BLAS and LAPACK#

Selecting BLAS and LAPACK libraries#

BLAS and LAPACK library selection, other than the OpenBLAS default, is implemented via Meson build options. For example, to select plain libblas and liblapack (this is typically Netlib BLAS/LAPACK on Linux distros, and can be dynamically switched between implementations on conda-forge), use:

$ # for a development build
$ spin build -S-Dblas=blas -S-Dlapack=lapack

$ # to build and install a wheel
$ python -m build -Csetup-args=-Dblas=blas -Csetup-args=-Dlapack=lapack
$ pip install dist/scipy*.whl

$ # Or, with pip>=23.1, this works too:
$ python -m pip -Csetup-args=-Dblas=blas -Csetup-args=-Dlapack=lapack

Other options that should work (as long as they’re installed with pkg-config or CMake support) include:

  • Accelerate: -Dblas=accelerate,

  • MKL: use the name of the relevant pkg-config file, e.g. -Dblas=mkl-dynamic-lp64-gomp,

  • BLIS: -Dblas=blis (is BLAS-only, so will need a -Dlapack= as well),

  • ATLAS: -Dblas=blas-atlas (name of the pkg-config file may vary)

Note that both Accelerate and scipy-openblas have flags in spin that are easier to remember, since they’re commonly used for development:

$ spin build --with-accelerate
$ spin build --with-scipy-openblas=32

The -Dlapack flag isn’t needed for Accelerate, MKL or scipy-openblas, since we can be sure that BLAS and LAPACK are the same for those options. E.g., to create a wheel with Accelerate, use:

$ python -m build -Csetup-args=-Dblas=accelerate

Using pkg-config to detect libraries in a nonstandard location#

The way BLAS and LAPACK detection works under the hood is that Meson tries to discover the specified libraries first with pkg-config, and then with CMake. If all you have is a standalone shared library file (e.g., armpl_lp64.so in /a/random/path/lib/ and a corresponding header file in /a/random/path/include/), then what you have to do is craft your own pkg-config file. It should have a matching name (so in this example, armpl_lp64.pc) and may be located anywhere. The PKG_CONFIG_PATH environment variable should be set to point to the location of the .pc file. The contents of that file should be:

libdir=/path/to/library-dir      # e.g., /a/random/path/lib
includedir=/path/to/include-dir  # e.g., /a/random/path/include
version=1.2.3                    # set to actual version
extralib=-lm -lpthread -lgfortran   # if needed, the flags to link in dependencies
Name: armpl_lp64
Description: ArmPL - Arm Performance Libraries
Version: ${version}
Libs: -L${libdir} -larmpl_lp64      # linker flags
Libs.private: ${extralib}
Cflags: -I${includedir}

To check that this works as expected, you should be able to run:

$ pkg-config --libs armpl_lp64
-L/path/to/library-dir -larmpl_lp64
$ pkg-config --cflags armpl_lp64
-I/path/to/include-dir

Specifying the Fortran ABI to use#

Some linear algebra libraries are built with the g77 ABI (also known as “the f2c calling convention”) and others with GFortran ABI, and these two ABIs are incompatible. Therefore, if you build SciPy with gfortran and link to a linear algebra library like MKL, which is built with a g77 ABI, there’ll be an exception or a segfault. SciPy fixes this by using ABI wrappers which rely on the CBLAS API for the few functions in the BLAS API that suffer from this issue.

Note that SciPy needs to know at build time what needs to be done. The build system will automatically check whether linear algebra library is MKL or Accelerate (which both always use the g77 ABI) and if so, use the CBLAS API instead of the BLAS API. If autodetection fails or if the user wants to override this autodetection mechanism for building against plain libblas/liblapack (this is what conda-forge does for example), use the -Duse-g77-abi=true build option. E.g.,:

$ python -m build -C-Duse-g77-abi=true -Csetup-args=-Dblas=blas -Csetup-args=-Dlapack=lapack

64-bit integer (ILP64) BLAS/LAPACK#

Support for ILP64 BLAS and LAPACK, as of version 1.18.0, requires that also LP64 symbols are available in the same library. Hence, ILP64 support is only available MKL and Accelerate. Note: after the deprecated scipy.odr (the last Fortran module) is removed, this restriction will be removed and OpenBLAS and other libraries will also be supported.

To build SciPy from source with ILP64 support, use -Duse-ilp64=true. E.g., to use Accelerate on macOS:

$ python -m build --wheel -Csetup-args=-Dblas=accelerate -C-Duse-ilp64=true -Dcython-blas-abi=lp64

And to use MKL on x86-64:

# Note the "lp64" in the `-Dblas=` argument is not a mistake; this is
# necessary as long as the cython_blas ABI is set to "lp64" (see next section)
$ python -m build --wheel -Csetup-args=-Dblas=mkl-dynamic-lp64-seq -C-Duse-ilp64=true -Dcython-blas-abi=lp64

Note

These instructions will change for scipy>=1.19.0, when the scipy.linalg.cython_blas ABI doesn’t need to be forced to LP64 anymore. Please see the next section for more details.

Building with -Duse-ilp64=true by default flips the low-level Python and Cython APIs in scipy.linalg to ILP64 as well. This may require downstream usage to be made compatible.

From Python, low-level BLAS and LAPACK functions are available from scipy.linalg.blas and scipy.linalg.lapack namespaces:

>>> from scipy.linalg.blas import dgemm  # this may be an LP64 or an ILP64 function

To choose the variant of a low-level routine, use get_blas_funcs and get_lapack_funcs functions:

>>> from scipy.linalg.blas import get_blas_funcs
>>> daxpy = get_blas_funcs('axpy', (np.ones(3),), ilp64='preferred')
>>> daxpy.int_dtype
dtype('int64')       # depends on the build option

High-level linear algebra functions (norm, solve and so on) should use this mechanism under the hood, so they work transparently with either LP64 or ILP64 routines.

Cython BLAS/LAPACK integer ABI#

The Cython BLAS and LAPACK API (scipy.linalg.cython_blas and scipy.linalg.cython_lapack) uses the blas_int type for all integer parameters. By default, blas_int follows the use-ilp64 setting: it is int (32-bit) for LP64 builds and int64_t (64-bit) for ILP64 builds.

This can be overridden with the -Dcython-blas-abi build option, which accepts three values:

  • auto (default): follows the use-ilp64 setting

  • lp64: always use 32-bit integers, even when use-ilp64=true

  • ilp64: always use 64-bit integers

For example, to build with ILP64 BLAS/LAPACK but keep the Cython API at LP64 (for downstream compatibility):

$ spin build -S-Dblas=accelerate -S-Duse-ilp64=true -S-Dcython-blas-abi=lp64

This is most useful for Accelerate, which has good support for using both LP64 and ILP64 at the same time.

MKL also supports this (it has ILP64 symbols with _64 symbol suffix available in mkl-dynamic-lp64-* shared libraries). MKL’s support will work fine for a Python-only stack, as long as you ensure to not link to different shared libraries - avoid mixing mkl_rt.so, mkl-dynamic-lp64-*.so and mkl-dynamic-ilp64-*.so.

It may be useful when starting to introduce ILP64 usage to have SciPy itself use ILP64 but keeping the Cython API at LP64, because downstream packages may not yet support 64-bit integers in their Cython BLAS/LAPACK calls.

The build configuration can be checked at runtime via scipy.show_config() — look for the 'blas cython ilp64' entry.

Work-in-progress#

These options are planned to be fully supported, but currently not usable out of the box:

  • ILP64 (64-bit integer size) builds: making -Duse-ilp64=true work without using another option to avoid issues with the deprecated scipy.odr. Those options are: -Dcython-blas-abi=lp64, or -D_without-fortran=true. The latter is a temporary option that doesn’t use a Fortran compiler at all and doesn’t install scipy.odr.

  • Automatically selecting from multiple possible BLAS and LAPACK options, with a user-provided order of precedence