[jira] [Updated] (SYSTEMML-1644) Incorrect results on multi-threaded dense-dense matrix multiply w/ short lhs

["Matthias Boehm (JIRA)" <ji...@apache.org>]

     [ https://issues.apache.org/jira/browse/SYSTEMML-1644?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ]

Matthias Boehm updated SYSTEMML-1644:
-------------------------------------
    Summary: Incorrect results on multi-threaded dense-dense matrix multiply w/ short lhs  (was: Incorrect results on multi-threaded dense-dense matrix multiply w/ wide rhs)

> Incorrect results on multi-threaded dense-dense matrix multiply w/ short lhs
> ----------------------------------------------------------------------------
>
>                 Key: SYSTEMML-1644
>                 URL: https://issues.apache.org/jira/browse/SYSTEMML-1644
>             Project: SystemML
>          Issue Type: Bug
>            Reporter: Matthias Boehm
>
> The following matrix multiply expression yields incorrect results compared to R when run with multi-threaded operations but correct results with single-threaded operations.
> {code}
> # X is dense 2023 x 1987 matrix
> # t(U) is dense 10 x  2023 matrix
> R = t(U) %*% X;
> {code}
> Note: This affects any dense-dense matrix multiplication with short left-hand-side (of <=16 rows) and wide right-hand-side (of #features > the cache block size of 1024 columns). Such operations occur for example in Mlogreg and Kmeans with a number of classes/centroids 1 < x <= 16 and a feature matrix with ncol(X) > 1024.



--
This message was sent by Atlassian JIRA
(v6.3.15#6346)