Previously, most of the benchmarks just measured time spent reading or
writing the array on the device. The transfer only happened on the first
iteration and was then cached on the device.

This change clears out the array every iteration so that the array has
to be transferred afresh.