[
https://issues.apache.org/jira/browse/SYSTEMML-1288?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Matthias Boehm updated SYSTEMML-1288:
-------------------------------------
Description:
This task aims to add further codegen templates besides the existing
SpoofCellwise, SpoofRowAggregate, and SpoofOuterProduct.
1) SpoofMultiOutput: There are quite a number of scenarios where we have
expressions like (a) sum(Y == check_min), sum(Y == check_max), or (b)
overlapping expressions with large common subexpression, i.e., fusable
operations over a shared input with multiple aggregation roots. Since
SystemML's compiler/runtime only supports single-output operations (with very
few exceptions), it would be beneficial to generate a single fused operator
with a combined aggregate output, which is subsequently split via traditional
right indexing operations.
2) Row Templates: In addition to our row aggregate template we want to be able
to fuse rowwise operations without aggregate in order to avoid unnecessary
intermediates for expressions like B = (A <= rowMins(A)); C = B / rowSums (B),
which should become a single operator instead of 4 scans and two matrix and two
vector intermediates.
was:
This task aims to add further codegen templates besides the existing
SpoofCellwise, SpoofRowAggregate, and SpoofOuterProduct.
1) SpoofMultiOutput: There are quite a number of scenarios where we have
expressions like (a) sum(Y == check_min), sum(Y == check_max), or (b)
overlapping expressions with large common subexpression, i.e., fusable
operations over a shared input with multiple aggregation roots. Since
SystemML's compiler/runtime only supports single-output operations (with very
few exceptions), it would be beneficial to generate a single fused operator
with a combined aggregate output, which is subsequently split via traditional
right indexing operations.
> Additional meta operator templates
> ----------------------------------
>
> Key: SYSTEMML-1288
> URL: https://issues.apache.org/jira/browse/SYSTEMML-1288
> Project: SystemML
> Issue Type: Sub-task
> Components: Compiler, Runtime
> Reporter: Matthias Boehm
> Assignee: Matthias Boehm
> Fix For: SystemML 1.0
>
>
> This task aims to add further codegen templates besides the existing
> SpoofCellwise, SpoofRowAggregate, and SpoofOuterProduct.
> 1) SpoofMultiOutput: There are quite a number of scenarios where we have
> expressions like (a) sum(Y == check_min), sum(Y == check_max), or (b)
> overlapping expressions with large common subexpression, i.e., fusable
> operations over a shared input with multiple aggregation roots. Since
> SystemML's compiler/runtime only supports single-output operations (with very
> few exceptions), it would be beneficial to generate a single fused operator
> with a combined aggregate output, which is subsequently split via traditional
> right indexing operations.
> 2) Row Templates: In addition to our row aggregate template we want to be
> able to fuse rowwise operations without aggregate in order to avoid
> unnecessary intermediates for expressions like B = (A <= rowMins(A)); C = B /
> rowSums (B), which should become a single operator instead of 4 scans and two
> matrix and two vector intermediates.
--
This message was sent by Atlassian JIRA
(v6.3.15#6346)
