[REVIEW] Introducing host API for PCG

[vinaydes]

This PR introduces following changes

1. Adds support for calling PCG RNG API from the host
2. Adds a test for verifying host API
3. Fixes a bug in Normal PDF deviate function for integers (custom_next)
4. Adds a wide 64-bit multiplication function for both host and device
5. Adds a test for 64-bit wide multiplication on host and device

[MatthiasKohl]

Thank you very much @vinaydes , LGTM

[MatthiasKohl]

Just referencing issue #1560 which should be closed by this (I cannot add it to the development dependencies since I don't seem to have privileges for this)

[MatthiasKohl]

I missed the fact that the file was moved, thanks for pointing it out.
However, I don't understand your comment b) @achirkin . The CUDA asm in the functions is protected by pre-processor already. However, the "HDI" macro is an issue, so regardless of CUDA asm, this function cannot be called from a host-compiler. I didn't know that this was the intention here, and it is in a .cuh file, so everything looked good to me.
If we want RAFT to actually care about host compilers, we need to significantly overhaul it.

[vinaydes]

@MatthiasKohl For now, going with option b) will enable us to call the wide multiply from .cu files for using in the host code.

[MatthiasKohl]

@MatthiasKohl For now, going with option b) will enable us to call the wide multiply from .cu files for using in host code.

Right, I don't understand the second part of Artem's comment though.
Basically, if these functions could go into .hpp then what file should they be in?
My point is that if we wanted them to be in .hpp, it would be very easy to allow them to be there by simply changing out the HDI macro to one that allows a host-compiler to compile as well.

[achirkin]

@MatthiasKohl , you're probably right. It didn't occur to me that __CUDA_ARCH__ would still be available in *.hpp included by a *.cu file :)
I guess, then it's natural to move it to the existing raft/util/integer_utils.hpp. HDI would need to be replaced with _RAFT_HOST_DEVICE from core/detail/macros.hpp.

[achirkin]

I wonder though. If we start to guard the whole kernels with pre-processor macros, we can end up with replacing all headers in raft with .hpp versions. Do we want that? Where's the boundary between what we want to have in .cuh and .hpp?..

[vinaydes]

I guess, then it's natural to move it to the existing raft/util/integer_utils.hpp.

I was going to say we create a integer_utils.cuh but then I took a quick glance at the other .hpp files in the project. Several of them have cuda code. I guess I'll put the function in the integer_utils.hpp.

[vinaydes]

@achirkin @MatthiasKohl I have tried to accommodate the changes. It should be non-breaking now 🤞.
Also Artem do you mind doing /ok to test one more time?

[achirkin]

/ok to test

[achirkin]

/ok to test

[vinaydes]

@achirkin Can we mark this as non-breaking again?

[vinaydes]

Can this be merged now?

[achirkin]

Apparently I don't have enough power to summon the merger bot :)

[achirkin]

/ok to test

[cjnolet]

/merge