CUDA FFT of reals in 2D

Accelerated Computing CUDA CUDA Programming and Performance
1

For a 1D FFT I can see in NUMERICAL RECIPES, how to interlave a real array in a complex array if I want to transform real data and save memory and speed of the calculation.

How to do this in 2D if I only want to transform real data.

Would it be feasible to have CUDA supporting real transformations in 1 and 2D.

Thanks for anyones help

Peter

2

Support for real transforms is coming in the next release.

3

:D Very good news! I’m looking forward that function!

4

Any dates?

5

Few weeks.

6

The CUFFT documentation in the 0.9 release states that CUDA does not implement specialized algorithms for real data (which can double performance), but doesn’t say if this fact is only temporary. So, will CUFFT eventually implement such specialized algorithms?

Another question is that the code example for the 2D C2R FFT, why is idata of dimensions Nx, Ny instead of Nx,(Ny/2 + 1)? Is there a performance reason for this?