Simply reuses gpu max pooling where instead we collect eval point with index
which corresponds to the maximum in the input.