A Systematic Approach for Acceleration of Matrix-Vector Operations in CGRA through Algorithm-Architecture Co-Design

摘要

Matrix-vector operations play pivotal role in engineering and scientific applications ranging from machine learning to computational finance. Matrix-vector operations have time complexity of O(n^2) and they are challenging to accelerate since these operations are memory bound operations where ratio of the arithmetic operations to the data movement is O(1). In this paper, we present a systematic methodology of algorithm-architecture co-design to accelerate matrix-vector operations where we emphasize on the matrix-vector multiplication (gemv) and the vector transpose-matrix multiplication (vtm). In our methodology, we perform a detailed analysis of directed acyclic graphs of the routines and identify macro operations that can be realized on a reconfigurable data-path that is tightly coupled to the pipeline of a processing element. It is shown that the PE clearly outperforms state-of-the-art realizations of gemv and vtm attaining 135% performance improvement over multicore and 200% over general purpose graphics processing units. In the parallel realization on REDEFINE coarse-grained reconfigurable architecture, it is shown that the solution is scalable.