[
https://issues.apache.org/jira/browse/SYSTEMML-1570?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15991142#comment-15991142
]
Niketan Pansare commented on SYSTEMML-1570:
-------------------------------------------
Before closing this PR, please ensure that the performance benefits for CNNs
(eg: Lenet) due fused sel+ operators (such as relu_maxpooling and
relu_maxpooling_backward) are preserved.
> Remove fused sel+ operator
> --------------------------
>
> Key: SYSTEMML-1570
> URL: https://issues.apache.org/jira/browse/SYSTEMML-1570
> Project: SystemML
> Issue Type: Task
> Reporter: Matthias Boehm
> Fix For: SystemML 1.0
>
>
> The fused operator sel+ (select positive values) is applied for patterns like
> (X>0)*X and max(X,0) in order to eliminate unnecessary intermediates. It
> stems from a time when max was sparse-unsafe and hence inefficient over
> sparse data. However, meanwhile we mark scalar operators as conditionally
> sparse-safe depending on the given scalar constant c, which applies for max
> if c<=0. Hence, this sel+ operator is meanwhile completely useless and should
> be removed.
> Furthermore, we should also generalize the rewrites to rewrite the selection
> of negative values (X<0)*X to min(X,0)
--
This message was sent by Atlassian JIRA
(v6.3.15#6346)
