[
https://issues.apache.org/jira/browse/SYSTEMML-1223?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15862282#comment-15862282
]
Matthias Boehm commented on SYSTEMML-1223:
------------------------------------------
In the sequence files, there is no explicit block size meta data but just keys
and values, i.e., matrix indexes and matrix blocks (although one could derive
the blocksizes from the block dimensions of full blocks, i.e., non-border
blocks). Generally, we force every matrix in a script into the global blocksize
of your configuration because otherwise one could create "misaligned" scenarios
where we could no longer simply join over block indexes for binary operations.
> 'write` Statements Ignore `rows_in_block`, `cols_in_block` Arguments
> --------------------------------------------------------------------
>
> Key: SYSTEMML-1223
> URL: https://issues.apache.org/jira/browse/SYSTEMML-1223
> Project: SystemML
> Issue Type: Bug
> Reporter: Mike Dusenberry
>
> Given a DML script that writes matrices in binary form to disk, the
> {{rows_in_block}} and {{cols_in_block}} arguments will be ignored, and the
> default of 1000 will be used.
> {code}
> write(X, "X", format="binary", rows_in_block=batchSize, cols_in_block=cols)
> {code}
--
This message was sent by Atlassian JIRA
(v6.3.15#6346)
