Supporting matrix operations in vector architectures

Abstract

Many elementary numerical algorithms involve not only vector operations but also matrix operations. Today's vector processors only support vector operations, and execute matrix operations in terms of vector operations, because they can not access matrix operands in one instruction. This will lead to poor sustained performances of vector machines. The paper discusses how to support both vector operations and matrix operations in vector architectures. At first subarray patterns for vector and matrix operations are introduced. Then it presents a set of accessing modes which can make vector architectures to access both vector and matrix operands. Finally the performance improvement for matrix multiplication and the FFT is demonstrated.<>