Force alignment of arrays in memory #20
Conversation
|
Neat, looks good. Feel free to merge. |
|
Numpy provides a setflags method on arrays which has an alignment option. Changing this doesn't actually change anything in the underlying array. It only changes how numpy looks at it. It also provides a require method which makes sure an array is aligned. However the size of alignment in this case is internal and quite arbitrary. Reading the discussions here it seems like the alignment is to 16 bytes. This seems like it might be aligning for improving SIMD performance. Alignment for optimal cache performance seems like a huge issue for numpy and they don't seem to have fixed this yet. |
|
Assuming we need to be able to align to any arbitrary number of bytes, this needs to be implemented ourselves. Implementing this in python is fairly trivial. To implement this in C, we would need to create our own data structures to implement the buffer protocol required by python, which seems to be a very tedious affair. Making these data structures interoperable with numpy might require additional work. |
|
I agree, the Python trick you linked might be the quickest and easiest to implement for now, and for our current benchmark purposes we might just want to do this. If we have more complicated data layouts on top of custom alignment I would hope that we might use some Cython magic to make this a little less tedious, but that would require adding a small Cython compilation setup. However, if we end up having to go through the buffer protocol and NumPy array API, I would recommend looking at petsc4py, in particular this code for an example. |
No description provided.