Info & discussion: NumPy and BLAS

[amitdo]

NumPy can use the BLAS library to do some of its operations.

On Debian 9, if you install numpy globally via apt-get, libblas3 will also be installed.
There are 2 other open source BLAS implementations which are generally more optimized for speed.
https://wiki.debian.org/DebianScience/LinearAlgebraLibraries

According to my testing you can get a nice speedup when using an accelerated BLAS lib with ocropy.

[Crabat]

I did some tests on a MacBook (late 2013, 2,8 GHz Intel Core i7, macOS High Sierra: 10.13.2, python-2.7.14, numpy-1.13.3, scipy-1.0.0).

Test suite: 50 MB real world training data, ocropus-rtrain -N 30000.

1. original code: 398min (100%)
2. with @amitdo Replace native code with regular functions #265 changes: 259min (65%)
3. numpy.einsum(): optimize=True: 266min
4. numpy.einsum(): optimize='optimal': 264min

Numpy is linked to the Accelerate Framework:

$ python2 -c 'import numpy; numpy.show_config()'
lapack_opt_info:
extra_link_args = ['-Wl,-framework', '-Wl,Accelerate']
extra_compile_args = ['-msse3']
define_macros = [('NO_ATLAS_INFO', 3), ('HAVE_CBLAS', None)]
openblas_lapack_info:
NOT AVAILABLE
atlas_3_10_blas_threads_info:
NOT AVAILABLE
atlas_threads_info:
NOT AVAILABLE
atlas_3_10_threads_info:
NOT AVAILABLE
atlas_blas_info:
NOT AVAILABLE
atlas_3_10_blas_info:
NOT AVAILABLE
atlas_blas_threads_info:
NOT AVAILABLE
openblas_info:
NOT AVAILABLE
blas_mkl_info:
NOT AVAILABLE
blas_opt_info:
extra_link_args = ['-Wl,-framework', '-Wl,Accelerate']
extra_compile_args = ['-msse3', '-I/System/Library/Frameworks/vecLib.framework/Headers']
define_macros = [('NO_ATLAS_INFO', 3), ('HAVE_CBLAS', None)]
blis_info:
NOT AVAILABLE
atlas_info:
NOT AVAILABLE
atlas_3_10_info:
NOT AVAILABLE
lapack_mkl_info:
NOT AVAILABLE

35% speed improvement. Thank you, @amitdo!

[amitdo]

You're welcome!

Apple's Accelerate Framework contains a highly tuned blas implementation.