numpy's QR appears faster than scipy's

For some reason, numpy's QR decomposition is always faster than scipy's which seems strange because scipy is supposed to include all of numpy plus more advanced features. Any idea why?

import numpy.linalg as nla
import scipy.linalg as la
A = np.random.randn(4000,4000)

%timeit -n 3 -r 3 Q,R = nla.qr(A)
%timeit -n 3 -r 3 Q,R = la.qr(A)
%timeit -n 3 -r 3 Q,R = nla.qr(A)
%timeit -n 3 -r 3 Q,R = la.qr(A)

3 loops, best of 3: 1.17 s per loop

3 loops, best of 3: 1.21 s per loop

3 loops, best of 3: 1.05 s per loop

3 loops, best of 3: 1.21 s per loop

The difference is more noticeable with mode="raw".

%timeit -n 3 -r 3 Q = nla.qr(A, mode='raw')
%timeit -n 3 -r 3 Q = la.qr(A, mode='raw')
%timeit -n 3 -r 3 Q = nla.qr(A, mode='raw')
%timeit -n 3 -r 3 Q = la.qr(A, mode='raw')

3 loops, best of 3: 440 ms per loop

3 loops, best of 3: 612 ms per loop

3 loops, best of 3: 436 ms per loop

3 loops, best of 3: 758 ms per loop

I'm using Intel MKL for both, and I have the latest dev versions

numpy version: 1.11.0.dev0+90a1a9f

scipy version: 0.17.0.dev0+8003fab

NumPy config:
blas_opt_info:
    library_dirs = ['/opt/intel/compilers_and_libraries_2016/linux/mkl/lib/intel64']
    include_dirs = ['/opt/intel/compilers_and_libraries_2016/linux/mkl/include']
    define_macros = [('SCIPY_MKL_H', None), ('HAVE_CBLAS', None)]
    libraries = ['mkl_rt', 'pthread']
mkl_info:
    library_dirs = ['/opt/intel/compilers_and_libraries_2016/linux/mkl/lib/intel64']
    include_dirs = ['/opt/intel/compilers_and_libraries_2016/linux/mkl/include']
    define_macros = [('SCIPY_MKL_H', None), ('HAVE_CBLAS', None)]
    libraries = ['mkl_rt', 'pthread']
lapack_opt_info:
    library_dirs = ['/opt/intel/compilers_and_libraries_2016/linux/mkl/lib/intel64']
    include_dirs = ['/opt/intel/compilers_and_libraries_2016/linux/mkl/include']
    define_macros = [('SCIPY_MKL_H', None), ('HAVE_CBLAS', None)]
    libraries = ['mkl_rt', 'pthread']
openblas_lapack_info:
  NOT AVAILABLE
lapack_mkl_info:
    library_dirs = ['/opt/intel/compilers_and_libraries_2016/linux/mkl/lib/intel64']
    include_dirs = ['/opt/intel/compilers_and_libraries_2016/linux/mkl/include']
    define_macros = [('SCIPY_MKL_H', None), ('HAVE_CBLAS', None)]
    libraries = ['mkl_rt', 'pthread']
blas_mkl_info:
    library_dirs = ['/opt/intel/compilers_and_libraries_2016/linux/mkl/lib/intel64']
    include_dirs = ['/opt/intel/compilers_and_libraries_2016/linux/mkl/include']
    define_macros = [('SCIPY_MKL_H', None), ('HAVE_CBLAS', None)]
    libraries = ['mkl_rt', 'pthread']

SciPy config:
blas_opt_info:
    libraries = ['mkl_rt', 'pthread']
    include_dirs = ['/opt/intel/compilers_and_libraries_2016/linux/mkl/include']
    define_macros = [('SCIPY_MKL_H', None), ('HAVE_CBLAS', None)]
    library_dirs = ['/opt/intel/compilers_and_libraries_2016/linux/mkl/lib/intel64']
mkl_info:
    libraries = ['mkl_rt', 'pthread']
    include_dirs = ['/opt/intel/compilers_and_libraries_2016/linux/mkl/include']
    define_macros = [('SCIPY_MKL_H', None), ('HAVE_CBLAS', None)]
    library_dirs = ['/opt/intel/compilers_and_libraries_2016/linux/mkl/lib/intel64']
blas_mkl_info:
    libraries = ['mkl_rt', 'pthread']
    include_dirs = ['/opt/intel/compilers_and_libraries_2016/linux/mkl/include']
    define_macros = [('SCIPY_MKL_H', None), ('HAVE_CBLAS', None)]
    library_dirs = ['/opt/intel/compilers_and_libraries_2016/linux/mkl/lib/intel64']
openblas_lapack_info:
  NOT AVAILABLE
lapack_mkl_info:
    libraries = ['mkl_rt', 'pthread']
    include_dirs = ['/opt/intel/compilers_and_libraries_2016/linux/mkl/include']
    define_macros = [('SCIPY_MKL_H', None), ('HAVE_CBLAS', None)]
    library_dirs = ['/opt/intel/compilers_and_libraries_2016/linux/mkl/lib/intel64']
lapack_opt_info:
    libraries = ['mkl_rt', 'pthread']
    include_dirs = ['/opt/intel/compilers_and_libraries_2016/linux/mkl/include']
    define_macros = [('SCIPY_MKL_H', None), ('HAVE_CBLAS', None)]
    library_dirs = ['/opt/intel/compilers_and_libraries_2016/linux/mkl/lib/intel64']

链接地址: http://www.djcxy.com/p/85702.html

上一篇: 由于类之间的循环依赖性，解决构建错误

下一篇: numpy的QR出现比scipy快