Using half precision with CUFFT

Accelerated Computing GPU-Accelerated Libraries
1

Hello,

I have a two part question regarding half precision transformations using CUFFT or CUFFTDX

I understood that only power of 2 signal size is support through CUFFT but what about CUFFTDX, from the documenation it seems that any FFT size is support between 2 and 32768

Also, can we run multiple FFTs concurrently with different plans (input sizes) in the same kernel using CUFFTDX?

Thank you.

2

Hi,

It is up to cuFFTDx user to combine FFTs of different sizes in one kernel. It may be not the most efficient as different sizes may have different optimal settings however it should be possible. It would be variant of block specialization technique. I’d suggest trying this with powers of 2 and padding and then adding more sizes.

Best regards,
L