Xtensor: can't reach numpy performance

Question

I'm learning xtensor and want to get the same or even higher performance then NumPy. But unfortunately, I can't and need help.

I did similar benchmark as here:

Performance of xtensor types vs. NumPy for simple reduction

This is C++ code, where I used pybind11 and 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);
}

I build this with CMake

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)

and the following command: cmake . && make which creates xtensor_basics.cpython-37m-x86_64-linux-gnu.so

Then I run the benchmark with this python file: 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()

Result:

Directory (after building)

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

Include dirrectory All folders containing headers (I didn't build these libraries, just copied headers)

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

System

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

Question: What flags, definitions etc. should I add to get the same performance?

Thanks in advance.

EDIT: 1 When I built with cmake -DCMAKE_BUILD_TYPE=Release ., i.e enabling optimisations, the result has improved, but still slower:

Answer 1

Change CMakeLists.txt a bit:

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)

and... Victory!