ITK/Release 4/Refactor Numerical Libraries

Background

ITK largely relies on third party libraries to provide support for numerical computations such as linear solvers, eigen analysis and non-linear optimization. These libraries have been integrated over the years of the project, based on content from other libraries, immediate needs and requirements of particular projects.

Goal

• The purpose of this project is to revise the numerical libraries currently used in ITK, and to determine whether they are the best (or a good enough) option to use, and whether they are being used properly in the toolkit.
  • When appropriate, some of these libraries will be replaced with better options.
  • When possible, the code will be restructured to make possible for users to pick other options of third party libraries that may provide better support for their particular applications.

Inventory

• First, we proceed to make an inventory of the numerical libraries currently used in ITK. This will serve as basis for the analysis of whether there are better options to use, and/or whether they are being used correctly.

Inventory of Numerical Libraries

Tcons

• 2011 08 24

Progress

• Remove LINPACK
  • Remove netlib code
    • Done. (Chuck Atkins)
  • Netlib lbfgsb code links to LINPACK for dpofa. Needs to be ported to LAPACK dpotrf.
    • Done. (Chuck Atkins)
  • Port vnl classes using LINPACK over to calling LAPACKE interface into LAPACK.
• Remove EISPACK
  • Remove netlib code
    • Done. (Chuck Atkins)
  • Port vnl classes using EISPACK over to calling LAPACKE interface into LAPACK.
• Replace f2c reference BLAS with the CBLAS, the C interface to BLAS linking against an externally supplied BLAS.
  • Remove netlib f2c code and replace with CBLAS
    • Done. (Chuck Atkins)
  • Port vnl classes using BLAS over to calling CBLAS interface into BLAS.
• Replace f2c LAPACK with LAPACKE, the C interface to LAPACK linking against an externally supplied LAPACK.
  • Remove netlib f2c code and replace with LAPACKE
    • Done. (Chuck Atkins)
      • lapacke.h slightly modified to better handle complex numbers in C++ and to remove prototypes of extended precision LAPACK functions that are not available in slightly older LAPACK interfaces.
  • Port vnl classes using LAPACK over to calling LAPACKE interface into LAPACK
• Optimize LBFGSB code using level 3 BLAS routines