I am trying to use Eigen 3.2.2 with OpenBLAS as a backend, but I get a fatal error during compilation when I define EIGEN_USE_BLAS before the Eigen headers are included:

Code: Select all

#define EIGEN_USE_BLAS
#include <Eigen/Dense>


This happens because the header file mkl.h cannot be found. It seems that Eigen still only provide support for Intel MKL's library, but in this thread https://forum.kde.org/viewtopic.php?f=74&t=110509, it is said that linking with OpenBLAS is possible.

Can Eigen really link to OpenBLAS (or any other BLAS library)? What about LAPACKE?

Indeed, after some checks this would require some minor and quite straightforward adjustments in Eigen's MKL wrapper to remove the strict dependence to MKL. Any volunteer?

Where are these minor changes necessary?

I am interested in this topic too, and I would like to use OpenBlas with Eigen, if possible.

I am not an expert on programming. However, I have been digging the Eigen files a bit and I have noticed some references to MKL and EIGEN_USE_BLAS. As default, the EIGEN_USE_BLAS define points to MKL functionalities. The file Core at Eigen's files shows the following:

Code: Select all

#ifdef EIGEN_USE_BLAS
#include "src/Core/products/GeneralMatrixMatrix_MKL.h"
#include "src/Core/products/GeneralMatrixVector_MKL.h"
#include "src/Core/products/GeneralMatrixMatrixTriangular_MKL.h"
#include "src/Core/products/SelfadjointMatrixMatrix_MKL.h"
#include "src/Core/products/SelfadjointMatrixVector_MKL.h"
#include "src/Core/products/TriangularMatrixMatrix_MKL.h"
#include "src/Core/products/TriangularMatrixVector_MKL.h"
#include "src/Core/products/TriangularSolverMatrix_MKL.h"
#endif // EIGEN_USE_BLAS


So, if I am not wrong, I guess these are the lines that need to be changed.

Basically, in src/util/Core/MklSupport.h we should not include mkl.h by default, then defines the MKL_* types, include misc/blas.h, add the "_" prefix to BLAS calls in Core/products/*MKL.h, and update the blas.h file wrt constness.

The coming version 3.3-beta2 seems to solve this problem:

ChangeLog:

"Relax dependency on MKL for EIGEN_USE_BLAS: any BLAS library can now be used as backend."
http://eigen.tuxfamily.org/index.php?ti ... _3.3-beta2


Right now I cannot find the download link for 3.3-beta2, but will try it whenever it's available.

Anyone compared different libraries which is faster on Eigen's benchmark?

Does MKL / OpenBLAS give any improvement over Eigen?

Thank You.

with multithreading MKL and OpenBLAS should be faster, for sequential use they all are close to the hardware maximal speed.

What about small to medium matrices (Dynamic and Static Wise)?

please, be more specific, "medium" size does tell much. Eventually, the best way to get your answer is to bench the two options for your own problem, it is just a matter of compiling with -DEIGEN_USE_BLAS -lopenblas

I'd say medium is matrices with up to 2500 Elements.

Small I'd say matrices with up to 250 Elements.