[ 
https://issues.apache.org/jira/browse/ARROW-5639?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
 ]

Liya Fan updated ARROW-5639:
----------------------------
    Description: 
Some getOffsetBufferValueCapacity methods uses floating point computation to 
calculate the capacity, which is not necessary.

(int) ((offsetBuffer.capacity() * 1.0) / OFFSET_WIDTH);

It is interesting to note that JIT cannot optimize away the floating point 
computations:

 !image-2019-06-18-20-30-17-826.png! 

So this has performance penalty:
Before:
VariableWidthVectorBenchmarks.getValueCapacity  avgt    5  6.570 ± 0.004  ns/op

After:
VariableWidthVectorBenchmarks.getValueCapacity  avgt    5  5.787 ± 0.575  ns/op


  was:
Method BaseVariableWidthVector#getOffsetBufferValueCapacity uses floating point 
computation to calculate the capacity, which is not necessary.

(int) ((offsetBuffer.capacity() * 1.0) / OFFSET_WIDTH);

It is interesting to note that JIT cannot optimize away the floating point 
computations:

 !image-2019-06-18-20-30-17-826.png! 

So this has performance penalty:
Before:
VariableWidthVectorBenchmarks.getValueCapacity  avgt    5  6.570 ± 0.004  ns/op

After:
VariableWidthVectorBenchmarks.getValueCapacity  avgt    5  5.787 ± 0.575  ns/op



> [Java] Remove floating point computation from getOffsetBufferValueCapacity
> --------------------------------------------------------------------------
>
>                 Key: ARROW-5639
>                 URL: https://issues.apache.org/jira/browse/ARROW-5639
>             Project: Apache Arrow
>          Issue Type: Improvement
>          Components: Java
>            Reporter: Liya Fan
>            Assignee: Liya Fan
>            Priority: Trivial
>              Labels: pull-request-available
>         Attachments: image-2019-06-18-20-30-17-826.png
>
>          Time Spent: 1h
>  Remaining Estimate: 0h
>
> Some getOffsetBufferValueCapacity methods uses floating point computation to 
> calculate the capacity, which is not necessary.
> (int) ((offsetBuffer.capacity() * 1.0) / OFFSET_WIDTH);
> It is interesting to note that JIT cannot optimize away the floating point 
> computations:
>  !image-2019-06-18-20-30-17-826.png! 
> So this has performance penalty:
> Before:
> VariableWidthVectorBenchmarks.getValueCapacity  avgt    5  6.570 ± 0.004  
> ns/op
> After:
> VariableWidthVectorBenchmarks.getValueCapacity  avgt    5  5.787 ± 0.575  
> ns/op



--
This message was sent by Atlassian JIRA
(v7.6.3#76005)

Reply via email to