eCommons

 

A Generic Programming System for Sparse Matrix Computations

dc.contributor.authorMateev, Nikolayen_US
dc.contributor.authorKotlyar, Vladimiren_US
dc.contributor.authorPingali, Keshaven_US
dc.contributor.authorStodghill, Paulen_US
dc.date.accessioned2007-04-23T18:18:06Z
dc.date.available2007-04-23T18:18:06Z
dc.date.issued1999-07en_US
dc.description.abstractSparse matrices are stored in compressed formats in which zeros are not stored explicitly. Writing high-performance sparse matrix libraries is a difficult and tedious job because there are many compressed formats in use and each of them requires specialized code. In this paper, we argue that (i) compressed formats should be viewed as {\em indexed-sequential access structures} (in the database sense), and (ii) efficient sparse codes exploit such indexing structures wherever possible. This point of view leads naturally to restructuring compiler technology that can be used to synthesize many sparse codes from high-level algorithms and specifications of sparse formats, exploiting indexing structures for efficiency. We show that appropriate abstractions of the indexing structures of commonly used formats can be provided to such a compiler through the type structure of a language like C++. Finally, we describe experimental results obtained from the {\em Bernoulli Sparse Compiler} which demonstrate that the performance of code generated by this compiler is comparable to the performance of programs in the NIST Sparse BLAS library. One view of this system is that it exploits restructuring compiler technology to perform a novel kind of template instantiation.en_US
dc.format.extent362264 bytes
dc.format.extent661523 bytes
dc.format.extent210947 bytes
dc.format.mimetypeapplication/pdf
dc.format.mimetypeapplication/postscript
dc.format.mimetypeapplication/postscript
dc.identifier.citationhttp://techreports.library.cornell.edu:8081/Dienst/UI/1.0/Display/cul.cs/TR99-1755en_US
dc.identifier.urihttps://hdl.handle.net/1813/7409
dc.language.isoen_USen_US
dc.publisherCornell Universityen_US
dc.subjectcomputer scienceen_US
dc.subjecttechnical reporten_US
dc.titleA Generic Programming System for Sparse Matrix Computationsen_US
dc.typetechnical reporten_US

Files

Original bundle
Now showing 1 - 3 of 3
Loading...
Thumbnail Image
Name:
99-1755.pdf
Size:
353.77 KB
Format:
Adobe Portable Document Format
No Thumbnail Available
Name:
99-1755.ps
Size:
646.02 KB
Format:
Postscript Files
No Thumbnail Available
Name:
old.ps
Size:
206 KB
Format:
Postscript Files