PETSc: How to accommodate library variants using 32-bit and 64-bit integers/floats

[KGHustad]

PETSc defines the integer type PetscInt that is used for indices and PetscReal that is used for floating-point values. Both of these can be set to various sizes at compile-time.
The default is float64 and int32, but I think all of the following combinations could also be interesting to use:

• float64 and int64
• float32 and int32
• float32 and int64

My question then is how we should handle this with the current structure? I suppose one solution is to suffix the version, e.g. 3.11.3-float64-int64, but with other suffixes like -cuda, this could become a bit messy.

@jamtrott, do you have any thoughts on this?

[jamtrott]

That's a good question. Are there any particular combinations that you need right now? I'm not sure that all of them are equally important.

In addition, there is another dimension if you want to support complex numbers. This immediately doubles the number of configurations. If you count two integer sizes, two floating-point sizes, real and complex, and with and without CUDA support, that makes sixteen configurations.

For the moment, I think the best option is to just build whatever versions that we need and append the details to the name of the module, like you suggested. I really don't see any better options, and I imagine that this is how PETSc is intended to be used, anyway.