为什么c比python快_为什么C比python快得多？

我的目标是在

Python中为频谱有限元编写一个小型库,为此我尝试使用Boost扩展python和C库,希望它能使我的代码更快.

class Quad {

public:

Quad(int, int);

double integrate(boost::function const&)> const&);

double integrate_wrapper(boost::python::object const&);

std::vector nodes;

std::vector weights;

};

...

namespace std {

typedef std::vector > cube;

typedef std::vector mat;

typedef std::vector vec;

}

...

double Quad::integrate(boost::function const& func) {

double result &＃61; 0.;

for (unsigned int i &＃61; 0; i

result &＃43;&＃61; func(nodes[i]) * weights[i];

}

return result;

}

// ---- PYTHON WRAPPER ----

double Quad::integrate_wrapper(boost::python::object const& func) {

std::function lambda;

switch (this->nodes[0].size()) {

case 1: lambda &＃61; [&func](vec const& v) -> double { return boost::python::extract(func (v[0])); }; break;

case 2: lambda &＃61; [&func](vec const& v) -> double { return boost::python::extract(func(v[0], v[1])); }; break;

case 3: lambda &＃61; [&func](vec const& v) -> double { return boost::python::extract(func(v[0], v[1], v[2])); }; break;

default: cout <<"Dimension must be 1, 2, or 3" <

}

return integrate(lambda);

}

// ---- EXPOSE TO PYTHON ----

BOOST_PYTHON_MODULE(hermite)

{

using namespace boost::python;

class_("double_vector")

.def(vector_indexing_suite())

;

class_("double_mat")

.def(vector_indexing_suite())

;

class_("Quad", init())

.def("integrate", &Quad::integrate_wrapper)

.def_readonly("nodes", &Quad::nodes)

.def_readonly("weights", &Quad::weights)

;

}

我比较了三种不同方法的性能来计算两个函数的积分.这两个功能是&＃xff1a;

>函数f1(x,y,z)&＃61; x * x

>一个更难评估的函数&＃xff1a;f2(x,y,z)&＃61; np.cos(2 * x 2 * y 2 * z)x * y np.exp(-z * z)np.cos(2 * x 2 * y 2 * z)x * y np.exp(-z * z)np.cos(2 * x 2 * y 2 * z)x * y np.exp(-z * z)np.cos (2 * x 2 * y 2 * z)x * y np.exp(-z * z)np.cos(2 * x 2 * y 2 * z)x * y np.exp(-z * z)

使用的方法是&＃xff1a;

>从C程序调用库&＃xff1a;

double func(vector v) {

return F1_OR_F2;

}

int main() {

hermite::Quad quadrature(100, 3);

double result &＃61; quadrature.integrate(func);

cout <<"Result &＃61; " <

}

>从Python脚本调用库&＃xff1a;

import hermite

def function(x, y, z): return F1_OR_F2

my_quad &＃61; hermite.Quad(100, 3)

result &＃61; my_quad.integrate(function)

>在Python中使用for循环&＃xff1a;

import hermite

def function(x, y, z): return F1_OR_F2

my_quad &＃61; hermite.Quad(100, 3)

weights &＃61; my_quad.weights

nodes &＃61; my_quad.nodes

result &＃61; 0.

for i in range(len(weights)):

result &＃43;&＃61; weights[i] * function(nodes[i][0], nodes[i][1], nodes[i][2])

以下是每种方法的执行时间(使用方法1的time命令测量时间,方法2和3使用python模块时间,使用Cmake和set编译C代码(CMAKE_BUILD_TYPE Release))

>对于f1&＃xff1a;

>方法1&＃xff1a;0.07s用户0.01s系统99&＃xff05;cpu 0.083总计

>方法2&＃xff1a;0.19s

>方法3&＃xff1a;3.06s

>对于f2&＃xff1a;

>方法1&＃xff1a;0.28s用户0.01s系统99&＃xff05;cpu 0.289总计

>方法2&＃xff1a;12.47s

>方法3&＃xff1a;16.31s

根据这些结果,我的问题如下&＃xff1a;

>为什么第一种方法比第二种方法快得多&＃xff1f;

>可以改进python包装器以达到方法1和方法2之间相当的性能吗&＃xff1f;

>为什么方法2比方法3更难以集成函数的难度&＃xff1f;

编辑&＃xff1a;我还尝试定义一个接受字符串作为参数的函数,将其写入文件,然后继续编译文件并动态加载生成的.so文件&＃xff1a;

double Quad::integrate_from_string(string const& function_body) {

// Write function to file

ofstream helper_file;

helper_file.open("/tmp/helper_function.cpp");

helper_file <<"#include \n#include \n";

helper_file <<"extern \"C\" double toIntegrate(std::vector v) {\n";

helper_file <<" return " <

helper_file.close();

// Compile file

system("c&＃43;&＃43; /tmp/helper_function.cpp -o /tmp/helper_function.so -shared -fPIC");

// Load function dynamically

typedef double (*vec_func)(vec);

void *function_so &＃61; dlopen("/tmp/helper_function.so", RTLD_NOW);

vec_func func &＃61; (vec_func) dlsym(function_so, "toIntegrate");

double result &＃61; integrate(func);

dlclose(function_so);

return result;

}

它非常脏,可能不太便携,所以我很乐意找到一个更好的解决方案,但它运行良好,并且与sympy的ccode功能很好地配合.

第二次编辑我用纯粹的Python使用Numpy重写了这个函数.

import numpy as np

import numpy.polynomial.hermite_e as herm

import time

def integrate(function, degrees):

dim &＃61; len(degrees)

nodes_multidim &＃61; []

weights_multidim &＃61; []

for i in range(dim):

nodes_1d, weights_1d &＃61; herm.hermegauss(degrees[i])

nodes_multidim.append(nodes_1d)

weights_multidim.append(weights_1d)

grid_nodes &＃61; np.meshgrid(*nodes_multidim)

grid_weights &＃61; np.meshgrid(*weights_multidim)

nodes_flattened &＃61; []

weights_flattened &＃61; []

for i in range(dim):

nodes_flattened.append(grid_nodes[i].flatten())

weights_flattened.append(grid_weights[i].flatten())

nodes &＃61; np.vstack(nodes_flattened)

weights &＃61; np.prod(np.vstack(weights_flattened), axis&＃61;0)

return np.dot(function(nodes), weights)

def function(v): return F1_OR_F2

result &＃61; integrate(function, [100,100,100])

print("-> Result &＃61; " &＃43; str(result) &＃43; ", Time &＃61; " &＃43; str(end-start))

有点令人惊讶(至少对我而言),此方法与纯C实现之间的性能没有显着差异.特别是,f1需要0.059s,f2需要0.36s.