DAG-Scheduled Linear Algebra Using Template-Based Building Blocks

ePubs

The open archive for STFC research publications

Persistent URL	http://purl.org/net/epubs/work/12417437
Record Status	Checked
Record Id	12417437
Title	DAG-Scheduled Linear Algebra Using Template-Based Building Blocks
Contributors	Dr JD Hogg (STFC Rutherford Appleton Lab.)
Abstract	We describe our experiences using DAG-driven algorithms built from templated BLAS-like building blocks to implement LAPACK-like functionality at the single kernel level. There will be a particular focus on strong scaling of multiple small dense factorizations, as required for sparse direct methods. The main objective is to overlap expensive latency-bound pivoting operations with highly parallel matrix-matrix multiplication operations. As the later are dependent on the output of previous pivoting decisions, a directed-acyclic graph (DAG) scheduler is implemented using global memory to manage fine-grained inter-block parallelism.
Organisation	STFC , SCI-COMP , SCI-COMP-CM
Keywords
Funding Information
Related Research Object(s):
Licence Information:
Language	English (EN)

Type	Details	URI(s)	Local file(s)	Year
Presentation	Presented at NVIDIA GPU Technology Conference (GTC 2015), San Jose, California, USA, 17-20 Mar 2015.		dag_sched_la.pdf	2015