An option is available, in all architectures, to apply convergent rounding to the data after the butterfly stage. However, selecting this option does not apply convergent rounding to all points in the datapath where wordlength reduction occurs.
In particular, the outputs of all complex multipliers in the FFT datapath are truncated to reduce datapath width (while still maintaining adequate precision) and a simple rounding constant added to the fractional bits. This constant implements non-symmetric, round-towards-minus-infinity rounding, and can introduce a small bias to the FFT results over a large number of samples.