[ 
https://issues.apache.org/jira/browse/HIVE-11521?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14681883#comment-14681883
 ] 

Teddy Choi commented on HIVE-11521:
-----------------------------------

Sorry. The benchmark result above shows the difference between whether AVX=0 
option is applied or not. The result bellow shows the difference between 
before/after applying this patch. 260% ~ 300% performance improvement in 
repeating cases.

Before this patch:
{noformat}
Benchmark                                                                       
Mode  Samples           Score   Error  Units
o.a.h.b.v.VectorizationBench.ColAndColBench.bench                               
avgt        2   169346433.000 ±   NaN  ns/op
o.a.h.b.v.VectorizationBench.ColAndRepeatingColBench.bench                      
avgt        2   503688769.000 ±   NaN  ns/op
o.a.h.b.v.VectorizationBench.ColOrColBench.bench                                
avgt        2   184679292.500 ±   NaN  ns/op
o.a.h.b.v.VectorizationBench.ColOrRepeatingColBench.bench                       
avgt        2   522471397.500 ±   NaN  ns/op
...
o.a.h.b.v.VectorizationBench.NotColBench.bench                                  
avgt        2   154808036.000 ±   NaN  ns/op
o.a.h.b.v.VectorizationBench.RepeatingColAndColBench.bench                      
avgt        2   478369669.500 ±   NaN  ns/op
o.a.h.b.v.VectorizationBench.RepeatingColOrColBench.bench                       
avgt        2   514816574.000 ±   NaN  ns/op
{noformat}

After this patch:
{noformat}
Benchmark                                                                       
Mode  Samples           Score   Error  Units
o.a.h.b.v.VectorizationBench.ColAndColBench.bench                               
avgt        2   171249531.500 ±   NaN  ns/op
o.a.h.b.v.VectorizationBench.ColAndRepeatingColBench.bench                      
avgt        2   130488848.500 ±   NaN  ns/op
o.a.h.b.v.VectorizationBench.ColOrColBench.bench                                
avgt        2   168669206.500 ±   NaN  ns/op
o.a.h.b.v.VectorizationBench.ColOrRepeatingColBench.bench                       
avgt        2   128768271.500 ±   NaN  ns/op
...
o.a.h.b.v.VectorizationBench.NotColBench.bench                                  
avgt        2   149107679.000 ±   NaN  ns/op
o.a.h.b.v.VectorizationBench.RepeatingColAndColBench.bench                      
avgt        2   138933203.000 ±   NaN  ns/op
o.a.h.b.v.VectorizationBench.RepeatingColOrColBench.bench                       
avgt        2   140216834.500 ±   NaN  ns/op
{noformat}

> Loop optimization for SIMD in logical operators
> -----------------------------------------------
>
>                 Key: HIVE-11521
>                 URL: https://issues.apache.org/jira/browse/HIVE-11521
>             Project: Hive
>          Issue Type: Sub-task
>            Reporter: Teddy Choi
>            Assignee: Teddy Choi
>            Priority: Minor
>         Attachments: HIVE-11521.patch
>
>
> JVM is quite strict on the code schema which may executed with SIMD 
> instructions, take a loop in ColOrCol.java for example,
> {code}
> for (int i = 0; i != n; i++) {
>   outputVector[i] = vector1[0] | vector2[i];
> }
> {code}
> The "vector1\[0\]" reference would prevent JVM to execute this part of code 
> with vectorized instructions, we need to assign the "vector1\[0\]" to a 
> variable outside of loop, and use that variable in loop.
> This issues covers AND, OR, NOT logical operators.



--
This message was sent by Atlassian JIRA
(v6.3.4#6332)

Reply via email to