Baseline setup

Lets start with the basics.

cd "$NUMPY_DIR"
micromamba create -f environment.yml
micromamba activate numpy-dev
micromamba remove openblas scipy numpy
micromamba install pyyaml # prettier show_config()

Now we can setup the OpenBLAS unencumbered by the tainetd conda variant.

Replace "$OBPATH" with the actual path to the OpenBLAS source.

NumPy setup

Having setup the .pc file with any of those above.. Which works with:

CFLAGS="-O0 -g -ggdb3" \
CXXFLAGS="-O0 -g -ggdb3" \
spin build --clean \
-- -Dblas="openblas" \
-Dblas-order=openblas,mkl,blis \
-Dlapack-order=openblas,mkl,lapack \
-Dallow-noblas=true \
-Dpkg_config_path="${OBPATH}" \
-Dbuildtype=debug \
-Ddisable-optimization=true

Where we need noblas to be allowed since LAPACK won’t be found otherwise.

Additionally, remember to set LD_LIBRARY_PATH to prevent inadvertently picking up system openblas.

make variant

The working make variant is basically built with:

make -j$(nproc)
make PREFIX=$(pwd)/tmpmake install
make clean # to remove .o files

With the following openblas.pc file:

Name: openblas
Description: OpenBLAS make
Version: 0.3.26.dev_make
prefix=${OBPATH}/tmpmake
includedir=${prefix}/include
libdir=${prefix}/lib
Cflags: -I${includedir}/
Libs: -L${libdir} -lopenblas

While running anything within spin run $SHELL we need to first setup the dynamic loading location:

export LD_LIBRARY_PATH="${OBPATH}/tmpmake/lib"

meson variant

Name: openblas
Description: OpenBLAS meson
Version: 0.3.26.dev_meson
prefix=${OBPATH}
includedir=${prefix}/tmpmake/include
libdir=${prefix}/local/lib
Cflags: -I${includedir}/
Libs: -L${libdir} -lopenblas