[DEBT] rename .h|.hpp|.cuh|.cu|.cpp files properly in our cuml C++ source base

[teju85]

Currently, they are just all over the place! I propose the following rule-of-thumb:

1. If a header file contains cuda kernels and/or device methods, then it should be a .cuh.
2. Else, it should be a .hpp (if the current header file is already named as a .h, feel free to leave it as it is, in order to minimize the amount of ripple changes this will cause).
3. If a header file is used to declare our C interface, it should be a .h
4. If the header file includes a .cuh, then it should be a .cuh as well
5. If a source includes a .cuh, then it should be a .cu, else it should be a .cpp
6. All source files defining our C interface should be .c

This will certainly help us in reducing compilation time as well (if we end up cleaning some of the .cu files and renaming them to .cpp).

[teju85]

By extension, this means that we'll also have to move cuda-runtime related calls (or wrappers) into a .hpp file and separate out the device-side utilities into a .cuh file. This will further aid in renaming certain .cu's to .cpp's.

[teju85]

@chaithyagr, how about this one?

[chaithyagr]

Makes sense, will give a good dive into the codebase.
This was also on my mind.

[chaithyagr]

So I could finally get code building!!
I see that code base is divied up into src and src_prims for easier usage.
For now I thought of starting the process from src_prims (A kind of bottom up approach)

Going by the rules described in #1675 (comment) , we would have a lot of file renaming, do we want this?

For now, I havent looked at any helpers for doing this, as I want to have a look into the library as I do this. But do let me know if there is any obvious fast way of doing this. Something which I thought could be used (not yet sure) is https://github.com/include-what-you-use/include-what-you-use to analyze the usage patterns

[teju85]

Glad to know that the build worked for you! In case you faced any issues during the build process, please do write about those in issue #1503 . We want to know such difficulties that the devs are having.

I'm pretty sure about all the rules I described above, except number 2, since that is mostly my personal preference. I'd like to have @JohnZed, @cjnolet and @dantegd to chime in here and get their opinion before proceeding any further.

BTW, @chaithyagr, do note that the rule number 1 above works the other way too. Meaning, let's assume that there is a .cuh which doesn't have any kernels or device methods in it. Even if it contains cuda runtime API calls, it should fall under rule number 2. In other words, we need nvcc only for compiling the kernels or device methods. Rest everything can be compiled via host compiler itself.

IWYU can be an interesting addition to our CI process. Can you file an issue about it against cuML?

[teju85]

yes, and it is better to start from src_prims. In fact I'd say start with one of the smaller subfolders in there, just to get a hang of this process and file a PR for just that change. It is better to get this change in small installments (or else everyone will face merge conflicts!)

[chaithyagr]

Sure, I will work on src_prims first.
However, I ended up making a major modification as I changed utils and cuda_utils first which has changed a lot of files, although the build works so I didnt break anything yet :P

However, I think I agree with you on having this PR in multiple parts, I will keep this change aside and work on smaller groups of folders and submit them rather than have a bulky PR.

I will go through the files and try to come up with a logical order of things and how to proceed in free time

[teju85]

@chaithyagr any updates on this?

[chaithyagr]

Oops, I seem to have forgotten to update here.
So I made one major change of changing the base cuda files as mentioned and I am about to make a more targetted change for small batched PRs.
However, for now, I seem to be limited by some other priorities and might get at this a little later, especially with the lockdown due to Corona.
Do we have any expected deadline, so that I can give an idea of what I can do before?
I can get the logical order in my opinion by end of tomorrow.