You are viewing a plain text version of this content. The canonical link for it is here.
Posted to issues@systemml.apache.org by "Dylan Hutchison (JIRA)" <ji...@apache.org> on 2017/11/03 04:58:02 UTC
[jira] [Updated] (SYSTEMML-1985) Sparsity estimates of constant
multiply
[ https://issues.apache.org/jira/browse/SYSTEMML-1985?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ]
Dylan Hutchison updated SYSTEMML-1985:
--------------------------------------
Description:
I've seen the following in the output of Explain recompile_hops:
{noformat}
--(505) TRead HV [327,1,1000,1000,327] [0,0,0 -> 0MB], CP
--(37959) b(*) (505,[-0.0053168...]) [327,1,1000,1000,-1] [0,0,0 -> 0MB], CP
{noformat}
(I altered the Explain output to show the value of the constant that HV is multiplied with, namely {{-0.0053168}})
The nnz of Hop 37959 here is 327, because multiplying a matrix by a non-zero constant does not change the non-zero structure. But the current code marks the nnz as -1 (unknown nnz).
Let's add a check to the {{refreshSizeInformation}} of BinaryOp to handle this and related cases like division by a nonzero.
Addition, subtraction, and mod might create zeros so they are tricky; however we can upper bound their nnz. Maybe this might be helpful in a different place?
was:
I've seen the following in the output of Explain recompile_hops:
{noformat}
--(505) TRead HV [327,1,1000,1000,327] [0,0,0 -> 0MB], CP
--(37959) b(*) (505,[-0.0053168...]) [327,1,1000,1000,-1] [0,0,0 -> 0MB], CP
{noformat}
The nnz of Hop 37959 here is 327, because multiplying a matrix by a non-zero constant does not change the non-zero structure. But the current code marks the nnz as -1 (unknown nnz).
Let's add a check to the {{refreshSizeInformation}} of BinaryOp to handle this and related cases like division by a nonzero.
Addition, subtraction, and mod might create zeros so they are tricky; however we can upper bound their nnz. Maybe this might be helpful in a different place?
> Sparsity estimates of constant multiply
> ---------------------------------------
>
> Key: SYSTEMML-1985
> URL: https://issues.apache.org/jira/browse/SYSTEMML-1985
> Project: SystemML
> Issue Type: Improvement
> Reporter: Dylan Hutchison
> Priority: Minor
>
> I've seen the following in the output of Explain recompile_hops:
> {noformat}
> --(505) TRead HV [327,1,1000,1000,327] [0,0,0 -> 0MB], CP
> --(37959) b(*) (505,[-0.0053168...]) [327,1,1000,1000,-1] [0,0,0 -> 0MB], CP
> {noformat}
>
> (I altered the Explain output to show the value of the constant that HV is multiplied with, namely {{-0.0053168}})
> The nnz of Hop 37959 here is 327, because multiplying a matrix by a non-zero constant does not change the non-zero structure. But the current code marks the nnz as -1 (unknown nnz).
> Let's add a check to the {{refreshSizeInformation}} of BinaryOp to handle this and related cases like division by a nonzero.
> Addition, subtraction, and mod might create zeros so they are tricky; however we can upper bound their nnz. Maybe this might be helpful in a different place?
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)