I know it's subject to rounding etc. but ReductionKernel often finds a
max value at a different position than both numpy.argmax and
thrust::max_element, the latter two agreeing always.
Thrust takes care to always return the first occurrence of the max--that
might be the reason: