[
https://issues.apache.org/jira/browse/ARROW-3701?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16717008#comment-16717008
]
Pindikura Ravindra commented on ARROW-3701:
-------------------------------------------
After running more benchmarks, I found the following for (divide a/b)
* The Decimal128 implementation has pretty consistent performance (between 300
to 400ms, for 10M iterations) whereas the i128 division varies widely (between
100ms to 2000ms) based on the value of b i.e the divisor.
* when a and b are both large random numbers, i128 divide (both clang and
llvm) is twice as fast as the Decimal128 implementation.
* when b is a small number (say, 10 or 100), Decimal128 implementation is
about 7 times faster. I think the reason is that the Decimal128 implementation
has short-cuts for small integers.
For decimal operations, there are several cases where we need to do small
divisions (to adjust scale). So, I'm using the Decimal128 implementation for
divides.
> [Gandiva] Add support for decimal operations
> --------------------------------------------
>
> Key: ARROW-3701
> URL: https://issues.apache.org/jira/browse/ARROW-3701
> Project: Apache Arrow
> Issue Type: Task
> Components: Gandiva
> Reporter: Pindikura Ravindra
> Assignee: Pindikura Ravindra
> Priority: Major
> Labels: pull-request-available
> Time Spent: 6h 50m
> Remaining Estimate: 0h
>
> To begin with, will add support for 128-bit decimals. There are two parts :
> # llvm_generator needs to understand decimal types (value, precision, scale)
> # code decimal operations : add/subtract/multiply/divide/mod/..
> ** This will be c++ code that can be pre-compiled to emit IR code
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
