This issue was really going to be about malloc pitch and 2D, but then I realized maybe the whole cust::memory needed a little refurbish.

Cuda gives better performance when rows are aligned in a specific manner. For 2D arrays (e.g. images) it is common to use cudaMallocPitch for allocation and cudaMemcpy2D for copying. In addition there exists corresponding functions for 3D which makes things a bit more complex as the cuda array type must be used.

The first question is: Should these be exposed, in some way, in cust? I think the answer is yes, or else there would be unnecessarily difficult to write code that is as performant as when using for host code C/C++.

The second question is: How should these functions be exposed in cust? Cuda is all over the place when it comes to naming, argument lists, return values and even how things really work. Should we follow the cuda runtime API to make it as similar as possible for the C++ crowd? Or should try to make cust::memory as coherent as possible?

I understand both sides, but it seems to me the goal for this project is to make Cuda in Rust as good as it can be without being afraid of diverging from how it works in C++. The logical solution might then be to try to improve on naming and function signatures.

When I started looking at this I realized that the current cust::memory module is not very unified either. The alloc and memcpy functions operate of different types (cust::DevicePointer vs CUdeviceptr), size specifications (bytes vs elements) and constraints (when allocing T must be DeviceCopy). How should the feature complete cust::memory look like?