systemml-issues mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Matthias Boehm (JIRA)" <>
Subject [jira] [Commented] (SYSTEMML-1782) Extend parfor block indexing analysis
Date Thu, 31 Aug 2017 21:53:00 GMT


Matthias Boehm commented on SYSTEMML-1782:

I would recommend to extend the parfor optimizer by this analysis. Note that parfor optimization
is invoked during runtime on entering the parfor loop. Hence, the range and increment of {{j}}
become known and can be used to perform this range indexing analysis. Specifically, we want
to determine if beg:end is constant for all assignments of {{j}}. For linear functions, we
could reuse the parfor dependency analysis framework, while for non-linear operations like
min and max, we might need to do a limit analysis and break it into piece-wise linear functions.

> Extend parfor block indexing analysis
> -------------------------------------
>                 Key: SYSTEMML-1782
>                 URL:
>             Project: SystemML
>          Issue Type: Task
>            Reporter: Matthias Boehm
>            Assignee: Fei Hu
> This task aims to better support arbitrary block indexing in a parfor context, for expressions
as this example from SYSTEMML-1774:
> {code}
>  parfor (j in 1:parallel_batches, log=DEBUG) {
>         # Get a mini-batch in this group
>         beg = ((j-1) * batch_size) %% nrow(X_group_batch) + 1
>         end = min(nrow(X_group_batch), beg + batch_size - 1)
>         X_batch = X_group_batch[beg:end,]
>         y_batch = y_group_batch[beg:end,]
> {code}
> Currently, we're unable to determine a constant size for X_batch and y_batch. However,
since the parfor body is anyway recompiled on parfor entry, we can improve upon this as follows:
> * Leverage known constants and index ranges to perform a limit analysis for linear functions
in indexing expressions. For example, the above expression for beg and end would simplify
to {{(j-1) * 32 + 1}} and  {{beg + batch_size -1}}.
> * Better handling of linear functions in right indexing HOPs - this would allow us to
determine batch_size as the number of rows in the output.

This message was sent by Atlassian JIRA

View raw message