Due to restrictions in Thrust, typically the lack of a vector-valued
reduction operation, some code here uses a parameter numberofchunks.
This quantity should be about the same as the anticipated number of
threads.  This may not be easy to determine, but here are guidlines:

1.  On a multicore system, take numberofchunks to be the number of
cores, e.g. 4 on a quad core machine.  With hyperthreading, you might
try double the number of cores.

2.  On a GPU, start with 100 threads, probably a conservative number.