compile opencv with CUDA support on windows 10




  • windows: 10
  • opencv: 3.1.0
  • nvidia driver: gtx 1060 382.05 (gtx 970m)
  • GPU arch(s): sm_61 (sm_52)
  • cuda: 8.0
  • cudnn: 5.0.5
  • cmake: 3.10.0
  • vs: vs2015 64

nvidia cuda CC

see cuda compute capacity

cuda-enabled nvidia GeForce cc

cpu vs gpu

for opencv functions

speed for cpu and gpu

get source

Get opencv 3.1.0 for git and fix some bugs

git clone
cd opencv
git checkout -b v3.1.0 3.1.0

# fix bugs for 3.1.0
git cherry-pick 10896
git cherry-pick cdb9c
git cherry-pick 24dbb

git branch 

* v3.1.0


mkdir build && cd build && cmake-gui ..


configure with VS 2015 win64 with options

CMAKE_CXX_FLAGS_RELEASE /MD /O2 /Ob2 /DNDEBUG /MP # for multiple processor

BUILD_PERF_TESTS OFF # if ON, build errors occur

CUDA_TOOLKIT_ROOT_DIR  C:/Program Files/NVIDIA GPU Computing Toolkit/CUDA/v8.0
#CUDA_ARCH_BIN  3.0 3.5 5.0 5.2 6.0 6.1 # very time-consuming 

for opencv
opencv cuda arch

CUDA_ARCH_BIN 3.0 3.5 5.0 5.2 6.0 6.1 relate with


CUDA_ARCH_PTX 3.0 relate with


for caffe
caffe cuda arch

the CUDA_ARCH_BIN parameter specifies multiple architectures so as to support a variety of GPU boards. otherwise, cuda programs will not run with other type of GPU boards.
为了支持在多个不同计算能力的GPU上运行可执行程序,opencv/caffe编译过程中需要支持多个不同架构,eg. CUDA_ARCH_BIN 3.0 3.5 5.0 5.2 6.0 6.1, 因此编译过程非常耗时。在编译的而过程中尽可能选择需要发布release版本的GPU架构进行配置编译。

configure and output:

Selecting Windows SDK version 10.0.14393.0 to target Windows 10.0.17134.
found IPP (ICV version): 9.0.1 [9.0.1]
at: C:/compile/opencv/3rdparty/ippicv/unpack/ippicv_win
CUDA detected: 8.0
CUDA NVCC target flags: -gencode;arch=compute_30,code=sm_30;-gencode;arch=compute_30,code=compute_30

General configuration for OpenCV 3.1.0 =====================================
  Version control:               3.1.0-3-g5e9beb8

    Host:                        Windows 10.0.17134 AMD64
    CMake:                       3.10.0
    CMake generator:             Visual Studio 14 2015 Win64
    CMake build tool:            C:/Program Files (x86)/MSBuild/14.0/bin/MSBuild.exe
    MSVC:                        1900

    Built as dynamic libs?:      YES
    C++ Compiler:                C:/Program Files (x86)/Microsoft Visual Studio 14.0/VC/bin/x86_amd64/cl.exe  (ver 19.0.24215.1)
    C++ flags (Release):         /DWIN32 /D_WINDOWS /W4 /GR /EHa  /D _CRT_SECURE_NO_DEPRECATE /D _CRT_NONSTDC_NO_DEPRECATE /D _SCL_SECURE_NO_WARNINGS /Gy /bigobj /Oi  /wd4251 /wd4324 /wd4275 /wd4589 /MP8  /MD /O2 /Ob2 /DNDEBUG /MP  /Zi
    C++ flags (Debug):           /DWIN32 /D_WINDOWS /W4 /GR /EHa  /D _CRT_SECURE_NO_DEPRECATE /D _CRT_NONSTDC_NO_DEPRECATE /D _SCL_SECURE_NO_WARNINGS /Gy /bigobj /Oi  /wd4251 /wd4324 /wd4275 /wd4589 /MP8  /MDd /Zi /Ob0 /Od /RTC1 
    C Compiler:                  C:/Program Files (x86)/Microsoft Visual Studio 14.0/VC/bin/x86_amd64/cl.exe
    C flags (Release):           /DWIN32 /D_WINDOWS /W3  /D _CRT_SECURE_NO_DEPRECATE /D _CRT_NONSTDC_NO_DEPRECATE /D _SCL_SECURE_NO_WARNINGS /Gy /bigobj /Oi    /MP8  /MD /O2 /Ob2 /DNDEBUG  /Zi
    C flags (Debug):             /DWIN32 /D_WINDOWS /W3  /D _CRT_SECURE_NO_DEPRECATE /D _CRT_NONSTDC_NO_DEPRECATE /D _SCL_SECURE_NO_WARNINGS /Gy /bigobj /Oi    /MP8  /MDd /Zi /Ob0 /Od /RTC1 
    Linker flags (Release):      /machine:x64  /INCREMENTAL:NO  /debug
    Linker flags (Debug):        /machine:x64  /debug /INCREMENTAL 
    Precompiled headers:         YES
    Extra dependencies:          comctl32 gdi32 ole32 setupapi ws2_32 vfw32 cudart nppc nppi npps cufft -LC:/Program Files/NVIDIA GPU Computing Toolkit/CUDA/v8.0/lib/x64
    3rdparty dependencies:       zlib libjpeg libwebp libpng libtiff libjasper IlmImf

  OpenCV modules:
    To be built:                 cudev core cudaarithm flann imgproc ml video cudabgsegm cudafilters cudaimgproc cudawarping imgcodecs photo shape videoio cudacodec highgui objdetect ts features2d calib3d cudafeatures2d cudalegacy cudaobjdetect cudaoptflow cudastereo stitching superres videostab python2
    Disabled:                    world
    Disabled by dependency:      -
    Unavailable:                 java python3 viz

  Windows RT support:            NO

    QT:                          NO
    Win32 UI:                    YES
    OpenGL support:              NO
    VTK support:                 NO

  Media I/O: 
    ZLib:                        build (ver 1.2.8)
    JPEG:                        build (ver 90)
    WEBP:                        build (ver 0.3.1)
    PNG:                         build (ver 1.6.19)
    TIFF:                        build (ver 42 - 4.0.2)
    JPEG 2000:                   build (ver 1.900.1)
    OpenEXR:                     build (ver 1.7.1)
    GDAL:                        NO

  Video I/O:
    Video for Windows:           YES
    DC1394 1.x:                  NO
    DC1394 2.x:                  NO
    FFMPEG:                      YES (prebuilt binaries)
      codec:                     YES (ver 56.41.100)
      format:                    YES (ver 56.36.101)
      util:                      YES (ver 54.27.100)
      swscale:                   YES (ver 3.1.101)
      resample:                  NO
      gentoo-style:              YES
    GStreamer:                   NO
    OpenNI:                      NO
    OpenNI PrimeSensor Modules:  NO
    OpenNI2:                     NO
    PvAPI:                       NO
    GigEVisionSDK:               NO
    DirectShow:                  YES
    Media Foundation:            NO
    XIMEA:                       NO
    Intel PerC:                  NO

  Parallel framework:            Concurrency

  Other third-party libraries:
    Use IPP:                     9.0.1 [9.0.1]
         at:                     C:/compile/opencv/3rdparty/ippicv/unpack/ippicv_win
    Use IPP Async:               NO
    Use Eigen:                   NO
    Use Cuda:                    YES (ver 8.0)
    Use OpenCL:                  YES
    Use custom HAL:              NO

    Use CUFFT:                   YES
    Use CUBLAS:                  NO
    USE NVCUVID:                 NO
    NVIDIA GPU arch:             30 35 50 52 60 61
    NVIDIA PTX archs:            30
    Use fast math:               NO

    Version:                     dynamic
    Include path:                C:/compile/opencv/3rdparty/include/opencl/1.2
    Use AMDFFT:                  NO
    Use AMDBLAS:                 NO

  Python 2:
    Interpreter:                 C:/Python27/python.exe (ver 2.7.13)
    Libraries:                   C:/Python27/libs/python27.lib (ver 2.7.13)
    numpy:                       C:/Python27/lib/site-packages/numpy/core/include (ver 1.11.3)
    packages path:               C:/Python27/Lib/site-packages

  Python 3:
    Interpreter:                 NO

  Python (for build):            C:/Python27/python.exe

    ant:                         NO
    JNI:                         C:/Program Files/Java/jdk1.8.0_161/include C:/Program Files/Java/jdk1.8.0_161/include/win32 C:/Program Files/Java/jdk1.8.0_161/include
    Java wrappers:               NO
    Java tests:                  NO

  Matlab:                        Matlab not found or implicitly disabled

    Doxygen:                     NO
    PlantUML:                    NO

  Tests and samples:
    Tests:                       YES
    Performance tests:           NO
    C/C++ Examples:              NO

  Install path:                  C:/compile/opencv/build/install

  cvconfig.h is in:              C:/compile/opencv/build

Configuring done
Generating done

Notice for gencode

CUDA NVCC target flags: -gencode;arch=compute_30,code=sm_30;-gencode;arch=compute_35,code=sm_35;-gencode;arch=compute_50,code=sm_50;-gencode;arch=compute_52,code=sm_52;-gencode;arch=compute_60,code=sm_60;-gencode;arch=compute_61,code=sm_61;-gencode;arch=compute_30,code=compute_30


Open OpenCV.sln with VS 2015 and build release version.

this may take hours to finish.

build success


opencv build errors

possible solutions

With BUILD_PERF_TESTS and BUILD_TESTS disabled, I managed to build OpenCV 3.1 with CUDA 8.0 on Windows 10 with VS2015 x64 arch target. Without building test/performance modules, the build process costs less time as well : )

I actually got it to work both on my laptop and my desktop (GTX960M and GTX970 respectively) running with OpenCV 3.2 and the latest version of CUDA 8.0 for Win10 in Visual Studio 15 Community! What I did was to enable WITH_CUBLAS aswell as WITH_CUDA. I also turned off BUILD_PERF_TESTS and BUILD_TESTS. The configuration was built using the Visual Studio 14 2015 C++ compiler.

my solution:


opencv disable build perf tests

configure and build again. this time cost only about 1 minutes.

after error fixed,build results
after errors fixed



OpenCV GPU module is written using CUDA, therefore it benefits from the CUDA ecosystem.

GPU modules includes class cv::cuda::GpuMat which is a primary container for data kept in GPU memory. It’s interface is very similar with cv::Mat, its CPU counterpart. All GPU functions receive GpuMat as input and output arguments. This allows to invoke several GPU algorithms without downloading data. GPU module API interface is also kept similar with CPU interface where possible. So developers who are familiar with Opencv on CPU could start using GPU straightaway.

The GPU module is designed as a host-level API. This means that if you have pre-compiled OpenCV GPU binaries, you are not required to have the CUDA Toolkit installed or write any extra code to make use of the GPU.


find_package(OpenCV REQUIRED COMPONENTS core highgui imgproc features2d calib3d 
    cudaarithm cudabgsegm cudafilters cudaimgproc cudawarping cudafeatures2d # for cuda-enabled
) #
MESSAGE( [Main] " OpenCV_LIBS = ${OpenCV_LIBS}")


In the sample below an image is loaded from local file, next it is uploaded to GPU, thresholded, downloaded and displayed.

#include <opencv2/cudaarithm.hpp>
#include <opencv2/cudabgsegm.hpp>
#include <opencv2/cudafilters.hpp>
#include <opencv2/cudaimgproc.hpp>
#include <opencv2/cudawarping.hpp>
#include <opencv2/cudafeatures2d.hpp>

int test_opencv_gpu()
        cv::Mat src_host = cv::imread("file.png", CV_LOAD_IMAGE_GRAYSCALE);
        cv::cuda::GpuMat dst, src;

        cv::cuda::threshold(src, dst, 128.0, 255.0, CV_THRESH_BINARY);

        cv::Mat result_host;;

        cv::imshow("Result", result_host);
    catch (const cv::Exception& ex)
        std::cout << "Error: " << ex.what() << std::endl;
    return 0;

cpu vs gpu time cost

  • (1)对于分辨率不特别大的图片间的ORB特征匹配,CPU运算得比GPU版的快(由于图像上传到GPU消耗了时间)
  • (2)但对于分辨率较大的图片,或者GPU比CPU好的机器(比如Nvidia Jetson系列),GPU版的ORB算法比CPU版的程序更高效。


(1) 使用cuda版本的opencv caffe网络的第一次创建非常耗时,后面的网络创建则非常快。
(2) opencv的gpu代码比cpu代码慢,初次启动多耗费20s左右。(事实是由于编译的caffe和GPU计算力不匹配导致的)


Your problem is that CUDA needs to initialize! And it will generally takes between serveral seconds

Why first function call is slow?
That is because of initialization overheads. On first GPU function call Cuda Runtime API is initialized implicitly.

The first gpu function call is always takes more time, because CUDA initialize context for device.
The following calls will be faster.

Not Reasons:
(1) CPU clockspeed is 10x faster than GPU clockspeed.
(2) memory transfer times between host (CPU) and device (GPU) (upload,downloa data)


runtime errors

gtx 1060 编译的opencv caffe在gtx 970m上运行出现错误 Check failed: error == cudaSuccess (8 vs. 0) invalid device function

    gtx 1060   sm_61
    gtx 970m   sm_52

im2col 是caffe的源文件,表明gtx 970m的计算能力不支持可执行文件的运行。


see what-is-the-purpose-of-using-multiple-arch-flags-in-nvidias-nvcc-compiler

Roughly speaking, the code compilation flow goes like this: CUDA C/C++ device code source --> PTX --> SASS
The virtual architecture (e.g. compute_20, whatever is specified by -arch compute…) determines what type of PTX code will be generated. The additional switches (e.g. -code sm_21) determine what type of SASS code will be generated. SASS is actually executable object code for a GPU (machine language). An executable can contain multiple versions of SASS and/or PTX, and there is a runtime loader mechanism that will pick appropriate versions based on the GPU actually being used.

win7/win10 deploy

  • compile opencv caffe on windows 10 for GTX 1060
  • deoply on windows 7 for GTX 1080 Ti successfully

for win7, if we install 398.82-desktop-win8-win7-64bit-international-whql.exe,errors may occur:

> nvidia-smi.exe 
Failed to initialize NVML: Unknown error

Solutions: use older drivers 385.69

linux/window performance

(1) api在linux平均耗时3ms;同样的代码在windows平均耗时14ms
(2) vs编译开启代码优化前后性能相差接近5倍,125ms vs 25ms
(3) cmake编译RELEASE选项默认已经开启了代码优化 -O3



  • 20180713: created.

Author: kezunlin
Reprint policy: All articles in this blog are used except for special statements CC BY 4.0 reprint polocy. If reproduced, please indicate source kezunlin !