merged model run slower

[lincolnhard]

model with batch normalization runs faster than merged

[Robert-JunWang]

How do you evaluate the speed? The speed may be the same if the framework merges bn layer automatically. But there is no reason that the model with bn is faster than merged.