[jira] [Commented] (ARROW-13121) [C++][Compute] Extract preallocation logic from KernelExecutor
[
https://issues.apache.org/jira/browse/ARROW-13121?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17366621#comment-17366621
]
Ben Kietzman commented on ARROW-13121:
--
Another point which we might like to optimize:
https://github.com/apache/arrow/pull/10550#discussion_r654521795
> [C++][Compute] Extract preallocation logic from KernelExecutor
> --
>
> Key: ARROW-13121
> URL: https://issues.apache.org/jira/browse/ARROW-13121
> Project: Apache Arrow
> Issue Type: Improvement
> Components: C++
>Reporter: Ben Kietzman
>Priority: Major
>
> Currently KernelExecutor handles preallocation of null bitmaps and other
> buffers based on simple flags on each Kernel. This is not very flexible and
> we end up leaving a lot of performance on the table in cases where we can
> preallocate but the behavior can't be captured in the available flags. For
> example, in the case of {{binary_string_join_element_wise}}, it would be
> possible to preallocate all buffers (even the character buffer) and write
> output into slices.
> Having this as a public function would enable us to unit test it directly
> (currently Executors are only tested indirectly through calling of
> compute::Functions) and reuse it, for example to correctly preallocate a
> small temporary for pipelined execution
> One way this could be added is as a new method on each Kernel:
> {code}
> // Output preallocated Datums sufficient for execution of the kernel on each
> ExecBatch.
> // The output Datums may not be identically chunked to the input batches, for
> example
> // kernels which support contiguous output preallocation will preallocate a
> single Datum
> // (and can then output into slices of that Datum).
> Result> Kernel::prepare_output(
> const Kernel*,
> KernelContext*,
> const std::vector& inputs)
> {code}
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
[jira] [Commented] (ARROW-13121) [C++][Compute] Extract preallocation logic from KernelExecutor
[
https://issues.apache.org/jira/browse/ARROW-13121?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17365633#comment-17365633
]
Antoine Pitrou commented on ARROW-13121:
I'd also mention that when I worked on ARROW-13042, I spent a lot of time
trying to figure out which code paths exactly got executed in {{exec.cc}} and I
never fully figured it out (one particularly case was implicit casting and
broadcasting with a NullScalar LHS and a ChunkedArray RHS on a scalar kernel).
I ended up trying to find clues in other places instead.
> [C++][Compute] Extract preallocation logic from KernelExecutor
> --
>
> Key: ARROW-13121
> URL: https://issues.apache.org/jira/browse/ARROW-13121
> Project: Apache Arrow
> Issue Type: Improvement
> Components: C++
>Reporter: Ben Kietzman
>Priority: Major
>
> Currently KernelExecutor handles preallocation of null bitmaps and other
> buffers based on simple flags on each Kernel. This is not very flexible and
> we end up leaving a lot of performance on the table in cases where we can
> preallocate but the behavior can't be captured in the available flags. For
> example, in the case of {{binary_string_join_element_wise}}, it would be
> possible to preallocate all buffers (even the character buffer) and write
> output into slices.
> Having this as a public function would enable us to unit test it directly
> (currently Executors are only tested indirectly through calling of
> compute::Functions) and reuse it, for example to correctly preallocate a
> small temporary for pipelined execution
> One way this could be added is as a new method on each Kernel:
> {code}
> // Output preallocated Datums sufficient for execution of the kernel on each
> ExecBatch.
> // The output Datums may not be identically chunked to the input batches, for
> example
> // kernels which support contiguous output preallocation will preallocate a
> single Datum
> // (and can then output into slices of that Datum).
> Result> Kernel::prepare_output(
> const Kernel*,
> KernelContext*,
> const std::vector& inputs)
> {code}
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
