Hi,

Eigen::Map<Eigen::MatrixXd> matA((double*)A.getDataPointer(), (Eigen::Index)A.getDimensions().getRows(), (Eigen::Index)A.getDimensions().getColumns());
Eigen::Map<Eigen::MatrixXd> matR((double*)R.getDataPointer(), (Eigen::Index)R.getDimensions().getRows(), (Eigen::Index)R.getDimensions().getColumns());
tic();
matR = matA.inverse();
toc();

On my pc (VS 2015 release) with a matrix 1000x1000
Elapsed time is 0.590442 seconds

With a pure lapack code based on DLANGE, DGETRF, DGECON, DGETRI with same matrix 1000x1000
Elapsed time is 0.378 seconds.

Maybe, inverse could use lapacke ?

Thanks for your great work about eigen

The problem might come from your compiler or compiling options. Here, using either gcc or clang, it takes 0.1s on a 2.6GHz i7 (haswell) and Eigen 3.3. If I disable AVX/FMA, it takes, as expected, 0.25s. So unless your computer is 10 years old, 0.59s seems indeed too slow.

Code: Select all

#include <iostream>
#include <Eigen/Dense>
#include <bench/BenchTimer.h>
using namespace Eigen;
using namespace std;

EIGEN_DONT_INLINE
void foo(const MatrixXd & A, MatrixXd & B)
{
  B = A.lu().matrixLU();
}

int main()
{
  int tries = 2;
  int rep = 1;
  int n = 1000;
  MatrixXd A(n,n), B(n,n);
  A.setRandom();
  B.setRandom();
  BenchTimer t;
  BENCH(t, tries, rep, foo(A,B));
  std::cout << "Time: " << t.best() << "s" << std::endl;
}


compiled with:

Code: Select all

clang++  -march=native -O3 -I ../eigen3.3 bench_inv.cpp && ./a.out


Moreover, do you really need to explicitly compute the inverse? It is usually much faster and more accurate to only factorize the matrix and call solve to apply the inverse:

Code: Select all

PartialPivLU<MatrixXd> lu(A);
v = lu.solve(u); // <-->  v = A^-1 * u


Hi,

I tested this on Windows with a core i7 4 cores 8 threads @ 1.87Ghz

Compiled with VS 2015 /GS /GL /W3 /Gy /Gd /Oi /MD /openmp

I will try with intel compilers to see if there is a better bench.

Thanks

You're lacking /O2 I think.

ggael wrote:You're lacking /O2 I think.

Thanks I checked on dual boot linux and bench are really better
It is a problem of flags on VS project

Thanks