You are viewing a plain text version of this content. The canonical link for it is here.
Posted to github@arrow.apache.org by "Dandandan (via GitHub)" <gi...@apache.org> on 2026/03/12 09:29:24 UTC
[PR] Unroll interleave -25% [arrow-rs]
Dandandan opened a new pull request, #9542:
URL: https://github.com/apache/arrow-rs/pull/9542
# Which issue does this PR close?
<!--
We generally require a GitHub issue to be filed for all bug fixes and enhancements and this helps us generate change logs for our releases. You can link an issue to this PR using the GitHub syntax.
-->
- Closes #NNN.
# Rationale for this change
<!--
Why are you proposing this change? If this is already explained clearly in the issue then this section is not needed.
Explaining clearly why changes are proposed helps reviewers understand your changes and offer better suggestions for fixes.
-->
# What changes are included in this PR?
<!--
There is no need to duplicate the description in the issue here but it is sometimes worth providing a summary of the individual changes in this PR.
-->
# Are these changes tested?
<!--
We typically require tests for all PRs in order to:
1. Prevent the code from being accidentally broken by subsequent changes
2. Serve as another way to document the expected behavior of the code
If tests are not included in your PR, please explain why (for example, are they covered by existing tests)?
-->
# Are there any user-facing changes?
<!--
If there are user-facing changes then we may require documentation to be updated before approving the PR.
If there are any breaking changes to public APIs, please call them out.
-->
--
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: github-unsubscribe@arrow.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
Re: [PR] Unroll interleave -25-30% [arrow-rs]
Posted by "alamb-ghbot (via GitHub)" <gi...@apache.org>.
alamb-ghbot commented on PR #9542:
URL: https://github.com/apache/arrow-rs/pull/9542#issuecomment-4050323223
🤖 `./gh_compare_arrow.sh` [gh_compare_arrow.sh](https://github.com/alamb/datafusion-benchmarking/blob/main/scripts/gh_compare_arrow.sh) Running
Linux aal-dev 6.14.0-1018-gcp #19~24.04.1-Ubuntu SMP Wed Sep 24 23:23:09 UTC 2025 x86_64 x86_64 x86_64 GNU/Linux
Comparing take_native (8ff8bbffaca4299842c81918071fdaa0aaf69672) to 9d0e8beae74fedb362d88cbc6e32d9760657c9de [diff](https://github.com/apache/arrow-rs/compare/9d0e8beae74fedb362d88cbc6e32d9760657c9de..8ff8bbffaca4299842c81918071fdaa0aaf69672)
BENCH_NAME=interleave_kernels
BENCH_COMMAND=cargo bench --features=arrow,async,test_common,experimental,object_store --bench interleave_kernels
BENCH_FILTER=
BENCH_BRANCH_NAME=take_native
Results will be posted here when complete
--
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: github-unsubscribe@arrow.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
Re: [PR] Unroll interleave -25-30% [arrow-rs]
Posted by "comphead (via GitHub)" <gi...@apache.org>.
comphead commented on code in PR #9542:
URL: https://github.com/apache/arrow-rs/pull/9542#discussion_r2928090821
##########
arrow-select/src/interleave.rs:
##########
@@ -154,13 +154,51 @@ fn interleave_primitive<T: ArrowPrimitiveType>(
data_type: &DataType,
) -> Result<ArrayRef, ArrowError> {
let interleaved = Interleave::<'_, PrimitiveArray<T>>::new(values, indices);
+ let arrays = &interleaved.arrays;
+ let len = indices.len();
+
+ let mut output = Vec::with_capacity(len);
+ let dst: *mut T::Native = output.as_mut_ptr();
+ let mut base = 0;
+
+ // Process 8 elements at a time to issue multiple independent loads
+ // and increase memory-level parallelism for random access patterns.
+ let chunks = indices.chunks_exact(8);
+ let remainder = chunks.remainder();
+ for chunk in chunks {
+ let v0 = arrays[chunk[0].0].value(chunk[0].1);
Review Comment:
yeah, thats what I was thinking if user may want to modify this value depending on their hardware/OS setup?
Or if someone would like to benchmark `take` or `interleave` with different `chunks` values, they would have to rewrite the method respectively? maybe some macro could help? it can be done as follow up if needed
--
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: github-unsubscribe@arrow.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
Re: [PR] Unroll take_native, interleave -25% [arrow-rs]
Posted by "alamb-ghbot (via GitHub)" <gi...@apache.org>.
alamb-ghbot commented on PR #9542:
URL: https://github.com/apache/arrow-rs/pull/9542#issuecomment-4045170163
🤖: Benchmark completed
<details><summary>Details</summary>
<p>
```
group main take_native
----- ---- -----------
take bool 1024 1.00 1334.0±36.78ns ? ?/sec 1.00 1328.9±7.22ns ? ?/sec
take bool 512 1.00 729.2±5.61ns ? ?/sec 1.00 727.1±5.66ns ? ?/sec
take bool null indices 1024 1.02 1658.4±92.98ns ? ?/sec 1.00 1632.9±33.60ns ? ?/sec
take bool null values 1024 1.00 2.6±0.01µs ? ?/sec 1.00 2.6±0.02µs ? ?/sec
take bool null values null indices 1024 1.00 3.2±0.08µs ? ?/sec 1.17 3.7±0.07µs ? ?/sec
take check bounds i32 1024 1.00 859.3±6.12ns ? ?/sec 1.00 860.0±9.54ns ? ?/sec
take check bounds i32 512 1.02 537.0±4.52ns ? ?/sec 1.00 528.6±5.18ns ? ?/sec
take i32 1024 1.08 720.4±6.62ns ? ?/sec 1.00 664.6±1.81ns ? ?/sec
take i32 512 1.00 386.3±2.20ns ? ?/sec 1.01 389.2±1.92ns ? ?/sec
take i32 null indices 1024 1.01 1002.5±8.12ns ? ?/sec 1.00 995.8±2.22ns ? ?/sec
take i32 null values 1024 1.03 2.0±0.02µs ? ?/sec 1.00 1956.5±33.84ns ? ?/sec
take i32 null values null indices 1024 1.00 2.6±0.03µs ? ?/sec 1.01 2.6±0.06µs ? ?/sec
take primitive fsb value len: 12, indices: 1024 1.00 3.6±0.01µs ? ?/sec 1.00 3.6±0.01µs ? ?/sec
take primitive fsb value len: 12, null values, indices: 1024 1.00 5.0±0.01µs ? ?/sec 1.00 5.0±0.01µs ? ?/sec
take primitive run logical len: 1024, physical len: 512, indices: 1024 1.00 20.1±0.16µs ? ?/sec 1.19 23.9±0.11µs ? ?/sec
take str 1024 1.02 11.2±0.30µs ? ?/sec 1.00 10.9±0.07µs ? ?/sec
take str 512 1.02 5.4±0.13µs ? ?/sec 1.00 5.3±0.03µs ? ?/sec
take str null indices 1024 1.02 7.0±0.12µs ? ?/sec 1.00 6.8±0.11µs ? ?/sec
take str null indices 512 1.03 3.4±0.09µs ? ?/sec 1.00 3.3±0.02µs ? ?/sec
take str null values 1024 1.01 8.8±0.11µs ? ?/sec 1.00 8.7±0.23µs ? ?/sec
take str null values null indices 1024 1.00 6.4±0.04µs ? ?/sec 1.01 6.5±0.05µs ? ?/sec
take stringview 1024 1.00 886.9±4.19ns ? ?/sec 1.56 1379.8±10.17ns ? ?/sec
take stringview 512 1.00 589.1±1.01ns ? ?/sec 1.22 719.1±3.59ns ? ?/sec
take stringview null indices 1024 1.01 1430.9±23.75ns ? ?/sec 1.00 1413.1±29.05ns ? ?/sec
take stringview null indices 512 1.00 732.0±10.19ns ? ?/sec 1.07 786.9±7.05ns ? ?/sec
take stringview null values 1024 1.00 2.1±0.01µs ? ?/sec 1.22 2.6±0.05µs ? ?/sec
take stringview null values null indices 1024 1.00 2.9±0.02µs ? ?/sec 1.00 2.9±0.03µs ? ?/sec
```
</p>
</details>
--
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: github-unsubscribe@arrow.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
Re: [PR] Unroll take_native, interleave -25% [arrow-rs]
Posted by "Dandandan (via GitHub)" <gi...@apache.org>.
Dandandan commented on PR #9542:
URL: https://github.com/apache/arrow-rs/pull/9542#issuecomment-4044842158
Hmm this looks actually worse on this machine (probably x86 vs ARM...)...
--
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: github-unsubscribe@arrow.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
Re: [PR] Unroll interleave -25-30% [arrow-rs]
Posted by "comphead (via GitHub)" <gi...@apache.org>.
comphead commented on code in PR #9542:
URL: https://github.com/apache/arrow-rs/pull/9542#discussion_r2926754944
##########
arrow-select/src/interleave.rs:
##########
@@ -154,13 +154,51 @@ fn interleave_primitive<T: ArrowPrimitiveType>(
data_type: &DataType,
) -> Result<ArrayRef, ArrowError> {
let interleaved = Interleave::<'_, PrimitiveArray<T>>::new(values, indices);
+ let arrays = &interleaved.arrays;
+ let len = indices.len();
+
+ let mut output = Vec::with_capacity(len);
+ let dst: *mut T::Native = output.as_mut_ptr();
+ let mut base = 0;
+
+ // Process 8 elements at a time to issue multiple independent loads
+ // and increase memory-level parallelism for random access patterns.
+ let chunks = indices.chunks_exact(8);
+ let remainder = chunks.remainder();
+ for chunk in chunks {
+ let v0 = arrays[chunk[0].0].value(chunk[0].1);
+ let v1 = arrays[chunk[1].0].value(chunk[1].1);
+ let v2 = arrays[chunk[2].0].value(chunk[2].1);
+ let v3 = arrays[chunk[3].0].value(chunk[3].1);
+ let v4 = arrays[chunk[4].0].value(chunk[4].1);
+ let v5 = arrays[chunk[5].0].value(chunk[5].1);
+ let v6 = arrays[chunk[6].0].value(chunk[6].1);
+ let v7 = arrays[chunk[7].0].value(chunk[7].1);
+
+ // SAFETY: base+7 < len == output capacity
+ unsafe {
+ dst.add(base).write(v0);
+ dst.add(base + 1).write(v1);
+ dst.add(base + 2).write(v2);
+ dst.add(base + 3).write(v3);
+ dst.add(base + 4).write(v4);
+ dst.add(base + 5).write(v5);
+ dst.add(base + 6).write(v6);
+ dst.add(base + 7).write(v7);
+ }
+ base += 8;
+ }
- let values = indices
- .iter()
- .map(|(a, b)| interleaved.arrays[*a].value(*b))
- .collect::<Vec<_>>();
+ for idx in remainder {
+ // SAFETY: base < len == output capacity
+ unsafe { dst.add(base).write(arrays[idx.0].value(idx.1)) };
Review Comment:
for unsafe calls, would it make sense to `debug_assert` safety requirements to catch unexpected things on CI debug builds?
--
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: github-unsubscribe@arrow.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
Re: [PR] Unroll take_native -25% [arrow-rs]
Posted by "alamb-ghbot (via GitHub)" <gi...@apache.org>.
alamb-ghbot commented on PR #9542:
URL: https://github.com/apache/arrow-rs/pull/9542#issuecomment-4044775461
🤖 `./gh_compare_arrow.sh` [gh_compare_arrow.sh](https://github.com/alamb/datafusion-benchmarking/blob/main/scripts/gh_compare_arrow.sh) Running
Linux aal-dev 6.14.0-1018-gcp #19~24.04.1-Ubuntu SMP Wed Sep 24 23:23:09 UTC 2025 x86_64 x86_64 x86_64 GNU/Linux
Comparing take_native (50c5da775dbc1a1b3d7538488d0f00d984cd6601) to 9d0e8beae74fedb362d88cbc6e32d9760657c9de [diff](https://github.com/apache/arrow-rs/compare/9d0e8beae74fedb362d88cbc6e32d9760657c9de..50c5da775dbc1a1b3d7538488d0f00d984cd6601)
BENCH_NAME=take_kernels
BENCH_COMMAND=cargo bench --features=arrow,async,test_common,experimental,object_store --bench take_kernels
BENCH_FILTER=
BENCH_BRANCH_NAME=take_native
Results will be posted here when complete
--
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: github-unsubscribe@arrow.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
Re: [PR] Unroll interleave -25-30% [arrow-rs]
Posted by "Dandandan (via GitHub)" <gi...@apache.org>.
Dandandan commented on PR #9542:
URL: https://github.com/apache/arrow-rs/pull/9542#issuecomment-4050125199
run benchmark interleave_kernels
--
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: github-unsubscribe@arrow.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
Re: [PR] Unroll take_native, interleave -25% [arrow-rs]
Posted by "Dandandan (via GitHub)" <gi...@apache.org>.
Dandandan closed pull request #9542: Unroll take_native, interleave -25%
URL: https://github.com/apache/arrow-rs/pull/9542
--
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: github-unsubscribe@arrow.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
Re: [PR] Unroll take_native [arrow-rs]
Posted by "Dandandan (via GitHub)" <gi...@apache.org>.
Dandandan commented on PR #9542:
URL: https://github.com/apache/arrow-rs/pull/9542#issuecomment-4044772198
run benchmarks take_kernels
--
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: github-unsubscribe@arrow.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
Re: [PR] Unroll interleave -25% [arrow-rs]
Posted by "alamb-ghbot (via GitHub)" <gi...@apache.org>.
alamb-ghbot commented on PR #9542:
URL: https://github.com/apache/arrow-rs/pull/9542#issuecomment-4045235691
🤖 `./gh_compare_arrow.sh` [gh_compare_arrow.sh](https://github.com/alamb/datafusion-benchmarking/blob/main/scripts/gh_compare_arrow.sh) Running
Linux aal-dev 6.14.0-1018-gcp #19~24.04.1-Ubuntu SMP Wed Sep 24 23:23:09 UTC 2025 x86_64 x86_64 x86_64 GNU/Linux
Comparing take_native (01a11614bf79048b8932a330558ae4378cc1b29e) to 9d0e8beae74fedb362d88cbc6e32d9760657c9de [diff](https://github.com/apache/arrow-rs/compare/9d0e8beae74fedb362d88cbc6e32d9760657c9de..01a11614bf79048b8932a330558ae4378cc1b29e)
BENCH_NAME=interleave_kernels
BENCH_COMMAND=cargo bench --features=arrow,async,test_common,experimental,object_store --bench interleave_kernels
BENCH_FILTER=
BENCH_BRANCH_NAME=take_native
Results will be posted here when complete
--
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: github-unsubscribe@arrow.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
Re: [PR] Unroll interleave -25-30% [arrow-rs]
Posted by "Dandandan (via GitHub)" <gi...@apache.org>.
Dandandan commented on code in PR #9542:
URL: https://github.com/apache/arrow-rs/pull/9542#discussion_r2927448171
##########
arrow-select/src/interleave.rs:
##########
@@ -154,13 +154,51 @@ fn interleave_primitive<T: ArrowPrimitiveType>(
data_type: &DataType,
) -> Result<ArrayRef, ArrowError> {
let interleaved = Interleave::<'_, PrimitiveArray<T>>::new(values, indices);
+ let arrays = &interleaved.arrays;
+ let len = indices.len();
+
+ let mut output = Vec::with_capacity(len);
+ let dst: *mut T::Native = output.as_mut_ptr();
+ let mut base = 0;
+
+ // Process 8 elements at a time to issue multiple independent loads
+ // and increase memory-level parallelism for random access patterns.
+ let chunks = indices.chunks_exact(8);
+ let remainder = chunks.remainder();
+ for chunk in chunks {
+ let v0 = arrays[chunk[0].0].value(chunk[0].1);
Review Comment:
It just a loop unrolling constant - the higher the more code it will generate inside the loop so it can hide more latency by running/pipelining the instructions not depending on the loads.
--
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: github-unsubscribe@arrow.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
Re: [PR] Unroll take_native, interleave -25% [arrow-rs]
Posted by "alamb-ghbot (via GitHub)" <gi...@apache.org>.
alamb-ghbot commented on PR #9542:
URL: https://github.com/apache/arrow-rs/pull/9542#issuecomment-4044827438
🤖 `./gh_compare_arrow.sh` [gh_compare_arrow.sh](https://github.com/alamb/datafusion-benchmarking/blob/main/scripts/gh_compare_arrow.sh) Running
Linux aal-dev 6.14.0-1018-gcp #19~24.04.1-Ubuntu SMP Wed Sep 24 23:23:09 UTC 2025 x86_64 x86_64 x86_64 GNU/Linux
Comparing take_native (01a11614bf79048b8932a330558ae4378cc1b29e) to 9d0e8beae74fedb362d88cbc6e32d9760657c9de [diff](https://github.com/apache/arrow-rs/compare/9d0e8beae74fedb362d88cbc6e32d9760657c9de..01a11614bf79048b8932a330558ae4378cc1b29e)
BENCH_NAME=interleave_kernels
BENCH_COMMAND=cargo bench --features=arrow,async,test_common,experimental,object_store --bench interleave_kernels
BENCH_FILTER=
BENCH_BRANCH_NAME=take_native
Results will be posted here when complete
--
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: github-unsubscribe@arrow.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
Re: [PR] Unroll interleave -25-30% [arrow-rs]
Posted by "Dandandan (via GitHub)" <gi...@apache.org>.
Dandandan commented on code in PR #9542:
URL: https://github.com/apache/arrow-rs/pull/9542#discussion_r2927409628
##########
arrow-select/src/interleave.rs:
##########
@@ -154,13 +154,51 @@ fn interleave_primitive<T: ArrowPrimitiveType>(
data_type: &DataType,
) -> Result<ArrayRef, ArrowError> {
let interleaved = Interleave::<'_, PrimitiveArray<T>>::new(values, indices);
+ let arrays = &interleaved.arrays;
+ let len = indices.len();
+
+ let mut output = Vec::with_capacity(len);
+ let dst: *mut T::Native = output.as_mut_ptr();
+ let mut base = 0;
+
+ // Process 8 elements at a time to issue multiple independent loads
+ // and increase memory-level parallelism for random access patterns.
+ let chunks = indices.chunks_exact(8);
+ let remainder = chunks.remainder();
+ for chunk in chunks {
+ let v0 = arrays[chunk[0].0].value(chunk[0].1);
+ let v1 = arrays[chunk[1].0].value(chunk[1].1);
+ let v2 = arrays[chunk[2].0].value(chunk[2].1);
+ let v3 = arrays[chunk[3].0].value(chunk[3].1);
+ let v4 = arrays[chunk[4].0].value(chunk[4].1);
+ let v5 = arrays[chunk[5].0].value(chunk[5].1);
+ let v6 = arrays[chunk[6].0].value(chunk[6].1);
+ let v7 = arrays[chunk[7].0].value(chunk[7].1);
+
+ // SAFETY: base+7 < len == output capacity
+ unsafe {
+ dst.add(base).write(v0);
+ dst.add(base + 1).write(v1);
+ dst.add(base + 2).write(v2);
+ dst.add(base + 3).write(v3);
+ dst.add(base + 4).write(v4);
+ dst.add(base + 5).write(v5);
+ dst.add(base + 6).write(v6);
+ dst.add(base + 7).write(v7);
+ }
+ base += 8;
+ }
- let values = indices
- .iter()
- .map(|(a, b)| interleaved.arrays[*a].value(*b))
- .collect::<Vec<_>>();
+ for idx in remainder {
+ // SAFETY: base < len == output capacity
+ unsafe { dst.add(base).write(arrays[idx.0].value(idx.1)) };
Review Comment:
Could do tha
##########
arrow-select/src/interleave.rs:
##########
@@ -154,13 +154,51 @@ fn interleave_primitive<T: ArrowPrimitiveType>(
data_type: &DataType,
) -> Result<ArrayRef, ArrowError> {
let interleaved = Interleave::<'_, PrimitiveArray<T>>::new(values, indices);
+ let arrays = &interleaved.arrays;
+ let len = indices.len();
+
+ let mut output = Vec::with_capacity(len);
+ let dst: *mut T::Native = output.as_mut_ptr();
+ let mut base = 0;
+
+ // Process 8 elements at a time to issue multiple independent loads
+ // and increase memory-level parallelism for random access patterns.
+ let chunks = indices.chunks_exact(8);
+ let remainder = chunks.remainder();
+ for chunk in chunks {
+ let v0 = arrays[chunk[0].0].value(chunk[0].1);
+ let v1 = arrays[chunk[1].0].value(chunk[1].1);
+ let v2 = arrays[chunk[2].0].value(chunk[2].1);
+ let v3 = arrays[chunk[3].0].value(chunk[3].1);
+ let v4 = arrays[chunk[4].0].value(chunk[4].1);
+ let v5 = arrays[chunk[5].0].value(chunk[5].1);
+ let v6 = arrays[chunk[6].0].value(chunk[6].1);
+ let v7 = arrays[chunk[7].0].value(chunk[7].1);
+
+ // SAFETY: base+7 < len == output capacity
+ unsafe {
+ dst.add(base).write(v0);
+ dst.add(base + 1).write(v1);
+ dst.add(base + 2).write(v2);
+ dst.add(base + 3).write(v3);
+ dst.add(base + 4).write(v4);
+ dst.add(base + 5).write(v5);
+ dst.add(base + 6).write(v6);
+ dst.add(base + 7).write(v7);
+ }
+ base += 8;
+ }
- let values = indices
- .iter()
- .map(|(a, b)| interleaved.arrays[*a].value(*b))
- .collect::<Vec<_>>();
+ for idx in remainder {
+ // SAFETY: base < len == output capacity
+ unsafe { dst.add(base).write(arrays[idx.0].value(idx.1)) };
Review Comment:
Could do that
--
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: github-unsubscribe@arrow.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
Re: [PR] Unroll interleave -25-30% [arrow-rs]
Posted by "Dandandan (via GitHub)" <gi...@apache.org>.
Dandandan commented on code in PR #9542:
URL: https://github.com/apache/arrow-rs/pull/9542#discussion_r2929185034
##########
arrow-select/src/interleave.rs:
##########
@@ -154,13 +154,51 @@ fn interleave_primitive<T: ArrowPrimitiveType>(
data_type: &DataType,
) -> Result<ArrayRef, ArrowError> {
let interleaved = Interleave::<'_, PrimitiveArray<T>>::new(values, indices);
+ let arrays = &interleaved.arrays;
+ let len = indices.len();
+
+ let mut output = Vec::with_capacity(len);
+ let dst: *mut T::Native = output.as_mut_ptr();
+ let mut base = 0;
+
+ // Process 8 elements at a time to issue multiple independent loads
+ // and increase memory-level parallelism for random access patterns.
+ let chunks = indices.chunks_exact(8);
+ let remainder = chunks.remainder();
+ for chunk in chunks {
+ let v0 = arrays[chunk[0].0].value(chunk[0].1);
Review Comment:
Ah like that - yeah that should be possible (would require some hardware setup) although I think it would not be really to a user setting compile time constants.
On the bencmarks it looks to be beneficial for both aarch64 and x86 CPUs - getting ~23% on M1 (my laptop) and -30% on intel (the @alamb-ghbot runner).
##########
arrow-select/src/interleave.rs:
##########
@@ -154,13 +154,51 @@ fn interleave_primitive<T: ArrowPrimitiveType>(
data_type: &DataType,
) -> Result<ArrayRef, ArrowError> {
let interleaved = Interleave::<'_, PrimitiveArray<T>>::new(values, indices);
+ let arrays = &interleaved.arrays;
+ let len = indices.len();
+
+ let mut output = Vec::with_capacity(len);
+ let dst: *mut T::Native = output.as_mut_ptr();
+ let mut base = 0;
+
+ // Process 8 elements at a time to issue multiple independent loads
+ // and increase memory-level parallelism for random access patterns.
+ let chunks = indices.chunks_exact(8);
+ let remainder = chunks.remainder();
+ for chunk in chunks {
+ let v0 = arrays[chunk[0].0].value(chunk[0].1);
Review Comment:
Ah like that - yeah that should be possible (would require some hardware setup) although I think it would not be really to a user setting compile time constants.
On the benchmarks it looks to be beneficial for both aarch64 and x86 CPUs - getting ~23% on M1 (my laptop) and -30% on intel (the @alamb-ghbot runner).
--
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: github-unsubscribe@arrow.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
Re: [PR] Unroll interleave -25-30% [arrow-rs]
Posted by "Dandandan (via GitHub)" <gi...@apache.org>.
Dandandan commented on PR #9542:
URL: https://github.com/apache/arrow-rs/pull/9542#issuecomment-4048688515
@comphead @andygrove FYI as you seem to be interested in sort performance (which is one of the consumers)
--
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: github-unsubscribe@arrow.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
Re: [PR] Unroll take_native -25% [arrow-rs]
Posted by "alamb-ghbot (via GitHub)" <gi...@apache.org>.
alamb-ghbot commented on PR #9542:
URL: https://github.com/apache/arrow-rs/pull/9542#issuecomment-4044772669
🤖 Hi @Dandandan, thanks for the request (https://github.com/apache/arrow-rs/pull/9542#issuecomment-4044772198).
[`scrape_comments.py`](https://github.com/alamb/datafusion-benchmarking/blob/main/scripts/scrape_comments.py) only supports whitelisted benchmarks.
- Standard: (none)
- Criterion: array_from, array_iter, arrow_reader, arrow_reader_clickbench, arrow_reader_row_filter, arrow_statistics, arrow_writer, bitwise_kernel, boolean_kernels, buffer_bit_ops, builder, cast_kernels, coalesce_kernels, comparison_kernels, concatenate_kernel, csv_writer, encoding, filter_kernels, interleave_kernels, ipc_reader, json-reader, metadata, row_format, sort_kernel, take_kernels, union_array, variant_builder, variant_kernels, variant_validation, view_types, zip_kernels
Please choose one or more of these with `run benchmark <name>` or `run benchmark <name1> <name2>...`
You can also set environment variables on subsequent lines:
```
run benchmark tpch_mem
DATAFUSION_RUNTIME_MEMORY_LIMIT=1G
```
--
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: github-unsubscribe@arrow.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
Re: [PR] Unroll take_native, interleave -25% [arrow-rs]
Posted by "alamb-ghbot (via GitHub)" <gi...@apache.org>.
alamb-ghbot commented on PR #9542:
URL: https://github.com/apache/arrow-rs/pull/9542#issuecomment-4044927756
🤖: Benchmark completed
<details><summary>Details</summary>
<p>
```
group main take_native
----- ---- -----------
interleave dict(20, 0.0) 100 [0..100, 100..230, 450..1000] 1.08 805.6±8.28ns ? ?/sec 1.00 748.5±14.05ns ? ?/sec
interleave dict(20, 0.0) 1024 [0..100, 100..230, 450..1000, 0..1000] 1.18 2.6±0.00µs ? ?/sec 1.00 2.2±0.01µs ? ?/sec
interleave dict(20, 0.0) 1024 [0..100, 100..230, 450..1000] 1.21 2.6±0.01µs ? ?/sec 1.00 2.2±0.02µs ? ?/sec
interleave dict(20, 0.0) 400 [0..100, 100..230, 450..1000] 1.16 1431.6±3.11ns ? ?/sec 1.00 1232.9±14.26ns ? ?/sec
interleave dict_distinct 100 1.03 2.9±0.12µs ? ?/sec 1.00 2.9±0.07µs ? ?/sec
interleave dict_distinct 1024 1.02 2.9±0.06µs ? ?/sec 1.00 2.8±0.03µs ? ?/sec
interleave dict_distinct 2048 1.03 2.9±0.02µs ? ?/sec 1.00 2.8±0.08µs ? ?/sec
interleave dict_sparse(20, 0.0) 100 [0..100, 100..230, 450..1000] 1.00 2.7±0.26µs ? ?/sec 1.02 2.8±0.21µs ? ?/sec
interleave dict_sparse(20, 0.0) 1024 [0..100, 100..230, 450..1000, 0..1000] 1.11 5.3±0.31µs ? ?/sec 1.00 4.8±0.40µs ? ?/sec
interleave dict_sparse(20, 0.0) 1024 [0..100, 100..230, 450..1000] 1.16 4.8±0.25µs ? ?/sec 1.00 4.1±0.23µs ? ?/sec
interleave dict_sparse(20, 0.0) 400 [0..100, 100..230, 450..1000] 1.05 3.5±0.31µs ? ?/sec 1.00 3.3±0.29µs ? ?/sec
interleave i32(0.0) 100 [0..100, 100..230, 450..1000] 1.21 313.8±1.03ns ? ?/sec 1.00 258.9±4.98ns ? ?/sec
interleave i32(0.0) 1024 [0..100, 100..230, 450..1000, 0..1000] 1.34 1856.5±17.40ns ? ?/sec 1.00 1385.9±32.73ns ? ?/sec
interleave i32(0.0) 1024 [0..100, 100..230, 450..1000] 1.34 1848.6±8.80ns ? ?/sec 1.00 1382.4±48.64ns ? ?/sec
interleave i32(0.0) 400 [0..100, 100..230, 450..1000] 1.37 843.3±7.37ns ? ?/sec 1.00 615.5±22.71ns ? ?/sec
interleave i32(0.5) 100 [0..100, 100..230, 450..1000] 1.09 604.2±5.60ns ? ?/sec 1.00 555.1±4.48ns ? ?/sec
interleave i32(0.5) 1024 [0..100, 100..230, 450..1000, 0..1000] 1.12 4.3±0.01µs ? ?/sec 1.00 3.8±0.04µs ? ?/sec
interleave i32(0.5) 1024 [0..100, 100..230, 450..1000] 1.13 4.4±0.06µs ? ?/sec 1.00 3.9±0.17µs ? ?/sec
interleave i32(0.5) 400 [0..100, 100..230, 450..1000] 1.12 1889.4±19.68ns ? ?/sec 1.00 1691.5±17.15ns ? ?/sec
interleave list<i64>(0.0,0.0,20) 100 [0..100, 100..230, 450..1000] 1.07 2.7±0.03µs ? ?/sec 1.00 2.5±0.03µs ? ?/sec
interleave list<i64>(0.0,0.0,20) 1024 [0..100, 100..230, 450..1000, 0..1000] 1.06 26.2±0.11µs ? ?/sec 1.00 24.6±0.31µs ? ?/sec
interleave list<i64>(0.0,0.0,20) 1024 [0..100, 100..230, 450..1000] 1.06 25.9±0.14µs ? ?/sec 1.00 24.5±0.29µs ? ?/sec
interleave list<i64>(0.0,0.0,20) 400 [0..100, 100..230, 450..1000] 1.07 10.5±0.21µs ? ?/sec 1.00 9.9±0.06µs ? ?/sec
interleave list<i64>(0.1,0.1,20) 100 [0..100, 100..230, 450..1000] 1.05 5.8±0.25µs ? ?/sec 1.00 5.5±0.06µs ? ?/sec
interleave list<i64>(0.1,0.1,20) 1024 [0..100, 100..230, 450..1000, 0..1000] 1.05 47.4±2.23µs ? ?/sec 1.00 45.2±0.14µs ? ?/sec
interleave list<i64>(0.1,0.1,20) 1024 [0..100, 100..230, 450..1000] 1.06 48.0±2.35µs ? ?/sec 1.00 45.5±0.64µs ? ?/sec
interleave list<i64>(0.1,0.1,20) 400 [0..100, 100..230, 450..1000] 1.05 19.2±0.90µs ? ?/sec 1.00 18.2±0.03µs ? ?/sec
interleave str(20, 0.0) 100 [0..100, 100..230, 450..1000] 1.01 786.8±1.50ns ? ?/sec 1.00 779.4±4.35ns ? ?/sec
interleave str(20, 0.0) 1024 [0..100, 100..230, 450..1000, 0..1000] 1.04 6.3±0.12µs ? ?/sec 1.00 6.0±0.02µs ? ?/sec
interleave str(20, 0.0) 1024 [0..100, 100..230, 450..1000] 1.04 6.2±0.08µs ? ?/sec 1.00 6.0±0.01µs ? ?/sec
interleave str(20, 0.0) 400 [0..100, 100..230, 450..1000] 1.09 2.7±0.01µs ? ?/sec 1.00 2.4±0.01µs ? ?/sec
interleave str(20, 0.5) 100 [0..100, 100..230, 450..1000] 1.04 1064.4±19.37ns ? ?/sec 1.00 1023.8±3.56ns ? ?/sec
interleave str(20, 0.5) 1024 [0..100, 100..230, 450..1000, 0..1000] 1.03 10.3±0.06µs ? ?/sec 1.00 10.1±0.13µs ? ?/sec
interleave str(20, 0.5) 1024 [0..100, 100..230, 450..1000] 1.02 10.3±0.05µs ? ?/sec 1.00 10.1±0.54µs ? ?/sec
interleave str(20, 0.5) 400 [0..100, 100..230, 450..1000] 1.04 3.7±0.03µs ? ?/sec 1.00 3.6±0.17µs ? ?/sec
interleave str_view(0.0) 100 [0..100, 100..230, 450..1000] 1.01 856.9±2.90ns ? ?/sec 1.00 849.1±7.00ns ? ?/sec
interleave str_view(0.0) 1024 [0..100, 100..230, 450..1000, 0..1000] 1.00 5.0±0.15µs ? ?/sec 1.02 5.1±0.02µs ? ?/sec
interleave str_view(0.0) 1024 [0..100, 100..230, 450..1000] 1.00 4.9±0.05µs ? ?/sec 1.04 5.1±0.02µs ? ?/sec
interleave str_view(0.0) 400 [0..100, 100..230, 450..1000] 1.00 2.2±0.05µs ? ?/sec 1.03 2.2±0.01µs ? ?/sec
interleave struct(i32(0.0), i32(0.0) 100 [0..100, 100..230, 450..1000] 1.20 874.3±4.12ns ? ?/sec 1.00 729.1±12.04ns ? ?/sec
interleave struct(i32(0.0), i32(0.0) 1024 [0..100, 100..230, 450..1000, 0..1000] 1.34 4.0±0.01µs ? ?/sec 1.00 3.0±0.02µs ? ?/sec
interleave struct(i32(0.0), i32(0.0) 1024 [0..100, 100..230, 450..1000] 1.31 4.0±0.04µs ? ?/sec 1.00 3.0±0.01µs ? ?/sec
interleave struct(i32(0.0), i32(0.0) 400 [0..100, 100..230, 450..1000] 1.24 1905.1±19.48ns ? ?/sec 1.00 1532.8±33.13ns ? ?/sec
interleave struct(i32(0.0), str(20, 0.0) 100 [0..100, 100..230, 450..1000] 1.00 1340.9±6.76ns ? ?/sec 1.01 1347.8±12.50ns ? ?/sec
interleave struct(i32(0.0), str(20, 0.0) 1024 [0..100, 100..230, 450..1000, 0..1000] 1.08 8.3±0.16µs ? ?/sec 1.00 7.7±0.02µs ? ?/sec
interleave struct(i32(0.0), str(20, 0.0) 1024 [0..100, 100..230, 450..1000] 1.08 8.3±0.06µs ? ?/sec 1.00 7.7±0.06µs ? ?/sec
interleave struct(i32(0.0), str(20, 0.0) 400 [0..100, 100..230, 450..1000] 1.09 3.7±0.13µs ? ?/sec 1.00 3.4±0.02µs ? ?/sec
interleave struct(str(20, 0.0), str(20, 0.0)) 100 [0..100, 100..230, 450..1000] 1.05 1927.3±9.31ns ? ?/sec 1.00 1842.2±18.19ns ? ?/sec
interleave struct(str(20, 0.0), str(20, 0.0)) 1024 [0..100, 100..230, 450..1000, 0..1000] 1.04 12.6±0.06µs ? ?/sec 1.00 12.1±0.08µs ? ?/sec
interleave struct(str(20, 0.0), str(20, 0.0)) 1024 [0..100, 100..230, 450..1000] 1.04 12.6±0.03µs ? ?/sec 1.00 12.1±0.14µs ? ?/sec
interleave struct(str(20, 0.0), str(20, 0.0)) 400 [0..100, 100..230, 450..1000] 1.04 5.4±0.07µs ? ?/sec 1.00 5.2±0.04µs ? ?/sec
```
</p>
</details>
--
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: github-unsubscribe@arrow.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
Re: [PR] Unroll interleave -25% [arrow-rs]
Posted by "Dandandan (via GitHub)" <gi...@apache.org>.
Dandandan commented on PR #9542:
URL: https://github.com/apache/arrow-rs/pull/9542#issuecomment-4045235132
run benchmark interleave_kernels
--
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: github-unsubscribe@arrow.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
Re: [PR] Unroll interleave -25-30% [arrow-rs]
Posted by "alamb-ghbot (via GitHub)" <gi...@apache.org>.
alamb-ghbot commented on PR #9542:
URL: https://github.com/apache/arrow-rs/pull/9542#issuecomment-4050427315
🤖: Benchmark completed
<details><summary>Details</summary>
<p>
```
group main take_native
----- ---- -----------
interleave dict(20, 0.0) 100 [0..100, 100..230, 450..1000] 1.00 757.1±3.84ns ? ?/sec 1.03 781.8±2.85ns ? ?/sec
interleave dict(20, 0.0) 1024 [0..100, 100..230, 450..1000, 0..1000] 1.00 2.2±0.00µs ? ?/sec 1.18 2.6±0.01µs ? ?/sec
interleave dict(20, 0.0) 1024 [0..100, 100..230, 450..1000] 1.00 2.2±0.00µs ? ?/sec 1.17 2.6±0.00µs ? ?/sec
interleave dict(20, 0.0) 400 [0..100, 100..230, 450..1000] 1.00 1242.5±3.19ns ? ?/sec 1.12 1396.2±5.93ns ? ?/sec
interleave dict_distinct 100 1.04 3.0±0.02µs ? ?/sec 1.00 2.8±0.02µs ? ?/sec
interleave dict_distinct 1024 1.05 3.0±0.02µs ? ?/sec 1.00 2.8±0.02µs ? ?/sec
interleave dict_distinct 2048 1.05 3.0±0.03µs ? ?/sec 1.00 2.8±0.04µs ? ?/sec
interleave dict_sparse(20, 0.0) 100 [0..100, 100..230, 450..1000] 1.00 2.5±0.20µs ? ?/sec 1.08 2.7±0.19µs ? ?/sec
interleave dict_sparse(20, 0.0) 1024 [0..100, 100..230, 450..1000, 0..1000] 1.01 5.0±0.36µs ? ?/sec 1.00 5.0±0.25µs ? ?/sec
interleave dict_sparse(20, 0.0) 1024 [0..100, 100..230, 450..1000] 1.00 4.2±0.22µs ? ?/sec 1.06 4.5±0.15µs ? ?/sec
interleave dict_sparse(20, 0.0) 400 [0..100, 100..230, 450..1000] 1.00 3.2±0.21µs ? ?/sec 1.03 3.3±0.17µs ? ?/sec
interleave i32(0.0) 100 [0..100, 100..230, 450..1000] 1.23 315.3±2.70ns ? ?/sec 1.00 256.8±3.27ns ? ?/sec
interleave i32(0.0) 1024 [0..100, 100..230, 450..1000, 0..1000] 1.35 1857.3±30.30ns ? ?/sec 1.00 1377.1±21.39ns ? ?/sec
interleave i32(0.0) 1024 [0..100, 100..230, 450..1000] 1.35 1845.9±16.19ns ? ?/sec 1.00 1362.8±49.93ns ? ?/sec
interleave i32(0.0) 400 [0..100, 100..230, 450..1000] 1.39 843.9±6.31ns ? ?/sec 1.00 605.4±3.22ns ? ?/sec
interleave i32(0.5) 100 [0..100, 100..230, 450..1000] 1.15 636.1±7.00ns ? ?/sec 1.00 555.5±2.24ns ? ?/sec
interleave i32(0.5) 1024 [0..100, 100..230, 450..1000, 0..1000] 1.13 4.3±0.06µs ? ?/sec 1.00 3.8±0.01µs ? ?/sec
interleave i32(0.5) 1024 [0..100, 100..230, 450..1000] 1.15 4.4±0.07µs ? ?/sec 1.00 3.8±0.01µs ? ?/sec
interleave i32(0.5) 400 [0..100, 100..230, 450..1000] 1.12 1901.5±13.38ns ? ?/sec 1.00 1693.7±6.90ns ? ?/sec
interleave list<i64>(0.0,0.0,20) 100 [0..100, 100..230, 450..1000] 1.15 2.9±0.01µs ? ?/sec 1.00 2.5±0.01µs ? ?/sec
interleave list<i64>(0.0,0.0,20) 1024 [0..100, 100..230, 450..1000, 0..1000] 1.18 29.0±0.14µs ? ?/sec 1.00 24.6±0.11µs ? ?/sec
interleave list<i64>(0.0,0.0,20) 1024 [0..100, 100..230, 450..1000] 1.18 28.9±0.29µs ? ?/sec 1.00 24.5±0.24µs ? ?/sec
interleave list<i64>(0.0,0.0,20) 400 [0..100, 100..230, 450..1000] 1.17 11.6±0.08µs ? ?/sec 1.00 10.0±0.22µs ? ?/sec
interleave list<i64>(0.1,0.1,20) 100 [0..100, 100..230, 450..1000] 1.08 6.0±0.10µs ? ?/sec 1.00 5.6±0.28µs ? ?/sec
interleave list<i64>(0.1,0.1,20) 1024 [0..100, 100..230, 450..1000, 0..1000] 1.07 48.8±0.19µs ? ?/sec 1.00 45.4±0.53µs ? ?/sec
interleave list<i64>(0.1,0.1,20) 1024 [0..100, 100..230, 450..1000] 1.08 49.5±0.84µs ? ?/sec 1.00 45.7±0.35µs ? ?/sec
interleave list<i64>(0.1,0.1,20) 400 [0..100, 100..230, 450..1000] 1.08 19.8±0.35µs ? ?/sec 1.00 18.3±0.24µs ? ?/sec
interleave str(20, 0.0) 100 [0..100, 100..230, 450..1000] 1.01 769.5±12.22ns ? ?/sec 1.00 761.8±1.21ns ? ?/sec
interleave str(20, 0.0) 1024 [0..100, 100..230, 450..1000, 0..1000] 1.00 5.9±0.03µs ? ?/sec 1.02 6.0±0.03µs ? ?/sec
interleave str(20, 0.0) 1024 [0..100, 100..230, 450..1000] 1.00 5.9±0.20µs ? ?/sec 1.01 6.0±0.05µs ? ?/sec
interleave str(20, 0.0) 400 [0..100, 100..230, 450..1000] 1.00 2.4±0.01µs ? ?/sec 1.00 2.4±0.00µs ? ?/sec
interleave str(20, 0.5) 100 [0..100, 100..230, 450..1000] 1.01 1063.5±9.67ns ? ?/sec 1.00 1053.0±3.78ns ? ?/sec
interleave str(20, 0.5) 1024 [0..100, 100..230, 450..1000, 0..1000] 1.00 9.9±0.04µs ? ?/sec 1.01 9.9±0.06µs ? ?/sec
interleave str(20, 0.5) 1024 [0..100, 100..230, 450..1000] 1.00 9.8±0.05µs ? ?/sec 1.02 9.9±0.11µs ? ?/sec
interleave str(20, 0.5) 400 [0..100, 100..230, 450..1000] 1.00 3.5±0.02µs ? ?/sec 1.00 3.5±0.02µs ? ?/sec
interleave str_view(0.0) 100 [0..100, 100..230, 450..1000] 1.00 830.7±19.98ns ? ?/sec 1.01 842.1±5.24ns ? ?/sec
interleave str_view(0.0) 1024 [0..100, 100..230, 450..1000, 0..1000] 1.00 4.8±0.06µs ? ?/sec 1.06 5.1±0.03µs ? ?/sec
interleave str_view(0.0) 1024 [0..100, 100..230, 450..1000] 1.00 4.7±0.01µs ? ?/sec 1.07 5.0±0.05µs ? ?/sec
interleave str_view(0.0) 400 [0..100, 100..230, 450..1000] 1.00 2.1±0.07µs ? ?/sec 1.07 2.2±0.02µs ? ?/sec
interleave struct(i32(0.0), i32(0.0) 100 [0..100, 100..230, 450..1000] 1.22 872.0±9.63ns ? ?/sec 1.00 716.9±2.32ns ? ?/sec
interleave struct(i32(0.0), i32(0.0) 1024 [0..100, 100..230, 450..1000, 0..1000] 1.33 4.1±0.16µs ? ?/sec 1.00 3.0±0.02µs ? ?/sec
interleave struct(i32(0.0), i32(0.0) 1024 [0..100, 100..230, 450..1000] 1.32 4.0±0.05µs ? ?/sec 1.00 3.0±0.01µs ? ?/sec
interleave struct(i32(0.0), i32(0.0) 400 [0..100, 100..230, 450..1000] 1.25 1900.5±20.74ns ? ?/sec 1.00 1525.3±11.31ns ? ?/sec
interleave struct(i32(0.0), str(20, 0.0) 100 [0..100, 100..230, 450..1000] 1.06 1365.6±3.87ns ? ?/sec 1.00 1292.1±9.39ns ? ?/sec
interleave struct(i32(0.0), str(20, 0.0) 1024 [0..100, 100..230, 450..1000, 0..1000] 1.07 8.1±0.09µs ? ?/sec 1.00 7.6±0.06µs ? ?/sec
interleave struct(i32(0.0), str(20, 0.0) 1024 [0..100, 100..230, 450..1000] 1.06 8.0±0.02µs ? ?/sec 1.00 7.6±0.04µs ? ?/sec
interleave struct(i32(0.0), str(20, 0.0) 400 [0..100, 100..230, 450..1000] 1.02 3.6±0.01µs ? ?/sec 1.00 3.5±0.01µs ? ?/sec
interleave struct(str(20, 0.0), str(20, 0.0)) 100 [0..100, 100..230, 450..1000] 1.02 1886.9±14.41ns ? ?/sec 1.00 1854.1±25.88ns ? ?/sec
interleave struct(str(20, 0.0), str(20, 0.0)) 1024 [0..100, 100..230, 450..1000, 0..1000] 1.03 12.6±0.20µs ? ?/sec 1.00 12.2±0.09µs ? ?/sec
interleave struct(str(20, 0.0), str(20, 0.0)) 1024 [0..100, 100..230, 450..1000] 1.01 12.3±0.08µs ? ?/sec 1.00 12.2±0.10µs ? ?/sec
interleave struct(str(20, 0.0), str(20, 0.0)) 400 [0..100, 100..230, 450..1000] 1.00 5.2±0.11µs ? ?/sec 1.02 5.3±0.09µs ? ?/sec
```
</p>
</details>
--
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: github-unsubscribe@arrow.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
Re: [PR] Unroll interleave -25-30% [arrow-rs]
Posted by "Dandandan (via GitHub)" <gi...@apache.org>.
Dandandan commented on PR #9542:
URL: https://github.com/apache/arrow-rs/pull/9542#issuecomment-4050094511
> I'm curious what a SIMD gather (AVX2, SVE) intrinsics' performance is like vs. unrolling manually on modern architectures. The former likely ends up as similar μops anyway. I don't love specializing the code like this (in either direction, SIMD intrinsics or unrolling especially with ARM increasingly prevalent in the cloud) but if today it's faster I'm okay with it. It's also a pretty well-understood technique at this point.
Agreed. Auto-vectorization is preferred and in an ideal scenario we don't need to specialize the code and write our APIs so LLVM can generate good code automatically.
Ideally I think, the `interleave` and `take` APIs would also have a way of not doing bounds checks - in that case the compiler could do a much better job at vectorizing the code and not having the branches.
I am thinking perhaps we can at least remove the bounds checks for `interleave` of the outer batch dimension by doing it upfront - I'll check that now.
--
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: github-unsubscribe@arrow.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
Re: [PR] Unroll take_native, interleave -25% [arrow-rs]
Posted by "alamb-ghbot (via GitHub)" <gi...@apache.org>.
alamb-ghbot commented on PR #9542:
URL: https://github.com/apache/arrow-rs/pull/9542#issuecomment-4045109369
🤖 `./gh_compare_arrow.sh` [gh_compare_arrow.sh](https://github.com/alamb/datafusion-benchmarking/blob/main/scripts/gh_compare_arrow.sh) Running
Linux aal-dev 6.14.0-1018-gcp #19~24.04.1-Ubuntu SMP Wed Sep 24 23:23:09 UTC 2025 x86_64 x86_64 x86_64 GNU/Linux
Comparing take_native (01a11614bf79048b8932a330558ae4378cc1b29e) to 9d0e8beae74fedb362d88cbc6e32d9760657c9de [diff](https://github.com/apache/arrow-rs/compare/9d0e8beae74fedb362d88cbc6e32d9760657c9de..01a11614bf79048b8932a330558ae4378cc1b29e)
BENCH_NAME=take_kernels
BENCH_COMMAND=cargo bench --features=arrow,async,test_common,experimental,object_store --bench take_kernels
BENCH_FILTER=
BENCH_BRANCH_NAME=take_native
Results will be posted here when complete
--
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: github-unsubscribe@arrow.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
Re: [PR] Unroll interleave -25% [arrow-rs]
Posted by "alamb-ghbot (via GitHub)" <gi...@apache.org>.
alamb-ghbot commented on PR #9542:
URL: https://github.com/apache/arrow-rs/pull/9542#issuecomment-4045350937
🤖: Benchmark completed
<details><summary>Details</summary>
<p>
```
group main take_native
----- ---- -----------
interleave dict(20, 0.0) 100 [0..100, 100..230, 450..1000] 1.04 805.2±7.13ns ? ?/sec 1.00 777.5±7.56ns ? ?/sec
interleave dict(20, 0.0) 1024 [0..100, 100..230, 450..1000, 0..1000] 1.17 2.6±0.02µs ? ?/sec 1.00 2.2±0.01µs ? ?/sec
interleave dict(20, 0.0) 1024 [0..100, 100..230, 450..1000] 1.19 2.6±0.01µs ? ?/sec 1.00 2.2±0.01µs ? ?/sec
interleave dict(20, 0.0) 400 [0..100, 100..230, 450..1000] 1.12 1418.5±8.89ns ? ?/sec 1.00 1264.8±6.99ns ? ?/sec
interleave dict_distinct 100 1.00 2.8±0.03µs ? ?/sec 1.03 2.9±0.05µs ? ?/sec
interleave dict_distinct 1024 1.00 2.8±0.02µs ? ?/sec 1.02 2.9±0.04µs ? ?/sec
interleave dict_distinct 2048 1.00 2.8±0.05µs ? ?/sec 1.02 2.9±0.02µs ? ?/sec
interleave dict_sparse(20, 0.0) 100 [0..100, 100..230, 450..1000] 1.00 2.7±0.21µs ? ?/sec 1.07 2.8±0.29µs ? ?/sec
interleave dict_sparse(20, 0.0) 1024 [0..100, 100..230, 450..1000, 0..1000] 1.17 5.5±0.30µs ? ?/sec 1.00 4.7±0.36µs ? ?/sec
interleave dict_sparse(20, 0.0) 1024 [0..100, 100..230, 450..1000] 1.14 4.8±0.28µs ? ?/sec 1.00 4.2±0.25µs ? ?/sec
interleave dict_sparse(20, 0.0) 400 [0..100, 100..230, 450..1000] 1.07 3.4±0.27µs ? ?/sec 1.00 3.2±0.20µs ? ?/sec
interleave i32(0.0) 100 [0..100, 100..230, 450..1000] 1.24 316.6±3.70ns ? ?/sec 1.00 256.1±0.49ns ? ?/sec
interleave i32(0.0) 1024 [0..100, 100..230, 450..1000, 0..1000] 1.35 1849.9±4.01ns ? ?/sec 1.00 1374.3±4.54ns ? ?/sec
interleave i32(0.0) 1024 [0..100, 100..230, 450..1000] 1.36 1843.1±18.43ns ? ?/sec 1.00 1354.8±1.82ns ? ?/sec
interleave i32(0.0) 400 [0..100, 100..230, 450..1000] 1.40 842.3±7.51ns ? ?/sec 1.00 603.6±1.46ns ? ?/sec
interleave i32(0.5) 100 [0..100, 100..230, 450..1000] 1.10 607.4±5.88ns ? ?/sec 1.00 554.6±35.14ns ? ?/sec
interleave i32(0.5) 1024 [0..100, 100..230, 450..1000, 0..1000] 1.13 4.3±0.06µs ? ?/sec 1.00 3.8±0.02µs ? ?/sec
interleave i32(0.5) 1024 [0..100, 100..230, 450..1000] 1.14 4.4±0.03µs ? ?/sec 1.00 3.8±0.03µs ? ?/sec
interleave i32(0.5) 400 [0..100, 100..230, 450..1000] 1.12 1893.6±12.56ns ? ?/sec 1.00 1687.5±23.57ns ? ?/sec
interleave list<i64>(0.0,0.0,20) 100 [0..100, 100..230, 450..1000] 1.07 2.7±0.03µs ? ?/sec 1.00 2.5±0.02µs ? ?/sec
interleave list<i64>(0.0,0.0,20) 1024 [0..100, 100..230, 450..1000, 0..1000] 1.06 26.1±0.12µs ? ?/sec 1.00 24.7±0.28µs ? ?/sec
interleave list<i64>(0.0,0.0,20) 1024 [0..100, 100..230, 450..1000] 1.05 25.9±0.10µs ? ?/sec 1.00 24.6±0.46µs ? ?/sec
interleave list<i64>(0.0,0.0,20) 400 [0..100, 100..230, 450..1000] 1.07 10.6±0.14µs ? ?/sec 1.00 9.9±0.04µs ? ?/sec
interleave list<i64>(0.1,0.1,20) 100 [0..100, 100..230, 450..1000] 1.03 5.7±0.02µs ? ?/sec 1.00 5.5±0.01µs ? ?/sec
interleave list<i64>(0.1,0.1,20) 1024 [0..100, 100..230, 450..1000, 0..1000] 1.03 46.4±0.36µs ? ?/sec 1.00 45.0±0.12µs ? ?/sec
interleave list<i64>(0.1,0.1,20) 1024 [0..100, 100..230, 450..1000] 1.04 47.0±1.18µs ? ?/sec 1.00 45.4±0.09µs ? ?/sec
interleave list<i64>(0.1,0.1,20) 400 [0..100, 100..230, 450..1000] 1.03 18.7±0.10µs ? ?/sec 1.00 18.2±0.10µs ? ?/sec
interleave str(20, 0.0) 100 [0..100, 100..230, 450..1000] 1.02 791.5±9.05ns ? ?/sec 1.00 778.5±1.54ns ? ?/sec
interleave str(20, 0.0) 1024 [0..100, 100..230, 450..1000, 0..1000] 1.04 6.3±0.03µs ? ?/sec 1.00 6.1±0.15µs ? ?/sec
interleave str(20, 0.0) 1024 [0..100, 100..230, 450..1000] 1.06 6.3±0.04µs ? ?/sec 1.00 6.0±0.02µs ? ?/sec
interleave str(20, 0.0) 400 [0..100, 100..230, 450..1000] 1.04 2.5±0.01µs ? ?/sec 1.00 2.4±0.00µs ? ?/sec
interleave str(20, 0.5) 100 [0..100, 100..230, 450..1000] 1.03 1060.4±10.47ns ? ?/sec 1.00 1024.7±11.65ns ? ?/sec
interleave str(20, 0.5) 1024 [0..100, 100..230, 450..1000, 0..1000] 1.03 10.4±0.20µs ? ?/sec 1.00 10.2±0.08µs ? ?/sec
interleave str(20, 0.5) 1024 [0..100, 100..230, 450..1000] 1.06 10.5±0.45µs ? ?/sec 1.00 10.0±0.10µs ? ?/sec
interleave str(20, 0.5) 400 [0..100, 100..230, 450..1000] 1.06 3.7±0.31µs ? ?/sec 1.00 3.5±0.04µs ? ?/sec
interleave str_view(0.0) 100 [0..100, 100..230, 450..1000] 1.05 898.7±7.81ns ? ?/sec 1.00 854.0±34.76ns ? ?/sec
interleave str_view(0.0) 1024 [0..100, 100..230, 450..1000, 0..1000] 1.00 5.1±0.01µs ? ?/sec 1.01 5.1±0.12µs ? ?/sec
interleave str_view(0.0) 1024 [0..100, 100..230, 450..1000] 1.00 5.0±0.02µs ? ?/sec 1.00 5.0±0.04µs ? ?/sec
interleave str_view(0.0) 400 [0..100, 100..230, 450..1000] 1.00 2.2±0.01µs ? ?/sec 1.00 2.2±0.01µs ? ?/sec
interleave struct(i32(0.0), i32(0.0) 100 [0..100, 100..230, 450..1000] 1.29 940.0±10.08ns ? ?/sec 1.00 730.4±9.49ns ? ?/sec
interleave struct(i32(0.0), i32(0.0) 1024 [0..100, 100..230, 450..1000, 0..1000] 1.34 4.1±0.06µs ? ?/sec 1.00 3.0±0.01µs ? ?/sec
interleave struct(i32(0.0), i32(0.0) 1024 [0..100, 100..230, 450..1000] 1.28 3.9±0.01µs ? ?/sec 1.00 3.0±0.01µs ? ?/sec
interleave struct(i32(0.0), i32(0.0) 400 [0..100, 100..230, 450..1000] 1.29 1964.3±19.31ns ? ?/sec 1.00 1528.5±18.37ns ? ?/sec
interleave struct(i32(0.0), str(20, 0.0) 100 [0..100, 100..230, 450..1000] 1.02 1342.7±6.89ns ? ?/sec 1.00 1318.3±12.41ns ? ?/sec
interleave struct(i32(0.0), str(20, 0.0) 1024 [0..100, 100..230, 450..1000, 0..1000] 1.09 8.3±0.05µs ? ?/sec 1.00 7.7±0.02µs ? ?/sec
interleave struct(i32(0.0), str(20, 0.0) 1024 [0..100, 100..230, 450..1000] 1.08 8.3±0.10µs ? ?/sec 1.00 7.7±0.02µs ? ?/sec
interleave struct(i32(0.0), str(20, 0.0) 400 [0..100, 100..230, 450..1000] 1.03 3.6±0.05µs ? ?/sec 1.00 3.5±0.01µs ? ?/sec
interleave struct(str(20, 0.0), str(20, 0.0)) 100 [0..100, 100..230, 450..1000] 1.00 1839.7±51.76ns ? ?/sec 1.00 1841.1±23.00ns ? ?/sec
interleave struct(str(20, 0.0), str(20, 0.0)) 1024 [0..100, 100..230, 450..1000, 0..1000] 1.03 12.7±0.16µs ? ?/sec 1.00 12.3±0.08µs ? ?/sec
interleave struct(str(20, 0.0), str(20, 0.0)) 1024 [0..100, 100..230, 450..1000] 1.03 12.6±0.13µs ? ?/sec 1.00 12.3±0.24µs ? ?/sec
interleave struct(str(20, 0.0), str(20, 0.0)) 400 [0..100, 100..230, 450..1000] 1.01 5.5±0.16µs ? ?/sec 1.00 5.4±0.04µs ? ?/sec
```
</p>
</details>
--
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: github-unsubscribe@arrow.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
Re: [PR] Unroll take_native, interleave -25% [arrow-rs]
Posted by "Dandandan (via GitHub)" <gi...@apache.org>.
Dandandan commented on PR #9542:
URL: https://github.com/apache/arrow-rs/pull/9542#issuecomment-4044970274
run benchmark take_kernels
--
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: github-unsubscribe@arrow.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
Re: [PR] Unroll interleave -25-30% [arrow-rs]
Posted by "mbutrovich (via GitHub)" <gi...@apache.org>.
mbutrovich commented on PR #9542:
URL: https://github.com/apache/arrow-rs/pull/9542#issuecomment-4049239586
I'm curious what a SIMD gather (AVX2, SVE) intrinsics' performance is like vs. unrolling manually on modern architectures. The former likely ends up as similar μops anyway. I don't love specializing the code like this (in either direction, SIMD intrinsics or unrolling especially with ARM increasingly prevalent in the cloud) but if today it's faster I'm okay with it. It's also a pretty well-understood technique at this point.
--
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: github-unsubscribe@arrow.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
Re: [PR] Unroll interleave -25-30% [arrow-rs]
Posted by "Dandandan (via GitHub)" <gi...@apache.org>.
Dandandan merged PR #9542:
URL: https://github.com/apache/arrow-rs/pull/9542
--
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: github-unsubscribe@arrow.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
Re: [PR] Unroll interleave -25-30% [arrow-rs]
Posted by "Dandandan (via GitHub)" <gi...@apache.org>.
Dandandan commented on PR #9542:
URL: https://github.com/apache/arrow-rs/pull/9542#issuecomment-4048744937
@mbutrovich as well :)
--
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: github-unsubscribe@arrow.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
Re: [PR] Unroll interleave -25-30% [arrow-rs]
Posted by "comphead (via GitHub)" <gi...@apache.org>.
comphead commented on code in PR #9542:
URL: https://github.com/apache/arrow-rs/pull/9542#discussion_r2926739616
##########
arrow-select/src/interleave.rs:
##########
@@ -154,13 +154,51 @@ fn interleave_primitive<T: ArrowPrimitiveType>(
data_type: &DataType,
) -> Result<ArrayRef, ArrowError> {
let interleaved = Interleave::<'_, PrimitiveArray<T>>::new(values, indices);
+ let arrays = &interleaved.arrays;
+ let len = indices.len();
+
+ let mut output = Vec::with_capacity(len);
+ let dst: *mut T::Native = output.as_mut_ptr();
+ let mut base = 0;
+
+ // Process 8 elements at a time to issue multiple independent loads
+ // and increase memory-level parallelism for random access patterns.
+ let chunks = indices.chunks_exact(8);
+ let remainder = chunks.remainder();
+ for chunk in chunks {
+ let v0 = arrays[chunk[0].0].value(chunk[0].1);
Review Comment:
thinking aloud, is 8 a constant? does it depend on number of cpu/threads, etc?
--
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: github-unsubscribe@arrow.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
Re: [PR] Unroll interleave -25-30% [arrow-rs]
Posted by "Dandandan (via GitHub)" <gi...@apache.org>.
Dandandan commented on PR #9542:
URL: https://github.com/apache/arrow-rs/pull/9542#issuecomment-4050156542
> I am thinking perhaps we can at least remove the bounds checks for interleave of the outer batch dimension by doing it upfront - I'll check that now.
Oh wait that is a not so useful idea as we get them for all of the rows (so it will be more expensive to check upfront than it is to do "on the go")
--
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: github-unsubscribe@arrow.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
Re: [PR] Unroll take_native, interleave -25% [arrow-rs]
Posted by "alamb-ghbot (via GitHub)" <gi...@apache.org>.
alamb-ghbot commented on PR #9542:
URL: https://github.com/apache/arrow-rs/pull/9542#issuecomment-4044827119
🤖: Benchmark completed
<details><summary>Details</summary>
<p>
```
group main take_native
----- ---- -----------
take bool 1024 1.00 1329.1±2.25ns ? ?/sec 1.00 1332.5±11.11ns ? ?/sec
take bool 512 1.00 730.0±11.00ns ? ?/sec 1.00 730.4±9.07ns ? ?/sec
take bool null indices 1024 1.01 1649.7±9.24ns ? ?/sec 1.00 1627.9±38.00ns ? ?/sec
take bool null values 1024 1.00 2.6±0.04µs ? ?/sec 1.00 2.6±0.02µs ? ?/sec
take bool null values null indices 1024 1.00 3.2±0.02µs ? ?/sec 1.16 3.6±0.06µs ? ?/sec
take check bounds i32 1024 1.00 844.2±9.82ns ? ?/sec 1.03 868.7±14.00ns ? ?/sec
take check bounds i32 512 1.01 535.4±3.41ns ? ?/sec 1.00 527.9±2.64ns ? ?/sec
take i32 1024 1.08 715.4±10.47ns ? ?/sec 1.00 665.0±2.75ns ? ?/sec
take i32 512 1.00 381.9±1.94ns ? ?/sec 1.02 388.8±2.61ns ? ?/sec
take i32 null indices 1024 1.00 995.2±2.99ns ? ?/sec 1.00 995.6±3.85ns ? ?/sec
take i32 null values 1024 1.03 2.0±0.01µs ? ?/sec 1.00 1951.0±3.34ns ? ?/sec
take i32 null values null indices 1024 1.00 2.6±0.02µs ? ?/sec 1.02 2.6±0.03µs ? ?/sec
take primitive fsb value len: 12, indices: 1024 1.00 3.6±0.01µs ? ?/sec 1.00 3.6±0.02µs ? ?/sec
take primitive fsb value len: 12, null values, indices: 1024 1.01 5.0±0.06µs ? ?/sec 1.00 5.0±0.10µs ? ?/sec
take primitive run logical len: 1024, physical len: 512, indices: 1024 1.00 20.2±0.13µs ? ?/sec 1.20 24.2±0.07µs ? ?/sec
take str 1024 1.01 11.0±0.10µs ? ?/sec 1.00 10.9±0.10µs ? ?/sec
take str 512 1.01 5.4±0.03µs ? ?/sec 1.00 5.3±0.04µs ? ?/sec
take str null indices 1024 1.02 6.9±0.04µs ? ?/sec 1.00 6.8±0.03µs ? ?/sec
take str null indices 512 1.01 3.4±0.32µs ? ?/sec 1.00 3.3±0.03µs ? ?/sec
take str null values 1024 1.01 8.7±0.16µs ? ?/sec 1.00 8.6±0.10µs ? ?/sec
take str null values null indices 1024 1.00 6.4±0.05µs ? ?/sec 1.02 6.5±0.09µs ? ?/sec
take stringview 1024 1.00 883.5±2.10ns ? ?/sec 1.56 1374.3±8.02ns ? ?/sec
take stringview 512 1.00 587.8±1.65ns ? ?/sec 1.23 721.3±5.73ns ? ?/sec
take stringview null indices 1024 1.00 1437.7±19.71ns ? ?/sec 1.00 1437.1±23.03ns ? ?/sec
take stringview null indices 512 1.00 737.0±3.13ns ? ?/sec 1.06 778.8±3.59ns ? ?/sec
take stringview null values 1024 1.00 2.1±0.02µs ? ?/sec 1.22 2.6±0.00µs ? ?/sec
take stringview null values null indices 1024 1.00 2.9±0.02µs ? ?/sec 1.01 2.9±0.04µs ? ?/sec
```
</p>
</details>
--
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: github-unsubscribe@arrow.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
Re: [PR] Unroll take_native, interleave -25% [arrow-rs]
Posted by "Dandandan (via GitHub)" <gi...@apache.org>.
Dandandan commented on PR #9542:
URL: https://github.com/apache/arrow-rs/pull/9542#issuecomment-4044813838
run benchmark interleave_kernels
--
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: github-unsubscribe@arrow.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
Re: [PR] Unroll interleave -25-30% [arrow-rs]
Posted by "mbutrovich (via GitHub)" <gi...@apache.org>.
mbutrovich commented on code in PR #9542:
URL: https://github.com/apache/arrow-rs/pull/9542#discussion_r2935498510
##########
arrow-select/src/interleave.rs:
##########
@@ -154,13 +154,51 @@ fn interleave_primitive<T: ArrowPrimitiveType>(
data_type: &DataType,
) -> Result<ArrayRef, ArrowError> {
let interleaved = Interleave::<'_, PrimitiveArray<T>>::new(values, indices);
+ let arrays = &interleaved.arrays;
+ let len = indices.len();
+
+ let mut output = Vec::with_capacity(len);
+ let dst: *mut T::Native = output.as_mut_ptr();
+ let mut base = 0;
+
+ // Process 8 elements at a time to issue multiple independent loads
+ // and increase memory-level parallelism for random access patterns.
+ let chunks = indices.chunks_exact(8);
+ let remainder = chunks.remainder();
+ for chunk in chunks {
+ let v0 = arrays[chunk[0].0].value(chunk[0].1);
Review Comment:
Not even knowing architecture specific stuff could help a user optimize this, necessarily. If all the targets (or many) being gathered are on the same cache line you won't get a huge benefit from this. Skew and cardinality per batch could affect the effectiveness. I think the current approach makes sense and don't see a good way to expose knobs here.
--
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: github-unsubscribe@arrow.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org