We have quite a number of math libraries in 3rdparty packages now. We should focus down and choose one or a few. As part of cleaning up 3rd party packages ROS/ThirdPartyDependencies

Options include:

• uBlas - (in boost), but only provides BLAS no higher level functions http://www.boost.org/doc/libs/1_37_0/libs/numeric/ublas/doc/overview.htm

  • Conclusion: not complete enough to recommend usage.

• newmat10 - an option (though we've had problems with svd) http://www.robertnz.net/nm10.htm

  • Easy to use, but not very efficient
  • SVD apparently doesn't converge properly
  • Conclusion: throw it out

• lapackpp - a c++ wrapper around LAPACK http://lapackpp.sourceforge.net/ (branched from http://math.nist.gov/lapack++/ ) http://www.netlib.org/lapack/

  • Incomplete, will require pushing back patches to expose everything we need.
  • lapack is the standard, with which more people are familiar

• eigen2 - used in the vision packages http://eigen.tuxfamily.org/index.php?title=Main_Page

  • Not as complete as lapack
  • Includes most decompositions plus regressions
  • Under active development, responsive community
  • Comparable performance to Intel MKL and specialized BLAS implementations

  • much nicer interface than lapack

  • sparse matrix operations are currently experimental
• openCV
  • Has some stuff, but will likely move to some BLAS implementation
  • Conclusion: don't use it.
• refblas
  • Not easy to use directly
  • Free implementations are generally not fast.
  • No reason to use refblas
• atlas, automatically tuned BLAS
  • Should be faster than refblas
  • Different versions available depending on SSE level -- linker problems.
  • Implements part of LAPACK (?)
• BFL has wrappers for uBLAS, NEWMAT, and liblti

Conclusions

• Recommend eigen2 for all new code, with uBLAS to fill in holes (e.g., sparse matrices)
• No NEWMAT
• No OpenCV
• No refblas
• No GSL CBLAS
• No lapackcpp
• Use atlas and/or lapack for existing code, but prefer to switch.