Detect CUDA install and configure OpenCV accordingly

[rajat2004]

There is a way to find the Compute Capability, not exactly a command-line option, need to compile the following and run

#include <stdio.h>
#include "cuda.h"
int main() {
    cudaDeviceProp prop;
    cudaGetDeviceProperties(&prop, 0);
    printf("%d", prop.major * 10 + prop.minor);
}

But this won't work inside Docker I think while building, since there's no device yet Could add something like pass an env variable to the script, to automate this

[rsnk96]

LGTM, would be nice to compile for the required Compute Capability only.

Also just need to verify if having comments in between the CMake command causes errors, I remember it did in the older versions of cmake. Doesn't seem to be causing an error in travis, will test on a system shortly and get back.

[rajat2004]

The comments are after the cmake command has ended, will be obvious to someone who's changing the scripts by hand The code for checking the CC could be added as a file in the extras folder, but suppose when building a docker image, the device won't be present which would cause a problem A workaround could be that an env variable CUDA_ARCH_BIN or something could be passed to the script, if it's set, use that else we build for all supported CC >= 5.3

[rajat2004]

Updated the flags to build for all supported architectures by default Some customization could be done, such as -

• Check if nvidia-smi is working, then maybe compile the above CUDA code and build for the specific arch
• User passes an env variable to build for a specific arch

[rsnk96]

Will leave those as upgrades to be made in the future

[rajat2004]

Thanks for the review! Merged