[jira] [Updated] (SYSTEMML-913) Performance matrix-vector multiplication w/ tall rhs vector

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

     [ https://issues.apache.org/jira/browse/SYSTEMML-913?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ]

Matthias Boehm updated SYSTEMML-913:
------------------------------------
    Description: So far, we compute matrix-vector multiplication with simple row-wise dot products. This works very well for the common case of tall&skinny matrices, where the right-hand-side vector is very small. However, for scenarios with many features and hence a tall rhs vector, this approach suffers from cache unfriendly behavior. Accordingly, this task tracks the introduction of a dedicated cache-conscious matrix-vector multiplication for both sparse and dense matrices.   (was: So far, we compute matrix-vector multiplication with simple row-wise dot products. This works very well for the common case of tall&skinny matrices, where the right-hand-side vector is very small. However, for scenarios with many features and hence a tall rhs vector, this approach suffers from cache unfriendly behavior. This tasks tracks the dedicated handling of cache-conscious matrix-vector multiplication for both sparse and dense matrices. )

> Performance matrix-vector multiplication w/ tall rhs vector
> -----------------------------------------------------------
>
>                 Key: SYSTEMML-913
>                 URL: https://issues.apache.org/jira/browse/SYSTEMML-913
>             Project: SystemML
>          Issue Type: Task
>            Reporter: Matthias Boehm
>
> So far, we compute matrix-vector multiplication with simple row-wise dot products. This works very well for the common case of tall&skinny matrices, where the right-hand-side vector is very small. However, for scenarios with many features and hence a tall rhs vector, this approach suffers from cache unfriendly behavior. Accordingly, this task tracks the introduction of a dedicated cache-conscious matrix-vector multiplication for both sparse and dense matrices. 



--
This message was sent by Atlassian JIRA
(v6.3.4#6332)