Autotuning GEMM kernels for the fermi GPU

Jakub Kurzak; Stanimire Tomov; Jack Dongarra

doi:10.1109/TPDS.2011.311

Autotuning GEMM kernels for the fermi GPU

Jakub Kurzak, Stanimire Tomov, Jack Dongarra

Research output: Contribution to journal › Article › peer-review

95 Scopus citations

Abstract

In recent years, the use of graphics chips has been recognized as a viable way of accelerating scientific and engineering applications, even more so since the introduction of the Fermi architecture by NVIDIA, with features essential to numerical computing, such as fast double precision arithmetic and memory protected with error correction codes. Being the crucial component of numerical software packages, such as LAPACK and ScaLAPACK, the general dense matrix multiplication routine is one of the more important workloads to be implemented on these devices. This paper presents a methodology for producing matrix multiplication kernels tuned for a specific architecture, through a canonical process of heuristic autotuning, based on generation of multiple code variants and selecting the fastest ones through benchmarking. The key contribution of this work is in the method for generating the search space; specifically, pruning it to a manageable size. Performance numbers match or exceed other available implementations.

Original language	English
Article number	6122021
Pages (from-to)	2045-2057
Number of pages	13
Journal	IEEE Transactions on Parallel and Distributed Systems
Volume	23
Issue number	11
DOIs	https://doi.org/10.1109/TPDS.2011.311
State	Published - 2012
Externally published	Yes

Funding

This work was supported by DOE grant #DE-SC0003852, “Architecture-aware Algorithms for Scalable Performance and Resilience on Heterogeneous Architectures,” DOE grant #DE-SC0004983, “Matrix Algebra for GPU and Multicore Architectures (MAGMA) for Large Petascale Systems,” Georgia Institute of Technology subcontract #RA241-G1 funded by NSF grant #OCI-0910735, “Keene-land: National Institute for Experimental Computing.” The authors would like to thank David Luebke, Steven Parker, and Massimiliano Fatica for their insightful comments about the Fermi architecture.

Funders	Funder number
National Science Foundation	-0910735
U.S. Department of Energy	-SC0003852, -SC0004983, 241-G1

Keywords

BLAS
CUDA
GEMM
Graphics processing unit
automatic tuning
code generation
matrix multiplication

Access to Document

10.1109/TPDS.2011.311

Cite this

@article{a9ec93e1c30b4fde959a5245478d4422,

title = "Autotuning GEMM kernels for the fermi GPU",

abstract = "In recent years, the use of graphics chips has been recognized as a viable way of accelerating scientific and engineering applications, even more so since the introduction of the Fermi architecture by NVIDIA, with features essential to numerical computing, such as fast double precision arithmetic and memory protected with error correction codes. Being the crucial component of numerical software packages, such as LAPACK and ScaLAPACK, the general dense matrix multiplication routine is one of the more important workloads to be implemented on these devices. This paper presents a methodology for producing matrix multiplication kernels tuned for a specific architecture, through a canonical process of heuristic autotuning, based on generation of multiple code variants and selecting the fastest ones through benchmarking. The key contribution of this work is in the method for generating the search space; specifically, pruning it to a manageable size. Performance numbers match or exceed other available implementations.",

keywords = "BLAS, CUDA, GEMM, Graphics processing unit, automatic tuning, code generation, matrix multiplication",

author = "Jakub Kurzak and Stanimire Tomov and Jack Dongarra",

year = "2012",

doi = "10.1109/TPDS.2011.311",

language = "English",

volume = "23",

pages = "2045--2057",

journal = "IEEE Transactions on Parallel and Distributed Systems",

issn = "1045-9219",

publisher = "Institute of Electrical and Electronics Engineers",

number = "11",

}

TY - JOUR

T1 - Autotuning GEMM kernels for the fermi GPU

AU - Kurzak, Jakub

AU - Tomov, Stanimire

AU - Dongarra, Jack

PY - 2012

Y1 - 2012

N2 - In recent years, the use of graphics chips has been recognized as a viable way of accelerating scientific and engineering applications, even more so since the introduction of the Fermi architecture by NVIDIA, with features essential to numerical computing, such as fast double precision arithmetic and memory protected with error correction codes. Being the crucial component of numerical software packages, such as LAPACK and ScaLAPACK, the general dense matrix multiplication routine is one of the more important workloads to be implemented on these devices. This paper presents a methodology for producing matrix multiplication kernels tuned for a specific architecture, through a canonical process of heuristic autotuning, based on generation of multiple code variants and selecting the fastest ones through benchmarking. The key contribution of this work is in the method for generating the search space; specifically, pruning it to a manageable size. Performance numbers match or exceed other available implementations.

AB - In recent years, the use of graphics chips has been recognized as a viable way of accelerating scientific and engineering applications, even more so since the introduction of the Fermi architecture by NVIDIA, with features essential to numerical computing, such as fast double precision arithmetic and memory protected with error correction codes. Being the crucial component of numerical software packages, such as LAPACK and ScaLAPACK, the general dense matrix multiplication routine is one of the more important workloads to be implemented on these devices. This paper presents a methodology for producing matrix multiplication kernels tuned for a specific architecture, through a canonical process of heuristic autotuning, based on generation of multiple code variants and selecting the fastest ones through benchmarking. The key contribution of this work is in the method for generating the search space; specifically, pruning it to a manageable size. Performance numbers match or exceed other available implementations.

KW - BLAS

KW - CUDA

KW - GEMM

KW - Graphics processing unit

KW - automatic tuning

KW - code generation

KW - matrix multiplication

UR - http://www.scopus.com/inward/record.url?scp=84867310932&partnerID=8YFLogxK

U2 - 10.1109/TPDS.2011.311

DO - 10.1109/TPDS.2011.311

M3 - Article

AN - SCOPUS:84867310932

SN - 1045-9219

VL - 23

SP - 2045

EP - 2057

JO - IEEE Transactions on Parallel and Distributed Systems

JF - IEEE Transactions on Parallel and Distributed Systems

IS - 11

M1 - 6122021

ER -

Autotuning GEMM kernels for the fermi GPU

Abstract

Funding

Keywords

Access to Document

Other files and links

Fingerprint

Cite this