Skip to content
Show report in:

UMINF 21.08

Comparing the Performance of Computing Eigenvectors from the Schur Matrix and by Inverse Iteration from the Hessenberg Matrix

Two standard approaches to the computation of eigenvectors corresponding to eigenvalues of a non-symmetric real matrix are (1) a backward substitution on the real Schur matrix followed by a matrix-matrix multiplication and (2) inverse iteration on the Hessenberg matrix. Both approaches have been realized as tiled, task-parallel solvers where a large proportion of the operations are matrix-matrix multiplications. The first contribution of this work addresses a memory bottleneck of the inverse iteration solver by improving the locality of memory accesses. The second contribution compares the two approaches quantitatively with numerical experiments.


No keywords specified


Angelika Schwarz

Back Edit this report
Entry responsible: Angelika Schwarz

Page Responsible: Frank Drewes