|
| 1 | +// SPDX-License-Identifier: Apache-2.0 |
| 2 | +// SPDX-FileCopyrightText: Copyright the Vortex contributors |
| 3 | + |
| 4 | +//! Whole-array sum dispatch and scalar partial construction. |
| 5 | +//! |
| 6 | +//! Run traversal and arithmetic are implemented in [`super::runs`]. |
| 7 | +
|
| 8 | +use std::ops::Range; |
| 9 | + |
| 10 | +use vortex_array::ArrayRef; |
| 11 | +use vortex_array::ExecutionCtx; |
| 12 | +use vortex_array::aggregate_fn::AggregateFnRef; |
| 13 | +use vortex_array::aggregate_fn::NumericalAggregateOpts; |
| 14 | +use vortex_array::aggregate_fn::fns::sum::Sum; |
| 15 | +use vortex_array::aggregate_fn::fns::sum_v2::SumV2; |
| 16 | +use vortex_array::aggregate_fn::kernels::DynAggregateKernel; |
| 17 | +use vortex_array::arrays::PrimitiveArray; |
| 18 | +use vortex_array::dtype::DType; |
| 19 | +use vortex_array::dtype::NativePType; |
| 20 | +use vortex_array::dtype::Nullability::Nullable; |
| 21 | +use vortex_array::dtype::UnsignedPType; |
| 22 | +use vortex_array::match_each_native_ptype; |
| 23 | +use vortex_array::match_each_unsigned_integer_ptype; |
| 24 | +use vortex_array::scalar::PValue; |
| 25 | +use vortex_array::scalar::Scalar; |
| 26 | +use vortex_error::VortexExpect; |
| 27 | +use vortex_error::VortexResult; |
| 28 | +use vortex_mask::Mask; |
| 29 | + |
| 30 | +use super::runs::add_float_run; |
| 31 | +use super::runs::add_signed_run; |
| 32 | +use super::runs::add_unsigned_run; |
| 33 | +use super::runs::sum_range; |
| 34 | +use crate::RunEnd; |
| 35 | +use crate::RunEndArrayExt; |
| 36 | +use crate::RunEndArraySlotsExt; |
| 37 | + |
| 38 | +/// Whole-array primitive sum kernel for [`RunEnd`]. |
| 39 | +#[derive(Debug)] |
| 40 | +pub(crate) struct RunEndSumKernel; |
| 41 | + |
| 42 | +impl DynAggregateKernel for RunEndSumKernel { |
| 43 | + fn aggregate( |
| 44 | + &self, |
| 45 | + aggregate_fn: &AggregateFnRef, |
| 46 | + batch: &ArrayRef, |
| 47 | + ctx: &mut ExecutionCtx, |
| 48 | + ) -> VortexResult<Option<Scalar>> { |
| 49 | + if !batch.dtype().is_primitive() { |
| 50 | + return Ok(None); |
| 51 | + } |
| 52 | + |
| 53 | + let Some(options) = aggregate_fn |
| 54 | + .as_opt::<Sum>() |
| 55 | + .or_else(|| aggregate_fn.as_opt::<SumV2>()) |
| 56 | + .copied() |
| 57 | + else { |
| 58 | + return Ok(None); |
| 59 | + }; |
| 60 | + |
| 61 | + let Some(array) = batch.as_opt::<RunEnd>() else { |
| 62 | + return Ok(None); |
| 63 | + }; |
| 64 | + |
| 65 | + if array.is_empty() { |
| 66 | + return Ok(Some(empty_partial(aggregate_fn, batch.dtype())?)); |
| 67 | + } |
| 68 | + |
| 69 | + let validity = array |
| 70 | + .values() |
| 71 | + .validity()? |
| 72 | + .execute_mask(array.values().len(), ctx)?; |
| 73 | + if validity.all_false() { |
| 74 | + return Ok(Some(empty_partial(aggregate_fn, batch.dtype())?)); |
| 75 | + } |
| 76 | + |
| 77 | + let ends = array.ends().clone().execute::<PrimitiveArray>(ctx)?; |
| 78 | + let values = array.values().clone().execute::<PrimitiveArray>(ctx)?; |
| 79 | + let range = array.offset()..array.offset() + batch.len(); |
| 80 | + let (sum, is_empty) = sum_primitive_runs(&ends, &values, &validity, range, options); |
| 81 | + |
| 82 | + Ok(Some(partial_scalar(aggregate_fn, sum, is_empty)?)) |
| 83 | + } |
| 84 | +} |
| 85 | + |
| 86 | +/// Select run-end and value types, widening sums to `u64`, `i64`, or `f64`. |
| 87 | +/// |
| 88 | +/// Returns `(sum, is_empty)` with overflow represented by a null scalar. Input requirements and |
| 89 | +/// empty-input semantics are defined in [`super::runs`]. |
| 90 | +fn sum_primitive_runs( |
| 91 | + ends: &PrimitiveArray, |
| 92 | + values: &PrimitiveArray, |
| 93 | + validity: &Mask, |
| 94 | + range: Range<usize>, |
| 95 | + options: NumericalAggregateOpts, |
| 96 | +) -> (Scalar, bool) { |
| 97 | + match_each_unsigned_integer_ptype!(ends.ptype(), |E| { |
| 98 | + let ends = ends.as_slice::<E>(); |
| 99 | + |
| 100 | + match_each_native_ptype!(values.ptype(), |
| 101 | + unsigned: |T| { |
| 102 | + sum_typed_runs(ends, values.as_slice::<T>(), validity, range, add_unsigned_run::<T>) |
| 103 | + }, |
| 104 | + signed: |T| { |
| 105 | + sum_typed_runs(ends, values.as_slice::<T>(), validity, range, add_signed_run::<T>) |
| 106 | + }, |
| 107 | + floating: |T| { |
| 108 | + sum_typed_runs(ends, values.as_slice::<T>(), validity, range, |sum, value, len| { |
| 109 | + add_float_run(sum, value, len, options.skip_nans) |
| 110 | + }) |
| 111 | + } |
| 112 | + ) |
| 113 | + }) |
| 114 | +} |
| 115 | + |
| 116 | +/// Convert the reduction's sum to a nullable scalar, preserving its empty-input flag. |
| 117 | +fn sum_typed_runs<E: UnsignedPType, T: NativePType, A: NativePType + Into<PValue>>( |
| 118 | + ends: &[E], |
| 119 | + values: &[T], |
| 120 | + validity: &Mask, |
| 121 | + range: Range<usize>, |
| 122 | + add_run: impl Fn(A, T, usize) -> Option<A>, |
| 123 | +) -> (Scalar, bool) { |
| 124 | + let (sum, is_empty) = sum_range(ends, values, validity, range, add_run); |
| 125 | + let sum = match sum { |
| 126 | + Some(sum) => Scalar::primitive(sum, Nullable), |
| 127 | + None => Scalar::null(DType::Primitive(A::PTYPE, Nullable)), |
| 128 | + }; |
| 129 | + |
| 130 | + (sum, is_empty) |
| 131 | +} |
| 132 | + |
| 133 | +fn empty_partial(aggregate_fn: &AggregateFnRef, dtype: &DType) -> VortexResult<Scalar> { |
| 134 | + let sum_dtype = aggregate_fn |
| 135 | + .return_dtype(dtype) |
| 136 | + .vortex_expect("The primitive sum kernel accepts only supported dtypes"); |
| 137 | + |
| 138 | + partial_scalar(aggregate_fn, Scalar::zero_value(&sum_dtype), true) |
| 139 | +} |
| 140 | + |
| 141 | +/// Wrap the sum in the aggregate's partial representation without finalizing empty inputs. |
| 142 | +fn partial_scalar( |
| 143 | + aggregate_fn: &AggregateFnRef, |
| 144 | + sum: Scalar, |
| 145 | + is_empty: bool, |
| 146 | +) -> VortexResult<Scalar> { |
| 147 | + if aggregate_fn.is::<SumV2>() { |
| 148 | + SumV2::partial_from_sum(sum, is_empty) |
| 149 | + } else { |
| 150 | + Ok(sum) |
| 151 | + } |
| 152 | +} |
0 commit comments