ESPResSo
Extensible Simulation Package for Research on Soft Matter Systems
Loading...
Searching...
No Matches
UpdateVelFromPDFDoublePrecisionAVX.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 UpdateVelFromPDFDoublePrecisionAVX.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 007e77e077ad9d22b5eed6f3d3118240993e553c
21
22#include <cmath>
23
24#include "UpdateVelFromPDFDoublePrecisionAVX.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_fb7a198ee8d4b87d3091e088a14673a2 {
51static FUNC_PREFIX void updatevelfrompdfdoubleprecisionavx_updatevelfrompdfdoubleprecisionavx(double *RESTRICT const _data_force, double *RESTRICT const _data_pdfs, double *RESTRICT _data_velocity, 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_velocity_1, int64_t const _stride_velocity_2, int64_t const _stride_velocity_3) {
52#pragma omp parallel
53 {
54#pragma omp for schedule(static)
55 for (int64_t ctr_2 = 1; ctr_2 < _size_force_2 - 1; ctr_2 += 1) {
56 for (int64_t ctr_1 = 1; ctr_1 < _size_force_1 - 1; ctr_1 += 1) {
57 {
58 for (int64_t ctr_0 = 1; ctr_0 < (int64_t)((_size_force_0 - 2) / (4)) * (4) + 1; ctr_0 += 4) {
59 const __m256d vel0Term = _mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_mm256_add_pd(_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_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_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_1 + _stride_pdfs_2 * ctr_2 + 10 * _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]));
60 const __m256d momdensity_0 = _mm256_add_pd(_mm256_add_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_2 * ctr_2 + 3 * _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 + 7 * _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_2 * ctr_2 - _stride_pdfs_2 + 13 * _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 + 9 * _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_2 * ctr_2 + _stride_pdfs_2 + 17 * _stride_pdfs_3 + ctr_0 + 1]))), vel0Term);
61 const __m256d vel1Term = _mm256_add_pd(_mm256_add_pd(_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])), _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]));
62 const __m256d momdensity_1 = _mm256_add_pd(_mm256_add_pd(_mm256_add_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 + 2 * _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 + 9 * _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 + 10 * _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 + 12 * _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 + 16 * _stride_pdfs_3 + ctr_0]))), vel1Term), _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]));
63 const __m256d vel2Term = _mm256_add_pd(_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_2 * ctr_2 - _stride_pdfs_2 + 13 * _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]));
64 const __m256d delta_rho = _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(vel0Term, vel1Term), vel2Term), _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 + 3 * _stride_pdfs_3 + ctr_0 + 1])), _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 + 6 * _stride_pdfs_3 + ctr_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_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]));
65 const __m256d momdensity_2 = _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(-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 + 17 * _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_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]))), _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]))), vel2Term), _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_1 + _stride_pdfs_2 * ctr_2 - _stride_pdfs_2 + 11 * _stride_pdfs_3 + ctr_0]));
66 const __m256d rho = _mm256_add_pd(_mm256_set_pd(1.0, 1.0, 1.0, 1.0), delta_rho);
67 const __m256d u_0 = _mm256_add_pd(_mm256_mul_pd(momdensity_0, _mm256_div_pd(_mm256_set_pd(1.0, 1.0, 1.0, 1.0), rho)), _mm256_mul_pd(_mm256_mul_pd(_mm256_set_pd(0.5, 0.5, 0.5, 0.5), _mm256_div_pd(_mm256_set_pd(1.0, 1.0, 1.0, 1.0), rho)), _mm256_load_pd(&_data_force[_stride_force_1 * ctr_1 + _stride_force_2 * ctr_2 + ctr_0])));
68 const __m256d u_1 = _mm256_add_pd(_mm256_mul_pd(momdensity_1, _mm256_div_pd(_mm256_set_pd(1.0, 1.0, 1.0, 1.0), rho)), _mm256_mul_pd(_mm256_mul_pd(_mm256_set_pd(0.5, 0.5, 0.5, 0.5), _mm256_div_pd(_mm256_set_pd(1.0, 1.0, 1.0, 1.0), rho)), _mm256_loadu_pd(&_data_force[_stride_force_1 * ctr_1 + _stride_force_2 * ctr_2 + _stride_force_3 + ctr_0])));
69 const __m256d u_2 = _mm256_add_pd(_mm256_mul_pd(momdensity_2, _mm256_div_pd(_mm256_set_pd(1.0, 1.0, 1.0, 1.0), rho)), _mm256_mul_pd(_mm256_mul_pd(_mm256_set_pd(0.5, 0.5, 0.5, 0.5), _mm256_div_pd(_mm256_set_pd(1.0, 1.0, 1.0, 1.0), rho)), _mm256_loadu_pd(&_data_force[_stride_force_1 * ctr_1 + _stride_force_2 * ctr_2 + 2 * _stride_force_3 + ctr_0])));
70 _mm256_store_pd(&_data_velocity[_stride_velocity_1 * ctr_1 + _stride_velocity_2 * ctr_2 + ctr_0], u_0);
71 _mm256_storeu_pd(&_data_velocity[_stride_velocity_1 * ctr_1 + _stride_velocity_2 * ctr_2 + _stride_velocity_3 + ctr_0], u_1);
72 _mm256_storeu_pd(&_data_velocity[_stride_velocity_1 * ctr_1 + _stride_velocity_2 * ctr_2 + 2 * _stride_velocity_3 + ctr_0], u_2);
73 }
74 for (int64_t ctr_0 = (int64_t)((_size_force_0 - 2) / (4)) * (4) + 1; ctr_0 < _size_force_0 - 1; ctr_0 += 1) {
75 const double vel0Term = _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_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 + 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] + _data_pdfs[_stride_pdfs_1 * ctr_1 - _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + 8 * _stride_pdfs_3 + ctr_0 - 1];
76 const double momdensity_0 = vel0Term - _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_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] - _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];
77 const double vel1Term = _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_2 + 15 * _stride_pdfs_3 + ctr_0] + _data_pdfs[_stride_pdfs_1 * ctr_1 - _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + _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];
78 const double momdensity_1 = vel1Term - _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 + 2 * _stride_pdfs_3 + ctr_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 + 16 * _stride_pdfs_3 + ctr_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] + _data_pdfs[_stride_pdfs_1 * ctr_1 - _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + 8 * _stride_pdfs_3 + ctr_0 - 1];
79 const double vel2Term = _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 + 13 * _stride_pdfs_3 + ctr_0 + 1] + _data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 - _stride_pdfs_2 + 5 * _stride_pdfs_3 + ctr_0];
80 const double delta_rho = vel0Term + vel1Term + vel2Term + _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 + 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 + 16 * _stride_pdfs_3 + ctr_0] + _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] + _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];
81 const double momdensity_2 = vel2Term - _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] - _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_2 * ctr_2 - _stride_pdfs_2 + 14 * _stride_pdfs_3 + ctr_0 - 1] - _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];
82 const double rho = delta_rho + 1.0;
83 const double u_0 = momdensity_0 * ((1.0) / (rho)) + 0.5 * ((1.0) / (rho)) * _data_force[_stride_force_1 * ctr_1 + _stride_force_2 * ctr_2 + ctr_0];
84 const double u_1 = momdensity_1 * ((1.0) / (rho)) + 0.5 * ((1.0) / (rho)) * _data_force[_stride_force_1 * ctr_1 + _stride_force_2 * ctr_2 + _stride_force_3 + ctr_0];
85 const double u_2 = momdensity_2 * ((1.0) / (rho)) + 0.5 * ((1.0) / (rho)) * _data_force[_stride_force_1 * ctr_1 + _stride_force_2 * ctr_2 + 2 * _stride_force_3 + ctr_0];
86 _data_velocity[_stride_velocity_1 * ctr_1 + _stride_velocity_2 * ctr_2 + ctr_0] = u_0;
87 _data_velocity[_stride_velocity_1 * ctr_1 + _stride_velocity_2 * ctr_2 + _stride_velocity_3 + ctr_0] = u_1;
88 _data_velocity[_stride_velocity_1 * ctr_1 + _stride_velocity_2 * ctr_2 + 2 * _stride_velocity_3 + ctr_0] = u_2;
89 }
90 }
91 }
92 }
93 }
94}
95} // namespace internal_fb7a198ee8d4b87d3091e088a14673a2
96
97void UpdateVelFromPDFDoublePrecisionAVX::run(IBlock *block) {
98
99 auto velocity = block->getData<field::GhostLayerField<double, 3>>(velocityID);
100 auto force = block->getData<field::GhostLayerField<double, 3>>(forceID);
101 auto pdfs = block->getData<field::GhostLayerField<double, 19>>(pdfsID);
102
103 WALBERLA_ASSERT_GREATER_EQUAL(-1, -int_c(force->nrOfGhostLayers()))
104 double *RESTRICT const _data_force = force->dataAt(-1, -1, -1, 0);
105 WALBERLA_ASSERT_EQUAL(force->layout(), field::fzyx)
106 WALBERLA_ASSERT_EQUAL((uintptr_t)force->dataAt(0, 0, 0, 0) % 32, 0)
107 WALBERLA_ASSERT_GREATER_EQUAL(-1, -int_c(pdfs->nrOfGhostLayers()))
108 double *RESTRICT const _data_pdfs = pdfs->dataAt(-1, -1, -1, 0);
109 WALBERLA_ASSERT_EQUAL(pdfs->layout(), field::fzyx)
110 WALBERLA_ASSERT_EQUAL((uintptr_t)pdfs->dataAt(0, 0, 0, 0) % 32, 0)
111 WALBERLA_ASSERT_GREATER_EQUAL(-1, -int_c(velocity->nrOfGhostLayers()))
112 double *RESTRICT _data_velocity = velocity->dataAt(-1, -1, -1, 0);
113 WALBERLA_ASSERT_EQUAL(velocity->layout(), field::fzyx)
114 WALBERLA_ASSERT_EQUAL((uintptr_t)velocity->dataAt(0, 0, 0, 0) % 32, 0)
115 WALBERLA_ASSERT_GREATER_EQUAL(force->xSizeWithGhostLayer(), int64_t(int64_c(force->xSize()) + 2))
116 const int64_t _size_force_0 = int64_t(int64_c(force->xSize()) + 2);
117 WALBERLA_ASSERT_EQUAL(force->layout(), field::fzyx)
118 WALBERLA_ASSERT_EQUAL((uintptr_t)force->dataAt(0, 0, 0, 0) % 32, 0)
119 WALBERLA_ASSERT_GREATER_EQUAL(force->ySizeWithGhostLayer(), int64_t(int64_c(force->ySize()) + 2))
120 const int64_t _size_force_1 = int64_t(int64_c(force->ySize()) + 2);
121 WALBERLA_ASSERT_EQUAL(force->layout(), field::fzyx)
122 WALBERLA_ASSERT_EQUAL((uintptr_t)force->dataAt(0, 0, 0, 0) % 32, 0)
123 WALBERLA_ASSERT_GREATER_EQUAL(force->zSizeWithGhostLayer(), int64_t(int64_c(force->zSize()) + 2))
124 const int64_t _size_force_2 = int64_t(int64_c(force->zSize()) + 2);
125 WALBERLA_ASSERT_EQUAL(force->layout(), field::fzyx)
126 WALBERLA_ASSERT_EQUAL((uintptr_t)force->dataAt(0, 0, 0, 0) % 32, 0)
127 const int64_t _stride_force_1 = int64_t(force->yStride());
128 const int64_t _stride_force_2 = int64_t(force->zStride());
129 const int64_t _stride_force_3 = int64_t(1 * int64_t(force->fStride()));
130 const int64_t _stride_pdfs_1 = int64_t(pdfs->yStride());
131 const int64_t _stride_pdfs_2 = int64_t(pdfs->zStride());
132 const int64_t _stride_pdfs_3 = int64_t(1 * int64_t(pdfs->fStride()));
133 const int64_t _stride_velocity_1 = int64_t(velocity->yStride());
134 const int64_t _stride_velocity_2 = int64_t(velocity->zStride());
135 const int64_t _stride_velocity_3 = int64_t(1 * int64_t(velocity->fStride()));
136 internal_fb7a198ee8d4b87d3091e088a14673a2::updatevelfrompdfdoubleprecisionavx_updatevelfrompdfdoubleprecisionavx(_data_force, _data_pdfs, _data_velocity, _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_velocity_1, _stride_velocity_2, _stride_velocity_3);
137}
138
139void UpdateVelFromPDFDoublePrecisionAVX::runOnCellInterval(const shared_ptr<StructuredBlockStorage> &blocks, const CellInterval &globalCellInterval, cell_idx_t ghostLayers, IBlock *block) {
140
141 CellInterval ci = globalCellInterval;
142 CellInterval blockBB = blocks->getBlockCellBB(*block);
143 blockBB.expand(ghostLayers);
144 ci.intersect(blockBB);
145 blocks->transformGlobalToBlockLocalCellInterval(ci, *block);
146 if (ci.empty())
147 return;
148
149 auto velocity = block->getData<field::GhostLayerField<double, 3>>(velocityID);
150 auto force = block->getData<field::GhostLayerField<double, 3>>(forceID);
151 auto pdfs = block->getData<field::GhostLayerField<double, 19>>(pdfsID);
152
153 WALBERLA_ASSERT_GREATER_EQUAL(ci.xMin() - 1, -int_c(force->nrOfGhostLayers()))
154 WALBERLA_ASSERT_GREATER_EQUAL(ci.yMin() - 1, -int_c(force->nrOfGhostLayers()))
155 WALBERLA_ASSERT_GREATER_EQUAL(ci.zMin() - 1, -int_c(force->nrOfGhostLayers()))
156 double *RESTRICT const _data_force = force->dataAt(ci.xMin() - 1, ci.yMin() - 1, ci.zMin() - 1, 0);
157 WALBERLA_ASSERT_EQUAL(force->layout(), field::fzyx)
158 WALBERLA_ASSERT_EQUAL((uintptr_t)force->dataAt(0, 0, 0, 0) % 32, 0)
159 WALBERLA_ASSERT_GREATER_EQUAL(ci.xMin() - 1, -int_c(pdfs->nrOfGhostLayers()))
160 WALBERLA_ASSERT_GREATER_EQUAL(ci.yMin() - 1, -int_c(pdfs->nrOfGhostLayers()))
161 WALBERLA_ASSERT_GREATER_EQUAL(ci.zMin() - 1, -int_c(pdfs->nrOfGhostLayers()))
162 double *RESTRICT const _data_pdfs = pdfs->dataAt(ci.xMin() - 1, ci.yMin() - 1, ci.zMin() - 1, 0);
163 WALBERLA_ASSERT_EQUAL(pdfs->layout(), field::fzyx)
164 WALBERLA_ASSERT_EQUAL((uintptr_t)pdfs->dataAt(0, 0, 0, 0) % 32, 0)
165 WALBERLA_ASSERT_GREATER_EQUAL(ci.xMin() - 1, -int_c(velocity->nrOfGhostLayers()))
166 WALBERLA_ASSERT_GREATER_EQUAL(ci.yMin() - 1, -int_c(velocity->nrOfGhostLayers()))
167 WALBERLA_ASSERT_GREATER_EQUAL(ci.zMin() - 1, -int_c(velocity->nrOfGhostLayers()))
168 double *RESTRICT _data_velocity = velocity->dataAt(ci.xMin() - 1, ci.yMin() - 1, ci.zMin() - 1, 0);
169 WALBERLA_ASSERT_EQUAL(velocity->layout(), field::fzyx)
170 WALBERLA_ASSERT_EQUAL((uintptr_t)velocity->dataAt(0, 0, 0, 0) % 32, 0)
171 WALBERLA_ASSERT_GREATER_EQUAL(force->xSizeWithGhostLayer(), int64_t(int64_c(ci.xSize()) + 2))
172 const int64_t _size_force_0 = int64_t(int64_c(ci.xSize()) + 2);
173 WALBERLA_ASSERT_EQUAL(force->layout(), field::fzyx)
174 WALBERLA_ASSERT_EQUAL((uintptr_t)force->dataAt(0, 0, 0, 0) % 32, 0)
175 WALBERLA_ASSERT_GREATER_EQUAL(force->ySizeWithGhostLayer(), int64_t(int64_c(ci.ySize()) + 2))
176 const int64_t _size_force_1 = int64_t(int64_c(ci.ySize()) + 2);
177 WALBERLA_ASSERT_EQUAL(force->layout(), field::fzyx)
178 WALBERLA_ASSERT_EQUAL((uintptr_t)force->dataAt(0, 0, 0, 0) % 32, 0)
179 WALBERLA_ASSERT_GREATER_EQUAL(force->zSizeWithGhostLayer(), int64_t(int64_c(ci.zSize()) + 2))
180 const int64_t _size_force_2 = int64_t(int64_c(ci.zSize()) + 2);
181 WALBERLA_ASSERT_EQUAL(force->layout(), field::fzyx)
182 WALBERLA_ASSERT_EQUAL((uintptr_t)force->dataAt(0, 0, 0, 0) % 32, 0)
183 const int64_t _stride_force_1 = int64_t(force->yStride());
184 const int64_t _stride_force_2 = int64_t(force->zStride());
185 const int64_t _stride_force_3 = int64_t(1 * int64_t(force->fStride()));
186 const int64_t _stride_pdfs_1 = int64_t(pdfs->yStride());
187 const int64_t _stride_pdfs_2 = int64_t(pdfs->zStride());
188 const int64_t _stride_pdfs_3 = int64_t(1 * int64_t(pdfs->fStride()));
189 const int64_t _stride_velocity_1 = int64_t(velocity->yStride());
190 const int64_t _stride_velocity_2 = int64_t(velocity->zStride());
191 const int64_t _stride_velocity_3 = int64_t(1 * int64_t(velocity->fStride()));
192 internal_fb7a198ee8d4b87d3091e088a14673a2::updatevelfrompdfdoubleprecisionavx_updatevelfrompdfdoubleprecisionavx(_data_force, _data_pdfs, _data_velocity, _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_velocity_1, _stride_velocity_2, _stride_velocity_3);
193}
194
195} // namespace pystencils
196} // namespace walberla
197
198#if (defined WALBERLA_CXX_COMPILER_IS_GNU) || (defined WALBERLA_CXX_COMPILER_IS_CLANG)
199#pragma GCC diagnostic pop
200#endif
201
202#if (defined WALBERLA_CXX_COMPILER_IS_INTEL)
203#pragma warning pop
204#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::UpdateVelFromPDFDoublePrecisionAVX::runOnCellInterval
void runOnCellInterval(const shared_ptr< StructuredBlockStorage > &blocks, const CellInterval &globalCellInterval, cell_idx_t ghostLayers, IBlock *block)
Definition UpdateVelFromPDFDoublePrecisionAVX.cpp:139
block
static double * block(double *p, std::size_t index, std::size_t size)
Definition elc.cpp:175
std
STL namespace.
walberla::pystencils::internal_fb7a198ee8d4b87d3091e088a14673a2::updatevelfrompdfdoubleprecisionavx_updatevelfrompdfdoubleprecisionavx
static FUNC_PREFIX void updatevelfrompdfdoubleprecisionavx_updatevelfrompdfdoubleprecisionavx(double *RESTRICT const _data_force, double *RESTRICT const _data_pdfs, double *RESTRICT _data_velocity, 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_velocity_1, int64_t const _stride_velocity_2, int64_t const _stride_velocity_3)
Definition UpdateVelFromPDFDoublePrecisionAVX.cpp:51
walberla
\file PackInfoPdfDoublePrecision.cpp \author pystencils
Definition EKWalberla.hpp:38
velocity
static Utils::Vector3d velocity(Particle const &p_ref, Particle const &p_vs)
Velocity of the virtual site.
Definition relative.cpp:64