Saved in:
Bibliographic Details
Main Authors: Steel, Thijs, Langou, Julien
Format: Preprint
Published: 2024
Subjects:
Online Access:https://arxiv.org/abs/2412.01852
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866929611361746944
author Steel, Thijs
Langou, Julien
author_facet Steel, Thijs
Langou, Julien
contents We present an efficient algorithm for the application of sequences of planar rotations to a matrix. Applying such sequences efficiently is important in many numerical linear algebra algorithms for eigenvalues. Our algorithm is novel in three main ways. First, we introduce a new kernel that is optimized for register reuse in a novel way. Second, we introduce a blocking and packing scheme that improves the cache efficiency of the algorithm. Finally, we thoroughly analyze the memory operations of the algorithm which leads to important theoretical insights and makes it easier to select good parameters. Numerical experiments show that our algorithm outperforms the state-of-the-art and achieves a flop rate close to the theoretical peak on modern hardware.
format Preprint
id arxiv_https___arxiv_org_abs_2412_01852
institution arXiv
publishDate 2024
record_format arxiv
spellingShingle Communication efficient application of sequences of planar rotations to a matrix
Steel, Thijs
Langou, Julien
Performance
Data Structures and Algorithms
65F15, 65Y05
G.4
We present an efficient algorithm for the application of sequences of planar rotations to a matrix. Applying such sequences efficiently is important in many numerical linear algebra algorithms for eigenvalues. Our algorithm is novel in three main ways. First, we introduce a new kernel that is optimized for register reuse in a novel way. Second, we introduce a blocking and packing scheme that improves the cache efficiency of the algorithm. Finally, we thoroughly analyze the memory operations of the algorithm which leads to important theoretical insights and makes it easier to select good parameters. Numerical experiments show that our algorithm outperforms the state-of-the-art and achieves a flop rate close to the theoretical peak on modern hardware.
title Communication efficient application of sequences of planar rotations to a matrix
topic Performance
Data Structures and Algorithms
65F15, 65Y05
G.4
url https://arxiv.org/abs/2412.01852