Am Wed, 27 Dec 2023 07:29:58 -0600
schrieb Jason Bacon <jtocino%gmx.com@localhost>:
Nothing's changed on my NetBSD 9 installation, even after my overnight
updates and successful pkg_chk run. I still don't get proper blas linkage:
Yeah, sorry … something changed on my end. I have tested this and forgot:
Index: Makefile
===================================================================
RCS file: /cvsroot/pkgsrc/math/py-scipy/Makefile,v
retrieving revision 1.74
diff -u -r1.74 Makefile
--- Makefile 7 Dec 2023 20:47:13 -0000 1.74
+++ Makefile 27 Dec 2023 13:45:07 -0000
@@ -53,10 +53,11 @@
LDFLAGS.Darwin+= -headerpad_max_install_names
-# Same as for py-numpy.
+# Same as for py-numpy. Or maybe not. Vendored superlu needs
+# direct linking with -lblas, not -lcblas.
BLAS_ACCEPTED= ${_BLAS_TYPES}
BLAS_C_INTERFACE= yes
-WHEEL_ARGS+= -Csetup-args=-Dblas=${CBLAS_PC}
+WHEEL_ARGS+= -Csetup-args=-Dblas=${BLAS_PC}
WHEEL_ARGS+= -Csetup-args=-Dlapack=${LAPACK_PC}
PYTHON_VERSIONS_INCOMPATIBLE= 27 38
It _seems_ that scipy doesn't actually call CBLAS API, as opposed to
numpy. Observe:
$ find /data/pkg/lib/python3.11/site-packages/scipy/ -name '*.so' | xargs -L 1 nm -g |grep cblas_
$ find /data/pkg/lib/python3.11/site-packages/numpy/ -name '*.so' | xargs -L 1 nm -g |grep cblas_
U cblas_caxpy
U cblas_cdotc_sub
U cblas_cdotu_sub
U cblas_cgemm
U cblas_cgemv
U cblas_csyrk
U cblas_daxpy
U cblas_ddot
U cblas_dgemm
U cblas_dgemv
U cblas_dsyrk
U cblas_saxpy
U cblas_sdot
U cblas_sgemm
U cblas_sgemv
U cblas_ssyrk
U cblas_zaxpy
U cblas_zdotc_sub
U cblas_zdotu_sub
U cblas_zgemm
U cblas_zgemv
U cblas_zsyrk
So the above could be a hotfix for the current version of scipy.
The wip/hmmlearn package will have to be fixed then, but has time for
after the branching.
Alrighty then,
Thomas