Xtensor：无法达到 numpy 性能

Question

我正在学习 xtensor 并希望获得与 NumPy 相同甚至更高的性能。但不幸的是，我不能也需要帮助。

我做了与here类似的基准测试：

Performance of xtensor types vs. NumPy for simple reduction

这是 C++ 代码，我在这里使用了 pybind11 和 xtensor-python

bench.cpp

#include <iostream>

#define XTENSOR_USE_XSIMD
#include "xtensor/xtensor.hpp"
#include "xtensor/xfixed.hpp"
#include "xtensor/xarray.hpp"
#include "xtensor/xio.hpp"
#include "xtensor/xview.hpp"
#define FORCE_IMPORT_ARRAY                // numpy C api loading
#include "xtensor-python/pytensor.hpp"
#include "xtensor-python/pyarray.hpp"

namespace py = pybind11;

inline double sum_pytensor(xt::pytensor<double, 1> &m)
{
  return xt::sum(m)();
}



inline double sum_pytensor_immediate(xt::pytensor<double, 1> &m)
{
  return xt::sum(m, xt::evaluation_strategy::immediate)();
}

PYBIND11_MODULE(xtensor_basics, m)
{
    xt::import_numpy();

    m.def("compute_xtensor", &sum_pytensor);
    m.def("compute_xtensor_immediate", &sum_pytensor_immediate);
}

我用 CMake 构建它

CMakeLists.txt

cmake_minimum_required(VERSION 2.8.12)
project(xtensor_basics)

add_definitions(-DXTENSOR_ENABLE_XSIMD) # <-- does this anything?
add_definitions(-DXTENSOR_USE_XSIMD)
add_subdirectory(pybind11)
pybind11_add_module(xtensor_basics bench.cpp)

include_directories(/home/--user--/include)
include_directories(/home/--user--/.miniconda3/lib/python3.7/site-packages/numpy/core/include)

和以下命令：cmake . && make 创建xtensor_basics.cpython-37m-x86_64-linux-gnu.so

然后我用这个 python 文件运行基准测试： bench.py

import timeit

def time_each(func_names, sizes):
    setup = f'''
import numpy; import xtensor_basics
arr = numpy.random.randn({sizes})
    '''
    tim = lambda func: min(timeit.Timer(f'{func}(arr)',
                                        setup=setup).repeat(3, 100))
    return [tim(func) for func in func_names]

from functools import partial

sizes = [10 ** i for i in range(7)]
funcs = ['numpy.sum',
         'xtensor_basics.compute_xtensor_immediate',
         'xtensor_basics.compute_xtensor']
sum_timer = partial(time_each, funcs)
times = list(map(sum_timer, sizes))
print(times)
from matplotlib import pyplot as plt 

plt.Figure(figsize=(5, 10))
plt.plot(times)
plt.legend(["numpy", "xtensor_immediate", "xtensor"])
plt.show()

结果：

目录（构建后）

bench.cpp
bench.py
CMakeCache.txt
CMakeFiles
cmake_install.cmake
CMakeLists.txt
Makefile
pybind11            <---clonned from the repo
xtensor_basics.cpython-37m-x86_64-linux-gnu.so

包括目录 所有包含标题的文件夹（我没有构建这些库，只是复制了标题）

$ ls /home/--user--/include -1
   xflens
   xsimd
   xtensor
   xtensor-blas
   xtensor-python
   xtl

系统

Ubuntu                         18.04
g++                            7.4.0
numpy                          1.16.4
openblas                       0.2.20
python                         3.7.3
xtensor                        0.20.8 

问题：我应该添加哪些标志、定义等才能获得相同的性能？

提前致谢。

编辑：1 当我使用cmake -DCMAKE_BUILD_TYPE=Release . 构建时，即启用优化，结果有所改善，但仍然较慢：

Answer 1

改一下CMakeLists.txt：

cmake_minimum_required(VERSION 2.8.12)
project(xtensor_basics)

add_definitions(-DXTENSOR_ENABLE_XSIMD)
add_definitions(-DXTENSOR_USE_XSIMD)\
set(CMAKE_CXX_FLAGS_RELEASE "${CMAKE_CXX_FLAGS_RELEASE} -O3 -mavx2 -ffast-math")
#                                                           ^^^^^^^^^^^^^^^^^^^

add_subdirectory(pybind11)
pybind11_add_module(xtensor_basics bench.cpp)

include_directories(/home/--user--/include)
include_directories(/home/--user--/.miniconda3/lib/python3.7/site-packages/numpy/core/include)

还有……胜利！