::imread(image_path);
    if (frame.empty())
    {
        std::cerr << "Input image " << image_path << " load failed\n";
        return;
    }
    cv::cuda::GpuMat gpu_frame;
    // upload image to GPU
    gpu_frame.upload(frame);

    auto input_width = dims.d[2];
    auto input_height = dims.d[1];
    auto channels = dims.d[0];
    auto input_size = cv::Size(input_width, input_height);
    // resize
    cv::cuda::GpuMat resized;
    cv::cuda::resize(gpu_frame, resized, input_size, 0, 0, cv::INTER_NEAREST);
    // normalize
    cv::cuda::GpuMat flt_image;
    resized.convertTo(flt_image, CV_32FC3, 1.f / 255.f);
    cv::cuda::subtract(flt_image, cv::Scalar(0.485f, 0.456f, 0.406f), flt_image, cv::noArray(), -1);
    cv::cuda::divide(flt_image, cv::Scalar(0.229f, 0.224f, 0.225f), flt_image, 1, -1);
    // to tensor
    std::vector<cv::cuda::GpuMat> chw;
    for (size_t i = 0; i < channels; ++i)
    {
        chw.emplace_back(cv::cuda::GpuMat(input_size, CV_32FC1, gpu_input + i * input_width * input_height));
    }
    cv::cuda::split(flt_image, chw);
}