tree 95eff92f0118cb295ee1fce9f9351885668a7971
parent ed8b5f5c725526f8414d2dbe9113d52260597352
author Matthias Boehm <mboehm7@gmail.com> 1608242658 +0100
committer Matthias Boehm <mboehm7@gmail.com> 1608242658 +0100

[SYSTEMDS-2760] Performance sparse-sparse transpose operations

This patch makes two minor performance improvements to multi-threaded
sparse-sparse transpose operations. The tasks get column groups of the
input and append outputs to the sparse output rows. For tall&skinny
sparse matrices there were two shortcomings.

First, we aligned the minimum column group size to 8, which is only
required for dense-dense transpose operations. This improved performance
on above scenario.

Second, if the number of cores is larger than the number of columns,
cache blocking and related position maintenance only adds unnecessary
overhead. Accordingly, we now use added a special case for these
scenarios.

On a scenario of a 2.5M x 50 matrix with sparsity = 0.1, the total
execution time of 10 transpose operations improved from 2.7s to 2.5s
(with 1) and 1.9s (with 1 and 2) respectively.