Scaling in Convolution/Correlation is controlled by the TP_SHIFT parameter, which describes the number of bits to shift the output to the right. Float, cfloat, and bfloat16 implementations do not support scaling. TP_SHIFT must be set to ‘0’.