[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index][Old Index]

**To**:**<tech-pkg%NetBSD.org@localhost>****Subject**:**math/py-numpy and math/py-numpy16 updated using math/cblas and could use some verification on Darwin at least****From**:**"Dr. Thomas Orgis" <thomas.orgis%uni-hamburg.de@localhost>**- Date: Tue, 20 Apr 2021 23:17:17 +0200

Hi all, I finally managed to import math/cblas and math/lapacke, the C interfaces to the BLAS and LAPACK libraries. During upstreaming the changes we do regarding that in math/py-numpy, I realized that NumPy really does need the C BLAS interface, hence the package now depends on math/cblas. Reminder: The math/cblas in pkgsrc is just the Netlib C interface that links to the actual BLAS lib that was configured at build time. We do support installing differing implementations of BLAS side-by-side, but math/cblas always installs just one libcblas that pulls in the configured (e.g. libcblas.so linked to libopenblas_openmp.so, selected via PKGSRC_BLAS_TYPES). There used to be a distinciton for OSX/Darwin to avoid our mk/blas.buildlink.mk machinery. The Acclerate framework _is_ handled by that, though. The special case is gone now, but I lack the system to verify that this really works. I also brought the PY27-variant of numpy (math/py-numpy16) up to date regarding BLAS usage. Since NumPy is a thing with some consequences, it would be good if people could verify on their platforms if this recent change truly works out for you. There are two paths: 1. Test the build without any BLAS-related configuration. This should build NumPy linking to math/blas (directly and via math/cblas), the standard Netlib implementation. 2. Test an optimized build (-framework Accelerate on Darwin). For that, you should set PKGSRC_BLAS_TYPES to your preferred variant (see mk/blas.buildlink3.mk) and re-build math/cblas and the numpy package(s). Things should just work … but of course you never know. If any issue crops up, chances are that we might have to updated the build system patch I managed to get upstream into NumPy. Would be good if that happens before they release version 1.21. Of course, verifying if the whole stack of scientific pythion stuff properly works with the updated NumPy is always a good idea. Nowadays, you'll want to ensure that the BLAS choice of NumPy is properly replicated in other packages (verifying the .so linkage of native objects). Alrighty then, Thomas -- Dr. Thomas Orgis HPC @ Universität Hamburg

- Prev by Date:
**daily pkgsrc CVS update output** - Next by Date:
**daily pkgsrc CVS update output** - Previous by Thread:
**problem installing kde4-l10n-en_GB-4.14.3** - Next by Thread:
**OS_VERSION and NetBSD 10.0** - Indexes: