Scaling in the FFT is controlled by the TP_SHIFT parameter which describes how many binary places by which to shift the result to the right, i.e. only power-of-2 scaling values are supported. The FFT implementation does not implement the 1/N scaling of an IFFT directly, but this may be configured via TP_SHIFT.
Internal to the FFT, for cint16 and cint32 data, an data type of cint32 is used for temporary value. After each rank, the values are scaled by only enough to normalize the bit growth caused by the twiddle multiplication (i.e., 15 bits), but there is no compensation for the bit growth of the adder in the butterfly operation.
No scaling is applied at any point when the data type is cfloat. Setting TP_SHIFT to any value other than 0 when TT_DATA is cfloat will result in an error.
In the case of TP_PARALLEL_POWER > 0 for cint16, the streams carrying data between subframe processors and the combiner stages carry cint16 data so as to allow for high performance. In this case, the scaling value applied to each subframe processor is (TP_SHIFT-TP_PARALLEL_POWER) (if positive and 0 if not). Each combiner stage will have a shift of 1 is applied, to compensate for the bit growth of 1 in the stage’s butterfly, if there is adequate TP_SHIFT to allow for this, or 0 if there is not.
For example, with an FFT configured to be POINT_SIZE=1024, DATA_TYPE=cint16, PARALLEL_POWER=2 and TP_SHIFT=10, there will be 4 subframe processors and 2 further ranks of 4 combiners. The 4 subframe processors will all have a local TP_SHIFT of 10-2 = 8 applied and each of the combiners will have a local TP_SHIFT of 1 applied.
This scheme is designed to preserve as much accuracy as possible without compromising performance.
If better accuracy or noise performance is required, this may be achieved at the expense of throughput by using TT_DATA=cint32.