Batched QR and SVD Algorithms on GPUs with Applications in Hierarchical Matrix Compression

Wagih Halim Boukaram; George Turkiyyah; Hatem Ltaief; David E. Keyes

doi:10.1016/j.parco.2017.09.001

Batched QR and SVD Algorithms on GPUs with Applications in Hierarchical Matrix Compression

Wagih Halim Boukaram, George Turkiyyah, Hatem Ltaief, David E. Keyes

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

35 Scopus citations

Abstract

We present high performance implementations of the QR and the singular value decomposition of a batch of small matrices hosted on the GPU with applications in the compression of hierarchical matrices. The one-sided Jacobi algorithm is used for its simplicity and inherent parallelism as a building block for the SVD of low rank blocks using randomized methods. We implement multiple kernels based on the level of the GPU memory hierarchy in which the matrices can reside and show substantial speedups against streamed cuSOLVER SVDs. The resulting batched routine is a key component of hierarchical matrix compression, opening up opportunities to perform H-matrix arithmetic efficiently on GPUs.

Original language	English (US)
Pages (from-to)	19-33
Number of pages	15
Journal	Parallel Computing
Volume	74
DOIs	https://doi.org/10.1016/j.parco.2017.09.001
State	Published - Sep 14 2017

Bibliographical note

KAUST Repository Item: Exported on 2020-10-01
Acknowledgements: The work of all four authors was supported by the Extreme Computing Research Center at the King Abdullah University of Science and Technology. We thank the NVIDIA Corporation for providing access to the P100 GPU used in this work.

Access to Document

10.1016/j.parco.2017.09.001

https://repository.kaust.edu.sa/bitstream/10754/625473/1/1-s2.0-S0167819117301461-main.pdf

Cite this

@article{547d0aca28fa4a38a2a2d31c1936d684,

title = "Batched QR and SVD Algorithms on GPUs with Applications in Hierarchical Matrix Compression",

abstract = "We present high performance implementations of the QR and the singular value decomposition of a batch of small matrices hosted on the GPU with applications in the compression of hierarchical matrices. The one-sided Jacobi algorithm is used for its simplicity and inherent parallelism as a building block for the SVD of low rank blocks using randomized methods. We implement multiple kernels based on the level of the GPU memory hierarchy in which the matrices can reside and show substantial speedups against streamed cuSOLVER SVDs. The resulting batched routine is a key component of hierarchical matrix compression, opening up opportunities to perform H-matrix arithmetic efficiently on GPUs.",

author = "Boukaram, {Wagih Halim} and George Turkiyyah and Hatem Ltaief and Keyes, {David E.}",

note = "KAUST Repository Item: Exported on 2020-10-01 Acknowledgements: The work of all four authors was supported by the Extreme Computing Research Center at the King Abdullah University of Science and Technology. We thank the NVIDIA Corporation for providing access to the P100 GPU used in this work.",

year = "2017",

month = sep,

day = "14",

doi = "10.1016/j.parco.2017.09.001",

language = "English (US)",

volume = "74",

pages = "19--33",

journal = "Parallel Computing",

issn = "0167-8191",

publisher = "Elsevier",

}

TY - JOUR

T1 - Batched QR and SVD Algorithms on GPUs with Applications in Hierarchical Matrix Compression

AU - Boukaram, Wagih Halim

AU - Turkiyyah, George

AU - Ltaief, Hatem

AU - Keyes, David E.

N1 - KAUST Repository Item: Exported on 2020-10-01 Acknowledgements: The work of all four authors was supported by the Extreme Computing Research Center at the King Abdullah University of Science and Technology. We thank the NVIDIA Corporation for providing access to the P100 GPU used in this work.

PY - 2017/9/14

Y1 - 2017/9/14

N2 - We present high performance implementations of the QR and the singular value decomposition of a batch of small matrices hosted on the GPU with applications in the compression of hierarchical matrices. The one-sided Jacobi algorithm is used for its simplicity and inherent parallelism as a building block for the SVD of low rank blocks using randomized methods. We implement multiple kernels based on the level of the GPU memory hierarchy in which the matrices can reside and show substantial speedups against streamed cuSOLVER SVDs. The resulting batched routine is a key component of hierarchical matrix compression, opening up opportunities to perform H-matrix arithmetic efficiently on GPUs.

AB - We present high performance implementations of the QR and the singular value decomposition of a batch of small matrices hosted on the GPU with applications in the compression of hierarchical matrices. The one-sided Jacobi algorithm is used for its simplicity and inherent parallelism as a building block for the SVD of low rank blocks using randomized methods. We implement multiple kernels based on the level of the GPU memory hierarchy in which the matrices can reside and show substantial speedups against streamed cuSOLVER SVDs. The resulting batched routine is a key component of hierarchical matrix compression, opening up opportunities to perform H-matrix arithmetic efficiently on GPUs.

UR - http://hdl.handle.net/10754/625473

UR - http://www.sciencedirect.com/science/article/pii/S0167819117301461

UR - http://www.scopus.com/inward/record.url?scp=85029565345&partnerID=8YFLogxK

U2 - 10.1016/j.parco.2017.09.001

DO - 10.1016/j.parco.2017.09.001

M3 - Article

SN - 0167-8191

VL - 74

SP - 19

EP - 33

JO - Parallel Computing

JF - Parallel Computing

ER -

Batched QR and SVD Algorithms on GPUs with Applications in Hierarchical Matrix Compression

Abstract

Bibliographical note

Access to Document

Other files and links

Fingerprint

Cite this