You are viewing a plain text version of this content. The canonical link for it is here.
Posted to issues@ignite.apache.org by "Artem Malykh (JIRA)" <ji...@apache.org> on 2017/05/29 14:56:04 UTC
[jira] [Commented] (IGNITE-5278) BLAS implementation
[ https://issues.apache.org/jira/browse/IGNITE-5278?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16028435#comment-16028435 ]
Artem Malykh commented on IGNITE-5278:
--------------------------------------
We have some problems here: the natural approach for this task is to delegate BLAS calls to any of existing BLAS libs. All BLAS libraries which I met during implementing this task (jBlas, netlib-java) require storing internal data for matrices flattened (i.e [row1, row2...rown] instead of [[row1], [row2], ... [rowN]]), therefore it seems that we have two options here:
1. Redesign storing;
2. keep storing, but do flatten array -> do blas call -> unflatten underlying arrays.
If we take second approach, we do not get any boost (rather, we'll get slowdowns) in BLAS levels 1 and 2 for array based matrix impls. We still can get boost in BLAS level 3 calls.
For first approach to be taken, separate research task need to be done.
> BLAS implementation
> -------------------
>
> Key: IGNITE-5278
> URL: https://issues.apache.org/jira/browse/IGNITE-5278
> Project: Ignite
> Issue Type: Sub-task
> Components: ml
> Reporter: Yury Babak
> Assignee: Artem Malykh
> Fix For: 2.1
>
>
> We need BLAS implementation for local computations.
--
This message was sent by Atlassian JIRA
(v6.3.15#6346)