[
https://issues.apache.org/jira/browse/SYSTEMML-2131?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Matthias Boehm updated SYSTEMML-2131:
-------------------------------------
Description:
The functions bias_add and bias_mult (see
http://apache.github.io/systemml/dml-language-reference for the semantics) are
language constructs and operators that aim to improve performance by fusing
element-wise operations with the replication of bias terms. For example,
bias_add is defined as follows:
{code}
ones = matrix(1, rows=1, cols=height*width);
output = input + matrix(bias %*% ones, rows=1, cols=numChannels*height*width)
{code}
The code generation frameworks will makes such custom operators obsolete.
However, because bias_add and bias_mult are mapped to specific
{{ConvolutionOp}} instances they are not yet supported by the codegen framework.
In detail this task entails:
1) Create new cell template tests in {{functions.codegen.CellwiseTmplTest}} for
both functions.
2) Modify the open/merge/close conditions for cell templates in
{{codegen.template.TemplateCell}}
3) Create a new codegen primitive for a mapped index lookups (for each cell in
input we want to access a specific cell in the bias vector)
4) Modify the CPlan construction in {{codegen.template.TemplateCell}} to create
binary plus or multiply and the mapped index lookup for all ConvolutionOp
instances of type {{BIAS_ADD}} or {{BIAS_MULTIPLY}}
> Add support for bias_add and bias_mult
> --------------------------------------
>
> Key: SYSTEMML-2131
> URL: https://issues.apache.org/jira/browse/SYSTEMML-2131
> Project: SystemML
> Issue Type: Sub-task
> Reporter: Matthias Boehm
> Priority: Major
>
> The functions bias_add and bias_mult (see
> http://apache.github.io/systemml/dml-language-reference for the semantics)
> are language constructs and operators that aim to improve performance by
> fusing element-wise operations with the replication of bias terms. For
> example, bias_add is defined as follows:
> {code}
> ones = matrix(1, rows=1, cols=height*width);
> output = input + matrix(bias %*% ones, rows=1, cols=numChannels*height*width)
> {code}
> The code generation frameworks will makes such custom operators obsolete.
> However, because bias_add and bias_mult are mapped to specific
> {{ConvolutionOp}} instances they are not yet supported by the codegen
> framework.
> In detail this task entails:
> 1) Create new cell template tests in {{functions.codegen.CellwiseTmplTest}}
> for both functions.
> 2) Modify the open/merge/close conditions for cell templates in
> {{codegen.template.TemplateCell}}
> 3) Create a new codegen primitive for a mapped index lookups (for each cell
> in input we want to access a specific cell in the bias vector)
> 4) Modify the CPlan construction in {{codegen.template.TemplateCell}} to
> create binary plus or multiply and the mapped index lookup for all
> ConvolutionOp instances of type {{BIAS_ADD}} or {{BIAS_MULTIPLY}}
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
