Hello Forum,

I am a Python Numpy user in the past. Right now I would like to do the same numerical linear algebra calculation in C++ so I select Eigen.
When I was doing some performance test with Eigen, I find some interesting thing regard to compilation of the source code.

I am writing a simple sigmoid fuction in a eigentest.cpp source code:

Code: Select all

#include <iostream>
#include <Eigen/Dense>
#include <time.h>

using Eigen::MatrixXd;
using namespace std;

int main() {
   MatrixXd m = MatrixXd::Random(3000,3000);
   time_t start = clock();
   m = -m;
   MatrixXd sig = 1 / (1 + m.array().exp());
   cout << double(clock() - start) / CLOCKS_PER_SEC * 1000 << "ms" << endl;
   return 0;
}


I do the compilation using the simple command:

Code: Select all

g++ -I. eigentest.cpp -o eigentest

When I run the executable it returns an elapsed time of roughly 900ms. It is even slower than the Numpy.
So I decide to do some research until I find this post on stackoverflow.
http://stackoverflow.com/questions/24566920/eigen-matrix-vs-numpy-array-multiplication-performance

Someone mentioned to add the compilation option like this:

Code: Select all

g++ -I. eigentest.cpp -o eigentest -O2 -DNDEBUG

Then I ran the program again. This time I got 170ms elapsed time. This is much faster than python Numpy.

I would like to ask what did the -O2 and -DNDEBUG switch do in the compilation that boost the performance here?
Do I need to apply this switch every time in my future compilation to get the best performance out of Eigen library?

Thank you very much!

It is a g++ related question, not an eigen one.

https://gcc.gnu.org/onlinedocs/gcc/Opti ... ze-Options

ctrl+f for -o2

mazz wrote:It is a g++ related question, not an eigen one.

https://gcc.gnu.org/onlinedocs/gcc/Opti ... ze-Options

ctrl+f for -o2

I think I should apply this optimization flag every time when compiling with g++.

Then what about compile on Windows platform?

According to the Configuring your project for optimal performance in this link: http://eigen.tuxfamily.org/index.php?title=Visual_Studio, I compiled the same code in visual studio 2010 professional using all the optimizer setting under Project property. This gave me an even worse result of 3500ms elapsed time.

I will try osx compile when I have chance.

Just a little bit confused on how to get best performance out of compiling on each platform.

Enabling compiler optimizations is indeed crucial in C++ to get rid of the deep function calls through inlining. -O2 is the minimal, -O3 is slightly better. This flag applies to both gcc and clang (default on OSX). For visual, the flag if /O2, or "release mode" through the interface.