Implement `alpaka::getOptimalWorkDiv` ?

[fwyzard]

Would it make sense to implement a variation of alpaka::getValidWorkDiv that uses the functionality described in the CUDA Occupancy section to try and provide a more optimal work division, at least for GPUs that support it ?

We could call it getOptimalWorkDiv, or getHeuristicWorkDiv, or anything else ?

[fwyzard]

On a related note: alpaka::getValidWorkDiv does a lot of querying and heuristics that are independent from the total work size, and then uses those results together with the work size to come up with a work division.
Can we think of a way to cache the first part of the work ?
So that calling alpaka::getValidWorkDiv for the same kernel, changing only the total work size, does not incur the overhead of querying the kernel attributes the second time (as long as the device is the same, of course) ?

[psychocoderHPC]

I would like that, as I know you need to querry the occupency API from cuda with different block sizes multiple times. I do not remember if the API provides a suggestion for the block size based on the kernel symbol.