I can’t seem to be able to do a complex-to-real FFT using half precision. I am able to use cufftXtMakePlanMany to do a complex-to-complex FFT half precision, and a complex-to-real FFT at full precision, but I can’t seem to do both.
There are some restrictions on FP16 transforms.
Thanks! Yes I’ve been trying to adhere to these restrictions, but perhaps I haven’t limited the input data to a size of a power of 2 correctly. However, I’ve also been meaning to ask: when padding input data with zeros so that the input data’s size is a power of 2, doesn’t this change ones result?