Summary: | math/suitesparse-cholmod do not build with math/atlas | ||||||
---|---|---|---|---|---|---|---|
Product: | Ports & Packages | Reporter: | gja822 | ||||
Component: | Individual Port(s) | Assignee: | Thierry Thomas <thierry> | ||||
Status: | Closed FIXED | ||||||
Severity: | Affects Only Me | CC: | dimpase+freebsd, fortran, thierry | ||||
Priority: | --- | ||||||
Version: | Latest | ||||||
Hardware: | amd64 | ||||||
OS: | Any | ||||||
Attachments: |
|
Description
gja822
2024-06-17 17:58:04 UTC
Created attachment 251567 [details]
Try to avoid a mix of different BLAS/LAPACK libraries
When math/blas and math/atlas are installed on the same machine, it seems that they can be mixed, causing a CHOLMOD failure.
The attached patch fixes the problem, at least on my work-station. Could you please check if it works for you?
(In reply to Thierry Thomas from comment #1) with your patch it compiles now. I do not see a way to leave only math/atlas, other ports pull in blas, lapack etc. Without an option to choose. (By the way, after splitting suitesparse, one should deinstall all in a bunch and then reinstall the components, as I can see it, after any version change, and simple portupgrade approach when ports are upgraded one by one could not always give expected results. Am I right?) Thanks for your feedback! What is your CPU? Actually ATLAS is almost abandoned, because it does not support recent CPU. Do you really measure better performance with ATLAS? If yes, you could send PRs for the ports where you need an ATLAS option: it is usually simple to add, and it is not a good thing to link a program with different implementations. Patch committed. A commit in branch main references this bug: URL: https://cgit.FreeBSD.org/ports/commit/?id=62af3b140f21f4f1eaf7adb3b469d1d276927602 commit 62af3b140f21f4f1eaf7adb3b469d1d276927602 Author: Thierry Thomas <thierry@FreeBSD.org> AuthorDate: 2024-06-19 11:26:46 +0000 Commit: Thierry Thomas <thierry@FreeBSD.org> CommitDate: 2024-06-21 20:18:10 +0000 math/suitesparse: avoid mixing various BLAS/LAPACK libraries When math/blas and math/atlas are installed on the same machine, it seems that they can be mixed, causing a CHOLMOD failure. PR: 279820 Reported by: gja822@narod.ru math/suitesparse/bsd.suitesparse.mk | 2 ++ 1 file changed, 2 insertions(+) (In reply to Thierry Thomas from comment #3) Thank you, I'll consider migrating to some other LA library. Not sure I know their differences, I stuck with ATLAS from a long time ago, when it was the main choice. By the way I do not see any other LA libraries to perform such an exhaustive tuning in order to find of optimal parameters during build like ATLAS do. (In reply to gja822 from comment #6) (I have AMD FX-8300, that is bdver2 in terms of GCC.) (In reply to gja822 from comment #6) OpenBLAS does this sort of tuning although less advanced than ATLAS. More and more frequent hardware patches (mitigations of hardware bugs such as Spectre, Meltdown, and similar) are rendering CPU tuning much harder to maintain, as such patches might require not only re-tuning of the patched host, but changes in the tuning code. Atlas is not only project affected, another one I know is MPIR (a fork of GMP with these sorts of tuning). (In reply to Thierry Thomas from comment #3) It was math/ipopt that pulled "plain" math/blas and math/lapack (and coin-or-metis and coin-or-mumps) for me. I'm almost sure that Ipopt also was some automatic dependency. So, LinAl libraries give some kind of mess. Now I've deleted Ipopt, blas, lapack and substituted openblas for atlas. Shall see. Thank you for you advice. |