Earticle

초록

영어

Sparse matrix-vector multiplication (SpMV) is a key linear algebra algorithm and is widely used in many application domains. Besides multi-core architecture, there is also extensive research focusing on accelerating SpMV on many-core Graphics Processing Units (GPUs). SpMV computations have many indirect and irregular memory accesses, and load imbalance could occur while mapping computations onto single-instruction, multiple-data (SIMD) GPUs. SpMV is highly memory bandwidth-bound, though GPUs have massive computational resources, the performance of SpMV on GPUs is still unsatisfying. In this paper, we present a new hybrid strategy for auto-tuning SpMV on GPUs. Our strategy combines the advantages of row-major storage and column-major storage. Like many other strategies, we reordered a given sparse matrix according to row lengths in decreasing order. In order to be more adaptive and efficient, we proposed a new hybrid Blocked CSR and JDS (BCJ) format based on original CSR and JDS. BCJ splits a sparse matrix into a denser part and a sparser part after reordering and uses different kernels to process the corresponding part. And we proposed corresponding auto-tuning framework to help transforming matrix and launching kernels according to the sparsity characteristics of the matrix. A CUDA implementation of BCJ outperforms the original formats significantly on a broad range of unstructured sparse matrices.

목차

Abstract
 1. Introduction
 2. Background
  2.1. GPUs and CUDA Programming Model
  2.2. Formats for Sparse Matrix
 3. Hybrid auto-tuning Framework
  3.1. Blocked and Padded CSR (B-CSR)
  3.2. Blocked and Padded JDS (B-JDS)
  3.3. Parameters to be Modeled
 4. Experimental Results
 5. Related Work
 6. Conclusions
 References

키워드

저자

• Zhaohui Wang [ University of Science and Technology of China ]
• Xiaojie Qiu [ University of Science and Technology of China ]
• Aimin Zhang [ University of Science and Technology of China ]
• Yichao Cheng [ University of Science and Technology of China ]
• Yi Peng [ University of Science and Technology of China ]
• Sun Sun [ University of Science and Technology of China ]

참고문헌

발행기관

간행물

표지보기 관심저널 등록

• 간행물명

  International Journal of Software Engineering and Its Applications

• 간기

  월간

• pISSN

  1738-9984

• 수록기간

  2008~2016

• 등재여부

  SCOPUS

• 십진분류

  KDC 505 DDC 605