C++14 is going away

[eddelbuettel]

Brian Ripley has already removed CXX11 and CXX14 from the Makeconf in R-devel so come April and R 4.6.0 we no longer can compile as C++14.

Which is now a problem as R has supported C++17 for three years and will attempt to default to C++20.

[coatless]

Thanks for handling the Bandicoot 3.1.00 bump + cherry picked items.

I'll be able to turn my attention back to the package come Friday.

[eddelbuettel]

It's done. Checks cleanly, builds, supports both backends if your machine has it.

(The main work was not the cherry pick. The main work was what I did bouncing back and forth with @rcurtin a few weeks ago and which you ignored for three weeks. I guess we all get busy, and have different priorities.)

[coatless]

@eddelbuettel so off to CRAN? Or?

[eddelbuettel]

I am wondering about that. It demands a lot. I also tossed it into my packages.json but it has not yet bubbled up on my r-universe page.

I want us to be damn sure we build if and when OpenCL is a given so may be worthwhile to check with other OpenCL packages how they build / what they assume / what they get. Then there is the can of worms called windows, plus of course your favourite OS with its three conflicting sources of compilers.

So I think we may start more slowly about getting the word out and let people test it first. I just set up a first 'horse race' for fastLm() between arma and coot ... only to find out that coot only solve()s square matrices. 😿

[coatless]

Okay, so I'll take some time on Friday looking at OpenCL. For Windows, I think we'll use the os_type and restrict to Linux + macOS as Ryan / Conrad / Marcus all have a disclaimer on Bandicoot that it isn't being tested there. (Plus, I can only imagine the hell that would be present detecting CUDA there.)

Did you want to announce the package as a pre-release now or after the OpenCL part is dug into?

[eddelbuettel]

Okay, so I'll take some time on Friday looking at OpenCL.

It should just work. It does here. It still has these quirks of no API accessor for the API variant, so I may have hardcoded 300 for now. But we do now co-exist with CUDA and when both are found, can switch. That is good. I'd say kick tires for a few days, see how (if) it builds at r-universe and then maybe start talking about it.

Would mac-builder work i.e. do you know if OpenCL is present there?

[coatless]

@eddelbuettel mac-builder will error at 300 due to my earlier note. We'll poke Simon likely to install OpenCL via recipes.

[eddelbuettel]

We can poke. I think I sketched out an approach. I just don't recall if I have done it. But Ryan and I found a function returning a string from which one get extract and transform the version. I think for me it returns something like 'OpenCL API 3.0' or so.

r-universe just came back and is utter crash and burn: https://github.com/r-universe/eddelbuettel/actions/runs/20325296275

[coatless]

@eddelbuettel I'll dig more into the failures on Friday when my calendar clears till end of year. In the interim, I'll write up an issue for simon on recipes repo and follow it with a patch.

[eddelbuettel]

I will try to ping Jeroen on this too.

[eddelbuettel]

Jeroen added the Linux libraries, and we have that platform built. He can use whatever recipes Simon provides, but we'd need to update those first per your other issue.

[coatless]

Out of pure curiosity, is it worth just trying to place the CLBlast binary during the configure? I think Simon doesn't really want the OpenCL part.

[eddelbuettel]

Can you rephrase the question? I don't think I grok what you are suggesting.

Simon is author of an OpenCL package: https://cran.r-project.org/package=OpenCL. What would he have against it?

[coatless]

Sorry for the brevity. The idea here is Simon seemed keen to want OpenCL v3 headers vendored in the package. However, we need clblast (newer/tuned) or clbas (old/not tuned) installed using the right OpenCL header, e.g. v3.0, when it's compiled. Hence, my inquiry on maybe just grabbing a binary as needed during configure.

I think there is prior CRAN precedent allowing this kind of approach as I'm remembering some kind of census package you worked on back in the day?

[eddelbuettel]

There are a few but CRAN does not like it so we may not get a new exemption.