high performance computing on graphics processing units: hgpu.org

Sparse Matrix Vector Multiplication (SpMV) is an important computational kernel in many scientific computing applications. Pipelining multiply-accumulate operations shifts SpMV from a computationally bounded kernel to an I/O bounded kernel. In this paper, we propose a design methodology and hardware architecture for SpMV that seeks to utilize system memory bandwidth as efficiently as possible, by Reducing the matrix element storage with on-chip decompression hardware, Reusing the vector data by mixing row and column matrix traversal, and Recycling data with matrix-dependent on-chip storage. Our experimental results with a Convey HC-1/HC-2 reconfigurable computing system indicate that for certain sparse matrices, our R^3 methodology performs twice as fast as previous reconfigurable implementations, and effectively competes against other platforms.