2

I would like to know (and ideally select) which blas implementation is used by numpy. On my system I have openblas and blis installed system wide. For numpy I created a new virtual environment and compiled numpy from source. Then I open python, import numpy and run numpy.show_config(), as suggested in answer. 

blas_mkl_info:
  NOT AVAILABLE
blis_info:
    libraries = ['blis', 'blis']
    library_dirs = ['/usr/lib64']
    define_macros = [('HAVE_CBLAS', None)]
    include_dirs = ['/usr/local/include', '/usr/include', '/home/wichmann/pythontests/blis_numpy/include']
    language = c
blas_opt_info:
    libraries = ['blis', 'blis']
    library_dirs = ['/usr/lib64']
    define_macros = [('HAVE_CBLAS', None)]
    include_dirs = ['/usr/local/include', '/usr/include', '/home/wichmann/pythontests/blis_numpy/include']
    language = c
lapack_mkl_info:
  NOT AVAILABLE
openblas_lapack_info:
    libraries = ['openblas', 'openblas']
    library_dirs = ['/usr/lib64']
    language = c
    define_macros = [('HAVE_CBLAS', None)]
lapack_opt_info:
    libraries = ['openblas', 'openblas']
    library_dirs = ['/usr/lib64']
    language = c
    define_macros = [('HAVE_CBLAS', None)]

However, this only shows against which blas implementations numpy was linked and not which are used during runtime. I also dug a bit deeper and used ldd, to no avail. In my virtual environment for blis_numpy/lib/python3.8/site-packages/numpy-1.19.0.dev0+8b81727-py3.8-linux-x86_64.egg/numpy/linalg/_umath_linalg.cpython-38-x86_64-linux-gnu.so I get

linux-vdso.so.1 (0x00007ffc68bce000)
libopenblas.so.0 => /lib64/libopenblas.so.0 (0x00007fd678cc5000)
libpthread.so.0 => /lib64/libpthread.so.0 (0x00007fd678ca3000)
libc.so.6 => /lib64/libc.so.6 (0x00007fd678ad9000)
libm.so.6 => /lib64/libm.so.6 (0x00007fd678993000)
libgfortran.so.5 => /lib64/libgfortran.so.5 (0x00007fd6786cb000)
/lib64/ld-linux-x86-64.so.2 (0x00007fd67afe1000)
libquadmath.so.0 => /lib64/libquadmath.so.0 (0x00007fd678681000)
libgcc_s.so.1 => /lib64/libgcc_s.so.1 (0x00007fd678664000)

and for blis_numpy/lib/python3.8/site-packages/numpy-1.19.0.dev0+8b81727-py3.8-linux-x86_64.egg/numpy/core/_multiarray_umath.cpython-38-x86_64-linux-gnu.so I get

linux-vdso.so.1 (0x00007ffff2da9000)
libblis.so.2 => /lib64/libblis.so.2 (0x00007fd4b62b3000)
libm.so.6 => /lib64/libm.so.6 (0x00007fd4b616d000)
libpthread.so.0 => /lib64/libpthread.so.0 (0x00007fd4b614b000)
libc.so.6 => /lib64/libc.so.6 (0x00007fd4b5f81000)
/lib64/ld-linux-x86-64.so.2 (0x00007fd4b73d3000)
librt.so.1 => /lib64/librt.so.1 (0x00007fd4b5f76000)

So will numpy use openblas or blis? Or both, depending on the function used? Mostly I am interested in which path numpy.linalg.eigvalsh takes, as I have the hope that one implementation or the other might be a bit faster.

laolux
  • 1,445
  • 1
  • 17
  • 28
  • 1
    Not an answer, but note that if speed is a concern, there is also [intel-numpy](https://pypi.org/project/intel-numpy/) which is linked with MKL. –  May 15 '20 at 05:45
  • Thanks, I tried, but it does not work. I am using python 3.8 which seems to be too new. I guess I would have to get python 3.6, but I am not willing to screw up my system just for that. I could also use the intel python distribution with everything included, but that's just huge and with many outdated packages which need to be updated and so on -> lots of work. – laolux May 15 '20 at 07:03

0 Answers0