Implement std::normal_distribution

[RAMitchell]

No description provided.

[copy-pr-bot]

Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.

[bernhardmgruber]

Side note: We also have some implementations of <random> in Thrust under thrust/thrust/random. I would love if we could replace those by new facilities in libcu++! You could try to replace the Thrust implementation by your new one for testing.

[RAMitchell]

#6595 is preventing us from making this constexpr. We can defer implementing constexpr or I will use another algorithm to generate the normal distribution.

[fbusato]

looks good. Just a couple of nits

[fbusato]

/ok to test 7c3c7bb

[miscco]

Why are those tests deleted?

[RAMitchell]

I've put all tests in test_distribution.h. I can now just specify a CDF and test param sets and it will test everything. This speeds up the 15 or so distributions still to implement, improves consistency and means over 100 source files less.

[RAMitchell]

Not sure what is up with CI.

Question for my reviewers:

I've been adding std::iostream operators so far. Does this make sense? I am not so sure any more, it might be better to just remove them all.

[fbusato]

CI had troubles. I reran it.

I've been adding std::iostream operators so far. Does this make sense? I am not so sure any more, it might be better to just remove them all.

I would really want to remove them, but they are part of the standard. @davebayer @miscco any thought?

[davebayer]

I would really want to remove them, but they are part of the standard. @davebayer @miscco any thought?

We implement them for all other cuda::std:: stuff, so I think we should stay consistent and implement them

[github-actions]

🥳 CI Workflow Results

🟩 Finished in 1h 50m: Pass: 100%/90 | Total: 2d 04h | Max: 1h 49m | Hits: 84%/204441

See results here.

[RAMitchell]

I would really want to remove them, but they are part of the standard. @davebayer @miscco any thought?

We implement them for all other cuda::std:: stuff, so I think we should stay consistent and implement them

I will continue implementing them then.

@miscco @fbusato any further changes?

This PR is a template for how I intend to do the remaining distributions, in particular with regards to testing. Once this is merged I will open several concurrent PRs for other distributions, which can be reviewed in parallel.