EN FR
EN FR


Section: New Results

Optimization of Sparse Triangular and Banded Matrix Codes

Participants : Vincent Loechner, Rachid Seghir, Toufik Baroudi.

This work is a collaboration between Vincent Loechner and Rachid Seghir from University of Batna (Algeria). Toufik Baroudi is a second year PhD student under his supervision. Rachid Seghir was visiting the CAMUS team from March 25th to April 8th, 2017.

In this work, we enabled static polyhedral optimization techniques to handle sparse matrix storage formats. When handling sparse triangular and banded matrices in their packed formats, such as in the LAPACK library band storage, loop nests bounds and array references of the resulting codes are not affine functions. We proposed to use a new 2d-packed layout and simple affine transformations to enable polyhedral optimization of sparse triangular and banded matrix operations. The effectiveness of our proposal was shown through an experimental study over a large set of linear algebra benchmarks.

These results were published in ACM TACO [8], and will be presented at the HiPEAC conference in January 2018.