Bilkent University
Department of Computer Engineering


Techniques for efficient usage of SIMD instructions in Sparse Matrix Vector Multiplication


Mehmet Başaran
PhD Student
Computer Engineering Department
Bilkent University

SpMV (sparse matrix vector multiplication) is a kernel operation in linear solvers in which a sparse matrix is multiplied with a dense vector repeatedly. Due to structure irregularities exhibited by SpMV operation, built-in specialized hardware components are almost always under-utilized. Consequently, limiting efficiency of parallel implementation. In this work, we explore algorithms and storage formats to use SIMD components more effectively which in turn makes SpMV kernel more robust and improves its performance in GPUs and architectures with advanced SIMD capability.


DATE: 17 October, 2016, Monday @ 16:00