| Message ID | 20210618161204.32542-1-ludo@gnu.org |
|---|---|
| State | Accepted |
| Headers | show |
| Series | [bug#49096] gnu: petsc: Link against OpenBLAS rather than LAPACK. | expand |
| Context | Check | Description |
|---|---|---|
| cbaines/comparison | success | View comparision |
| cbaines/git branch | success | View Git branch |
| cbaines/applying patch | success | View Laminar job |
| cbaines/issue | success | View issue |
On Fri, Jun 18, 2021 at 06:12:04PM +0200, Ludovic Courtès wrote: > From: Ludovic Courtès <ludovic.courtes@inria.fr> > > OpenBLAS performs better than LAPACK, the reference implementation. > This is also consistent with the choice of BLAS/LAPACK implementations > made in most other packages. > > * gnu/packages/maths.scm (petsc)[inputs]: Replace LAPACK by OPENBLAS. > [arguments]: Pass "--with-openmp=1". > --- > gnu/packages/maths.scm | 3 ++- > 1 file changed, 2 insertions(+), 1 deletion(-) > > Hi! > > Usually there’s no reason to depend on the reference (“Netlib”) BLAS/LAPACK > implementations because they perform poorly. This change is one way to > bring consistency here. > > Thoughts? > > Ludo’. It sounds reasonable to me. As long as you're open to thoughts, perhaps we want to have a discussion (in a new thread) about renaming the input labels for lapack/openblas so it's easier to do 'guix build foo --with-inputs=blas=openblas-tuned-for-my-machine' > > diff --git a/gnu/packages/maths.scm b/gnu/packages/maths.scm > index a79d546abd..8d47f91310 100644 > --- a/gnu/packages/maths.scm > +++ b/gnu/packages/maths.scm > @@ -2417,7 +2417,7 @@ September 2004}") > `(("python" ,python-2))) > (inputs > `(("gfortran" ,gfortran) > - ("lapack" ,lapack) > + ("openblas" ,openblas) > ("superlu" ,superlu) > ;; leaving out hdf5 and fftw, as petsc expects them to be built with mpi > ;; leaving out opengl, as configuration seems to only be for mac > @@ -2428,6 +2428,7 @@ September 2004}") > #:configure-flags > `("--with-mpi=0" > "--with-openmp=1" > + "--with-openblas=1" > "--with-superlu=1") > #:make-flags > ;; Honor (parallel-job-count) for build. Do not use --with-make-np, > -- > 2.32.0 > > > >
Hi, Efraim Flashner <efraim@flashner.co.il> skribis: > On Fri, Jun 18, 2021 at 06:12:04PM +0200, Ludovic Courtès wrote: [...] >> Usually there’s no reason to depend on the reference (“Netlib”) BLAS/LAPACK >> implementations because they perform poorly. This change is one way to >> bring consistency here. >> >> Thoughts? >> >> Ludo’. > > It sounds reasonable to me. Cool, thanks for taking a look. > As long as you're open to thoughts, perhaps we want to have a discussion > (in a new thread) about renaming the input labels for lapack/openblas so > it's easier to do 'guix build foo > --with-inputs=blas=openblas-tuned-for-my-machine' We’ve very much interested in that. However, input labels don’t have a role here since ‘--with-input’ matches packages by name. So most likely, we’d (1) use OpenBLAS by default everywhere instead of netlib blas/lapack, and (2) you’d use: --with-input=openblas=openblas-tuned WDYT? Ludo’.
Efraim Flashner <efraim@flashner.co.il> skribis: > On Fri, Jun 18, 2021 at 06:12:04PM +0200, Ludovic Courtès wrote: >> From: Ludovic Courtès <ludovic.courtes@inria.fr> >> >> OpenBLAS performs better than LAPACK, the reference implementation. >> This is also consistent with the choice of BLAS/LAPACK implementations >> made in most other packages. >> >> * gnu/packages/maths.scm (petsc)[inputs]: Replace LAPACK by OPENBLAS. >> [arguments]: Pass "--with-openmp=1". [...] > It sounds reasonable to me. Pushed as c7a5c3e0bba2ac3d197b219d21527863308b75dd. Ludo'.
Hi, On Tue, 22 Jun 2021 at 09:31, Ludovic Courtès <ludo@gnu.org> wrote: > Efraim Flashner <efraim@flashner.co.il> skribis: > > On Fri, Jun 18, 2021 at 06:12:04PM +0200, Ludovic Courtès wrote: > > [...] > > >> Usually there’s no reason to depend on the reference (“Netlib”) BLAS/LAPACK > >> implementations because they perform poorly. This change is one way to > >> bring consistency here. > >> > >> Thoughts? > >> > >> Ludo’. > > > > It sounds reasonable to me. > > Cool, thanks for taking a look. LGTM. > > As long as you're open to thoughts, perhaps we want to have a discussion > > (in a new thread) about renaming the input labels for lapack/openblas so > > it's easier to do 'guix build foo > > --with-inputs=blas=openblas-tuned-for-my-machine' > > We’ve very much interested in that. However, input labels don’t have a > role here since ‘--with-input’ matches packages by name. > > So most likely, we’d (1) use OpenBLAS by default everywhere instead of > netlib blas/lapack, and (2) you’d use: > > --with-input=openblas=openblas-tuned Yeah, it make sense. :-) Cheers, simon
diff --git a/gnu/packages/maths.scm b/gnu/packages/maths.scm index a79d546abd..8d47f91310 100644 --- a/gnu/packages/maths.scm +++ b/gnu/packages/maths.scm @@ -2417,7 +2417,7 @@ September 2004}") `(("python" ,python-2))) (inputs `(("gfortran" ,gfortran) - ("lapack" ,lapack) + ("openblas" ,openblas) ("superlu" ,superlu) ;; leaving out hdf5 and fftw, as petsc expects them to be built with mpi ;; leaving out opengl, as configuration seems to only be for mac @@ -2428,6 +2428,7 @@ September 2004}") #:configure-flags `("--with-mpi=0" "--with-openmp=1" + "--with-openblas=1" "--with-superlu=1") #:make-flags ;; Honor (parallel-job-count) for build. Do not use --with-make-np,
From: Ludovic Courtès <ludovic.courtes@inria.fr> OpenBLAS performs better than LAPACK, the reference implementation. This is also consistent with the choice of BLAS/LAPACK implementations made in most other packages. * gnu/packages/maths.scm (petsc)[inputs]: Replace LAPACK by OPENBLAS. [arguments]: Pass "--with-openmp=1". --- gnu/packages/maths.scm | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) Hi! Usually there’s no reason to depend on the reference (“Netlib”) BLAS/LAPACK implementations because they perform poorly. This change is one way to bring consistency here. Thoughts? Ludo’.