cuFFT on cudaSurfaceObject?

Accelerated Computing CUDA CUDA Programming and Performance
1

Hi, so far I’ve used cuFFT for 2D & 3D transforms on larger inputs (up to GB) of linear device memory, but now I might need inputs/outputs as 2D/3D surface memory. To conserve limited memory and to avoid copying to/from linear memory, are there FFT libraries that work directly on CUDA surfaces? I didn’t see any option for this in the cuFFT docs. But maybe that’s a bad idea anyway as FFTs are nicely separable and the 3D coherence would not only not help, but perhaps make it worse. Still I’m curious if anyone has tried this.

2

Use cuFFTDx

There you get the computation and can implement the memory accesses yourself.

They are meant as building blocks.