Adaptive libraries for multicore architectures with explicitly-managed memory hierarchies

Programming of commodity multicore processors is a challenging task and it becomes even harder when the processor has an explicitly-managed memory hierarchy (EMMA). Software libraries in the field of matrix algebra try to keep pace with this challenge by using the dataflow model of computation and constructing tiled algorithms. A new approach to high-performance software library construction is proposed, which moves scheduling decisions to compile-time and is portable between different EMMA platforms. Performance and scalability analyses both demonstrate promising results. Experiments demonstrate near linear speedup on a synthetic multicore architecture, incorporating up to 16 working computational cores. Performance of a generated code is competitive with vendor BLAS implementations for the Cell processor.