@@ -10,90 +10,17 @@ The full license is in the LICENSE file, distributed with this software.
1010#pragma once
1111
1212#include " pandas/moments.h"
13+ #include " pandas/simd/moments_common.hpp"
1314#include " xsimd/xsimd.hpp" // IWYU pragma: keep
14- #include < algorithm>
1515#include < array>
1616#include < cassert>
17- #include < cmath>
1817#include < optional>
1918#include < span>
2019
2120namespace pandas ::moments {
2221
2322namespace detail {
2423
25- template <typename T, typename U> struct MeanAcc {
26- T sum{};
27- U count{};
28- };
29-
30- template <typename T> struct CentralDiffs {
31- T m1{}, m2{}, m3{}, m4{};
32- };
33-
34- // / Compute Moments struct using two-pass
35- // / with correction presented in Pébay (2016).
36- // /
37- // / References:
38- // / Pébay, P., Terriberry, T. B., Kolla, H., & Bennett, J. (2016).
39- // / Numerically stable, scalable formulas for parallel
40- // / and online computation of higher-order multivariate central moments with
41- // / arbitrary weights. Comput. Stat., 31(4), 1305–1325.
42- // / https://doi.org/10.1007/s00180-015-0637-z
43- static inline Moments
44- compute_moments_with_correction (MeanAcc<double , std::size_t > total_acc,
45- CentralDiffs<double > central_diffs,
46- int max_moment) {
47- const auto count_double = static_cast <double >(total_acc.count );
48- const double trial_mean = total_acc.sum / count_double;
49- const double correction_term = central_diffs.m1 / count_double;
50- const double correction_term2 = correction_term * correction_term;
51- const double term1 = central_diffs.m1 * correction_term;
52-
53- Moments result{};
54-
55- result.n = total_acc.count ;
56- result.mean = trial_mean + correction_term;
57- result.m2 = central_diffs.m2 - term1;
58- result.m2 = std::max (result.m2 , 0.0 );
59-
60- if (max_moment >= 3 ) {
61- result.m3 = central_diffs.m3 - (3.0 * central_diffs.m2 * correction_term) +
62- (2.0 * correction_term * term1);
63- }
64- if (max_moment >= 4 ) {
65- result.m4 = central_diffs.m4 - (4.0 * central_diffs.m3 * correction_term) +
66- (6.0 * central_diffs.m2 * correction_term2) -
67- (3.0 * term1 * correction_term2);
68- result.m4 = std::max (result.m4 , 0.0 );
69- }
70-
71- return result;
72- }
73-
74- static inline std::optional<detail::MeanAcc<double , std::size_t >>
75- accumulate_mean_scalar_direct (
76- const std::span<const double > values,
77- const std::optional<std::span<const uint8_t >> mask, const bool skipna) {
78- double sum = 0.0 ;
79- std::size_t count = 0 ;
80- for (std::size_t i = 0 ; i < values.size (); ++i) {
81- const double val = values[i];
82- const bool isna = mask.has_value () ? (*mask)[i] != 0 : std::isnan (val);
83- if (isna) {
84- if (!skipna) {
85- return std::nullopt ;
86- }
87- continue ;
88- }
89- sum += val;
90- count++;
91- }
92- // Explicit template arguments: CTAD for aggregates (P1816) is unavailable
93- // in Apple clang < 17 (Xcode < 16).
94- return MeanAcc<double , std::size_t >{.sum = sum, .count = count};
95- }
96-
9724template <class Arch >
9825std::optional<
9926 MeanAcc<xsimd::batch<double , Arch>, xsimd::batch<std::uint64_t , Arch>>>
@@ -185,35 +112,6 @@ accumulate_mean_direct(std::span<const double> values,
185112 .sum = sum[0 ], .count = count};
186113}
187114
188- static inline CentralDiffs<double > accumulate_central_diffs_scalar_direct (
189- std::span<const double > values,
190- std::optional<std::span<const uint8_t >> mask, double mean, int max_moment) {
191- assert (!mask.has_value () || mask->size () == values.size ());
192-
193- CentralDiffs<double > acc{};
194- for (std::size_t i = 0 ; i < values.size (); i++) {
195- const auto isna = mask ? ((*mask)[i] != 0 ) : std::isnan (values[i]);
196-
197- if (isna) {
198- continue ;
199- }
200-
201- const auto diff = values[i] - mean;
202- const auto diff2 = diff * diff;
203-
204- acc.m1 += diff;
205- acc.m2 += diff2;
206- if (max_moment >= 3 ) {
207- acc.m3 += diff2 * diff;
208- }
209- if (max_moment >= 4 ) {
210- acc.m4 += diff2 * diff2;
211- }
212- }
213-
214- return acc;
215- }
216-
217115template <class Arch >
218116CentralDiffs<xsimd::batch<double , Arch>>
219117accumulate_central_diffs_direct (std::span<const double > values,
@@ -446,30 +344,6 @@ accumulate_central_diffs(std::span<const double> values,
446344 };
447345}
448346
449- template <class Arch >
450- Moments
451- accumulate_moments_simd_impl (std::span<const double > values,
452- std::optional<std::span<const uint8_t >> mask,
453- bool skipna, int max_moment) {
454- const auto total_acc = accumulate_mean<Arch>(values, mask, skipna);
455-
456- if (!total_acc.has_value ()) {
457- return {.mean = NAN , .m2 = NAN , .m3 = NAN , .m4 = NAN , .n = values.size ()};
458- }
459-
460- if (total_acc->count == 0 ) {
461- return {};
462- }
463-
464- const auto count_double = static_cast <double >(total_acc->count );
465- double trial_mean = total_acc->sum / count_double;
466-
467- const auto central_diffs =
468- accumulate_central_diffs<Arch>(values, mask, trial_mean, max_moment);
469-
470- return compute_moments_with_correction (*total_acc, central_diffs, max_moment);
471- }
472-
473347} // namespace detail
474348
475349struct accumulate_moments_simd {
@@ -483,8 +357,8 @@ template <class Arch>
483357Moments accumulate_moments_simd::operator ()(
484358 Arch /* arch*/ , std::span<const double > values, bool skipna,
485359 std::optional<std::span<const uint8_t >> mask, int max_moment) noexcept {
486- return detail::accumulate_moments_simd_impl <Arch>(values, mask, skipna,
487- max_moment);
360+ return detail::accumulate_moments_impl <Arch>(values, mask, skipna,
361+ max_moment);
488362}
489363
490364} // namespace pandas::moments
0 commit comments