Performance engineering for real and complex tall & skinny matrix multiplication kernels on GPUs

Ernst D, Hager G, Thies J, Wellein G (2020)


Publication Type: Journal article

Publication year: 2020

Journal

DOI: 10.1177/1094342020965661

Abstract

General matrix-matrix multiplications with double-precision real and complex entries (DGEMM and ZGEMM) in vendor-supplied BLAS libraries are best optimized for square matrices but often show bad performance for tall & skinny matrices, which are much taller than wide. NVIDIA’s current CUBLAS implementation delivers only a fraction of the potential performance as indicated by the roofline model in this case. We describe the challenges and key characteristics of an implementation that can achieve close to optimal performance. We further evaluate different strategies of parallelization and thread distribution and devise a flexible, configurable mapping scheme. To ensure flexibility and allow for highly tailored implementations we use code generation combined with autotuning. For a large range of matrix sizes in the domain of interest we achieve at least 2/3 of the roofline performance and often substantially outperform state-of-the art CUBLAS results on an NVIDIA Volta GPGPU.

Authors with CRIS profile

Involved external institutions

How to cite

APA:

Ernst, D., Hager, G., Thies, J., & Wellein, G. (2020). Performance engineering for real and complex tall & skinny matrix multiplication kernels on GPUs. International Journal of High Performance Computing Applications. https://doi.org/10.1177/1094342020965661

MLA:

Ernst, Dominik, et al. "Performance engineering for real and complex tall & skinny matrix multiplication kernels on GPUs." International Journal of High Performance Computing Applications (2020).

BibTeX: Download