Efficient Implementation of Matrix-Matrix Multiplication for Contemporary Multi-Core Processors

Art der Arbeit:
Adresse: Johannes Hofmann
Lehrstuhl für Informatik 3 (Rechnerarchitektur)
Martensstraße 3
91058 Erlangen
Raum: 07.158
Telefon: +49 9131 85 27913
Fax: +49 9131 85 27912
Homepage: http://www3.informatik.uni-erlangen.de/Persons/hofmann/
E-Mail: johannes.hofmann@fau.de
Beschreibung der Arbeit:

The goal of this work is to devise an optimized implementation of the (small) matrix-matrix multiplication for contemporary multi-core processors. Matrix-matrix multiplication is the performance-critical component in many deep learning applications. The student is supposed to start with a naive matrix-matrix multiplication implementation in C and iteratively apply optimizations, such as SIMD vectorization and cache blocking, to improve the implementation's performance.