ESPResSo
Extensible Simulation Package for Research on Soft Matter Systems
Loading...
Searching...
No Matches
StreamCollideSweepLeesEdwardsDoublePrecisionAVX.cpp
Go to the documentation of this file.
1//======================================================================================================================
2//
3// This file is part of waLBerla. waLBerla is free software: you can
4// redistribute it and/or modify it under the terms of the GNU General Public
5// License as published by the Free Software Foundation, either version 3 of
6// the License, or (at your option) any later version.
7//
8// waLBerla is distributed in the hope that it will be useful, but WITHOUT
9// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
10// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
11// for more details.
12//
13// You should have received a copy of the GNU General Public License along
14// with waLBerla (see COPYING.txt). If not, see <http://www.gnu.org/licenses/>.
15//
16//! \\file StreamCollideSweepLeesEdwardsDoublePrecisionAVX.cpp
17//! \\author pystencils
18//======================================================================================================================
19
20// kernel generated with pystencils v1.4+1.ge851f4e, lbmpy v1.4+1.ge9efe34, sympy v1.12.1, lbmpy_walberla/pystencils_walberla from waLBerla commit 272d4a09ec35da50685afc9586645e1b9984b423
21
22#include <cmath>
23
24#include "StreamCollideSweepLeesEdwardsDoublePrecisionAVX.h"
25#include "core/DataTypes.h"
26#include "core/Macros.h"
27
28#include <immintrin.h>
29
30#define FUNC_PREFIX
31
32#if (defined WALBERLA_CXX_COMPILER_IS_GNU) || (defined WALBERLA_CXX_COMPILER_IS_CLANG)
33#pragma GCC diagnostic push
34#pragma GCC diagnostic ignored "-Wfloat-equal"
35#pragma GCC diagnostic ignored "-Wshadow"
36#pragma GCC diagnostic ignored "-Wconversion"
37#pragma GCC diagnostic ignored "-Wunused-variable"
38#endif
39
40#if (defined WALBERLA_CXX_COMPILER_IS_INTEL)
41#pragma warning push
42#pragma warning(disable : 1599)
43#endif
44
45using namespace std;
46
47namespace walberla {
48namespace pystencils {
49
50namespace internal_08be08b6e7ea45132a735524ef494de6 {
51static FUNC_PREFIX void streamcollidesweepleesedwardsdoubleprecisionavx_streamcollidesweepleesedwardsdoubleprecisionavx(double *RESTRICT const _data_force, double *RESTRICT const _data_pdfs, double *RESTRICT _data_pdfs_tmp, int64_t const _size_force_0, int64_t const _size_force_1, int64_t const _size_force_2, int64_t const _stride_force_1, int64_t const _stride_force_2, int64_t const _stride_force_3, int64_t const _stride_pdfs_1, int64_t const _stride_pdfs_2, int64_t const _stride_pdfs_3, int64_t const _stride_pdfs_tmp_1, int64_t const _stride_pdfs_tmp_2, int64_t const _stride_pdfs_tmp_3, int64_t lebc_bot_index, int64_t lebc_top_index, double omega_bulk, double omega_even, double omega_odd, double omega_shear, double v_s) {
52#ifdef _OPENMP
53#pragma omp parallel
54#endif
55 {
56 const double xi_20 = omega_bulk * 0.5;
57 const double xi_47 = omega_shear * 0.041666666666666664;
58 const double xi_51 = omega_bulk * 0.041666666666666664;
59 const double xi_62 = omega_shear * 0.125;
60 const double xi_127 = omega_odd * 0.25;
61 const double xi_132 = omega_odd * 0.083333333333333329;
62 const double xi_158 = omega_shear * 0.25;
63 const double xi_173 = omega_odd * 0.041666666666666664;
64 const double xi_175 = omega_odd * 0.125;
65 const double rr_0 = 0.0;
66 const double xi_45 = rr_0 * 0.041666666666666664;
67#ifdef _OPENMP
68#pragma omp for schedule(static)
69#endif
70 for (int64_t ctr_2 = 1; ctr_2 < _size_force_2 - 1; ctr_2 += 1) {
71 for (int64_t ctr_1 = 1; ctr_1 < _size_force_1 - 1; ctr_1 += 1) {
72 {
73 for (int64_t ctr_0 = 1; ctr_0 < (int64_t)((_size_force_0 - 2) / (4)) * (4) + 1; ctr_0 += 4) {
74 const __m256d xi_2 = _mm256_add_pd(_mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 - _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + 8 * _stride_pdfs_3 + ctr_0 - 1]), _mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + 10 * _stride_pdfs_3 + ctr_0 - 1]));
75 const __m256d xi_3 = _mm256_add_pd(_mm256_add_pd(xi_2, _mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 - _stride_pdfs_2 + 14 * _stride_pdfs_3 + ctr_0 - 1])), _mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 + _stride_pdfs_2 + 18 * _stride_pdfs_3 + ctr_0 - 1]));
76 const __m256d xi_4 = _mm256_add_pd(_mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 - _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + _stride_pdfs_3 + ctr_0]), _mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 - _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + 7 * _stride_pdfs_3 + ctr_0 + 1]));
77 const __m256d xi_5 = _mm256_add_pd(_mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 - _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 - _stride_pdfs_2 + 11 * _stride_pdfs_3 + ctr_0]), _mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 - _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + _stride_pdfs_2 + 15 * _stride_pdfs_3 + ctr_0]));
78 const __m256d xi_6 = _mm256_add_pd(_mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 - _stride_pdfs_2 + 5 * _stride_pdfs_3 + ctr_0]), _mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 - _stride_pdfs_2 + 12 * _stride_pdfs_3 + ctr_0]));
79 const __m256d xi_7 = _mm256_add_pd(_mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 + 3 * _stride_pdfs_3 + ctr_0 + 1]), _mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 + _stride_pdfs_2 + 17 * _stride_pdfs_3 + ctr_0 + 1]));
80 const __m256d xi_8 = _mm256_add_pd(xi_7, _mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + 9 * _stride_pdfs_3 + ctr_0 + 1]));
81 const __m256d xi_9 = _mm256_add_pd(_mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + 2 * _stride_pdfs_3 + ctr_0]), _mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + _stride_pdfs_2 + 16 * _stride_pdfs_3 + ctr_0]));
82 const __m256d xi_11 = _mm256_add_pd(_mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 - _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + 7 * _stride_pdfs_3 + ctr_0 + 1]), _mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 - _stride_pdfs_2 + 13 * _stride_pdfs_3 + ctr_0 + 1]));
83 const __m256d xi_12 = _mm256_add_pd(_mm256_mul_pd(_mm256_set_pd(-1.0, -1.0, -1.0, -1.0), _mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 - _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + 8 * _stride_pdfs_3 + ctr_0 - 1])), _mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + 10 * _stride_pdfs_3 + ctr_0 - 1]));
84 const __m256d xi_13 = _mm256_add_pd(_mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + 9 * _stride_pdfs_3 + ctr_0 + 1]), _mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 - _stride_pdfs_2 + 12 * _stride_pdfs_3 + ctr_0]));
85 const __m256d xi_14 = _mm256_mul_pd(_mm256_set_pd(-1.0, -1.0, -1.0, -1.0), _mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 - _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 - _stride_pdfs_2 + 11 * _stride_pdfs_3 + ctr_0]));
86 const __m256d xi_15 = _mm256_mul_pd(_mm256_set_pd(-1.0, -1.0, -1.0, -1.0), _mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 - _stride_pdfs_2 + 14 * _stride_pdfs_3 + ctr_0 - 1]));
87 const __m256d xi_16 = _mm256_add_pd(_mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 + _stride_pdfs_2 + 17 * _stride_pdfs_3 + ctr_0 + 1]), _mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + _stride_pdfs_2 + 16 * _stride_pdfs_3 + ctr_0]));
88 const __m256d xi_21 = _mm256_mul_pd(_mm256_set_pd(0.16666666666666666, 0.16666666666666666, 0.16666666666666666, 0.16666666666666666), _mm256_loadu_pd(&_data_force[_stride_force_1 * ctr_1 + _stride_force_2 * ctr_2 + _stride_force_3 + ctr_0]));
89 const __m256d xi_22 = _mm256_mul_pd(_mm256_set_pd(0.083333333333333329, 0.083333333333333329, 0.083333333333333329, 0.083333333333333329), _mm256_loadu_pd(&_data_force[_stride_force_1 * ctr_1 + _stride_force_2 * ctr_2 + _stride_force_3 + ctr_0]));
90 const __m256d xi_33 = _mm256_mul_pd(_mm256_set_pd(0.16666666666666666, 0.16666666666666666, 0.16666666666666666, 0.16666666666666666), _mm256_load_pd(&_data_force[_stride_force_1 * ctr_1 + _stride_force_2 * ctr_2 + ctr_0]));
91 const __m256d xi_34 = _mm256_mul_pd(_mm256_set_pd(0.083333333333333329, 0.083333333333333329, 0.083333333333333329, 0.083333333333333329), _mm256_load_pd(&_data_force[_stride_force_1 * ctr_1 + _stride_force_2 * ctr_2 + ctr_0]));
92 const __m256d xi_39 = _mm256_mul_pd(_mm256_set_pd(0.16666666666666666, 0.16666666666666666, 0.16666666666666666, 0.16666666666666666), _mm256_loadu_pd(&_data_force[_stride_force_1 * ctr_1 + _stride_force_2 * ctr_2 + 2 * _stride_force_3 + ctr_0]));
93 const __m256d xi_40 = _mm256_mul_pd(_mm256_set_pd(0.083333333333333329, 0.083333333333333329, 0.083333333333333329, 0.083333333333333329), _mm256_loadu_pd(&_data_force[_stride_force_1 * ctr_1 + _stride_force_2 * ctr_2 + 2 * _stride_force_3 + ctr_0]));
94 const __m256d xi_58 = _mm256_mul_pd(_mm256_set_pd(0.25, 0.25, 0.25, 0.25), _mm256_loadu_pd(&_data_force[_stride_force_1 * ctr_1 + _stride_force_2 * ctr_2 + _stride_force_3 + ctr_0]));
95 const __m256d xi_63 = _mm256_mul_pd(_mm256_set_pd(xi_62, xi_62, xi_62, xi_62), _mm256_loadu_pd(&_data_force[_stride_force_1 * ctr_1 + _stride_force_2 * ctr_2 + _stride_force_3 + ctr_0]));
96 const __m256d xi_97 = _mm256_mul_pd(_mm256_set_pd(-1.0, -1.0, -1.0, -1.0), _mm256_load_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 + ctr_0]));
97 const __m256d xi_102 = _mm256_add_pd(xi_11, xi_3);
98 const __m256d xi_104 = _mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_mul_pd(_mm256_set_pd(2.0, 2.0, 2.0, 2.0), _mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 - _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 - _stride_pdfs_2 + 11 * _stride_pdfs_3 + ctr_0])), _mm256_mul_pd(_mm256_set_pd(2.0, 2.0, 2.0, 2.0), _mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 - _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + _stride_pdfs_2 + 15 * _stride_pdfs_3 + ctr_0]))), _mm256_mul_pd(_mm256_set_pd(2.0, 2.0, 2.0, 2.0), _mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 - _stride_pdfs_2 + 12 * _stride_pdfs_3 + ctr_0]))), _mm256_mul_pd(_mm256_set_pd(2.0, 2.0, 2.0, 2.0), _mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + _stride_pdfs_2 + 16 * _stride_pdfs_3 + ctr_0]))), _mm256_mul_pd(_mm256_set_pd(5.0, 5.0, 5.0, 5.0), _mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 + 3 * _stride_pdfs_3 + ctr_0 + 1]))), _mm256_mul_pd(_mm256_set_pd(5.0, 5.0, 5.0, 5.0), _mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 + 4 * _stride_pdfs_3 + ctr_0 - 1])));
99 const __m256d xi_107 = _mm256_mul_pd(_mm256_set_pd(2.0, 2.0, 2.0, 2.0), _mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 - _stride_pdfs_2 + 13 * _stride_pdfs_3 + ctr_0 + 1]));
100 const __m256d xi_108 = _mm256_mul_pd(_mm256_set_pd(2.0, 2.0, 2.0, 2.0), _mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 - _stride_pdfs_2 + 14 * _stride_pdfs_3 + ctr_0 - 1]));
101 const __m256d xi_109 = _mm256_add_pd(_mm256_mul_pd(_mm256_set_pd(2.0, 2.0, 2.0, 2.0), _mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 + _stride_pdfs_2 + 17 * _stride_pdfs_3 + ctr_0 + 1])), _mm256_mul_pd(_mm256_set_pd(2.0, 2.0, 2.0, 2.0), _mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 + _stride_pdfs_2 + 18 * _stride_pdfs_3 + ctr_0 - 1])));
102 const __m256d xi_113 = _mm256_mul_pd(_mm256_set_pd(-1.0, -1.0, -1.0, -1.0), _mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 - _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + _stride_pdfs_3 + ctr_0]));
103 const __m256d xi_114 = _mm256_mul_pd(_mm256_set_pd(-1.0, -1.0, -1.0, -1.0), _mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 + _stride_pdfs_2 + 17 * _stride_pdfs_3 + ctr_0 + 1]));
104 const __m256d xi_115 = _mm256_mul_pd(_mm256_set_pd(-1.0, -1.0, -1.0, -1.0), _mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 + _stride_pdfs_2 + 18 * _stride_pdfs_3 + ctr_0 - 1]));
105 const __m256d xi_116 = _mm256_mul_pd(_mm256_set_pd(-1.0, -1.0, -1.0, -1.0), _mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 - _stride_pdfs_2 + 5 * _stride_pdfs_3 + ctr_0]));
106 const __m256d xi_117 = _mm256_mul_pd(_mm256_set_pd(-1.0, -1.0, -1.0, -1.0), _mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 + _stride_pdfs_2 + 6 * _stride_pdfs_3 + ctr_0]));
107 const __m256d xi_122 = _mm256_mul_pd(_mm256_set_pd(-1.0, -1.0, -1.0, -1.0), _mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 - _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + _stride_pdfs_2 + 15 * _stride_pdfs_3 + ctr_0]));
108 const __m256d xi_123 = _mm256_add_pd(xi_14, _mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 - _stride_pdfs_2 + 12 * _stride_pdfs_3 + ctr_0]));
109 const __m256d xi_124 = _mm256_mul_pd(_mm256_set_pd(-1.0, -1.0, -1.0, -1.0), _mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + 2 * _stride_pdfs_3 + ctr_0]));
110 const __m256d xi_125 = _mm256_add_pd(xi_124, _mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 - _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + _stride_pdfs_3 + ctr_0]));
111 const __m256d xi_126 = _mm256_add_pd(_mm256_add_pd(_mm256_add_pd(xi_122, xi_123), xi_125), _mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + _stride_pdfs_2 + 16 * _stride_pdfs_3 + ctr_0]));
112 const __m256d xi_128 = _mm256_mul_pd(_mm256_set_pd(-1.0, -1.0, -1.0, -1.0), _mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + _stride_pdfs_2 + 16 * _stride_pdfs_3 + ctr_0]));
113 const __m256d xi_129 = _mm256_mul_pd(_mm256_set_pd(2.0, 2.0, 2.0, 2.0), _mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 - _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + 8 * _stride_pdfs_3 + ctr_0 - 1]));
114 const __m256d xi_130 = _mm256_add_pd(_mm256_mul_pd(_mm256_set_pd(2.0, 2.0, 2.0, 2.0), _mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + 10 * _stride_pdfs_3 + ctr_0 - 1])), _mm256_mul_pd(_mm256_set_pd(-2.0, -2.0, -2.0, -2.0), _mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 - _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + 7 * _stride_pdfs_3 + ctr_0 + 1])));
115 const __m256d xi_131 = _mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_mul_pd(_mm256_set_pd(2.0, 2.0, 2.0, 2.0), _mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + 9 * _stride_pdfs_3 + ctr_0 + 1])), _mm256_mul_pd(xi_129, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0))), _mm256_mul_pd(_mm256_set_pd(-1.0, -1.0, -1.0, -1.0), _mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 - _stride_pdfs_2 + 12 * _stride_pdfs_3 + ctr_0]))), xi_125), xi_128), xi_130), xi_5);
116 const __m256d xi_133 = _mm256_mul_pd(xi_131, _mm256_set_pd(xi_132, xi_132, xi_132, xi_132));
117 const __m256d xi_134 = _mm256_add_pd(_mm256_mul_pd(xi_126, _mm256_set_pd(xi_127, xi_127, xi_127, xi_127)), xi_133);
118 const __m256d xi_144 = _mm256_add_pd(xi_15, _mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 - _stride_pdfs_2 + 13 * _stride_pdfs_3 + ctr_0 + 1]));
119 const __m256d xi_145 = _mm256_add_pd(xi_115, xi_144);
120 const __m256d xi_146 = _mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_mul_pd(_mm256_set_pd(-1.0, -1.0, -1.0, -1.0), _mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 + 3 * _stride_pdfs_3 + ctr_0 + 1])), xi_145), _mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 + 4 * _stride_pdfs_3 + ctr_0 - 1])), _mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 + _stride_pdfs_2 + 17 * _stride_pdfs_3 + ctr_0 + 1]));
121 const __m256d xi_147 = _mm256_mul_pd(_mm256_set_pd(-1.0, -1.0, -1.0, -1.0), _mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 + 4 * _stride_pdfs_3 + ctr_0 - 1]));
122 const __m256d xi_148 = _mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_mul_pd(_mm256_set_pd(2.0, 2.0, 2.0, 2.0), _mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + 9 * _stride_pdfs_3 + ctr_0 + 1])), _mm256_mul_pd(xi_129, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0))), _mm256_mul_pd(xi_130, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0))), _mm256_mul_pd(xi_145, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0))), _mm256_mul_pd(xi_147, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0))), _mm256_mul_pd(xi_7, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0)));
123 const __m256d xi_149 = _mm256_mul_pd(xi_148, _mm256_set_pd(xi_132, xi_132, xi_132, xi_132));
124 const __m256d xi_150 = _mm256_add_pd(_mm256_mul_pd(xi_146, _mm256_set_pd(xi_127, xi_127, xi_127, xi_127)), xi_149);
125 const __m256d xi_152 = _mm256_add_pd(_mm256_add_pd(xi_122, xi_128), _mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 - _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 - _stride_pdfs_2 + 11 * _stride_pdfs_3 + ctr_0]));
126 const __m256d xi_153 = _mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_mul_pd(xi_116, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0)), _mm256_mul_pd(xi_152, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0))), _mm256_mul_pd(_mm256_set_pd(-1.0, -1.0, -1.0, -1.0), _mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 + _stride_pdfs_2 + 6 * _stride_pdfs_3 + ctr_0]))), _mm256_mul_pd(_mm256_set_pd(-1.0, -1.0, -1.0, -1.0), _mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 - _stride_pdfs_2 + 12 * _stride_pdfs_3 + ctr_0])));
127 const __m256d xi_154 = _mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_mul_pd(xi_107, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0)), _mm256_mul_pd(xi_108, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0))), xi_109), xi_117), xi_152), xi_6);
128 const __m256d xi_155 = _mm256_mul_pd(xi_154, _mm256_set_pd(xi_132, xi_132, xi_132, xi_132));
129 const __m256d xi_156 = _mm256_add_pd(_mm256_mul_pd(xi_153, _mm256_set_pd(xi_127, xi_127, xi_127, xi_127)), xi_155);
130 const __m256d xi_161 = _mm256_mul_pd(xi_149, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0));
131 const __m256d xi_174 = _mm256_mul_pd(xi_154, _mm256_set_pd(xi_173, xi_173, xi_173, xi_173));
132 const __m256d xi_176 = _mm256_mul_pd(xi_153, _mm256_set_pd(xi_175, xi_175, xi_175, xi_175));
133 const __m256d xi_177 = _mm256_add_pd(_mm256_mul_pd(xi_174, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0)), xi_176);
134 const __m256d xi_178 = _mm256_mul_pd(xi_131, _mm256_set_pd(xi_173, xi_173, xi_173, xi_173));
135 const __m256d xi_179 = _mm256_mul_pd(xi_126, _mm256_set_pd(xi_175, xi_175, xi_175, xi_175));
136 const __m256d xi_180 = _mm256_add_pd(_mm256_mul_pd(xi_178, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0)), xi_179);
137 const __m256d xi_181 = _mm256_add_pd(_mm256_mul_pd(xi_179, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0)), xi_178);
138 const __m256d xi_188 = _mm256_mul_pd(xi_146, _mm256_set_pd(xi_175, xi_175, xi_175, xi_175));
139 const __m256d xi_189 = _mm256_mul_pd(xi_148, _mm256_set_pd(xi_173, xi_173, xi_173, xi_173));
140 const __m256d xi_190 = _mm256_add_pd(_mm256_mul_pd(xi_188, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0)), xi_189);
141 const __m256d xi_191 = _mm256_add_pd(_mm256_mul_pd(xi_189, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0)), xi_188);
142 const __m256d xi_192 = _mm256_add_pd(_mm256_mul_pd(xi_176, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0)), xi_174);
143 const __m256d xi_23 = _mm256_mul_pd(xi_22, _mm256_set_pd(rr_0, rr_0, rr_0, rr_0));
144 const __m256d xi_35 = _mm256_mul_pd(xi_34, _mm256_set_pd(rr_0, rr_0, rr_0, rr_0));
145 const __m256d xi_41 = _mm256_mul_pd(xi_40, _mm256_set_pd(rr_0, rr_0, rr_0, rr_0));
146 const __m256d xi_46 = _mm256_mul_pd(_mm256_set_pd(xi_45, xi_45, xi_45, xi_45), _mm256_load_pd(&_data_force[_stride_force_1 * ctr_1 + _stride_force_2 * ctr_2 + ctr_0]));
147 const __m256d xi_50 = _mm256_mul_pd(_mm256_set_pd(xi_45, xi_45, xi_45, xi_45), _mm256_loadu_pd(&_data_force[_stride_force_1 * ctr_1 + _stride_force_2 * ctr_2 + _stride_force_3 + ctr_0]));
148 const __m256d xi_72 = _mm256_mul_pd(_mm256_set_pd(xi_45, xi_45, xi_45, xi_45), _mm256_loadu_pd(&_data_force[_stride_force_1 * ctr_1 + _stride_force_2 * ctr_2 + 2 * _stride_force_3 + ctr_0]));
149 const __m256d vel0Term = _mm256_add_pd(xi_3, _mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 + 4 * _stride_pdfs_3 + ctr_0 - 1]));
150 const __m256d vel1Term = _mm256_add_pd(xi_4, xi_5);
151 const __m256d vel2Term = _mm256_add_pd(xi_6, _mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 - _stride_pdfs_2 + 13 * _stride_pdfs_3 + ctr_0 + 1]));
152 const __m256d delta_rho = _mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_add_pd(vel0Term, vel1Term), vel2Term), xi_8), xi_9), _mm256_load_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 + ctr_0])), _mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 + _stride_pdfs_2 + 6 * _stride_pdfs_3 + ctr_0]));
153 const __m256d rho = _mm256_add_pd(_mm256_set_pd(1.0, 1.0, 1.0, 1.0), delta_rho);
154 const __m256d xi_0 = _mm256_div_pd(_mm256_set_pd(1.0, 1.0, 1.0, 1.0), rho);
155 const __m256d xi_10 = _mm256_mul_pd(xi_0, _mm256_set_pd(0.5, 0.5, 0.5, 0.5));
156 const __m256d u_0 = _mm256_add_pd(_mm256_mul_pd(xi_0, _mm256_add_pd(_mm256_add_pd(_mm256_mul_pd(xi_11, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0)), _mm256_mul_pd(xi_8, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0))), vel0Term)), _mm256_mul_pd(xi_10, _mm256_load_pd(&_data_force[_stride_force_1 * ctr_1 + _stride_force_2 * ctr_2 + ctr_0])));
157 const __m256d xi_17 = _mm256_mul_pd(u_0, _mm256_load_pd(&_data_force[_stride_force_1 * ctr_1 + _stride_force_2 * ctr_2 + ctr_0]));
158 const __m256d xi_28 = _mm256_mul_pd(xi_17, _mm256_set_pd(0.16666666666666666, 0.16666666666666666, 0.16666666666666666, 0.16666666666666666));
159 const __m256d xi_29 = _mm256_mul_pd(xi_28, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0));
160 const __m256d xi_30 = _mm256_mul_pd(xi_17, _mm256_set_pd(0.083333333333333329, 0.083333333333333329, 0.083333333333333329, 0.083333333333333329));
161 const __m256d xi_31 = _mm256_add_pd(_mm256_mul_pd(xi_30, _mm256_set_pd(omega_shear, omega_shear, omega_shear, omega_shear)), xi_29);
162 const __m256d xi_48 = _mm256_add_pd(_mm256_mul_pd(xi_17, _mm256_set_pd(xi_47, xi_47, xi_47, xi_47)), xi_29);
163 const __m256d xi_49 = _mm256_add_pd(_mm256_add_pd(_mm256_mul_pd(xi_46, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0)), xi_34), xi_48);
164 const __m256d xi_52 = _mm256_mul_pd(xi_17, _mm256_set_pd(xi_51, xi_51, xi_51, xi_51));
165 const __m256d xi_59 = _mm256_mul_pd(u_0, xi_58);
166 const __m256d xi_64 = _mm256_mul_pd(u_0, xi_63);
167 const __m256d xi_68 = _mm256_add_pd(_mm256_add_pd(_mm256_mul_pd(xi_34, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0)), xi_46), xi_48);
168 const __m256d xi_75 = _mm256_mul_pd(_mm256_mul_pd(_mm256_mul_pd(u_0, _mm256_set_pd(-0.083333333333333329, -0.083333333333333329, -0.083333333333333329, -0.083333333333333329)), _mm256_set_pd(omega_shear, omega_shear, omega_shear, omega_shear)), _mm256_load_pd(&_data_force[_stride_force_1 * ctr_1 + _stride_force_2 * ctr_2 + ctr_0]));
169 const __m256d xi_85 = _mm256_mul_pd(u_0, _mm256_loadu_pd(&_data_force[_stride_force_1 * ctr_1 + _stride_force_2 * ctr_2 + 2 * _stride_force_3 + ctr_0]));
170 const __m256d xi_86 = _mm256_mul_pd(xi_85, _mm256_set_pd(0.25, 0.25, 0.25, 0.25));
171 const __m256d xi_89 = _mm256_mul_pd(xi_85, _mm256_set_pd(xi_62, xi_62, xi_62, xi_62));
172 const __m256d xi_96 = _mm256_mul_pd(u_0, u_0);
173 const __m256d u_1 = _mm256_add_pd(_mm256_mul_pd(xi_0, _mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_mul_pd(xi_12, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0)), _mm256_mul_pd(xi_13, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0))), _mm256_mul_pd(xi_9, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0))), vel1Term)), _mm256_mul_pd(xi_10, _mm256_loadu_pd(&_data_force[_stride_force_1 * ctr_1 + _stride_force_2 * ctr_2 + _stride_force_3 + ctr_0])));
174 const __m256d xi_18 = _mm256_mul_pd(u_1, _mm256_loadu_pd(&_data_force[_stride_force_1 * ctr_1 + _stride_force_2 * ctr_2 + _stride_force_3 + ctr_0]));
175 const __m256d xi_26 = _mm256_mul_pd(xi_18, _mm256_set_pd(0.16666666666666666, 0.16666666666666666, 0.16666666666666666, 0.16666666666666666));
176 const __m256d xi_36 = _mm256_mul_pd(_mm256_mul_pd(_mm256_mul_pd(u_1, _mm256_set_pd(-0.083333333333333329, -0.083333333333333329, -0.083333333333333329, -0.083333333333333329)), _mm256_set_pd(omega_shear, omega_shear, omega_shear, omega_shear)), _mm256_loadu_pd(&_data_force[_stride_force_1 * ctr_1 + _stride_force_2 * ctr_2 + _stride_force_3 + ctr_0]));
177 const __m256d xi_42 = _mm256_mul_pd(xi_26, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0));
178 const __m256d xi_43 = _mm256_mul_pd(xi_18, _mm256_set_pd(0.083333333333333329, 0.083333333333333329, 0.083333333333333329, 0.083333333333333329));
179 const __m256d xi_53 = _mm256_mul_pd(xi_18, _mm256_set_pd(xi_51, xi_51, xi_51, xi_51));
180 const __m256d xi_60 = _mm256_mul_pd(u_1, _mm256_set_pd(0.25, 0.25, 0.25, 0.25));
181 const __m256d xi_61 = _mm256_mul_pd(xi_60, _mm256_load_pd(&_data_force[_stride_force_1 * ctr_1 + _stride_force_2 * ctr_2 + ctr_0]));
182 const __m256d xi_65 = _mm256_mul_pd(u_1, _mm256_set_pd(xi_62, xi_62, xi_62, xi_62));
183 const __m256d xi_66 = _mm256_mul_pd(xi_65, _mm256_load_pd(&_data_force[_stride_force_1 * ctr_1 + _stride_force_2 * ctr_2 + ctr_0]));
184 const __m256d xi_67 = _mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_mul_pd(xi_64, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0)), _mm256_mul_pd(xi_66, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0))), xi_59), xi_61);
185 const __m256d xi_69 = _mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_mul_pd(xi_59, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0)), _mm256_mul_pd(xi_61, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0))), xi_64), xi_66);
186 const __m256d xi_77 = _mm256_mul_pd(xi_60, _mm256_loadu_pd(&_data_force[_stride_force_1 * ctr_1 + _stride_force_2 * ctr_2 + 2 * _stride_force_3 + ctr_0]));
187 const __m256d xi_79 = _mm256_mul_pd(xi_65, _mm256_loadu_pd(&_data_force[_stride_force_1 * ctr_1 + _stride_force_2 * ctr_2 + 2 * _stride_force_3 + ctr_0]));
188 const __m256d xi_95 = _mm256_mul_pd(rho, _mm256_mul_pd(u_1, u_1));
189 const __m256d xi_101 = _mm256_mul_pd(xi_95, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0));
190 const __m256d xi_157 = _mm256_mul_pd(rho, u_1);
191 const __m256d xi_159 = _mm256_mul_pd(_mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_mul_pd(_mm256_set_pd(-1.0, -1.0, -1.0, -1.0), _mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + 9 * _stride_pdfs_3 + ctr_0 + 1])), _mm256_mul_pd(u_0, xi_157)), xi_12), _mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 - _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + 7 * _stride_pdfs_3 + ctr_0 + 1])), _mm256_set_pd(xi_158, xi_158, xi_158, xi_158));
192 const __m256d xi_160 = _mm256_mul_pd(xi_159, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0));
193 const __m256d u_2 = _mm256_add_pd(_mm256_mul_pd(xi_0, _mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_mul_pd(xi_14, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0)), _mm256_mul_pd(xi_15, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0))), _mm256_mul_pd(xi_16, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0))), _mm256_mul_pd(_mm256_set_pd(-1.0, -1.0, -1.0, -1.0), _mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 + _stride_pdfs_2 + 6 * _stride_pdfs_3 + ctr_0]))), _mm256_mul_pd(_mm256_set_pd(-1.0, -1.0, -1.0, -1.0), _mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 + _stride_pdfs_2 + 18 * _stride_pdfs_3 + ctr_0 - 1]))), _mm256_mul_pd(_mm256_set_pd(-1.0, -1.0, -1.0, -1.0), _mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 - _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + _stride_pdfs_2 + 15 * _stride_pdfs_3 + ctr_0]))), vel2Term)), _mm256_mul_pd(xi_10, _mm256_loadu_pd(&_data_force[_stride_force_1 * ctr_1 + _stride_force_2 * ctr_2 + 2 * _stride_force_3 + ctr_0])));
194 const __m256d xi_19 = _mm256_mul_pd(u_2, _mm256_loadu_pd(&_data_force[_stride_force_1 * ctr_1 + _stride_force_2 * ctr_2 + 2 * _stride_force_3 + ctr_0]));
195 const __m256d xi_24 = _mm256_mul_pd(xi_19, _mm256_set_pd(0.16666666666666666, 0.16666666666666666, 0.16666666666666666, 0.16666666666666666));
196 const __m256d xi_25 = _mm256_mul_pd(xi_24, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0));
197 const __m256d xi_27 = _mm256_mul_pd(xi_19, _mm256_set_pd(0.083333333333333329, 0.083333333333333329, 0.083333333333333329, 0.083333333333333329));
198 const __m256d xi_32 = _mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_mul_pd(xi_18, _mm256_set_pd(0.33333333333333331, 0.33333333333333331, 0.33333333333333331, 0.33333333333333331)), _mm256_mul_pd(xi_27, _mm256_set_pd(omega_shear, omega_shear, omega_shear, omega_shear))), _mm256_mul_pd(_mm256_mul_pd(xi_26, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0)), _mm256_set_pd(omega_shear, omega_shear, omega_shear, omega_shear))), xi_25), xi_31);
199 const __m256d xi_37 = _mm256_mul_pd(_mm256_mul_pd(_mm256_mul_pd(u_2, _mm256_set_pd(-0.083333333333333329, -0.083333333333333329, -0.083333333333333329, -0.083333333333333329)), _mm256_set_pd(omega_shear, omega_shear, omega_shear, omega_shear)), _mm256_loadu_pd(&_data_force[_stride_force_1 * ctr_1 + _stride_force_2 * ctr_2 + 2 * _stride_force_3 + ctr_0]));
200 const __m256d xi_38 = _mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_mul_pd(xi_28, _mm256_set_pd(omega_shear, omega_shear, omega_shear, omega_shear)), _mm256_mul_pd(_mm256_mul_pd(u_0, _mm256_set_pd(-0.33333333333333331, -0.33333333333333331, -0.33333333333333331, -0.33333333333333331)), _mm256_load_pd(&_data_force[_stride_force_1 * ctr_1 + _stride_force_2 * ctr_2 + ctr_0]))), xi_24), xi_26), xi_36), xi_37);
201 const __m256d xi_44 = _mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_mul_pd(xi_19, _mm256_set_pd(0.33333333333333331, 0.33333333333333331, 0.33333333333333331, 0.33333333333333331)), _mm256_mul_pd(xi_43, _mm256_set_pd(omega_shear, omega_shear, omega_shear, omega_shear))), _mm256_mul_pd(_mm256_mul_pd(xi_24, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0)), _mm256_set_pd(omega_shear, omega_shear, omega_shear, omega_shear))), xi_31), xi_42);
202 const __m256d xi_54 = _mm256_mul_pd(xi_19, _mm256_set_pd(xi_51, xi_51, xi_51, xi_51));
203 const __m256d xi_55 = _mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_mul_pd(xi_18, _mm256_set_pd(xi_47, xi_47, xi_47, xi_47)), xi_42), xi_52), xi_53), xi_54);
204 const __m256d xi_56 = _mm256_add_pd(_mm256_add_pd(_mm256_mul_pd(xi_22, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0)), xi_50), xi_55);
205 const __m256d xi_57 = _mm256_add_pd(_mm256_add_pd(xi_27, xi_37), xi_56);
206 const __m256d xi_70 = _mm256_add_pd(_mm256_add_pd(_mm256_mul_pd(xi_50, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0)), xi_22), xi_55);
207 const __m256d xi_71 = _mm256_add_pd(_mm256_add_pd(xi_27, xi_37), xi_70);
208 const __m256d xi_73 = _mm256_add_pd(_mm256_mul_pd(xi_19, _mm256_set_pd(xi_47, xi_47, xi_47, xi_47)), xi_25);
209 const __m256d xi_74 = _mm256_add_pd(_mm256_add_pd(_mm256_mul_pd(xi_40, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0)), xi_72), xi_73);
210 const __m256d xi_76 = _mm256_add_pd(_mm256_add_pd(xi_30, xi_56), xi_75);
211 const __m256d xi_78 = _mm256_mul_pd(u_2, xi_58);
212 const __m256d xi_80 = _mm256_mul_pd(u_2, xi_63);
213 const __m256d xi_81 = _mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_mul_pd(xi_77, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0)), _mm256_mul_pd(xi_78, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0))), xi_79), xi_80);
214 const __m256d xi_82 = _mm256_add_pd(_mm256_add_pd(xi_30, xi_70), xi_75);
215 const __m256d xi_83 = _mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_mul_pd(xi_79, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0)), _mm256_mul_pd(xi_80, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0))), xi_77), xi_78);
216 const __m256d xi_84 = _mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_add_pd(xi_36, xi_43), xi_52), xi_53), xi_54), xi_74);
217 const __m256d xi_87 = _mm256_mul_pd(u_2, _mm256_load_pd(&_data_force[_stride_force_1 * ctr_1 + _stride_force_2 * ctr_2 + ctr_0]));
218 const __m256d xi_88 = _mm256_mul_pd(xi_87, _mm256_set_pd(0.25, 0.25, 0.25, 0.25));
219 const __m256d xi_90 = _mm256_mul_pd(xi_87, _mm256_set_pd(xi_62, xi_62, xi_62, xi_62));
220 const __m256d xi_91 = _mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_mul_pd(xi_89, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0)), _mm256_mul_pd(xi_90, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0))), xi_86), xi_88);
221 const __m256d xi_92 = _mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_mul_pd(xi_86, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0)), _mm256_mul_pd(xi_88, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0))), xi_89), xi_90);
222 const __m256d xi_93 = _mm256_add_pd(_mm256_add_pd(_mm256_mul_pd(xi_72, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0)), xi_40), xi_73);
223 const __m256d xi_94 = _mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_add_pd(xi_36, xi_43), xi_52), xi_53), xi_54), xi_93);
224 const __m256d xi_98 = _mm256_mul_pd(rho, _mm256_mul_pd(u_2, u_2));
225 const __m256d xi_99 = _mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_mul_pd(_mm256_set_pd(3.0, 3.0, 3.0, 3.0), _mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 - _stride_pdfs_2 + 5 * _stride_pdfs_3 + ctr_0])), _mm256_mul_pd(_mm256_set_pd(3.0, 3.0, 3.0, 3.0), _mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 + _stride_pdfs_2 + 6 * _stride_pdfs_3 + ctr_0]))), _mm256_mul_pd(xi_98, _mm256_set_pd(0.66666666666666663, 0.66666666666666663, 0.66666666666666663, 0.66666666666666663))), xi_97);
226 const __m256d xi_100 = _mm256_mul_pd(_mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_mul_pd(_mm256_set_pd(3.0, 3.0, 3.0, 3.0), _mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 - _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + _stride_pdfs_3 + ctr_0])), _mm256_mul_pd(_mm256_set_pd(3.0, 3.0, 3.0, 3.0), _mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + 2 * _stride_pdfs_3 + ctr_0]))), _mm256_mul_pd(xi_95, _mm256_set_pd(0.66666666666666663, 0.66666666666666663, 0.66666666666666663, 0.66666666666666663))), _mm256_mul_pd(_mm256_set_pd(-3.0, -3.0, -3.0, -3.0), _mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 - _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 - _stride_pdfs_2 + 11 * _stride_pdfs_3 + ctr_0]))), _mm256_mul_pd(_mm256_set_pd(-3.0, -3.0, -3.0, -3.0), _mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 - _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + _stride_pdfs_2 + 15 * _stride_pdfs_3 + ctr_0]))), _mm256_mul_pd(_mm256_set_pd(-3.0, -3.0, -3.0, -3.0), _mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 - _stride_pdfs_2 + 12 * _stride_pdfs_3 + ctr_0]))), _mm256_mul_pd(_mm256_set_pd(-3.0, -3.0, -3.0, -3.0), _mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + _stride_pdfs_2 + 16 * _stride_pdfs_3 + ctr_0]))), _mm256_mul_pd(_mm256_mul_pd(rho, xi_96), _mm256_set_pd(1.6666666666666667, 1.6666666666666667, 1.6666666666666667, 1.6666666666666667))), xi_99), _mm256_set_pd(omega_even, omega_even, omega_even, omega_even));
227 const __m256d xi_103 = _mm256_mul_pd(_mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_mul_pd(xi_101, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0)), _mm256_mul_pd(xi_102, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0))), _mm256_mul_pd(xi_13, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0))), _mm256_mul_pd(xi_16, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0))), _mm256_mul_pd(xi_5, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0))), _mm256_mul_pd(xi_97, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0))), _mm256_mul_pd(rho, xi_96)), xi_98), _mm256_set_pd(omega_bulk, omega_bulk, omega_bulk, omega_bulk));
228 const __m256d xi_105 = _mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_mul_pd(xi_95, _mm256_set_pd(2.3333333333333335, 2.3333333333333335, 2.3333333333333335, 2.3333333333333335)), _mm256_mul_pd(_mm256_set_pd(-2.0, -2.0, -2.0, -2.0), _mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 - _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + _stride_pdfs_3 + ctr_0]))), _mm256_mul_pd(_mm256_set_pd(-2.0, -2.0, -2.0, -2.0), _mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + 2 * _stride_pdfs_3 + ctr_0]))), _mm256_mul_pd(_mm256_set_pd(-5.0, -5.0, -5.0, -5.0), _mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 - _stride_pdfs_2 + 13 * _stride_pdfs_3 + ctr_0 + 1]))), _mm256_mul_pd(_mm256_set_pd(-5.0, -5.0, -5.0, -5.0), _mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 + _stride_pdfs_2 + 17 * _stride_pdfs_3 + ctr_0 + 1]))), _mm256_mul_pd(_mm256_set_pd(-5.0, -5.0, -5.0, -5.0), _mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 - _stride_pdfs_2 + 14 * _stride_pdfs_3 + ctr_0 - 1]))), _mm256_mul_pd(_mm256_set_pd(-5.0, -5.0, -5.0, -5.0), _mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 + _stride_pdfs_2 + 18 * _stride_pdfs_3 + ctr_0 - 1]))), xi_104), xi_99);
229 const __m256d xi_106 = _mm256_mul_pd(xi_105, _mm256_set_pd(omega_even, omega_even, omega_even, omega_even));
230 const __m256d xi_110 = _mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_mul_pd(xi_98, _mm256_set_pd(3.0, 3.0, 3.0, 3.0)), _mm256_mul_pd(_mm256_set_pd(5.0, 5.0, 5.0, 5.0), _mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 - _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + _stride_pdfs_3 + ctr_0]))), _mm256_mul_pd(_mm256_set_pd(5.0, 5.0, 5.0, 5.0), _mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + 2 * _stride_pdfs_3 + ctr_0]))), _mm256_mul_pd(_mm256_set_pd(-4.0, -4.0, -4.0, -4.0), _mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 - _stride_pdfs_2 + 5 * _stride_pdfs_3 + ctr_0]))), _mm256_mul_pd(_mm256_set_pd(-4.0, -4.0, -4.0, -4.0), _mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 + _stride_pdfs_2 + 6 * _stride_pdfs_3 + ctr_0]))), _mm256_mul_pd(_mm256_set_pd(-7.0, -7.0, -7.0, -7.0), _mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 - _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + 7 * _stride_pdfs_3 + ctr_0 + 1]))), _mm256_mul_pd(_mm256_set_pd(-7.0, -7.0, -7.0, -7.0), _mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + 9 * _stride_pdfs_3 + ctr_0 + 1]))), _mm256_mul_pd(_mm256_set_pd(-7.0, -7.0, -7.0, -7.0), _mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 - _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + 8 * _stride_pdfs_3 + ctr_0 - 1]))), _mm256_mul_pd(_mm256_set_pd(-7.0, -7.0, -7.0, -7.0), _mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + 10 * _stride_pdfs_3 + ctr_0 - 1]))), xi_104), xi_107), xi_108), xi_109), xi_97);
231 const __m256d xi_111 = _mm256_mul_pd(xi_110, _mm256_set_pd(omega_even, omega_even, omega_even, omega_even));
232 const __m256d xi_112 = _mm256_mul_pd(xi_111, _mm256_set_pd(0.01984126984126984, 0.01984126984126984, 0.01984126984126984, 0.01984126984126984));
233 const __m256d xi_118 = _mm256_add_pd(_mm256_add_pd(_mm256_add_pd(xi_116, xi_117), xi_98), _mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + 9 * _stride_pdfs_3 + ctr_0 + 1]));
234 const __m256d xi_119 = _mm256_mul_pd(_mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_mul_pd(xi_101, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0)), _mm256_mul_pd(xi_114, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0))), _mm256_mul_pd(xi_115, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0))), _mm256_mul_pd(xi_118, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0))), _mm256_mul_pd(xi_15, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0))), _mm256_mul_pd(xi_2, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0))), _mm256_mul_pd(xi_4, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0))), _mm256_mul_pd(_mm256_set_pd(-1.0, -1.0, -1.0, -1.0), _mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + 2 * _stride_pdfs_3 + ctr_0]))), _mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 - _stride_pdfs_2 + 13 * _stride_pdfs_3 + ctr_0 + 1])), _mm256_set_pd(omega_shear, omega_shear, omega_shear, omega_shear));
235 const __m256d xi_120 = _mm256_mul_pd(xi_119, _mm256_set_pd(0.125, 0.125, 0.125, 0.125));
236 const __m256d xi_121 = _mm256_mul_pd(xi_120, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0));
237 const __m256d xi_135 = _mm256_mul_pd(xi_100, _mm256_set_pd(0.050000000000000003, 0.050000000000000003, 0.050000000000000003, 0.050000000000000003));
238 const __m256d xi_136 = _mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_mul_pd(_mm256_set_pd(2.0, 2.0, 2.0, 2.0), _mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 - _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 - _stride_pdfs_2 + 11 * _stride_pdfs_3 + ctr_0])), _mm256_mul_pd(_mm256_set_pd(2.0, 2.0, 2.0, 2.0), _mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 - _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + _stride_pdfs_2 + 15 * _stride_pdfs_3 + ctr_0]))), _mm256_mul_pd(_mm256_set_pd(2.0, 2.0, 2.0, 2.0), _mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 - _stride_pdfs_2 + 12 * _stride_pdfs_3 + ctr_0]))), _mm256_mul_pd(_mm256_set_pd(2.0, 2.0, 2.0, 2.0), _mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + _stride_pdfs_2 + 16 * _stride_pdfs_3 + ctr_0]))), _mm256_mul_pd(xi_102, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0))), _mm256_mul_pd(xi_113, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0))), _mm256_mul_pd(xi_118, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0))), _mm256_mul_pd(xi_124, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0))), _mm256_mul_pd(xi_95, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0))), _mm256_mul_pd(_mm256_set_pd(-1.0, -1.0, -1.0, -1.0), _mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 + _stride_pdfs_2 + 17 * _stride_pdfs_3 + ctr_0 + 1]))), _mm256_mul_pd(_mm256_set_pd(-2.0, -2.0, -2.0, -2.0), _mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 + 3 * _stride_pdfs_3 + ctr_0 + 1]))), _mm256_mul_pd(_mm256_set_pd(-2.0, -2.0, -2.0, -2.0), _mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 + 4 * _stride_pdfs_3 + ctr_0 - 1]))), _mm256_mul_pd(_mm256_mul_pd(rho, xi_96), _mm256_set_pd(2.0, 2.0, 2.0, 2.0)));
239 const __m256d xi_137 = _mm256_mul_pd(xi_136, _mm256_set_pd(omega_shear, omega_shear, omega_shear, omega_shear));
240 const __m256d xi_138 = _mm256_mul_pd(xi_137, _mm256_set_pd(0.041666666666666664, 0.041666666666666664, 0.041666666666666664, 0.041666666666666664));
241 const __m256d xi_139 = _mm256_add_pd(xi_135, xi_138);
242 const __m256d xi_140 = _mm256_mul_pd(xi_112, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0));
243 const __m256d xi_141 = _mm256_mul_pd(xi_138, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0));
244 const __m256d xi_142 = _mm256_add_pd(_mm256_mul_pd(xi_135, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0)), xi_141);
245 const __m256d xi_143 = _mm256_mul_pd(xi_106, _mm256_set_pd(0.035714285714285712, 0.035714285714285712, 0.035714285714285712, 0.035714285714285712));
246 const __m256d xi_151 = _mm256_mul_pd(xi_106, _mm256_set_pd(0.021428571428571429, 0.021428571428571429, 0.021428571428571429, 0.021428571428571429));
247 const __m256d xi_162 = _mm256_mul_pd(xi_119, _mm256_set_pd(0.0625, 0.0625, 0.0625, 0.0625));
248 const __m256d xi_163 = _mm256_mul_pd(xi_111, _mm256_set_pd(0.013888888888888888, 0.013888888888888888, 0.013888888888888888, 0.013888888888888888));
249 const __m256d xi_164 = _mm256_mul_pd(xi_103, _mm256_set_pd(0.041666666666666664, 0.041666666666666664, 0.041666666666666664, 0.041666666666666664));
250 const __m256d xi_165 = _mm256_add_pd(_mm256_mul_pd(xi_137, _mm256_set_pd(0.020833333333333332, 0.020833333333333332, 0.020833333333333332, 0.020833333333333332)), xi_164);
251 const __m256d xi_166 = _mm256_add_pd(_mm256_add_pd(_mm256_add_pd(xi_133, xi_162), xi_163), xi_165);
252 const __m256d xi_167 = _mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_mul_pd(xi_133, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0)), xi_162), xi_163), xi_165);
253 const __m256d xi_168 = _mm256_mul_pd(xi_111, _mm256_set_pd(-0.003968253968253968, -0.003968253968253968, -0.003968253968253968, -0.003968253968253968));
254 const __m256d xi_169 = _mm256_mul_pd(xi_106, _mm256_set_pd(-0.0071428571428571426, -0.0071428571428571426, -0.0071428571428571426, -0.0071428571428571426));
255 const __m256d xi_170 = _mm256_mul_pd(_mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_mul_pd(u_2, xi_157), xi_123), xi_128), _mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 - _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + _stride_pdfs_2 + 15 * _stride_pdfs_3 + ctr_0])), _mm256_set_pd(xi_158, xi_158, xi_158, xi_158));
256 const __m256d xi_171 = _mm256_mul_pd(xi_100, _mm256_set_pd(0.025000000000000001, 0.025000000000000001, 0.025000000000000001, 0.025000000000000001));
257 const __m256d xi_172 = _mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_add_pd(xi_141, xi_164), xi_168), xi_169), xi_170), xi_171);
258 const __m256d xi_182 = _mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_mul_pd(xi_170, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0)), xi_141), xi_164), xi_168), xi_169), xi_171);
259 const __m256d xi_183 = _mm256_mul_pd(_mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_mul_pd(_mm256_mul_pd(rho, u_0), u_2), xi_114), xi_144), _mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 + _stride_pdfs_2 + 18 * _stride_pdfs_3 + ctr_0 - 1])), _mm256_set_pd(xi_158, xi_158, xi_158, xi_158));
260 const __m256d xi_184 = _mm256_mul_pd(xi_183, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0));
261 const __m256d xi_185 = _mm256_mul_pd(xi_162, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0));
262 const __m256d xi_186 = _mm256_mul_pd(xi_106, _mm256_set_pd(0.017857142857142856, 0.017857142857142856, 0.017857142857142856, 0.017857142857142856));
263 const __m256d xi_187 = _mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_add_pd(xi_155, xi_165), xi_168), xi_185), xi_186);
264 const __m256d xi_193 = _mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_mul_pd(xi_155, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0)), xi_165), xi_168), xi_185), xi_186);
265 const __m256d forceTerm_0 = _mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_mul_pd(xi_17, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0)), _mm256_mul_pd(xi_18, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0))), _mm256_mul_pd(xi_19, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0))), _mm256_mul_pd(xi_17, _mm256_set_pd(xi_20, xi_20, xi_20, xi_20))), _mm256_mul_pd(xi_18, _mm256_set_pd(xi_20, xi_20, xi_20, xi_20))), _mm256_mul_pd(xi_19, _mm256_set_pd(xi_20, xi_20, xi_20, xi_20)));
266 const __m256d forceTerm_1 = _mm256_add_pd(_mm256_add_pd(_mm256_mul_pd(xi_23, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0)), xi_21), xi_32);
267 const __m256d forceTerm_2 = _mm256_add_pd(_mm256_add_pd(_mm256_mul_pd(xi_21, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0)), xi_23), xi_32);
268 const __m256d forceTerm_3 = _mm256_add_pd(_mm256_add_pd(_mm256_mul_pd(xi_33, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0)), _mm256_mul_pd(xi_38, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0))), xi_35);
269 const __m256d forceTerm_4 = _mm256_add_pd(_mm256_add_pd(_mm256_mul_pd(xi_35, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0)), _mm256_mul_pd(xi_38, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0))), xi_33);
270 const __m256d forceTerm_5 = _mm256_add_pd(_mm256_add_pd(_mm256_mul_pd(xi_41, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0)), xi_39), xi_44);
271 const __m256d forceTerm_6 = _mm256_add_pd(_mm256_add_pd(_mm256_mul_pd(xi_39, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0)), xi_41), xi_44);
272 const __m256d forceTerm_7 = _mm256_add_pd(_mm256_add_pd(_mm256_mul_pd(xi_49, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0)), _mm256_mul_pd(xi_57, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0))), _mm256_mul_pd(xi_67, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0)));
273 const __m256d forceTerm_8 = _mm256_add_pd(_mm256_add_pd(_mm256_mul_pd(xi_57, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0)), _mm256_mul_pd(xi_68, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0))), _mm256_mul_pd(xi_69, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0)));
274 const __m256d forceTerm_9 = _mm256_add_pd(_mm256_add_pd(_mm256_mul_pd(xi_49, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0)), _mm256_mul_pd(xi_69, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0))), _mm256_mul_pd(xi_71, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0)));
275 const __m256d forceTerm_10 = _mm256_add_pd(_mm256_add_pd(_mm256_mul_pd(xi_67, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0)), _mm256_mul_pd(xi_68, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0))), _mm256_mul_pd(xi_71, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0)));
276 const __m256d forceTerm_11 = _mm256_add_pd(_mm256_add_pd(_mm256_mul_pd(xi_74, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0)), _mm256_mul_pd(xi_76, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0))), _mm256_mul_pd(xi_81, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0)));
277 const __m256d forceTerm_12 = _mm256_add_pd(_mm256_add_pd(_mm256_mul_pd(xi_74, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0)), _mm256_mul_pd(xi_82, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0))), _mm256_mul_pd(xi_83, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0)));
278 const __m256d forceTerm_13 = _mm256_add_pd(_mm256_add_pd(_mm256_mul_pd(xi_49, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0)), _mm256_mul_pd(xi_84, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0))), _mm256_mul_pd(xi_91, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0)));
279 const __m256d forceTerm_14 = _mm256_add_pd(_mm256_add_pd(_mm256_mul_pd(xi_68, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0)), _mm256_mul_pd(xi_84, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0))), _mm256_mul_pd(xi_92, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0)));
280 const __m256d forceTerm_15 = _mm256_add_pd(_mm256_add_pd(_mm256_mul_pd(xi_76, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0)), _mm256_mul_pd(xi_83, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0))), _mm256_mul_pd(xi_93, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0)));
281 const __m256d forceTerm_16 = _mm256_add_pd(_mm256_add_pd(_mm256_mul_pd(xi_81, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0)), _mm256_mul_pd(xi_82, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0))), _mm256_mul_pd(xi_93, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0)));
282 const __m256d forceTerm_17 = _mm256_add_pd(_mm256_add_pd(_mm256_mul_pd(xi_49, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0)), _mm256_mul_pd(xi_92, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0))), _mm256_mul_pd(xi_94, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0)));
283 const __m256d forceTerm_18 = _mm256_add_pd(_mm256_add_pd(_mm256_mul_pd(xi_68, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0)), _mm256_mul_pd(xi_91, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0))), _mm256_mul_pd(xi_94, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0)));
284 _mm256_store_pd(&_data_pdfs_tmp[_stride_pdfs_tmp_1 * ctr_1 + _stride_pdfs_tmp_2 * ctr_2 + ctr_0], _mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_mul_pd(xi_111, _mm256_set_pd(0.023809523809523808, 0.023809523809523808, 0.023809523809523808, 0.023809523809523808)), _mm256_mul_pd(xi_106, _mm256_set_pd(0.042857142857142858, 0.042857142857142858, 0.042857142857142858, 0.042857142857142858))), _mm256_mul_pd(xi_100, _mm256_set_pd(0.10000000000000001, 0.10000000000000001, 0.10000000000000001, 0.10000000000000001))), _mm256_mul_pd(xi_103, _mm256_set_pd(-0.5, -0.5, -0.5, -0.5))), forceTerm_0), _mm256_load_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 + ctr_0])));
285 _mm256_storeu_pd(&_data_pdfs_tmp[_stride_pdfs_tmp_1 * ctr_1 + _stride_pdfs_tmp_2 * ctr_2 + _stride_pdfs_tmp_3 + ctr_0], _mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_mul_pd(xi_112, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0)), _mm256_mul_pd(xi_113, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0))), _mm256_mul_pd(xi_121, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0))), _mm256_mul_pd(xi_134, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0))), _mm256_mul_pd(xi_139, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0))), _mm256_mul_pd(_mm256_mul_pd(xi_105, _mm256_set_pd(0.014285714285714285, 0.014285714285714285, 0.014285714285714285, 0.014285714285714285)), _mm256_set_pd(omega_even, omega_even, omega_even, omega_even))), _mm256_blendv_pd(_mm256_set_pd(0.0, 0.0, 0.0, 0.0), _mm256_mul_pd(_mm256_mul_pd(_mm256_mul_pd(rho, _mm256_add_pd(_mm256_mul_pd(u_0, _mm256_set_pd(2.0, 2.0, 2.0, 2.0)), _mm256_set_pd(v_s, v_s, v_s, v_s))), _mm256_set_pd(0.16666666666666666, 0.16666666666666666, 0.16666666666666666, 0.16666666666666666)), _mm256_set_pd(v_s, v_s, v_s, v_s)), _mm256_cmp_pd(_mm256_set_pd(((double)(ctr_1)), ((double)(ctr_1)), ((double)(ctr_1)), ((double)(ctr_1))), _mm256_add_pd(_mm256_set_pd(-0.10000000000000001, -0.10000000000000001, -0.10000000000000001, -0.10000000000000001), _mm256_set_pd(((double)(lebc_top_index)), ((double)(lebc_top_index)), ((double)(lebc_top_index)), ((double)(lebc_top_index)))), _CMP_GE_OQ))), forceTerm_1));
286 _mm256_storeu_pd(&_data_pdfs_tmp[_stride_pdfs_tmp_1 * ctr_1 + _stride_pdfs_tmp_2 * ctr_2 + 2 * _stride_pdfs_tmp_3 + ctr_0], _mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_mul_pd(xi_106, _mm256_set_pd(0.014285714285714285, 0.014285714285714285, 0.014285714285714285, 0.014285714285714285)), _mm256_blendv_pd(_mm256_set_pd(0.0, 0.0, 0.0, 0.0), _mm256_mul_pd(_mm256_mul_pd(_mm256_mul_pd(rho, _mm256_add_pd(_mm256_mul_pd(u_0, _mm256_set_pd(-2.0, -2.0, -2.0, -2.0)), _mm256_set_pd(v_s, v_s, v_s, v_s))), _mm256_set_pd(0.16666666666666666, 0.16666666666666666, 0.16666666666666666, 0.16666666666666666)), _mm256_set_pd(v_s, v_s, v_s, v_s)), _mm256_cmp_pd(_mm256_set_pd(((double)(ctr_1)), ((double)(ctr_1)), ((double)(ctr_1)), ((double)(ctr_1))), _mm256_add_pd(_mm256_set_pd(1.1000000000000001, 1.1000000000000001, 1.1000000000000001, 1.1000000000000001), _mm256_set_pd(((double)(lebc_bot_index)), ((double)(lebc_bot_index)), ((double)(lebc_bot_index)), ((double)(lebc_bot_index)))), _CMP_LE_OQ))), forceTerm_2), xi_120), xi_134), xi_140), xi_142), _mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + 2 * _stride_pdfs_3 + ctr_0])));
287 _mm256_storeu_pd(&_data_pdfs_tmp[_stride_pdfs_tmp_1 * ctr_1 + _stride_pdfs_tmp_2 * ctr_2 + 3 * _stride_pdfs_tmp_3 + ctr_0], _mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_mul_pd(xi_137, _mm256_set_pd(0.083333333333333329, 0.083333333333333329, 0.083333333333333329, 0.083333333333333329)), _mm256_mul_pd(xi_143, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0))), forceTerm_3), xi_140), xi_150), _mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 + 3 * _stride_pdfs_3 + ctr_0 + 1])));
288 _mm256_store_pd(&_data_pdfs_tmp[_stride_pdfs_tmp_1 * ctr_1 + _stride_pdfs_tmp_2 * ctr_2 + 4 * _stride_pdfs_tmp_3 + ctr_0], _mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_mul_pd(xi_112, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0)), _mm256_mul_pd(xi_143, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0))), _mm256_mul_pd(xi_147, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0))), _mm256_mul_pd(xi_150, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0))), _mm256_mul_pd(_mm256_mul_pd(xi_136, _mm256_set_pd(0.083333333333333329, 0.083333333333333329, 0.083333333333333329, 0.083333333333333329)), _mm256_set_pd(omega_shear, omega_shear, omega_shear, omega_shear))), forceTerm_4));
289 _mm256_storeu_pd(&_data_pdfs_tmp[_stride_pdfs_tmp_1 * ctr_1 + _stride_pdfs_tmp_2 * ctr_2 + 5 * _stride_pdfs_tmp_3 + ctr_0], _mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_mul_pd(xi_116, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0)), _mm256_mul_pd(xi_120, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0))), _mm256_mul_pd(xi_139, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0))), _mm256_mul_pd(xi_151, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0))), _mm256_mul_pd(xi_156, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0))), _mm256_mul_pd(_mm256_mul_pd(xi_110, _mm256_set_pd(0.015873015873015872, 0.015873015873015872, 0.015873015873015872, 0.015873015873015872)), _mm256_set_pd(omega_even, omega_even, omega_even, omega_even))), forceTerm_5));
290 _mm256_storeu_pd(&_data_pdfs_tmp[_stride_pdfs_tmp_1 * ctr_1 + _stride_pdfs_tmp_2 * ctr_2 + 6 * _stride_pdfs_tmp_3 + ctr_0], _mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_mul_pd(xi_111, _mm256_set_pd(0.015873015873015872, 0.015873015873015872, 0.015873015873015872, 0.015873015873015872)), _mm256_mul_pd(xi_151, _mm256_set_pd(-1.0, -1.0, -1.0, -1.0))), forceTerm_6), xi_121), xi_142), xi_156), _mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 + _stride_pdfs_2 + 6 * _stride_pdfs_3 + ctr_0])));
291 _mm256_storeu_pd(&_data_pdfs_tmp[_stride_pdfs_tmp_1 * ctr_1 + _stride_pdfs_tmp_2 * ctr_2 + 7 * _stride_pdfs_tmp_3 + ctr_0], _mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_blendv_pd(_mm256_set_pd(0.0, 0.0, 0.0, 0.0), _mm256_mul_pd(_mm256_mul_pd(_mm256_mul_pd(rho, _mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_mul_pd(_mm256_set_pd(-1.0, -1.0, -1.0, -1.0), _mm256_set_pd(v_s, v_s, v_s, v_s)), _mm256_mul_pd(u_1, _mm256_set_pd(3.0, 3.0, 3.0, 3.0))), _mm256_mul_pd(u_0, _mm256_set_pd(-2.0, -2.0, -2.0, -2.0))), _mm256_set_pd(1.0, 1.0, 1.0, 1.0))), _mm256_set_pd(0.083333333333333329, 0.083333333333333329, 0.083333333333333329, 0.083333333333333329)), _mm256_set_pd(v_s, v_s, v_s, v_s)), _mm256_cmp_pd(_mm256_set_pd(((double)(ctr_1)), ((double)(ctr_1)), ((double)(ctr_1)), ((double)(ctr_1))), _mm256_add_pd(_mm256_set_pd(-0.10000000000000001, -0.10000000000000001, -0.10000000000000001, -0.10000000000000001), _mm256_set_pd(((double)(lebc_top_index)), ((double)(lebc_top_index)), ((double)(lebc_top_index)), ((double)(lebc_top_index)))), _CMP_GE_OQ)), forceTerm_7), xi_160), xi_161), xi_166), _mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 - _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + 7 * _stride_pdfs_3 + ctr_0 + 1])));
292 _mm256_store_pd(&_data_pdfs_tmp[_stride_pdfs_tmp_1 * ctr_1 + _stride_pdfs_tmp_2 * ctr_2 + 8 * _stride_pdfs_tmp_3 + ctr_0], _mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_blendv_pd(_mm256_set_pd(0.0, 0.0, 0.0, 0.0), _mm256_mul_pd(_mm256_mul_pd(_mm256_mul_pd(rho, _mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_mul_pd(_mm256_set_pd(-1.0, -1.0, -1.0, -1.0), _mm256_set_pd(v_s, v_s, v_s, v_s)), _mm256_mul_pd(u_0, _mm256_set_pd(-2.0, -2.0, -2.0, -2.0))), _mm256_mul_pd(u_1, _mm256_set_pd(-3.0, -3.0, -3.0, -3.0))), _mm256_set_pd(-1.0, -1.0, -1.0, -1.0))), _mm256_set_pd(0.083333333333333329, 0.083333333333333329, 0.083333333333333329, 0.083333333333333329)), _mm256_set_pd(v_s, v_s, v_s, v_s)), _mm256_cmp_pd(_mm256_set_pd(((double)(ctr_1)), ((double)(ctr_1)), ((double)(ctr_1)), ((double)(ctr_1))), _mm256_add_pd(_mm256_set_pd(-0.10000000000000001, -0.10000000000000001, -0.10000000000000001, -0.10000000000000001), _mm256_set_pd(((double)(lebc_top_index)), ((double)(lebc_top_index)), ((double)(lebc_top_index)), ((double)(lebc_top_index)))), _CMP_GE_OQ)), forceTerm_8), xi_149), xi_159), xi_166), _mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 - _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + 8 * _stride_pdfs_3 + ctr_0 - 1])));
293 _mm256_storeu_pd(&_data_pdfs_tmp[_stride_pdfs_tmp_1 * ctr_1 + _stride_pdfs_tmp_2 * ctr_2 + 9 * _stride_pdfs_tmp_3 + ctr_0], _mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_blendv_pd(_mm256_set_pd(0.0, 0.0, 0.0, 0.0), _mm256_mul_pd(_mm256_mul_pd(_mm256_mul_pd(rho, _mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_mul_pd(_mm256_set_pd(-1.0, -1.0, -1.0, -1.0), _mm256_set_pd(v_s, v_s, v_s, v_s)), _mm256_mul_pd(u_0, _mm256_set_pd(2.0, 2.0, 2.0, 2.0))), _mm256_mul_pd(u_1, _mm256_set_pd(3.0, 3.0, 3.0, 3.0))), _mm256_set_pd(-1.0, -1.0, -1.0, -1.0))), _mm256_set_pd(0.083333333333333329, 0.083333333333333329, 0.083333333333333329, 0.083333333333333329)), _mm256_set_pd(v_s, v_s, v_s, v_s)), _mm256_cmp_pd(_mm256_set_pd(((double)(ctr_1)), ((double)(ctr_1)), ((double)(ctr_1)), ((double)(ctr_1))), _mm256_add_pd(_mm256_set_pd(1.1000000000000001, 1.1000000000000001, 1.1000000000000001, 1.1000000000000001), _mm256_set_pd(((double)(lebc_bot_index)), ((double)(lebc_bot_index)), ((double)(lebc_bot_index)), ((double)(lebc_bot_index)))), _CMP_LE_OQ)), forceTerm_9), xi_159), xi_161), xi_167), _mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + 9 * _stride_pdfs_3 + ctr_0 + 1])));
294 _mm256_storeu_pd(&_data_pdfs_tmp[_stride_pdfs_tmp_1 * ctr_1 + _stride_pdfs_tmp_2 * ctr_2 + 10 * _stride_pdfs_tmp_3 + ctr_0], _mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_blendv_pd(_mm256_set_pd(0.0, 0.0, 0.0, 0.0), _mm256_mul_pd(_mm256_mul_pd(_mm256_mul_pd(rho, _mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_mul_pd(_mm256_set_pd(-1.0, -1.0, -1.0, -1.0), _mm256_set_pd(v_s, v_s, v_s, v_s)), _mm256_mul_pd(u_0, _mm256_set_pd(2.0, 2.0, 2.0, 2.0))), _mm256_mul_pd(u_1, _mm256_set_pd(-3.0, -3.0, -3.0, -3.0))), _mm256_set_pd(1.0, 1.0, 1.0, 1.0))), _mm256_set_pd(0.083333333333333329, 0.083333333333333329, 0.083333333333333329, 0.083333333333333329)), _mm256_set_pd(v_s, v_s, v_s, v_s)), _mm256_cmp_pd(_mm256_set_pd(((double)(ctr_1)), ((double)(ctr_1)), ((double)(ctr_1)), ((double)(ctr_1))), _mm256_add_pd(_mm256_set_pd(1.1000000000000001, 1.1000000000000001, 1.1000000000000001, 1.1000000000000001), _mm256_set_pd(((double)(lebc_bot_index)), ((double)(lebc_bot_index)), ((double)(lebc_bot_index)), ((double)(lebc_bot_index)))), _CMP_LE_OQ)), forceTerm_10), xi_149), xi_160), xi_167), _mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + 10 * _stride_pdfs_3 + ctr_0 - 1])));
295 _mm256_storeu_pd(&_data_pdfs_tmp[_stride_pdfs_tmp_1 * ctr_1 + _stride_pdfs_tmp_2 * ctr_2 + 11 * _stride_pdfs_tmp_3 + ctr_0], _mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_add_pd(forceTerm_11, xi_172), xi_177), xi_180), _mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 - _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 - _stride_pdfs_2 + 11 * _stride_pdfs_3 + ctr_0])));
296 _mm256_store_pd(&_data_pdfs_tmp[_stride_pdfs_tmp_1 * ctr_1 + _stride_pdfs_tmp_2 * ctr_2 + 12 * _stride_pdfs_tmp_3 + ctr_0], _mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_add_pd(forceTerm_12, xi_177), xi_181), xi_182), _mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 - _stride_pdfs_2 + 12 * _stride_pdfs_3 + ctr_0])));
297 _mm256_storeu_pd(&_data_pdfs_tmp[_stride_pdfs_tmp_1 * ctr_1 + _stride_pdfs_tmp_2 * ctr_2 + 13 * _stride_pdfs_tmp_3 + ctr_0], _mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_add_pd(forceTerm_13, xi_184), xi_187), xi_190), _mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 - _stride_pdfs_2 + 13 * _stride_pdfs_3 + ctr_0 + 1])));
298 _mm256_storeu_pd(&_data_pdfs_tmp[_stride_pdfs_tmp_1 * ctr_1 + _stride_pdfs_tmp_2 * ctr_2 + 14 * _stride_pdfs_tmp_3 + ctr_0], _mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_add_pd(forceTerm_14, xi_183), xi_187), xi_191), _mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 - _stride_pdfs_2 + 14 * _stride_pdfs_3 + ctr_0 - 1])));
299 _mm256_storeu_pd(&_data_pdfs_tmp[_stride_pdfs_tmp_1 * ctr_1 + _stride_pdfs_tmp_2 * ctr_2 + 15 * _stride_pdfs_tmp_3 + ctr_0], _mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_add_pd(forceTerm_15, xi_180), xi_182), xi_192), _mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 - _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + _stride_pdfs_2 + 15 * _stride_pdfs_3 + ctr_0])));
300 _mm256_store_pd(&_data_pdfs_tmp[_stride_pdfs_tmp_1 * ctr_1 + _stride_pdfs_tmp_2 * ctr_2 + 16 * _stride_pdfs_tmp_3 + ctr_0], _mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_add_pd(forceTerm_16, xi_172), xi_181), xi_192), _mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + _stride_pdfs_2 + 16 * _stride_pdfs_3 + ctr_0])));
301 _mm256_storeu_pd(&_data_pdfs_tmp[_stride_pdfs_tmp_1 * ctr_1 + _stride_pdfs_tmp_2 * ctr_2 + 17 * _stride_pdfs_tmp_3 + ctr_0], _mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_add_pd(forceTerm_17, xi_183), xi_190), xi_193), _mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 + _stride_pdfs_2 + 17 * _stride_pdfs_3 + ctr_0 + 1])));
302 _mm256_storeu_pd(&_data_pdfs_tmp[_stride_pdfs_tmp_1 * ctr_1 + _stride_pdfs_tmp_2 * ctr_2 + 18 * _stride_pdfs_tmp_3 + ctr_0], _mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_add_pd(forceTerm_18, xi_184), xi_191), xi_193), _mm256_loadu_pd(&_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 + _stride_pdfs_2 + 18 * _stride_pdfs_3 + ctr_0 - 1])));
303 }
304 for (int64_t ctr_0 = (int64_t)((_size_force_0 - 2) / (4)) * (4) + 1; ctr_0 < _size_force_0 - 1; ctr_0 += 1) {
305 const double xi_2 = _data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + 10 * _stride_pdfs_3 + ctr_0 - 1] + _data_pdfs[_stride_pdfs_1 * ctr_1 - _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + 8 * _stride_pdfs_3 + ctr_0 - 1];
306 const double xi_3 = xi_2 + _data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 + _stride_pdfs_2 + 18 * _stride_pdfs_3 + ctr_0 - 1] + _data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 - _stride_pdfs_2 + 14 * _stride_pdfs_3 + ctr_0 - 1];
307 const double xi_4 = _data_pdfs[_stride_pdfs_1 * ctr_1 - _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + 7 * _stride_pdfs_3 + ctr_0 + 1] + _data_pdfs[_stride_pdfs_1 * ctr_1 - _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + _stride_pdfs_3 + ctr_0];
308 const double xi_5 = _data_pdfs[_stride_pdfs_1 * ctr_1 - _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + _stride_pdfs_2 + 15 * _stride_pdfs_3 + ctr_0] + _data_pdfs[_stride_pdfs_1 * ctr_1 - _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 - _stride_pdfs_2 + 11 * _stride_pdfs_3 + ctr_0];
309 const double xi_6 = _data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 - _stride_pdfs_2 + 12 * _stride_pdfs_3 + ctr_0] + _data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 - _stride_pdfs_2 + 5 * _stride_pdfs_3 + ctr_0];
310 const double xi_7 = _data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 + 3 * _stride_pdfs_3 + ctr_0 + 1] + _data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 + _stride_pdfs_2 + 17 * _stride_pdfs_3 + ctr_0 + 1];
311 const double xi_8 = xi_7 + _data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + 9 * _stride_pdfs_3 + ctr_0 + 1];
312 const double xi_9 = _data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + 2 * _stride_pdfs_3 + ctr_0] + _data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + _stride_pdfs_2 + 16 * _stride_pdfs_3 + ctr_0];
313 const double xi_11 = _data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 - _stride_pdfs_2 + 13 * _stride_pdfs_3 + ctr_0 + 1] + _data_pdfs[_stride_pdfs_1 * ctr_1 - _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + 7 * _stride_pdfs_3 + ctr_0 + 1];
314 const double xi_12 = _data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + 10 * _stride_pdfs_3 + ctr_0 - 1] - _data_pdfs[_stride_pdfs_1 * ctr_1 - _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + 8 * _stride_pdfs_3 + ctr_0 - 1];
315 const double xi_13 = _data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + 9 * _stride_pdfs_3 + ctr_0 + 1] + _data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 - _stride_pdfs_2 + 12 * _stride_pdfs_3 + ctr_0];
316 const double xi_14 = -_data_pdfs[_stride_pdfs_1 * ctr_1 - _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 - _stride_pdfs_2 + 11 * _stride_pdfs_3 + ctr_0];
317 const double xi_15 = -_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 - _stride_pdfs_2 + 14 * _stride_pdfs_3 + ctr_0 - 1];
318 const double xi_16 = _data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + _stride_pdfs_2 + 16 * _stride_pdfs_3 + ctr_0] + _data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 + _stride_pdfs_2 + 17 * _stride_pdfs_3 + ctr_0 + 1];
319 const double xi_21 = 0.16666666666666666 * _data_force[_stride_force_1 * ctr_1 + _stride_force_2 * ctr_2 + _stride_force_3 + ctr_0];
320 const double xi_22 = 0.083333333333333329 * _data_force[_stride_force_1 * ctr_1 + _stride_force_2 * ctr_2 + _stride_force_3 + ctr_0];
321 const double xi_33 = 0.16666666666666666 * _data_force[_stride_force_1 * ctr_1 + _stride_force_2 * ctr_2 + ctr_0];
322 const double xi_34 = 0.083333333333333329 * _data_force[_stride_force_1 * ctr_1 + _stride_force_2 * ctr_2 + ctr_0];
323 const double xi_39 = 0.16666666666666666 * _data_force[_stride_force_1 * ctr_1 + _stride_force_2 * ctr_2 + 2 * _stride_force_3 + ctr_0];
324 const double xi_40 = 0.083333333333333329 * _data_force[_stride_force_1 * ctr_1 + _stride_force_2 * ctr_2 + 2 * _stride_force_3 + ctr_0];
325 const double xi_58 = 0.25 * _data_force[_stride_force_1 * ctr_1 + _stride_force_2 * ctr_2 + _stride_force_3 + ctr_0];
326 const double xi_63 = xi_62 * _data_force[_stride_force_1 * ctr_1 + _stride_force_2 * ctr_2 + _stride_force_3 + ctr_0];
327 const double xi_97 = -_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 + ctr_0];
328 const double xi_102 = xi_11 + xi_3;
329 const double xi_104 = 2.0 * _data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + _stride_pdfs_2 + 16 * _stride_pdfs_3 + ctr_0] + 2.0 * _data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 - _stride_pdfs_2 + 12 * _stride_pdfs_3 + ctr_0] + 2.0 * _data_pdfs[_stride_pdfs_1 * ctr_1 - _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + _stride_pdfs_2 + 15 * _stride_pdfs_3 + ctr_0] + 2.0 * _data_pdfs[_stride_pdfs_1 * ctr_1 - _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 - _stride_pdfs_2 + 11 * _stride_pdfs_3 + ctr_0] + 5.0 * _data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 + 3 * _stride_pdfs_3 + ctr_0 + 1] + 5.0 * _data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 + 4 * _stride_pdfs_3 + ctr_0 - 1];
330 const double xi_107 = 2.0 * _data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 - _stride_pdfs_2 + 13 * _stride_pdfs_3 + ctr_0 + 1];
331 const double xi_108 = 2.0 * _data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 - _stride_pdfs_2 + 14 * _stride_pdfs_3 + ctr_0 - 1];
332 const double xi_109 = 2.0 * _data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 + _stride_pdfs_2 + 17 * _stride_pdfs_3 + ctr_0 + 1] + 2.0 * _data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 + _stride_pdfs_2 + 18 * _stride_pdfs_3 + ctr_0 - 1];
333 const double xi_113 = -_data_pdfs[_stride_pdfs_1 * ctr_1 - _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + _stride_pdfs_3 + ctr_0];
334 const double xi_114 = -_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 + _stride_pdfs_2 + 17 * _stride_pdfs_3 + ctr_0 + 1];
335 const double xi_115 = -_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 + _stride_pdfs_2 + 18 * _stride_pdfs_3 + ctr_0 - 1];
336 const double xi_116 = -_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 - _stride_pdfs_2 + 5 * _stride_pdfs_3 + ctr_0];
337 const double xi_117 = -_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 + _stride_pdfs_2 + 6 * _stride_pdfs_3 + ctr_0];
338 const double xi_122 = -_data_pdfs[_stride_pdfs_1 * ctr_1 - _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + _stride_pdfs_2 + 15 * _stride_pdfs_3 + ctr_0];
339 const double xi_123 = xi_14 + _data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 - _stride_pdfs_2 + 12 * _stride_pdfs_3 + ctr_0];
340 const double xi_124 = -_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + 2 * _stride_pdfs_3 + ctr_0];
341 const double xi_125 = xi_124 + _data_pdfs[_stride_pdfs_1 * ctr_1 - _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + _stride_pdfs_3 + ctr_0];
342 const double xi_126 = xi_122 + xi_123 + xi_125 + _data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + _stride_pdfs_2 + 16 * _stride_pdfs_3 + ctr_0];
343 const double xi_128 = -_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + _stride_pdfs_2 + 16 * _stride_pdfs_3 + ctr_0];
344 const double xi_129 = 2.0 * _data_pdfs[_stride_pdfs_1 * ctr_1 - _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + 8 * _stride_pdfs_3 + ctr_0 - 1];
345 const double xi_130 = -2.0 * _data_pdfs[_stride_pdfs_1 * ctr_1 - _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + 7 * _stride_pdfs_3 + ctr_0 + 1] + 2.0 * _data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + 10 * _stride_pdfs_3 + ctr_0 - 1];
346 const double xi_131 = xi_125 + xi_128 - xi_129 + xi_130 + xi_5 + 2.0 * _data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + 9 * _stride_pdfs_3 + ctr_0 + 1] - _data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 - _stride_pdfs_2 + 12 * _stride_pdfs_3 + ctr_0];
347 const double xi_133 = xi_131 * xi_132;
348 const double xi_134 = xi_126 * xi_127 + xi_133;
349 const double xi_144 = xi_15 + _data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 - _stride_pdfs_2 + 13 * _stride_pdfs_3 + ctr_0 + 1];
350 const double xi_145 = xi_115 + xi_144;
351 const double xi_146 = xi_145 - _data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 + 3 * _stride_pdfs_3 + ctr_0 + 1] + _data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 + 4 * _stride_pdfs_3 + ctr_0 - 1] + _data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 + _stride_pdfs_2 + 17 * _stride_pdfs_3 + ctr_0 + 1];
352 const double xi_147 = -_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 + 4 * _stride_pdfs_3 + ctr_0 - 1];
353 const double xi_148 = -xi_129 - xi_130 - xi_145 - xi_147 - xi_7 + 2.0 * _data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + 9 * _stride_pdfs_3 + ctr_0 + 1];
354 const double xi_149 = xi_132 * xi_148;
355 const double xi_150 = xi_127 * xi_146 + xi_149;
356 const double xi_152 = xi_122 + xi_128 + _data_pdfs[_stride_pdfs_1 * ctr_1 - _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 - _stride_pdfs_2 + 11 * _stride_pdfs_3 + ctr_0];
357 const double xi_153 = -xi_116 - xi_152 - _data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 - _stride_pdfs_2 + 12 * _stride_pdfs_3 + ctr_0] - _data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 + _stride_pdfs_2 + 6 * _stride_pdfs_3 + ctr_0];
358 const double xi_154 = -xi_107 - xi_108 + xi_109 + xi_117 + xi_152 + xi_6;
359 const double xi_155 = xi_132 * xi_154;
360 const double xi_156 = xi_127 * xi_153 + xi_155;
361 const double xi_161 = -xi_149;
362 const double xi_174 = xi_154 * xi_173;
363 const double xi_176 = xi_153 * xi_175;
364 const double xi_177 = -xi_174 + xi_176;
365 const double xi_178 = xi_131 * xi_173;
366 const double xi_179 = xi_126 * xi_175;
367 const double xi_180 = -xi_178 + xi_179;
368 const double xi_181 = xi_178 - xi_179;
369 const double xi_188 = xi_146 * xi_175;
370 const double xi_189 = xi_148 * xi_173;
371 const double xi_190 = -xi_188 + xi_189;
372 const double xi_191 = xi_188 - xi_189;
373 const double xi_192 = xi_174 - xi_176;
374 const double xi_23 = rr_0 * xi_22;
375 const double xi_35 = rr_0 * xi_34;
376 const double xi_41 = rr_0 * xi_40;
377 const double xi_46 = xi_45 * _data_force[_stride_force_1 * ctr_1 + _stride_force_2 * ctr_2 + ctr_0];
378 const double xi_50 = xi_45 * _data_force[_stride_force_1 * ctr_1 + _stride_force_2 * ctr_2 + _stride_force_3 + ctr_0];
379 const double xi_72 = xi_45 * _data_force[_stride_force_1 * ctr_1 + _stride_force_2 * ctr_2 + 2 * _stride_force_3 + ctr_0];
380 const double vel0Term = xi_3 + _data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 + 4 * _stride_pdfs_3 + ctr_0 - 1];
381 const double vel1Term = xi_4 + xi_5;
382 const double vel2Term = xi_6 + _data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 - _stride_pdfs_2 + 13 * _stride_pdfs_3 + ctr_0 + 1];
383 const double delta_rho = vel0Term + vel1Term + vel2Term + xi_8 + xi_9 + _data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 + _stride_pdfs_2 + 6 * _stride_pdfs_3 + ctr_0] + _data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 + ctr_0];
384 const double rho = delta_rho + 1.0;
385 const double xi_0 = ((1.0) / (rho));
386 const double xi_10 = xi_0 * 0.5;
387 const double u_0 = xi_0 * (vel0Term - xi_11 - xi_8) + xi_10 * _data_force[_stride_force_1 * ctr_1 + _stride_force_2 * ctr_2 + ctr_0];
388 const double xi_17 = u_0 * _data_force[_stride_force_1 * ctr_1 + _stride_force_2 * ctr_2 + ctr_0];
389 const double xi_28 = xi_17 * 0.16666666666666666;
390 const double xi_29 = -xi_28;
391 const double xi_30 = xi_17 * 0.083333333333333329;
392 const double xi_31 = omega_shear * xi_30 + xi_29;
393 const double xi_48 = xi_17 * xi_47 + xi_29;
394 const double xi_49 = xi_34 - xi_46 + xi_48;
395 const double xi_52 = xi_17 * xi_51;
396 const double xi_59 = u_0 * xi_58;
397 const double xi_64 = u_0 * xi_63;
398 const double xi_68 = -xi_34 + xi_46 + xi_48;
399 const double xi_75 = omega_shear * u_0 * -0.083333333333333329 * _data_force[_stride_force_1 * ctr_1 + _stride_force_2 * ctr_2 + ctr_0];
400 const double xi_85 = u_0 * _data_force[_stride_force_1 * ctr_1 + _stride_force_2 * ctr_2 + 2 * _stride_force_3 + ctr_0];
401 const double xi_86 = xi_85 * 0.25;
402 const double xi_89 = xi_62 * xi_85;
403 const double xi_96 = (u_0 * u_0);
404 const double u_1 = xi_0 * (vel1Term - xi_12 - xi_13 - xi_9) + xi_10 * _data_force[_stride_force_1 * ctr_1 + _stride_force_2 * ctr_2 + _stride_force_3 + ctr_0];
405 const double xi_18 = u_1 * _data_force[_stride_force_1 * ctr_1 + _stride_force_2 * ctr_2 + _stride_force_3 + ctr_0];
406 const double xi_26 = xi_18 * 0.16666666666666666;
407 const double xi_36 = omega_shear * u_1 * -0.083333333333333329 * _data_force[_stride_force_1 * ctr_1 + _stride_force_2 * ctr_2 + _stride_force_3 + ctr_0];
408 const double xi_42 = -xi_26;
409 const double xi_43 = xi_18 * 0.083333333333333329;
410 const double xi_53 = xi_18 * xi_51;
411 const double xi_60 = u_1 * 0.25;
412 const double xi_61 = xi_60 * _data_force[_stride_force_1 * ctr_1 + _stride_force_2 * ctr_2 + ctr_0];
413 const double xi_65 = u_1 * xi_62;
414 const double xi_66 = xi_65 * _data_force[_stride_force_1 * ctr_1 + _stride_force_2 * ctr_2 + ctr_0];
415 const double xi_67 = xi_59 + xi_61 - xi_64 - xi_66;
416 const double xi_69 = -xi_59 - xi_61 + xi_64 + xi_66;
417 const double xi_77 = xi_60 * _data_force[_stride_force_1 * ctr_1 + _stride_force_2 * ctr_2 + 2 * _stride_force_3 + ctr_0];
418 const double xi_79 = xi_65 * _data_force[_stride_force_1 * ctr_1 + _stride_force_2 * ctr_2 + 2 * _stride_force_3 + ctr_0];
419 const double xi_95 = rho * (u_1 * u_1);
420 const double xi_101 = -xi_95;
421 const double xi_157 = rho * u_1;
422 const double xi_159 = xi_158 * (u_0 * xi_157 + xi_12 - _data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + 9 * _stride_pdfs_3 + ctr_0 + 1] + _data_pdfs[_stride_pdfs_1 * ctr_1 - _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + 7 * _stride_pdfs_3 + ctr_0 + 1]);
423 const double xi_160 = -xi_159;
424 const double u_2 = xi_0 * (vel2Term - xi_14 - xi_15 - xi_16 - _data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 + _stride_pdfs_2 + 18 * _stride_pdfs_3 + ctr_0 - 1] - _data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 + _stride_pdfs_2 + 6 * _stride_pdfs_3 + ctr_0] - _data_pdfs[_stride_pdfs_1 * ctr_1 - _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + _stride_pdfs_2 + 15 * _stride_pdfs_3 + ctr_0]) + xi_10 * _data_force[_stride_force_1 * ctr_1 + _stride_force_2 * ctr_2 + 2 * _stride_force_3 + ctr_0];
425 const double xi_19 = u_2 * _data_force[_stride_force_1 * ctr_1 + _stride_force_2 * ctr_2 + 2 * _stride_force_3 + ctr_0];
426 const double xi_24 = xi_19 * 0.16666666666666666;
427 const double xi_25 = -xi_24;
428 const double xi_27 = xi_19 * 0.083333333333333329;
429 const double xi_32 = -omega_shear * xi_26 + omega_shear * xi_27 + xi_18 * 0.33333333333333331 + xi_25 + xi_31;
430 const double xi_37 = omega_shear * u_2 * -0.083333333333333329 * _data_force[_stride_force_1 * ctr_1 + _stride_force_2 * ctr_2 + 2 * _stride_force_3 + ctr_0];
431 const double xi_38 = omega_shear * xi_28 + u_0 * -0.33333333333333331 * _data_force[_stride_force_1 * ctr_1 + _stride_force_2 * ctr_2 + ctr_0] + xi_24 + xi_26 + xi_36 + xi_37;
432 const double xi_44 = -omega_shear * xi_24 + omega_shear * xi_43 + xi_19 * 0.33333333333333331 + xi_31 + xi_42;
433 const double xi_54 = xi_19 * xi_51;
434 const double xi_55 = xi_18 * xi_47 + xi_42 + xi_52 + xi_53 + xi_54;
435 const double xi_56 = -xi_22 + xi_50 + xi_55;
436 const double xi_57 = xi_27 + xi_37 + xi_56;
437 const double xi_70 = xi_22 - xi_50 + xi_55;
438 const double xi_71 = xi_27 + xi_37 + xi_70;
439 const double xi_73 = xi_19 * xi_47 + xi_25;
440 const double xi_74 = -xi_40 + xi_72 + xi_73;
441 const double xi_76 = xi_30 + xi_56 + xi_75;
442 const double xi_78 = u_2 * xi_58;
443 const double xi_80 = u_2 * xi_63;
444 const double xi_81 = -xi_77 - xi_78 + xi_79 + xi_80;
445 const double xi_82 = xi_30 + xi_70 + xi_75;
446 const double xi_83 = xi_77 + xi_78 - xi_79 - xi_80;
447 const double xi_84 = xi_36 + xi_43 + xi_52 + xi_53 + xi_54 + xi_74;
448 const double xi_87 = u_2 * _data_force[_stride_force_1 * ctr_1 + _stride_force_2 * ctr_2 + ctr_0];
449 const double xi_88 = xi_87 * 0.25;
450 const double xi_90 = xi_62 * xi_87;
451 const double xi_91 = xi_86 + xi_88 - xi_89 - xi_90;
452 const double xi_92 = -xi_86 - xi_88 + xi_89 + xi_90;
453 const double xi_93 = xi_40 - xi_72 + xi_73;
454 const double xi_94 = xi_36 + xi_43 + xi_52 + xi_53 + xi_54 + xi_93;
455 const double xi_98 = rho * (u_2 * u_2);
456 const double xi_99 = xi_97 + xi_98 * 0.66666666666666663 + 3.0 * _data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 + _stride_pdfs_2 + 6 * _stride_pdfs_3 + ctr_0] + 3.0 * _data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 - _stride_pdfs_2 + 5 * _stride_pdfs_3 + ctr_0];
457 const double xi_100 = omega_even * (rho * xi_96 * 1.6666666666666667 + xi_95 * 0.66666666666666663 + xi_99 - 3.0 * _data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + _stride_pdfs_2 + 16 * _stride_pdfs_3 + ctr_0] - 3.0 * _data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 - _stride_pdfs_2 + 12 * _stride_pdfs_3 + ctr_0] - 3.0 * _data_pdfs[_stride_pdfs_1 * ctr_1 - _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + _stride_pdfs_2 + 15 * _stride_pdfs_3 + ctr_0] - 3.0 * _data_pdfs[_stride_pdfs_1 * ctr_1 - _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 - _stride_pdfs_2 + 11 * _stride_pdfs_3 + ctr_0] + 3.0 * _data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + 2 * _stride_pdfs_3 + ctr_0] + 3.0 * _data_pdfs[_stride_pdfs_1 * ctr_1 - _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + _stride_pdfs_3 + ctr_0]);
458 const double xi_103 = omega_bulk * (rho * xi_96 - xi_101 - xi_102 - xi_13 - xi_16 - xi_5 - xi_97 + xi_98);
459 const double xi_105 = xi_104 + xi_95 * 2.3333333333333335 + xi_99 - 2.0 * _data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + 2 * _stride_pdfs_3 + ctr_0] - 2.0 * _data_pdfs[_stride_pdfs_1 * ctr_1 - _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + _stride_pdfs_3 + ctr_0] - 5.0 * _data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 + _stride_pdfs_2 + 17 * _stride_pdfs_3 + ctr_0 + 1] - 5.0 * _data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 + _stride_pdfs_2 + 18 * _stride_pdfs_3 + ctr_0 - 1] - 5.0 * _data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 - _stride_pdfs_2 + 13 * _stride_pdfs_3 + ctr_0 + 1] - 5.0 * _data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 - _stride_pdfs_2 + 14 * _stride_pdfs_3 + ctr_0 - 1];
460 const double xi_106 = omega_even * xi_105;
461 const double xi_110 = xi_104 + xi_107 + xi_108 + xi_109 + xi_97 + xi_98 * 3.0 - 4.0 * _data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 + _stride_pdfs_2 + 6 * _stride_pdfs_3 + ctr_0] - 4.0 * _data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 - _stride_pdfs_2 + 5 * _stride_pdfs_3 + ctr_0] - 7.0 * _data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + 10 * _stride_pdfs_3 + ctr_0 - 1] - 7.0 * _data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + 9 * _stride_pdfs_3 + ctr_0 + 1] - 7.0 * _data_pdfs[_stride_pdfs_1 * ctr_1 - _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + 7 * _stride_pdfs_3 + ctr_0 + 1] - 7.0 * _data_pdfs[_stride_pdfs_1 * ctr_1 - _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + 8 * _stride_pdfs_3 + ctr_0 - 1] + 5.0 * _data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + 2 * _stride_pdfs_3 + ctr_0] + 5.0 * _data_pdfs[_stride_pdfs_1 * ctr_1 - _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + _stride_pdfs_3 + ctr_0];
462 const double xi_111 = omega_even * xi_110;
463 const double xi_112 = xi_111 * 0.01984126984126984;
464 const double xi_118 = xi_116 + xi_117 + xi_98 + _data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + 9 * _stride_pdfs_3 + ctr_0 + 1];
465 const double xi_119 = omega_shear * (-xi_101 - xi_114 - xi_115 - xi_118 - xi_15 - xi_2 - xi_4 - _data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + 2 * _stride_pdfs_3 + ctr_0] + _data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 - _stride_pdfs_2 + 13 * _stride_pdfs_3 + ctr_0 + 1]);
466 const double xi_120 = xi_119 * 0.125;
467 const double xi_121 = -xi_120;
468 const double xi_135 = xi_100 * 0.050000000000000003;
469 const double xi_136 = rho * xi_96 * 2.0 - xi_102 - xi_113 - xi_118 - xi_124 - xi_95 - 2.0 * _data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 + 3 * _stride_pdfs_3 + ctr_0 + 1] - 2.0 * _data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 + 4 * _stride_pdfs_3 + ctr_0 - 1] + 2.0 * _data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + _stride_pdfs_2 + 16 * _stride_pdfs_3 + ctr_0] + 2.0 * _data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 - _stride_pdfs_2 + 12 * _stride_pdfs_3 + ctr_0] + 2.0 * _data_pdfs[_stride_pdfs_1 * ctr_1 - _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + _stride_pdfs_2 + 15 * _stride_pdfs_3 + ctr_0] + 2.0 * _data_pdfs[_stride_pdfs_1 * ctr_1 - _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 - _stride_pdfs_2 + 11 * _stride_pdfs_3 + ctr_0] - _data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 + _stride_pdfs_2 + 17 * _stride_pdfs_3 + ctr_0 + 1];
470 const double xi_137 = omega_shear * xi_136;
471 const double xi_138 = xi_137 * 0.041666666666666664;
472 const double xi_139 = xi_135 + xi_138;
473 const double xi_140 = -xi_112;
474 const double xi_141 = -xi_138;
475 const double xi_142 = -xi_135 + xi_141;
476 const double xi_143 = xi_106 * 0.035714285714285712;
477 const double xi_151 = xi_106 * 0.021428571428571429;
478 const double xi_162 = xi_119 * 0.0625;
479 const double xi_163 = xi_111 * 0.013888888888888888;
480 const double xi_164 = xi_103 * 0.041666666666666664;
481 const double xi_165 = xi_137 * 0.020833333333333332 + xi_164;
482 const double xi_166 = xi_133 + xi_162 + xi_163 + xi_165;
483 const double xi_167 = -xi_133 + xi_162 + xi_163 + xi_165;
484 const double xi_168 = xi_111 * -0.003968253968253968;
485 const double xi_169 = xi_106 * -0.0071428571428571426;
486 const double xi_170 = xi_158 * (u_2 * xi_157 + xi_123 + xi_128 + _data_pdfs[_stride_pdfs_1 * ctr_1 - _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + _stride_pdfs_2 + 15 * _stride_pdfs_3 + ctr_0]);
487 const double xi_171 = xi_100 * 0.025000000000000001;
488 const double xi_172 = xi_141 + xi_164 + xi_168 + xi_169 + xi_170 + xi_171;
489 const double xi_182 = xi_141 + xi_164 + xi_168 + xi_169 - xi_170 + xi_171;
490 const double xi_183 = xi_158 * (rho * u_0 * u_2 + xi_114 + xi_144 + _data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 + _stride_pdfs_2 + 18 * _stride_pdfs_3 + ctr_0 - 1]);
491 const double xi_184 = -xi_183;
492 const double xi_185 = -xi_162;
493 const double xi_186 = xi_106 * 0.017857142857142856;
494 const double xi_187 = xi_155 + xi_165 + xi_168 + xi_185 + xi_186;
495 const double xi_193 = -xi_155 + xi_165 + xi_168 + xi_185 + xi_186;
496 const double forceTerm_0 = xi_17 * xi_20 - xi_17 + xi_18 * xi_20 - xi_18 + xi_19 * xi_20 - xi_19;
497 const double forceTerm_1 = xi_21 - xi_23 + xi_32;
498 const double forceTerm_2 = -xi_21 + xi_23 + xi_32;
499 const double forceTerm_3 = -xi_33 + xi_35 - xi_38;
500 const double forceTerm_4 = xi_33 - xi_35 - xi_38;
501 const double forceTerm_5 = xi_39 - xi_41 + xi_44;
502 const double forceTerm_6 = -xi_39 + xi_41 + xi_44;
503 const double forceTerm_7 = -xi_49 - xi_57 - xi_67;
504 const double forceTerm_8 = -xi_57 - xi_68 - xi_69;
505 const double forceTerm_9 = -xi_49 - xi_69 - xi_71;
506 const double forceTerm_10 = -xi_67 - xi_68 - xi_71;
507 const double forceTerm_11 = -xi_74 - xi_76 - xi_81;
508 const double forceTerm_12 = -xi_74 - xi_82 - xi_83;
509 const double forceTerm_13 = -xi_49 - xi_84 - xi_91;
510 const double forceTerm_14 = -xi_68 - xi_84 - xi_92;
511 const double forceTerm_15 = -xi_76 - xi_83 - xi_93;
512 const double forceTerm_16 = -xi_81 - xi_82 - xi_93;
513 const double forceTerm_17 = -xi_49 - xi_92 - xi_94;
514 const double forceTerm_18 = -xi_68 - xi_91 - xi_94;
515 _data_pdfs_tmp[_stride_pdfs_tmp_1 * ctr_1 + _stride_pdfs_tmp_2 * ctr_2 + ctr_0] = forceTerm_0 + xi_100 * 0.10000000000000001 + xi_103 * -0.5 + xi_106 * 0.042857142857142858 + xi_111 * 0.023809523809523808 + _data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 + ctr_0];
516 _data_pdfs_tmp[_stride_pdfs_tmp_1 * ctr_1 + _stride_pdfs_tmp_2 * ctr_2 + _stride_pdfs_tmp_3 + ctr_0] = forceTerm_1 + omega_even * xi_105 * 0.014285714285714285 - xi_112 - xi_113 - xi_121 - xi_134 - xi_139 + ((((double)(ctr_1)) >= -0.10000000000000001 + ((double)(lebc_top_index))) ? (rho * v_s * (u_0 * 2.0 + v_s) * 0.16666666666666666) : (0.0));
517 _data_pdfs_tmp[_stride_pdfs_tmp_1 * ctr_1 + _stride_pdfs_tmp_2 * ctr_2 + 2 * _stride_pdfs_tmp_3 + ctr_0] = forceTerm_2 + xi_106 * 0.014285714285714285 + xi_120 + xi_134 + xi_140 + xi_142 + ((((double)(ctr_1)) <= 1.1000000000000001 + ((double)(lebc_bot_index))) ? (rho * v_s * (u_0 * -2.0 + v_s) * 0.16666666666666666) : (0.0)) + _data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + 2 * _stride_pdfs_3 + ctr_0];
518 _data_pdfs_tmp[_stride_pdfs_tmp_1 * ctr_1 + _stride_pdfs_tmp_2 * ctr_2 + 3 * _stride_pdfs_tmp_3 + ctr_0] = forceTerm_3 + xi_137 * 0.083333333333333329 + xi_140 - xi_143 + xi_150 + _data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 + 3 * _stride_pdfs_3 + ctr_0 + 1];
519 _data_pdfs_tmp[_stride_pdfs_tmp_1 * ctr_1 + _stride_pdfs_tmp_2 * ctr_2 + 4 * _stride_pdfs_tmp_3 + ctr_0] = forceTerm_4 + omega_shear * xi_136 * 0.083333333333333329 - xi_112 - xi_143 - xi_147 - xi_150;
520 _data_pdfs_tmp[_stride_pdfs_tmp_1 * ctr_1 + _stride_pdfs_tmp_2 * ctr_2 + 5 * _stride_pdfs_tmp_3 + ctr_0] = forceTerm_5 + omega_even * xi_110 * 0.015873015873015872 - xi_116 - xi_120 - xi_139 - xi_151 - xi_156;
521 _data_pdfs_tmp[_stride_pdfs_tmp_1 * ctr_1 + _stride_pdfs_tmp_2 * ctr_2 + 6 * _stride_pdfs_tmp_3 + ctr_0] = forceTerm_6 + xi_111 * 0.015873015873015872 + xi_121 + xi_142 - xi_151 + xi_156 + _data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 + _stride_pdfs_2 + 6 * _stride_pdfs_3 + ctr_0];
522 _data_pdfs_tmp[_stride_pdfs_tmp_1 * ctr_1 + _stride_pdfs_tmp_2 * ctr_2 + 7 * _stride_pdfs_tmp_3 + ctr_0] = forceTerm_7 + xi_160 + xi_161 + xi_166 + ((((double)(ctr_1)) >= -0.10000000000000001 + ((double)(lebc_top_index))) ? (rho * v_s * (u_0 * -2.0 + u_1 * 3.0 - v_s + 1.0) * 0.083333333333333329) : (0.0)) + _data_pdfs[_stride_pdfs_1 * ctr_1 - _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + 7 * _stride_pdfs_3 + ctr_0 + 1];
523 _data_pdfs_tmp[_stride_pdfs_tmp_1 * ctr_1 + _stride_pdfs_tmp_2 * ctr_2 + 8 * _stride_pdfs_tmp_3 + ctr_0] = forceTerm_8 + xi_149 + xi_159 + xi_166 + ((((double)(ctr_1)) >= -0.10000000000000001 + ((double)(lebc_top_index))) ? (rho * v_s * (u_0 * -2.0 + u_1 * -3.0 - v_s - 1.0) * 0.083333333333333329) : (0.0)) + _data_pdfs[_stride_pdfs_1 * ctr_1 - _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + 8 * _stride_pdfs_3 + ctr_0 - 1];
524 _data_pdfs_tmp[_stride_pdfs_tmp_1 * ctr_1 + _stride_pdfs_tmp_2 * ctr_2 + 9 * _stride_pdfs_tmp_3 + ctr_0] = forceTerm_9 + xi_159 + xi_161 + xi_167 + ((((double)(ctr_1)) <= 1.1000000000000001 + ((double)(lebc_bot_index))) ? (rho * v_s * (u_0 * 2.0 + u_1 * 3.0 - v_s - 1.0) * 0.083333333333333329) : (0.0)) + _data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + 9 * _stride_pdfs_3 + ctr_0 + 1];
525 _data_pdfs_tmp[_stride_pdfs_tmp_1 * ctr_1 + _stride_pdfs_tmp_2 * ctr_2 + 10 * _stride_pdfs_tmp_3 + ctr_0] = forceTerm_10 + xi_149 + xi_160 + xi_167 + ((((double)(ctr_1)) <= 1.1000000000000001 + ((double)(lebc_bot_index))) ? (rho * v_s * (u_0 * 2.0 + u_1 * -3.0 - v_s + 1.0) * 0.083333333333333329) : (0.0)) + _data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + 10 * _stride_pdfs_3 + ctr_0 - 1];
526 _data_pdfs_tmp[_stride_pdfs_tmp_1 * ctr_1 + _stride_pdfs_tmp_2 * ctr_2 + 11 * _stride_pdfs_tmp_3 + ctr_0] = forceTerm_11 + xi_172 + xi_177 + xi_180 + _data_pdfs[_stride_pdfs_1 * ctr_1 - _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 - _stride_pdfs_2 + 11 * _stride_pdfs_3 + ctr_0];
527 _data_pdfs_tmp[_stride_pdfs_tmp_1 * ctr_1 + _stride_pdfs_tmp_2 * ctr_2 + 12 * _stride_pdfs_tmp_3 + ctr_0] = forceTerm_12 + xi_177 + xi_181 + xi_182 + _data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 - _stride_pdfs_2 + 12 * _stride_pdfs_3 + ctr_0];
528 _data_pdfs_tmp[_stride_pdfs_tmp_1 * ctr_1 + _stride_pdfs_tmp_2 * ctr_2 + 13 * _stride_pdfs_tmp_3 + ctr_0] = forceTerm_13 + xi_184 + xi_187 + xi_190 + _data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 - _stride_pdfs_2 + 13 * _stride_pdfs_3 + ctr_0 + 1];
529 _data_pdfs_tmp[_stride_pdfs_tmp_1 * ctr_1 + _stride_pdfs_tmp_2 * ctr_2 + 14 * _stride_pdfs_tmp_3 + ctr_0] = forceTerm_14 + xi_183 + xi_187 + xi_191 + _data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 - _stride_pdfs_2 + 14 * _stride_pdfs_3 + ctr_0 - 1];
530 _data_pdfs_tmp[_stride_pdfs_tmp_1 * ctr_1 + _stride_pdfs_tmp_2 * ctr_2 + 15 * _stride_pdfs_tmp_3 + ctr_0] = forceTerm_15 + xi_180 + xi_182 + xi_192 + _data_pdfs[_stride_pdfs_1 * ctr_1 - _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + _stride_pdfs_2 + 15 * _stride_pdfs_3 + ctr_0];
531 _data_pdfs_tmp[_stride_pdfs_tmp_1 * ctr_1 + _stride_pdfs_tmp_2 * ctr_2 + 16 * _stride_pdfs_tmp_3 + ctr_0] = forceTerm_16 + xi_172 + xi_181 + xi_192 + _data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + _stride_pdfs_2 + 16 * _stride_pdfs_3 + ctr_0];
532 _data_pdfs_tmp[_stride_pdfs_tmp_1 * ctr_1 + _stride_pdfs_tmp_2 * ctr_2 + 17 * _stride_pdfs_tmp_3 + ctr_0] = forceTerm_17 + xi_183 + xi_190 + xi_193 + _data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 + _stride_pdfs_2 + 17 * _stride_pdfs_3 + ctr_0 + 1];
533 _data_pdfs_tmp[_stride_pdfs_tmp_1 * ctr_1 + _stride_pdfs_tmp_2 * ctr_2 + 18 * _stride_pdfs_tmp_3 + ctr_0] = forceTerm_18 + xi_184 + xi_191 + xi_193 + _data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 + _stride_pdfs_2 + 18 * _stride_pdfs_3 + ctr_0 - 1];
534 }
535 }
536 }
537 }
538 }
539}
540} // namespace internal_08be08b6e7ea45132a735524ef494de6
541
542void StreamCollideSweepLeesEdwardsDoublePrecisionAVX::run(IBlock *block) {
543
544 auto force = block->getData<field::GhostLayerField<double, 3>>(forceID);
545 auto pdfs = block->getData<field::GhostLayerField<double, 19>>(pdfsID);
546 field::GhostLayerField<double, 19> *pdfs_tmp;
547 {
548 if (cache_pdfs_.find(block) == cache_pdfs_.end()) {
549 pdfs_tmp = pdfs->cloneUninitialized();
550 cache_pdfs_[block] = pdfs_tmp;
551 } else {
552 pdfs_tmp = cache_pdfs_[block];
553 }
554 }
555
556 auto &v_s = this->v_s_;
557 auto &omega_bulk = this->omega_bulk_;
558 auto &omega_even = this->omega_even_;
559 auto &omega_shear = this->omega_shear_;
560 auto &omega_odd = this->omega_odd_;
561 auto &lebc_top_index = this->lebc_top_index_;
562 auto &lebc_bot_index = this->lebc_bot_index_;
563 WALBERLA_ASSERT_GREATER_EQUAL(-1, -int_c(force->nrOfGhostLayers()))
564 double *RESTRICT const _data_force = force->dataAt(-1, -1, -1, 0);
565 WALBERLA_ASSERT_EQUAL(force->layout(), field::fzyx)
566 WALBERLA_ASSERT_EQUAL((uintptr_t)force->dataAt(0, 0, 0, 0) % 32, 0)
567 WALBERLA_ASSERT_GREATER_EQUAL(-1, -int_c(pdfs->nrOfGhostLayers()))
568 double *RESTRICT const _data_pdfs = pdfs->dataAt(-1, -1, -1, 0);
569 WALBERLA_ASSERT_EQUAL(pdfs->layout(), field::fzyx)
570 WALBERLA_ASSERT_EQUAL((uintptr_t)pdfs->dataAt(0, 0, 0, 0) % 32, 0)
571 WALBERLA_ASSERT_GREATER_EQUAL(-1, -int_c(pdfs_tmp->nrOfGhostLayers()))
572 double *RESTRICT _data_pdfs_tmp = pdfs_tmp->dataAt(-1, -1, -1, 0);
573 WALBERLA_ASSERT_EQUAL(pdfs_tmp->layout(), field::fzyx)
574 WALBERLA_ASSERT_EQUAL((uintptr_t)pdfs_tmp->dataAt(0, 0, 0, 0) % 32, 0)
575 WALBERLA_ASSERT_GREATER_EQUAL(force->xSizeWithGhostLayer(), int64_t(int64_c(force->xSize()) + 2))
576 const int64_t _size_force_0 = int64_t(int64_c(force->xSize()) + 2);
577 WALBERLA_ASSERT_EQUAL(force->layout(), field::fzyx)
578 WALBERLA_ASSERT_EQUAL((uintptr_t)force->dataAt(0, 0, 0, 0) % 32, 0)
579 WALBERLA_ASSERT_GREATER_EQUAL(force->ySizeWithGhostLayer(), int64_t(int64_c(force->ySize()) + 2))
580 const int64_t _size_force_1 = int64_t(int64_c(force->ySize()) + 2);
581 WALBERLA_ASSERT_EQUAL(force->layout(), field::fzyx)
582 WALBERLA_ASSERT_EQUAL((uintptr_t)force->dataAt(0, 0, 0, 0) % 32, 0)
583 WALBERLA_ASSERT_GREATER_EQUAL(force->zSizeWithGhostLayer(), int64_t(int64_c(force->zSize()) + 2))
584 const int64_t _size_force_2 = int64_t(int64_c(force->zSize()) + 2);
585 WALBERLA_ASSERT_EQUAL(force->layout(), field::fzyx)
586 WALBERLA_ASSERT_EQUAL((uintptr_t)force->dataAt(0, 0, 0, 0) % 32, 0)
587 const int64_t _stride_force_1 = int64_t(force->yStride());
588 const int64_t _stride_force_2 = int64_t(force->zStride());
589 const int64_t _stride_force_3 = int64_t(1 * int64_t(force->fStride()));
590 const int64_t _stride_pdfs_1 = int64_t(pdfs->yStride());
591 const int64_t _stride_pdfs_2 = int64_t(pdfs->zStride());
592 const int64_t _stride_pdfs_3 = int64_t(1 * int64_t(pdfs->fStride()));
593 const int64_t _stride_pdfs_tmp_1 = int64_t(pdfs_tmp->yStride());
594 const int64_t _stride_pdfs_tmp_2 = int64_t(pdfs_tmp->zStride());
595 const int64_t _stride_pdfs_tmp_3 = int64_t(1 * int64_t(pdfs_tmp->fStride()));
596 internal_08be08b6e7ea45132a735524ef494de6::streamcollidesweepleesedwardsdoubleprecisionavx_streamcollidesweepleesedwardsdoubleprecisionavx(_data_force, _data_pdfs, _data_pdfs_tmp, _size_force_0, _size_force_1, _size_force_2, _stride_force_1, _stride_force_2, _stride_force_3, _stride_pdfs_1, _stride_pdfs_2, _stride_pdfs_3, _stride_pdfs_tmp_1, _stride_pdfs_tmp_2, _stride_pdfs_tmp_3, lebc_bot_index, lebc_top_index, omega_bulk, omega_even, omega_odd, omega_shear, v_s);
597 pdfs->swapDataPointers(pdfs_tmp);
598}
599
600void StreamCollideSweepLeesEdwardsDoublePrecisionAVX::runOnCellInterval(const shared_ptr<StructuredBlockStorage> &blocks, const CellInterval &globalCellInterval, cell_idx_t ghostLayers, IBlock *block) {
601
602 CellInterval ci = globalCellInterval;
603 CellInterval blockBB = blocks->getBlockCellBB(*block);
604 blockBB.expand(ghostLayers);
605 ci.intersect(blockBB);
606 blocks->transformGlobalToBlockLocalCellInterval(ci, *block);
607 if (ci.empty())
608 return;
609
610 auto force = block->getData<field::GhostLayerField<double, 3>>(forceID);
611 auto pdfs = block->getData<field::GhostLayerField<double, 19>>(pdfsID);
612 field::GhostLayerField<double, 19> *pdfs_tmp;
613 {
614 if (cache_pdfs_.find(block) == cache_pdfs_.end()) {
615 pdfs_tmp = pdfs->cloneUninitialized();
616 cache_pdfs_[block] = pdfs_tmp;
617 } else {
618 pdfs_tmp = cache_pdfs_[block];
619 }
620 }
621
622 auto &v_s = this->v_s_;
623 auto &omega_bulk = this->omega_bulk_;
624 auto &omega_even = this->omega_even_;
625 auto &omega_shear = this->omega_shear_;
626 auto &omega_odd = this->omega_odd_;
627 auto &lebc_top_index = this->lebc_top_index_;
628 auto &lebc_bot_index = this->lebc_bot_index_;
629 WALBERLA_ASSERT_GREATER_EQUAL(ci.xMin() - 1, -int_c(force->nrOfGhostLayers()))
630 WALBERLA_ASSERT_GREATER_EQUAL(ci.yMin() - 1, -int_c(force->nrOfGhostLayers()))
631 WALBERLA_ASSERT_GREATER_EQUAL(ci.zMin() - 1, -int_c(force->nrOfGhostLayers()))
632 double *RESTRICT const _data_force = force->dataAt(ci.xMin() - 1, ci.yMin() - 1, ci.zMin() - 1, 0);
633 WALBERLA_ASSERT_EQUAL(force->layout(), field::fzyx)
634 WALBERLA_ASSERT_EQUAL((uintptr_t)force->dataAt(0, 0, 0, 0) % 32, 0)
635 WALBERLA_ASSERT_GREATER_EQUAL(ci.xMin() - 1, -int_c(pdfs->nrOfGhostLayers()))
636 WALBERLA_ASSERT_GREATER_EQUAL(ci.yMin() - 1, -int_c(pdfs->nrOfGhostLayers()))
637 WALBERLA_ASSERT_GREATER_EQUAL(ci.zMin() - 1, -int_c(pdfs->nrOfGhostLayers()))
638 double *RESTRICT const _data_pdfs = pdfs->dataAt(ci.xMin() - 1, ci.yMin() - 1, ci.zMin() - 1, 0);
639 WALBERLA_ASSERT_EQUAL(pdfs->layout(), field::fzyx)
640 WALBERLA_ASSERT_EQUAL((uintptr_t)pdfs->dataAt(0, 0, 0, 0) % 32, 0)
641 WALBERLA_ASSERT_GREATER_EQUAL(ci.xMin() - 1, -int_c(pdfs_tmp->nrOfGhostLayers()))
642 WALBERLA_ASSERT_GREATER_EQUAL(ci.yMin() - 1, -int_c(pdfs_tmp->nrOfGhostLayers()))
643 WALBERLA_ASSERT_GREATER_EQUAL(ci.zMin() - 1, -int_c(pdfs_tmp->nrOfGhostLayers()))
644 double *RESTRICT _data_pdfs_tmp = pdfs_tmp->dataAt(ci.xMin() - 1, ci.yMin() - 1, ci.zMin() - 1, 0);
645 WALBERLA_ASSERT_EQUAL(pdfs_tmp->layout(), field::fzyx)
646 WALBERLA_ASSERT_EQUAL((uintptr_t)pdfs_tmp->dataAt(0, 0, 0, 0) % 32, 0)
647 WALBERLA_ASSERT_GREATER_EQUAL(force->xSizeWithGhostLayer(), int64_t(int64_c(ci.xSize()) + 2))
648 const int64_t _size_force_0 = int64_t(int64_c(ci.xSize()) + 2);
649 WALBERLA_ASSERT_EQUAL(force->layout(), field::fzyx)
650 WALBERLA_ASSERT_EQUAL((uintptr_t)force->dataAt(0, 0, 0, 0) % 32, 0)
651 WALBERLA_ASSERT_GREATER_EQUAL(force->ySizeWithGhostLayer(), int64_t(int64_c(ci.ySize()) + 2))
652 const int64_t _size_force_1 = int64_t(int64_c(ci.ySize()) + 2);
653 WALBERLA_ASSERT_EQUAL(force->layout(), field::fzyx)
654 WALBERLA_ASSERT_EQUAL((uintptr_t)force->dataAt(0, 0, 0, 0) % 32, 0)
655 WALBERLA_ASSERT_GREATER_EQUAL(force->zSizeWithGhostLayer(), int64_t(int64_c(ci.zSize()) + 2))
656 const int64_t _size_force_2 = int64_t(int64_c(ci.zSize()) + 2);
657 WALBERLA_ASSERT_EQUAL(force->layout(), field::fzyx)
658 WALBERLA_ASSERT_EQUAL((uintptr_t)force->dataAt(0, 0, 0, 0) % 32, 0)
659 const int64_t _stride_force_1 = int64_t(force->yStride());
660 const int64_t _stride_force_2 = int64_t(force->zStride());
661 const int64_t _stride_force_3 = int64_t(1 * int64_t(force->fStride()));
662 const int64_t _stride_pdfs_1 = int64_t(pdfs->yStride());
663 const int64_t _stride_pdfs_2 = int64_t(pdfs->zStride());
664 const int64_t _stride_pdfs_3 = int64_t(1 * int64_t(pdfs->fStride()));
665 const int64_t _stride_pdfs_tmp_1 = int64_t(pdfs_tmp->yStride());
666 const int64_t _stride_pdfs_tmp_2 = int64_t(pdfs_tmp->zStride());
667 const int64_t _stride_pdfs_tmp_3 = int64_t(1 * int64_t(pdfs_tmp->fStride()));
668 internal_08be08b6e7ea45132a735524ef494de6::streamcollidesweepleesedwardsdoubleprecisionavx_streamcollidesweepleesedwardsdoubleprecisionavx(_data_force, _data_pdfs, _data_pdfs_tmp, _size_force_0, _size_force_1, _size_force_2, _stride_force_1, _stride_force_2, _stride_force_3, _stride_pdfs_1, _stride_pdfs_2, _stride_pdfs_3, _stride_pdfs_tmp_1, _stride_pdfs_tmp_2, _stride_pdfs_tmp_3, lebc_bot_index, lebc_top_index, omega_bulk, omega_even, omega_odd, omega_shear, v_s);
669 pdfs->swapDataPointers(pdfs_tmp);
670}
671
672} // namespace pystencils
673} // namespace walberla
674
675#if (defined WALBERLA_CXX_COMPILER_IS_GNU) || (defined WALBERLA_CXX_COMPILER_IS_CLANG)
676#pragma GCC diagnostic pop
677#endif
678
679#if (defined WALBERLA_CXX_COMPILER_IS_INTEL)
680#pragma warning pop
681#endif
FUNC_PREFIX
#define FUNC_PREFIX
\file AdvectiveFluxKernel_double_precision.cpp \author pystencils
Definition AdvectiveFluxKernel_double_precision.cpp:28
RESTRICT
#define RESTRICT
\file AdvectiveFluxKernel_double_precision.h \author pystencils
Definition AdvectiveFluxKernel_double_precision.h:40
walberla::pystencils::StreamCollideSweepLeesEdwardsDoublePrecisionAVX::runOnCellInterval
void runOnCellInterval(const shared_ptr< StructuredBlockStorage > &blocks, const CellInterval &globalCellInterval, cell_idx_t ghostLayers, IBlock *block)
Definition StreamCollideSweepLeesEdwardsDoublePrecisionAVX.cpp:600
stream
cudaStream_t stream[1]
CUDA streams for parallel computing on CPU and GPU.
Definition common_cuda.cu:34
block
static double * block(double *p, std::size_t index, std::size_t size)
Definition elc.cpp:177
std
STL namespace.
walberla::pystencils::internal_08be08b6e7ea45132a735524ef494de6::streamcollidesweepleesedwardsdoubleprecisionavx_streamcollidesweepleesedwardsdoubleprecisionavx
static FUNC_PREFIX void streamcollidesweepleesedwardsdoubleprecisionavx_streamcollidesweepleesedwardsdoubleprecisionavx(double *RESTRICT const _data_force, double *RESTRICT const _data_pdfs, double *RESTRICT _data_pdfs_tmp, int64_t const _size_force_0, int64_t const _size_force_1, int64_t const _size_force_2, int64_t const _stride_force_1, int64_t const _stride_force_2, int64_t const _stride_force_3, int64_t const _stride_pdfs_1, int64_t const _stride_pdfs_2, int64_t const _stride_pdfs_3, int64_t const _stride_pdfs_tmp_1, int64_t const _stride_pdfs_tmp_2, int64_t const _stride_pdfs_tmp_3, int64_t lebc_bot_index, int64_t lebc_top_index, double omega_bulk, double omega_even, double omega_odd, double omega_shear, double v_s)
Definition StreamCollideSweepLeesEdwardsDoublePrecisionAVX.cpp:51
walberla
\file PackInfoPdfDoublePrecision.cpp \author pystencils
Definition EKWalberla.hpp:38