Optimizing Matrix Multiplication with a Classifier Learning System

* Final gross prices may vary according to local VAT.

Get Access


Compilers have been very successful on automating the process of program optimization, but there is still a significant difference in performance between the code generated by the compiler and the hand-optimized code. Library generators such as ATLAS, SPIRAL, and FFTW address this problem by using empirical search to find the parameter values of certain optimization such as degree of unroll. We have recently developed a generator of sorting routines. Sorting differs from the algorithms implemented by other library generators in that performance of sorting depends not only on the target platform but also on the characteristics of the input data. In our work we used a classifier learning system to generate sorting routines that are capable of adapting to the input data. In this paper we follow a similar approach and use a classifier learning system to generate high performance libraries for matrix-matrix multiplication. Our library generator produces matrix multiplication routines that use recursive layouts and several levels of tiling. Our approach is to use a classifier learning system to search in the space of the different ways to partition the input matrices the one that performs the best. As a result, our system will determine the number of levels of tiling and tile size for each level depending on the target platform and the dimensions of the input matrices.

This work was supported in part by the National Science Foundation under grant CCR 01-21401 ITR; by DARPA under contract NBCH30390004; and by gifts from INTEL and IBM. This work is not necessarily representative of the positions or policies of the Army or Government.