fix: Don't instantiate the scalar composition code quadratically for dictionaries #2391
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
added 2 commits
August 10, 2022 13:59
…dictionaries Instead, re-use the ones normal function. Reduces how much code `datafusion-physical-expr` generated significantly (since the functions are generic, and not instantiated in `arrow` itself, it only shows up downstream). https://github.com/apache/arrow-datafusion There is technically an extra indirect call now as the recursive call to `eq_dyn_scalar` etc coerces to a `dyn Array` again but that seems unlikely to matter. ## cargo llvm-lines -p datafusion-physical-expr ### Before ``` Lines Copies Function name ----- ------ ------------- 2270242 (100%) 38377 (100%) (TOTAL) 245854 (10.8%) 5580 (14.5%) core::option::Option<T>::ok_or_else 58690 (2.6%) 10 (0.0%) arrow::compute::kernels::comparison::eq_dyn_scalar 58690 (2.6%) 10 (0.0%) arrow::compute::kernels::comparison::gt_dyn_scalar 58690 (2.6%) 10 (0.0%) arrow::compute::kernels::comparison::gt_eq_dyn_scalar 58690 (2.6%) 10 (0.0%) arrow::compute::kernels::comparison::lt_dyn_scalar 58690 (2.6%) 10 (0.0%) arrow::compute::kernels::comparison::lt_eq_dyn_scalar 58690 (2.6%) 10 (0.0%) arrow::compute::kernels::comparison::neq_dyn_scalar 55800 (2.5%) 900 (2.3%) arrow::compute::kernels::comparison::eq_dyn_scalar::{{closure}} 55800 (2.5%) 900 (2.3%) arrow::compute::kernels::comparison::gt_dyn_scalar::{{closure}} 55800 (2.5%) 900 (2.3%) arrow::compute::kernels::comparison::gt_eq_dyn_scalar::{{closure}} 55800 (2.5%) 900 (2.3%) arrow::compute::kernels::comparison::lt_dyn_scalar::{{closure}} 55800 (2.5%) 900 (2.3%) arrow::compute::kernels::comparison::lt_eq_dyn_scalar::{{closure}} 55800 (2.5%) 900 (2.3%) arrow::compute::kernels::comparison::neq_dyn_scalar::{{closure}} 44929 (2.0%) 900 (2.3%) core::option::Option<T>::map 40986 (1.8%) 162 (0.4%) <arrow::array::array_boolean::BooleanArray as core::iter::traits::collect::FromIterator<Ptr>>::from_iter 37528 (1.7%) 508 (1.3%) core::iter::traits::iterator::Iterator::fold 30595 (1.3%) 245 (0.6%) <alloc::vec::Vec<T> as alloc::vec::spec_from_iter_nested::SpecFromIterNested<T,I>>::from_iter 29272 (1.3%) 46 (0.1%) <core::iter::adapters::flatten::FlattenCompat<I,U> as core::iter::traits::iterator::Iterator>::size_hint 27815 (1.2%) 285 (0.7%) core::iter::traits::iterator::Iterator::try_fold 26014 (1.1%) 1 (0.0%) datafusion_physical_expr::expressions::binary::BinaryExpr::evaluate_array_scalar 25095 (1.1%) 441 (1.1%) core::iter::adapters::map::map_fold::{{closure}} 22849 (1.0%) 174 (0.5%) <core::iter::adapters::GenericShunt<I,R> as core::iter::traits::iterator::Iterator>::try_fold::{{closure}} 21888 (1.0%) 96 (0.3%) arrow::compute::kernels::comparison::compare_op_scalar 21464 (0.9%) 56 (0.1%) <arrow::array::array_string::GenericStringArray<OffsetSize> as core::iter::traits::collect::FromIterator<core::option::Option<Ptr>>>::from_iter 21461 (0.9%) 441 (1.1%) <core::iter::adapters::map::Map<I,F> as core::iter::traits::iterator::Iterator>::fold 19918 (0.9%) 118 (0.3%) arrow::buffer::mutable::MutableBuffer::from_trusted_len_iter 16916 (0.7%) 246 (0.6%) <alloc::vec::Vec<T,A> as alloc::vec::spec_extend::SpecExtend<T,I>>::spec_extend ``` ### After ``` Lines Copies Function name ----- ------ ------------- 1475122 (100%) 28777 (100%) (TOTAL) 44929 (3.0%) 900 (3.1%) core::option::Option<T>::map 40986 (2.8%) 162 (0.6%) <arrow::array::array_boolean::BooleanArray as core::iter::traits::collect::FromIterator<Ptr>>::from_iter 37528 (2.5%) 508 (1.8%) core::iter::traits::iterator::Iterator::fold 34174 (2.3%) 780 (2.7%) core::option::Option<T>::ok_or_else 30595 (2.1%) 245 (0.9%) <alloc::vec::Vec<T> as alloc::vec::spec_from_iter_nested::SpecFromIterNested<T,I>>::from_iter 29272 (2.0%) 46 (0.2%) <core::iter::adapters::flatten::FlattenCompat<I,U> as core::iter::traits::iterator::Iterator>::size_hint 27815 (1.9%) 285 (1.0%) core::iter::traits::iterator::Iterator::try_fold 26014 (1.8%) 1 (0.0%) datafusion_physical_expr::expressions::binary::BinaryExpr::evaluate_array_scalar 25095 (1.7%) 441 (1.5%) core::iter::adapters::map::map_fold::{{closure}} 22849 (1.5%) 174 (0.6%) <core::iter::adapters::GenericShunt<I,R> as core::iter::traits::iterator::Iterator>::try_fold::{{closure}} 21888 (1.5%) 96 (0.3%) arrow::compute::kernels::comparison::compare_op_scalar 21464 (1.5%) 56 (0.2%) <arrow::array::array_string::GenericStringArray<OffsetSize> as core::iter::traits::collect::FromIterator<core::option::Option<Ptr>>>::from_iter 21461 (1.5%) 441 (1.5%) <core::iter::adapters::map::Map<I,F> as core::iter::traits::iterator::Iterator>::fold 19918 (1.4%) 118 (0.4%) arrow::buffer::mutable::MutableBuffer::from_trusted_len_iter 16916 (1.1%) 246 (0.9%) <alloc::vec::Vec<T,A> as alloc::vec::spec_extend::SpecExtend<T,I>>::spec_extend 16146 (1.1%) 960 (3.3%) core::iter::adapters::map::Map<I,F>::new 15492 (1.1%) 427 (1.5%) core::iter::traits::iterator::Iterator::for_each 14921 (1.0%) 111 (0.4%) alloc::vec::Vec<T,A>::extend_desugared 14670 (1.0%) 126 (0.4%) core::iter::adapters::try_process 13918 (0.9%) 1 (0.0%) datafusion_physical_expr::expressions::binary::BinaryExpr::evaluate_scalar_array 13120 (0.9%) 64 (0.2%) <arrow::array::array_primitive::PrimitiveArray<T> as core::iter::traits::collect::FromIterator<Ptr>>::from_iter 12963 (0.9%) 52 (0.2%) <core::iter::adapters::flatten::FlattenCompat<I,U> as core::iter::traits::iterator::Iterator>::try_fold 12245 (0.8%) 180 (0.6%) <core::iter::adapters::enumerate::Enumerate<I> as core::iter::traits::iterator::Iterator>::fold::enumerate::{{closure}} 12201 (0.8%) 81 (0.3%) arrow::buffer::mutable::MutableBuffer::extend_from_iter 11826 (0.8%) 162 (0.6%) <arrow::array::array_boolean::BooleanArray as core::iter::traits::collect::FromIterator<Ptr>>::from_iter::{{closure}} 11536 (0.8%) 960 (3.3%) core::iter::traits::iterator::Iterator::map 11200 (0.8%) 32 (0.1%) alloc::raw_vec::RawVec<T,A>::grow_amortized ```
…ry_to_type in comparisons (-4%) Reduces the number of llvm-lines in datafusion-physical-expr by another 4%
alamb
approved these changes
Aug 11, 2022
| @@ -1186,7 +1163,7 @@ where | |||
| { | |||
| match left.data_type() { | |||
| DataType::Dictionary(key_type, _value_type) => { | |||
| dyn_compare_scalar!(left, right, key_type, eq_scalar) | |||
| dyn_compare_scalar!(left, right, key_type, eq_dyn_scalar) | |||
|
Benchmark runs are scheduled for baseline = 6b1369e and contender = 6948373. 6948373 is a master commit associated with this PR. Results will be available as each benchmark for each run completes. |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Instead, re-use the ones normal function. Reduces how much code
datafusion-physical-exprgenerated significantly (since the functions are generic, and not instantiated inarrowitself, it only shows up downstream).https://github.com/apache/arrow-datafusion
There is technically an extra indirect call now as the recursive call to
eq_dyn_scalaretc coerces to adyn Arrayagain but that seems unlikely to matter.cargo llvm-lines -p datafusion-physical-expr
Before
After
Rationale for this change
Cuts down (release) compile times when these functions are used (in datafusion-physical-expr)
Before
After
Debug compile times remain pretty much unchanged, phyiscal-expr itself gets ~4s faster (21 -> 17) but it ends up having to wait for arrow to complete instead.