The original Libsvm project by Chih-Chung Chang and Chih-Jen Lin is available at http://www.csie.ntu.edu.tw/~cjlin/libsvm along with sample preformated datasets and extensive documentation.

Work done as part of the simd4libsvm project is aimed at eventually being contributed back to the original project once I get something stable and useful enough and provided the original authors are interested in it :)

As of now (5th of Apr. 2008) only few of the afore mentioned goals are implemented:

  - it is possible to choose at runtime between sparse (as in the original libsvm) or dense (new) representation of the input vectors ;
  - it is possible to switch at compile time between single or double precision floats
  - implementation of SSE2 optimized kernel evaluations including SIMD evaluations of exp in single precision RBF
  - pthread based support of multi-core CPUs

Planned feature include:

  - Altivec optimised kernels (using intrinsincs)
  - libspe2 enabled kernels (for IBM Cell/BE processors that can be found in the Playstation3)

On a longer term, I plan to add support for OpenGL-2 based GPUs. I haven't started to think about which tools to use to do that and
which hardware to target (though the Nvidia RSX chip of the PS3 is probably the first I will have look at).

Please feel free to contact me (olivier.grisel on-the-server ensta.org) should you be interested in working on any of those topics. I personnaly use the Bazaar decentralised version control system to manage the source code: please branch my code ( https://code.launchpad.net/~olivier-grisel/simd4libsvm/libsvm.og.main ) and send me an email to request merge of contribution of any kind.