[ARROW-13121] [C++][Compute] Extract preallocation logic from KernelExecutor

• Type: Improvement
• Status: Open
• Priority: Major
• Resolution: Unresolved
• Component/s: C++
• External issue URL: https://github.com/apache/arrow/issues/28823
• Assignee: Unassigned
• Reporter: Ben Kietzman
• Created: 18/Jun/21 17:19
• Updated: 11/Jan/23 08:31

Description

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:

// 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<std::vector<Datum>> Kernel::prepare_output(
  const Kernel*,
  KernelContext*,
  const std::vector<ExecBatch>& inputs)

Issue Links

• is a child of: ARROW-8894 [C++] C++ array kernels framework and execution buildout (umbrella issue)
• is related to: ARROW-16758 [C++] Rewrite ExecuteScalarExpression to not use ScalarExecutor
• relates to: ARROW-16755 [C++] Improve array expression and kernel evaluation performance on small inputs
• relates to: ARROW-11647 [C++][Compute] CastFromNull does not use preallocated buffers