cv_cuda_sum.cu

#include <opencv2/opencv.hpp>
#include <opencv2/cudaarithm.hpp>

void initialize_cuda() {
    int runtime_version = 0;
    cudaError_t error = cudaRuntimeGetVersion(&runtime_version);
    printf("Cuda runtime version = %d\n", runtime_version);

    int driver_version = 0;
    cudaDriverGetVersion(&driver_version);
    printf("Cuda driver version = %d\n", driver_version);

    int device_count = 0;
    cudaGetDeviceCount(&device_count);
    printf("Device count = %d\n", device_count);

    // Pick the first device
    error = cudaSetDevice(0);
    if(error) {
        printf("There was an error setting the Cuda device\n");
    } else {
        printf("CUDA initialized and ready to go\n");
    }
}

int main() {
    // Set the correct device
    initialize_cuda();

    // Create an identity matrix:
    // 1 0 0
    // 0 1 0
    // 0 0 1
    cv::Mat identity = cv::Mat::eye(3, 3, CV_32FC1);

    // Upload it to the GPU
    cv::cuda::GpuMat identity_gpu(identity);

    // Use the OpenCV cuda method for summation (crashes)
    cv::cuda::sum(identity_gpu);

    // Be a good citizen
    return 0;
}

//////////////////////////////////////////////////////////////
// Compiled with the following Makefile
/*
LIBRARIES=`pkg-config --cflags --libs opencv` -lopencv_cudev -lopencv_cudaarithm -lopencv_cudabgsegm -lopencv_cudacodec -lopencv_cudafeatures2d -lopencv_cudafilters -lopencv_cudalegacy -lopencv_cudaobjdetect -lopencv_cudaoptflow -lopencv_cudastereo -lopencv_cudawarping -lopencv_cudev -lm
all:
	nvcc main.cu -o sum.o $(LIBRARIES)
*/
//////////////////////////////////////////////////////////////
// To execute this: `./sum.o`

//////////////////////////////////////////////////////////////
// System details:
// opencv-3.1.0
// nvidia 960m / cuda 7.5
// It is only the cv::cuda::sum function that fails. The output looks like this:
// Cuda runtime version = 7050
// Cuda driver version = 7050
// Device count = 1
// CUDA initialized and ready to go
// *** Error in `./sum.o': malloc(): memory corruption: 0x0000000004ecdc90 ***
// Aborted (core dumped)

//////////////////////////////////////////////////////////////////
// Things i've tried
// - Changing the matrix type from CV_32FC1 to CV_8UC1 (doesn't work)
// - Passing a host memory matrix (`identity`, doesn't work)
// - Removing the `initialize_cuda` call (hoping OpenCV does that, doesn' work)