Debian and derivatives: Fortran compiler, Python headers, lapack
-$ sudo apt install python3-dev gfortran liblapack-dev
+Fortran compiler, Python headers, lapack, meson on Debian/Ubuntu:
+$ sudo apt install python3-dev gfortran liblapack-dev meson
Then
@@ -191,13 +190,6 @@ Dependenciesase) won’t fail at import time, but later at runtime when e.g.
crys.Structure.get_ase_atoms() is called. What packages are optional might
change depending on usage.
-You also need the following to compile extensions. On a Debian-ish system
-# apt
-python3-dev # for compiling extensions
-gfortran # or ifort, see src/Makefile
-liblapack-dev
-
-
Also note that you can get most Python packages via your system’s package
manager. Debian and derivatives:
python3-ase
@@ -240,16 +232,9 @@ Optional dependenciesRunning tests¶
See test/README. Actually, all of these are good examples, too!
-
-Python versions¶
-Only Python3 is supported, tested: Python 3.6-3.11
-The package was developed mostly with Python 2.5-2.7 and ported using 2to3 +
-manual changes. Therefore, you might find some crufty Python 2 style code
-fragments in lesser used parts of the code base.
-
Fortran extensions and OpenMP notes¶
-Use src/Makefile:
+Use src/_ext_src/Makefile:
$ make help
make gfortran # gfortran, default
make gfortran-omp # gfortran + OpenMP
@@ -259,21 +244,36 @@ Python versions ifort-mkl # ifort, Intel MKL lapack, set MKL_LIB
-Generates *.so and *.pyf (f2py interface) files.
-You need:
-
-numpy for f2py
a Fortran compiler
Python headers (Debian/Ubuntu: python-dev)
Lapack (Debian: liblapack-dev)
-The module is compiled with f2py (currently part of numpy, tested with numpy
-1.1.0 .. 1.21.x).
+Generates *.so files
+src/_ext_src/_dcd.so # will be copied to src/pwtools/_dcd.so
+src/_ext_src/_flib.so # will be copied to src/pwtools/_flib.so
+
+
+You need the following to compile extensions. On Debian/Ubuntu
+# apt
+python3-numpy # for f2py
+python3-dev # Python headers
+gfortran # or ifort, see src/_ext_src/Makefile
+liblapack-dev # see below
+meson # f2py build backend as of Python 3.12, will also install ninja
+
+
+liblapack-dev is only one option, maybe the slowest. On Debian/Ubuntu,
+there are other options, i.e. anything that provides a liblapack.so works.
+$ apt-file search -x 'liblapack.so$'
+libatlas-base-dev: /usr/lib/x86_64-linux-gnu/atlas/liblapack.so
+liblapack-dev: /usr/lib/x86_64-linux-gnu/lapack/liblapack.so
+libmkl-dev: /usr/lib/x86_64-linux-gnu/mkl/liblapack.so
+libopenblas-openmp-dev: /usr/lib/x86_64-linux-gnu/openblas-openmp/liblapack.so
+libopenblas-pthread-dev: /usr/lib/x86_64-linux-gnu/openblas-pthread/liblapack.so
+libopenblas-serial-dev: /usr/lib/x86_64-linux-gnu/openblas-serial/liblapack.so
+
+
The default make target is “gfortran” which tries to build a serial version
-using system BLAS and LAPACK (e.g. from liblapack-dev). If you want another
+using system LAPACK (e.g. from liblapack-dev). If you want another
target (e.g. ifort-mkl), then
-$ cd src
+$ cd src/_ext_src
$ make clean
$ make ifort-mkl
@@ -284,9 +284,9 @@ Python versions$ PWTOOLS_EXT_MAKE_TARGET=ifort-mkl pip install ...
-This will use the Intel ifort compiler instead fo the default gfortran and
+
This will use the Intel ifort compiler instead of the default gfortran and
link against the MKL.
-In the MKL case, the Makefile uses the env var $MKL_LIB which sould point
+
In the MKL case, the Makefile uses the env var $MKL_LIB which should point
to the location where things like libmkl_core.so live. You may need to set
this. On a HPC cluster, that could look like this.
# module only sets MKL_ROOT=/path/to/intel/20.4/mkl
@@ -294,24 +294,7 @@ Python versions MKL_LIB=$MKL_ROOT/lib/intel64 PWTOOLS_EXT_MAKE_TARGET=ifort-mkl pip install ...
-See src/Makefile for more details.
-
-Compiler / f2py¶
-Instead of letting numpy.distutils pick a compiler + special flags, which is
-not trivial and therefore almost never works, it is much easier to simply
-define the compiler to use + architecture-specific flags. See F90 and ARCH in
-the Makefile.
-Also, numpy.distutils has default -03 for fcompiler. --f90flags="-02" does NOT
-override this. We get -O3 -O2 and a compiler warning. We have to use f2py’s
---opt= flag.
-On some systems (Debian), you may have:
-/usr/bin/f2py -> f2py2.6
-/usr/bin/f2py2.5
-/usr/bin/f2py2.6
-
-
-and such. But usually F2PY=f2py is fine.
-
+See src/_ext_src/Makefile for more details.
OpenMP¶
We managed to speed up the calculations by sprinkling some OpenMP
@@ -330,14 +313,8 @@
OpenMP¶
If this env var is NOT set, then OpenMP uses all available cores (e.g. 4 on a
quad-core box).
-
-- IMPORTANT:
Note that we may have found a f2py bug (see test/test_f2py_flib_openmp.py)
-re. OMP_NUM_THREADS. We have a workaround for that in pydos.fvacf().
-
-There is also an optional arg ‘nthreads’ to _flib.vacf(). If this is
-supplied, then it will override OMP_NUM_THREADS. Currently, this is the
-safest way to set the number of threads.
+There is also an optional arg ‘nthreads’ to _flib.vacf(). If this is
+supplied, then it will override OMP_NUM_THREADS.
Tests¶