Anaconda 中的默认 numpy 与加速

Question

我刚刚在 Windows 的 Python 2.7.10 环境中通过 Anaconda 安装了 numpy-1.10.1。令我惊讶的是，我发现它有开箱即用的 MKL（请参阅下面的配置）。我运行a benchmark against a "manual" Python 2.7.10 install with numpy+mkl-1.10.1 from Gohlke，和他们显示的数字一样。而且配置是一样的。

我想知道Anaconda加速包给numpy带来了什么？

Anaconda numpy-1.10.1 配置

>>> np.__config__.show()
lapack_opt_info:
    libraries = ['mkl_lapack95_lp64', 'mkl_blas95_lp64', 'mkl_intel_lp64', 'mkl_intel_thread', 'mkl_core', 'libiomp5md', 'libifportmd', 'mkl_lapack95_lp64', 'mkl_blas95_lp64', 'mkl_intel_lp64', 'mkl_intel_thread', 'mkl_core', 'libiomp5md', 'libifportmd']
    library_dirs = ['C:/Program Files (x86)/Intel/Composer XE/mkl/lib/intel64']
    define_macros = [('SCIPY_MKL_H', None), ('HAVE_CBLAS', None)]
    include_dirs = ['C:/Program Files (x86)/Intel/Composer XE/mkl/include']
blas_opt_info:
    libraries = ['mkl_lapack95_lp64', 'mkl_blas95_lp64', 'mkl_intel_lp64', 'mkl_intel_thread', 'mkl_core', 'libiomp5md', 'libifportmd']
    library_dirs = ['C:/Program Files (x86)/Intel/Composer XE/mkl/lib/intel64']
    define_macros = [('SCIPY_MKL_H', None), ('HAVE_CBLAS', None)]
    include_dirs = ['C:/Program Files (x86)/Intel/Composer XE/mkl/include']
openblas_lapack_info:
  NOT AVAILABLE
lapack_mkl_info:
    libraries = ['mkl_lapack95_lp64', 'mkl_blas95_lp64', 'mkl_intel_lp64', 'mkl_intel_thread', 'mkl_core', 'libiomp5md', 'libifportmd', 'mkl_lapack95_lp64', 'mkl_blas95_lp64', 'mkl_intel_lp64', 'mkl_intel_thread', 'mkl_core', 'libiomp5md', 'libifportmd']
    library_dirs = ['C:/Program Files (x86)/Intel/Composer XE/mkl/lib/intel64']
    define_macros = [('SCIPY_MKL_H', None), ('HAVE_CBLAS', None)]
    include_dirs = ['C:/Program Files (x86)/Intel/Composer XE/mkl/include']
blas_mkl_info:
    libraries = ['mkl_lapack95_lp64', 'mkl_blas95_lp64', 'mkl_intel_lp64', 'mkl_intel_thread', 'mkl_core', 'libiomp5md', 'libifportmd']
    library_dirs = ['C:/Program Files (x86)/Intel/Composer XE/mkl/lib/intel64']
    define_macros = [('SCIPY_MKL_H', None), ('HAVE_CBLAS', None)]
    include_dirs = ['C:/Program Files (x86)/Intel/Composer XE/mkl/include']
mkl_info:
    libraries = ['mkl_lapack95_lp64', 'mkl_blas95_lp64', 'mkl_intel_lp64', 'mkl_intel_thread', 'mkl_core', 'libiomp5md', 'libifportmd']
    library_dirs = ['C:/Program Files (x86)/Intel/Composer XE/mkl/lib/intel64']
    define_macros = [('SCIPY_MKL_H', None), ('HAVE_CBLAS', None)]
    include_dirs = ['C:/Program Files (x86)/Intel/Composer XE/mkl/include']

（有趣的是，Link MKL to an installed Numpy in Anaconda? 报告了不同的配置。）

Gohlke numpy-1.10.1+mkl 配置

>>> np.__config__.show()
lapack_opt_info:
    libraries = ['mkl_lapack95_lp64', 'mkl_blas95_lp64', 'mkl_intel_lp64', 'mkl_intel_thread', 'mkl_core', 'libiomp5md', 'libifportmd', 'mkl_lapack95_lp64', 'mkl_blas95_lp64', 'mkl_intel_lp64', 'mkl_intel_thread', 'mkl_core', 'libiomp5md', 'libifportmd']
    library_dirs = ['C:/Program Files (x86)/Intel/Composer XE/mkl/lib/intel64']
    define_macros = [('SCIPY_MKL_H', None), ('HAVE_CBLAS', None)]
    include_dirs = ['C:/Program Files (x86)/Intel/Composer XE/mkl/include']
blas_opt_info:
    libraries = ['mkl_lapack95_lp64', 'mkl_blas95_lp64', 'mkl_intel_lp64', 'mkl_intel_thread', 'mkl_core', 'libiomp5md', 'libifportmd']
    library_dirs = ['C:/Program Files (x86)/Intel/Composer XE/mkl/lib/intel64']
    define_macros = [('SCIPY_MKL_H', None), ('HAVE_CBLAS', None)]
    include_dirs = ['C:/Program Files (x86)/Intel/Composer XE/mkl/include']
openblas_lapack_info:
  NOT AVAILABLE
lapack_mkl_info:
    libraries = ['mkl_lapack95_lp64', 'mkl_blas95_lp64', 'mkl_intel_lp64', 'mkl_intel_thread', 'mkl_core', 'libiomp5md', 'libifportmd', 'mkl_lapack95_lp64', 'mkl_blas95_lp64', 'mkl_intel_lp64', 'mkl_intel_thread', 'mkl_core', 'libiomp5md', 'libifportmd']
    library_dirs = ['C:/Program Files (x86)/Intel/Composer XE/mkl/lib/intel64']
    define_macros = [('SCIPY_MKL_H', None), ('HAVE_CBLAS', None)]
    include_dirs = ['C:/Program Files (x86)/Intel/Composer XE/mkl/include']
blas_mkl_info:
    libraries = ['mkl_lapack95_lp64', 'mkl_blas95_lp64', 'mkl_intel_lp64', 'mkl_intel_thread', 'mkl_core', 'libiomp5md', 'libifportmd']
    library_dirs = ['C:/Program Files (x86)/Intel/Composer XE/mkl/lib/intel64']
    define_macros = [('SCIPY_MKL_H', None), ('HAVE_CBLAS', None)]
    include_dirs = ['C:/Program Files (x86)/Intel/Composer XE/mkl/include']
mkl_info:
    libraries = ['mkl_lapack95_lp64', 'mkl_blas95_lp64', 'mkl_intel_lp64', 'mkl_intel_thread', 'mkl_core', 'libiomp5md', 'libifportmd']
    library_dirs = ['C:/Program Files (x86)/Intel/Composer XE/mkl/lib/intel64']
    define_macros = [('SCIPY_MKL_H', None), ('HAVE_CBLAS', None)]
    include_dirs = ['C:/Program Files (x86)/Intel/Composer XE/mkl/include']

Answer 1

使用问题中链接的基准测试脚本，我可以看到加速中的 numpy 和 Anaconda 中的 numpy 提供了相同的性能。

比较是在 Python 2.7.10 64 位 i7-4790K @ 4 GHz 和 32GB RAM 下进行的。由于accelerate目前会将numpy降级到1.9.3，所以我将它与numpy 1.9.3和numpy 1.10.1进行了比较以供参考

我的结论是，当只使用 numpy 时，加速不会带来额外的性能。

为了完整起见，这里是 numpy-1.9.3 在 accelerate 中的配置：

>>> np.__config__.show()
lapack_opt_info:
    libraries = ['mkl_lapack95_lp64', 'mkl_core_dll', 'mkl_intel_lp64_dll', 'mkl_intel_thread_dll']
    library_dirs = ['C:\aroot\stage\libs']
    define_macros = [('SCIPY_MKL_H', None)]
    include_dirs = ['C:\aroot\stage\include']
blas_opt_info:
    libraries = ['mkl_core_dll', 'mkl_intel_lp64_dll', 'mkl_intel_thread_dll']
    library_dirs = ['C:\aroot\stage\libs']
    define_macros = [('SCIPY_MKL_H', None)]
    include_dirs = ['C:\aroot\stage\include']
openblas_lapack_info:
  NOT AVAILABLE
lapack_mkl_info:
    libraries = ['mkl_lapack95_lp64', 'mkl_core_dll', 'mkl_intel_lp64_dll', 'mkl_intel_thread_dll']
    library_dirs = ['C:\aroot\stage\libs']
    define_macros = [('SCIPY_MKL_H', None)]
    include_dirs = ['C:\aroot\stage\include']
blas_mkl_info:
    libraries = ['mkl_core_dll', 'mkl_intel_lp64_dll', 'mkl_intel_thread_dll']
    library_dirs = ['C:\aroot\stage\libs']
    define_macros = [('SCIPY_MKL_H', None)]
    include_dirs = ['C:\aroot\stage\include']
mkl_info:
    libraries = ['mkl_core_dll', 'mkl_intel_lp64_dll', 'mkl_intel_thread_dll']
    library_dirs = ['C:\aroot\stage\libs']
    define_macros = [('SCIPY_MKL_H', None)]
    include_dirs = ['C:\aroot\stage\include']

Anaconda 中的默认 numpy 与加速

Default numpy vs accelerate in Anaconda

numpy

intel-mkl

anaconda