This pull request allows for roundoff error in the ResetStateOnTriggerGPU test.
Since inputs vary in magnitude, the quantity tested is relative error, that is, (GPU - CPU)/CPU. The tolerance is set to 10^-6, to allow for single-precision floating point roundoff errors accumulated over 100 timesteps.
To prevent division-by-zero exceptions, the input layer runs from
4 to 256 instead of from 0 to 252.
This pull request allows for roundoff error in the ResetStateOnTriggerGPU test.
Since inputs vary in magnitude, the quantity tested is relative error, that is, (GPU - CPU)/CPU. The tolerance is set to 10^-6, to allow for single-precision floating point roundoff errors accumulated over 100 timesteps.
To prevent division-by-zero exceptions, the input layer runs from 4 to 256 instead of from 0 to 252.