Support for cudaGetDeviceProperties

[mxmlnkn]

I have some code, which queries cudaGetDeviceProperties for multiProcessorCount and maxThreadsPerMultiProcessor in order to determine how many blocks, threads and streams to start.

It would be cool if cupla could emulate the device properties to some extent for this kind of use-case by mapping it onto corresponding features. E.g. maxThreadsPerMultiProcessor = OMP_NUM_CORES, multiProcessorCount=1;

[ax3l]

Also, as in alpaka-group/alpaka#846, these should be replaced with cudaDeviceGetAttribute for the CUDA backend.

[jyoung3131]

I realize this is an older issue but I would suggest having some documentation on how to replace calls for cudaGetDeviceProp in the porting guide since it's pretty common to query at least a few things - I realize with a non-CUDA backend it's likely to be a NOP but some warnings might be nice.

The runtime backend page suggests using "alpaka::getAccDevProps" but there doesn't seem to be any notes on how to use this.
https://alpaka.readthedocs.io/en/latest/dev/backends.html#cuda-runtime-api

[sbastrakov]

So alpaka's getAccDevProps is a reduced subset of what CUDA has as device properties. It returns this structure. In practice we call the native version inside guards when we need more specific info.