I compiled OpenCV with cuda support and it works fine with a program that does not use ROS.
I created a node in ROS Kinetic that uses OpenCV CUDA implemented algorithms and used catkin_make to compile it, pointing to the OpenCV installation in /usr/local. The output of catkin_make is shown here:
-- The C compiler identification is GNU 5.4.0
-- The CXX compiler identification is GNU 5.4.0
-- Check for working C compiler: /usr/bin/gcc
-- Check for working C compiler: /usr/bin/gcc -- works
-- Detecting C compiler ABI info
-- Detecting C compiler ABI info - done
-- Detecting C compile features
-- Detecting C compile features - done
-- Check for working CXX compiler: /usr/bin/g++
-- Check for working CXX compiler: /usr/bin/g++ -- works
-- Detecting CXX compiler ABI info
-- Detecting CXX compiler ABI info - done
-- Detecting CXX compile features
-- Detecting CXX compile features - done
-- Using CATKIN_DEVEL_PREFIX: /home/degraw/catkin_ws/devel
-- Using CMAKE_PREFIX_PATH: /opt/ros/kinetic
-- This workspace overlays: /opt/ros/kinetic
-- Found PythonInterp: /usr/bin/python (found version "2.7.12")
-- Using PYTHON_EXECUTABLE: /usr/bin/python
-- Using Debian Python package layout
-- Using empy: /usr/bin/empy
-- Using CATKIN_ENABLE_TESTING: ON
-- Call enable_testing()
-- Using CATKIN_TEST_RESULTS_DIR: /home/degraw/catkin_ws/build/test_results
-- Looking for pthread.h
-- Looking for pthread.h - found
-- Looking for pthread_create
-- Looking for pthread_create - not found
-- Looking for pthread_create in pthreads
-- Looking for pthread_create in pthreads - not found
-- Looking for pthread_create in pthread
-- Looking for pthread_create in pthread - found
-- Found Threads: TRUE
-- Found gtest sources under '/usr/src/gtest': gtests will be built
-- Using Python nosetests: /usr/bin/nosetests-2.7
-- catkin 0.7.6
-- BUILD_SHARED_LIBS is on
-- ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-- ~~ traversing 2 packages in topological order:
-- ~~ - depth_calculator
-- ~~ - zed_wrapper
-- ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-- +++ processing catkin package: 'depth_calculator'
-- ==> add_subdirectory(depth_calculator)
-- Found CUDA: /usr/local/cuda (found suitable exact version "8.0")
-- Found OpenCV: /usr/local (found version "3.2.0")
During runtime i get the error:
OpenCV Error: No CUDA support (The library is compiled without CUDA support) in throw_no_cuda, file /tmp/binarydeb/ros-kinetic-opencv3-3.2.0/modules/core/include/opencv2/core/private.cuda.hpp, line 97
terminate called after throwing an instance of 'cv::Exception'
what(): /tmp/binarydeb/ros-kinetic-opencv3-3.2.0/modules/core/include/opencv2/core/private.cuda.hpp:97: error: (-216) The library is compiled without CUDA support in function throw_no_cuda
[1] 21085 abort (core dumped) rosrun depth_calculator double_image_acq
even though there are no files in /tmp (i deleted everything, deleted build and devel folders in catkin_ws and rerun catkin_make, which had the output shown above).
Could this be a caching issue from previous runs? Are there any runtime Environment Variables that point to the OpenCV installed in /opt/ros/kinetic instead of the one in /usr/local, with which the node was compiled and linked against?
catkin_make by default use installed with ROS OpenCV library version, which comes without CUDA support. In order to use external version of OpenCV library you have to manually force CMake to use it with following instructions:
find_package(OpenCV REQUIRED
NO_MODULE # should be optional, tells CMake to use config mode
PATHS /usr/local # look here
NO_DEFAULT_PATH) # and don't look anywhere else
In general, it is said that mixing OpenCV library versions is a bad thing. Thus probably you should re-build with your CUDA-enabled version all other ROS packages which use default OpenCV version (image_geometry, cv_bridge). Although personally I haven't tested this yet. I base my answer on this post.
You can find some information about OpenCV in ROS kinetic here. Better place to ask such question is a ROS Answers site.