[
https://issues.apache.org/jira/browse/SYSTEMML-1965?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16208746#comment-16208746
]
Matthias Boehm commented on SYSTEMML-1965:
------------------------------------------
I agree with [~dusenberrymw] here, let's take this as a use case and see why
the multiple rounds of IPA (incl constant propagation and folding) are not
sufficient here. Furthermore, code motion (moving loop-invariant operators)
automatically outside loops, is on our TODO list for a while now - from a
mechanic perspective, it's trivial but on a first attempt, we encountered
scenarios where it was counter-productive due to moving (and materializing)
intermediates that otherwise would not have been computed at all.
> Refactor nn layers to move the computation in forward/backward function known
> at compile time to init function
> --------------------------------------------------------------------------------------------------------------
>
> Key: SYSTEMML-1965
> URL: https://issues.apache.org/jira/browse/SYSTEMML-1965
> Project: SystemML
> Issue Type: Bug
> Reporter: Niketan Pansare
>
> Ideally, we should move the computation known at compile time to init layer,
> rather than keep in the forward function. This reduces recompilation time and
> also potentially unnecessary instructions. Here is an example snippet from
> our conv2d layer:
> {code}
> Hout = as.integer(floor((Hin + 2*padh - Hf)/strideh + 1))
> Wout = as.integer(floor((Win + 2*padw - Wf)/stridew + 1))
> {code}
> [~prithvi_r_s] [~dusenberrymw] [~reinwald] do you have any comments or
> concerns ?
> [~dusenberrymw] do you have free cycles to take this over ?
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
