add openmp on windows

[xoviat]

Checklist

• Used a personal fork of the feedstock to propose changes
• Bumped the build number (if the version is unchanged)
• Reset the build number to 0 (if the version changed)
• Re-rendered with the latest conda-smithy (Use the phrase @conda-forge-admin, please rerender in a comment in this PR for automated rerendering)
• Ensured the license file is being packaged.

[conda-forge-linter]

Hi! This is the friendly automated conda-forge-linting service.

I just wanted to let you know that I linted all conda-recipes in your PR (recipe) and found it was in an excellent condition.

[xoviat]

@conda-forge-admin rerender

[xoviat]

@h-vetinari Do you know why flang 5 was installed? It should have been flang 11.

https://dev.azure.com/conda-forge/feedstock-builds/_build/results?buildId=271791&view=logs&j=3f36b4b8-4169-5a3e-5488-fb35002959b2&t=2688110d-0d85-5239-ab56-20c0d3b88554&l=63

[xoviat]

I guess there's cache in conda-forge-admin that's waiting to catch up.

[h-vetinari]

It can take up to 90min until the finished package has made it through to the content delivery network (CDN), from which it is downloaded. As soon as you see download numbers >1 for the newest build here, that should be the case.

[isuruf]

We pin to flang 5 at the moment. We'll need to rebuild all the packages with flang 11. We have a bot that can do that. Just needs to find time to do that.

[xoviat]

@isuruf where is that flang-5 defined? In some repository?

[h-vetinari]

Here: https://github.com/conda-forge/conda-forge-pinning-feedstock/

Wasn't aware that flang was pinned as well - but didn't @xoviat say that the ABI is compatible? Would it make sense to just relax the pin, instead of having a migrator? Or would that be too risky? @isuruf

[xoviat]

@h-vetinari flang ABI is NOT compatible! see conda-forge/flang-feedstock#9

[h-vetinari]

Ah, sorry, my bad - brainfart.

[h-vetinari]

Here's some docs about the kind of migration that would be necessary, as well as the currently open ones

[xoviat]

come to think of it, when f18 eventually comes out in a year or two, we'll have the same situation where another patch to the packages will be required. Oh well.

[h-vetinari]

Good news @conda-forge/openblas: with the in-progress flang 19, we can build OpenBLAS and pass the test suite, both with/without OpenMP. 🥳

For more background: llvm/llvm-project#86463 & OpenMathLib/OpenBLAS#4768

There's a positively shocking performance difference in the test suite with/without OpenMP

pthreads build (i.e. without OpenMP)

100% tests passed, 0 tests failed out of 120

Total Test time (real) = 8389.82 sec

and the OpenMP-enabled one:

100% tests passed, 0 tests failed out of 120

Total Test time (real) =  93.20 sec

By way of comparison, the current pthread build on main (using the old, non-llvm flang) has:

100% tests passed, 0 tests failed out of 120

Total Test time (real) = 8150.37 sec

I honestly don't know why this difference is so large, but I'm assuming the pthreads variant has some issues (c.f. #160).

Furthermore, there are some issues with AVX512 instructions (see OpenMathLib/OpenBLAS#4789) which will have to be fixed in LLVM, but for now the workaround that OpenBLAS is using itself (NO_AVX512=1) works.

All in all, it looks like with flang 19 in ~September, we should finally be able to switch our Fortran compilers on windows over to llvm-flang.

CC @martin-frbg @conda-forge/core

[conda-forge-webservices]

Hi! This is the friendly automated conda-forge-linting service.

I just wanted to let you know that I linted all conda-recipes in your PR (recipe/meta.yaml) and found it was in an excellent condition.

[h-vetinari]

@conda-forge/openblas, this is ready for review (aside from reverting 153ce91 and waiting for the release of LLVM 19, but that's just a question of ~2 weeks), PTAL! :)

[h-vetinari]

OK, the momentous occasion has arrived - if you're still receiving notifications here, thanks for starting this effort @xoviat 🙏, sorry it took so long, but I'm glad we could finally get it across the finish line!