Department of Computer Engineering
CS 690 SEMINAR
Efficient Vectorization of General Sparse Matrix-Matrix Multiplication
Mustafa Ozan Karsavuran
Computer Engineering Department
In many applications like linear programing, multiple source bread first search, molecular dynamics simulation, and fluid simulation, General sparse matrix-matrix multiplication (SpGEMM) operation is an important kernel. In recent CPU or accelerator architectures powerful vector processing units exists. However, vectorization of SpGEMM operation is not trivial due to irregular memory accesses caused by sparsity pattern of the matrices. We propose a method that is efficient and effective in terms of vectorization and parallelization. Our approach is valid on many modern many core architectures. We conduct experiments on the Intel Xeon Phi coprocessor which has 60 cores.
DATE: 14 November, 2016, Monday @ 16:00