Optimization of Sparse Matrix - Vector Multiplication on Heterogeneous Multi-Core

Provided by: AICIT
Topic: Data Centers
Format: PDF
Sparse matrix-vector multiplication is one of the most important kernels in scientific and engineering computing. This paper proposes a strategy which facilitates parallel sparse matrix-vector multiplication computations on heterogeneous multi-core. The strategy exploits an auto-tuning framework which considers not only the structure of the sparse matrices but also the architectural characteristics of the heterogeneous multi-core. The auto-tuning strategy adaptively partitions the sparse matrix at the boundary of the rows and roughly evenly splits the nonzero elements among threads, which improves the spatial locality in accessing the vector and achieves load balance as well.

Find By Topic