Towards Dynamic Autotuning of SpMV in CUSP Library

Kód výsledku v IS VaVaI

<a href="https://www.isvavai.cz/riv?ss=detail&h=RIV%2F00216224%3A14610%2F24%3A00137324" target="_blank" >RIV/00216224:14610/24:00137324 - isvavai.cz</a>

Výsledek na webu

<a href="http://dx.doi.org/10.1109/IPDPSW63119.2024.00012" target="_blank" >http://dx.doi.org/10.1109/IPDPSW63119.2024.00012</a>

DOI - Digital Object Identifier

<a href="http://dx.doi.org/10.1109/IPDPSW63119.2024.00012" target="_blank" >10.1109/IPDPSW63119.2024.00012</a>

Jazyk výsledku

angličtina

Název v původním jazyce

Towards Dynamic Autotuning of SpMV in CUSP Library

Popis výsledku v původním jazyce

Sparse matrix-vector product (SpMV) is a central operation in many iterative methods for solving linear systems and, as such, is an attractive candidate for acceleration on the GPU. However, the performance of the SpMV kernel can vary depending both on the target architecture as well as on the sparsity pattern of the matrix. Thus, to achieve optimal performance, the implementation might need to be adjusted for each particular matrix and architecture. This can be achieved through dynamic autotuning, a method that can optimize a source code during program runtime. In this paper, we present a dynamic autotuning of SpMV kernel included in a production-quality CUSP library. We identify and implement tuning parameters and use the Kernel Tuning Toolkit framework for autotuning of SpMV working with the DIA and ELL sparse matrix formats. The dynamic autotuning integration is fully transparent to the library users - it can be activated just by re-compiling software using our tunable version of the CUSP. The proposed autotuned library is evaluated by comparing it with the original CUSP kernels on a set of representative matrices and by examining the contribution of autotuning. The results show that the autotuned kernels can reach up to about 16.9 × speedup compared to a fixed implementation.

Název v anglickém jazyce

Towards Dynamic Autotuning of SpMV in CUSP Library

Popis výsledku anglicky

Sparse matrix-vector product (SpMV) is a central operation in many iterative methods for solving linear systems and, as such, is an attractive candidate for acceleration on the GPU. However, the performance of the SpMV kernel can vary depending both on the target architecture as well as on the sparsity pattern of the matrix. Thus, to achieve optimal performance, the implementation might need to be adjusted for each particular matrix and architecture. This can be achieved through dynamic autotuning, a method that can optimize a source code during program runtime. In this paper, we present a dynamic autotuning of SpMV kernel included in a production-quality CUSP library. We identify and implement tuning parameters and use the Kernel Tuning Toolkit framework for autotuning of SpMV working with the DIA and ELL sparse matrix formats. The dynamic autotuning integration is fully transparent to the library users - it can be activated just by re-compiling software using our tunable version of the CUSP. The proposed autotuned library is evaluated by comparing it with the original CUSP kernels on a set of representative matrices and by examining the contribution of autotuning. The results show that the autotuned kernels can reach up to about 16.9 × speedup compared to a fixed implementation.

Druh

D - Stať ve sborníku

CEP obor

—

OECD FORD obor

10201 - Computer sciences, information science, bioinformathics (hardware development to be 2.2, social aspect to be 5.8)

Projekt

<a href="/cs/project/LM2023054" target="_blank" >LM2023054: e-Infrastruktura CZ</a><br>

Návaznosti

P - Projekt vyzkumu a vyvoje financovany z verejnych zdroju (s odkazem do CEP)

Rok uplatnění

2024

Kód důvěrnosti údajů

S - Úplné a pravdivé údaje o projektu nepodléhají ochraně podle zvláštních právních předpisů

Název statě ve sborníku

IEEE International Parallel and Distributed Processing Symposium Workshops (IPDPSW)

ISBN

9798350364613

ISSN

—

e-ISSN

—

Počet stran výsledku

9

Strana od-do

14-22

Název nakladatele

IEEE

Místo vydání

San Francisco, USA

Místo konání akce

Mauritius

Datum konání akce

1. 1. 2024

Typ akce podle státní příslušnosti

WRD - Celosvětová akce

Kód UT WoS článku

001284697300084