Bug 246939

Summary: math/armadillo: do not link against 2 different blas libraries
Product: Ports & Packages Reporter: Oleg Sidorkin <osidorkin>
Component: Individual Port(s)Assignee: Danilo Egea Gondolfo <danilo>
Status: Open ---    
Severity: Affects Only Me Keywords: buildisok
Priority: --- Flags: bugzilla: maintainer-feedback? (danilo)
Version: Latest   
Hardware: Any   
OS: Any   
Attachments:
Description Flags
svn diff patch to use single blas implementation in math/armadillo none

Description Oleg Sidorkin 2020-06-02 15:19:42 UTC
Created attachment 215169 [details]
svn diff patch to use single blas implementation in math/armadillo
Comment 1 Danilo Egea Gondolfo freebsd_committer 2020-06-03 21:39:44 UTC
Hi, thank you for your patch.

While I agree that the correct way to define openblas as dependency is through USES=blaslapack:openblas, it doesn't prevent math/blas to be installed and linked as it's a math/arpack-ng dependency.

We can switch to math/blas though (and remove openblas).

Are you an armadillo user? Do you know if one "blas" is faster than the other?

I would remove openblas but if it's faster than math/blas it would be good to continue using it.
Comment 2 Danilo Egea Gondolfo freebsd_committer 2020-06-03 22:02:13 UTC
I just tried running the armadillo tests with both math/openblas and math/blas.

With math/blas the tests take 12.8 seconds to run. With math/openblas they take 3.5 seconds to run. There are several tests failing, so I'm not entirely sure if openblas is way faster than blas or if there are more tests failing with openblas. It's probably faster I suppose.
Comment 3 Oleg Sidorkin 2020-06-04 09:43:50 UTC
(In reply to Danilo Egea Gondolfo from comment #1)

I tried to debug hang on arpack-ng check during math/scilab compiling (with openblas option) and found out that it does not happen if arpack-ng is linked against openblas (I've reported it as https://bugs.freebsd.org/bugzilla/show_bug.cgi?id=246938). Then I started to look around for another libraries that can couse the problem.

OpenBLAS is known to be the fastest implementation (Intel MKL is even faster, but it's not the case for freebsd).

I'll check armadillo tests also.
Comment 4 Automation User 2020-08-06 02:36:48 UTC
Build info is available at https://gitlab.com/swills/freebsd-ports/pipelines/174779395