CUDAdrv & future development

[samuelpowell]

I'm assume that the long term plan is that GPU backed FFTs will be implemented through GPUArrays.jl, CUArrays.jl, CLArrays.jl, et al., using the new abstract FFT interface.

In the mean time, I needed to use CUFFT.jl as part of an application that uses CUDAdrv.jl, and ported it accordingly (removing the functionality that required pitched arrays, and the in-place transforms).

My intention is to strip out the convenience functions from this fork, and maintain it as a low-level interface to the CUFFT.jl library, as this is sufficient for my purposes.

However, since CUDArt.jl is essentially deprecated, I'm happy to open a PR before I start stripping it down.

[timholy]

Why not just implement support for pitched pointers somewhere? My understanding is that used appropriately they can enhance performance considerably, though I have not read anything about them in a number of years and the technology is advancing rapidly.

[samuelpowell]

Hi Tim. In my current application, transforms are always of a size that gives good alignment, so I had no need to pursue this.

I believe that newer devices are less sensitive to the alignment issues which motivate the use of pitched pointers, but indeed this could offer improved performance. A basic implementation is straightforward, but a nice interface will require a little more thought. I don't have enough time to tackle this right now, but I will open an issue in the CUDAdrv.jl repo, and if/when time allows, have a crack at it myself.