tustvold commented on code in PR #3622:
URL: https://github.com/apache/arrow-rs/pull/3622#discussion_r1089825671
##########
arrow-select/src/take.rs:
##########
@@ -810,6 +816,70 @@ where
Ok(DictionaryArray::<T>::from(data))
}
+macro_rules! primitive_run_take {
+ ($t:ty, $o:ty, $indices:ident, $value:ident) => {
+ take_primitive_run_values::<$o, $t>(
+ $indices,
+ as_primitive_array::<$t>($value.values()),
+ )
+ };
+}
+
+/// `take` implementation for run arrays
+///
+/// performs binary search on `run_ends` to get physical indices for the given
logical indices.
+/// builds output run array by taking values in the input run array at the
physical indices.
+/// for e.g. an input `RunArray{ run_ends = [2,4,6,8], values=[1,2,1,2] }` and
`indices=[2,7]`
+/// would be converted to `physical_indices=[1,3]` which will be used to build
+/// output `RunArray{ run_ends=[2], values=[2] }`
+
+pub fn take_run<T, I>(
Review Comment:
The issue lies mainly with generating code for every type of value, the fact
there are multiple index types makes it worse, but even with a single index
type as written this would double the codegen for this kernel.
Operating solely on the run ends should be both faster, and significantly
less codegen
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
