A0537
Title: Tricks that accelerate matrix multiplication on CPUs
Authors: Martin Schlather - Universitat Mannheim (Germany) [presenting]
Alexander Freudenberg - University of Mannheim (Germany)
Abstract: General approaches to accelerate matrix multiplication are parallel computing, SIMD programming and the Strassen algorithm. It is surprising that, although SNP matrices are large, the Strassen algorithm might be significantly outperformed by fine-tuned standard algorithms, when calculating the genomic relationship matrix. We give an introduction to low-level matrix multiplication and show the acceleration when SIMD commands, cache and registers are used exhaustively. Clearly, the results highly depend on the compression level.
