UpdateVelFromPDFSinglePrecisionAVX.cpp

Go to the documentation of this file.

//======================================================================================================================
//
//  This file is part of waLBerla. waLBerla is free software: you can
//  redistribute it and/or modify it under the terms of the GNU General Public
//  License as published by the Free Software Foundation, either version 3 of
//  the License, or (at your option) any later version.
//
//  waLBerla is distributed in the hope that it will be useful, but WITHOUT
//  ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
//  FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
//  for more details.
//
//  You should have received a copy of the GNU General Public License along
//  with waLBerla (see COPYING.txt). If not, see <http://www.gnu.org/licenses/>.
//
//! \\file UpdateVelFromPDFSinglePrecisionAVX.cpp
//! \\author pystencils
//======================================================================================================================
 
// 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
 
#include <cmath>
 
#include "UpdateVelFromPDFSinglePrecisionAVX.h"
#include "core/DataTypes.h"
#include "core/Macros.h"
 
#include <immintrin.h>
 
#define FUNC_PREFIX
 
#if (defined WALBERLA_CXX_COMPILER_IS_GNU) || (defined WALBERLA_CXX_COMPILER_IS_CLANG)
#pragma GCC diagnostic push
#pragma GCC diagnostic ignored "-Wfloat-equal"
#pragma GCC diagnostic ignored "-Wshadow"
#pragma GCC diagnostic ignored "-Wconversion"
#pragma GCC diagnostic ignored "-Wunused-variable"
#endif
 
#if (defined WALBERLA_CXX_COMPILER_IS_INTEL)
#pragma warning push
#pragma warning(disable : 1599)
#endif
 
using namespace std;
 
namespace walberla {
namespace pystencils {
 
namespace internal_9f3cf46adc24d4ebfb8e08168d5e7562 {
static FUNC_PREFIX void updatevelfrompdfsingleprecisionavx_updatevelfrompdfsingleprecisionavx(float *RESTRICT const _data_force, float *RESTRICT const _data_pdfs, float *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) {
#ifdef _OPENMP
#pragma omp parallel
#endif
  {
#ifdef _OPENMP
#pragma omp for schedule(static)
#endif
    for (int64_t ctr_2 = 1; ctr_2 < _size_force_2 - 1; ctr_2 += 1) {
      for (int64_t ctr_1 = 1; ctr_1 < _size_force_1 - 1; ctr_1 += 1) {
        {
          for (int64_t ctr_0 = 1; ctr_0 < (int64_t)((_size_force_0 - 2) / (8)) * (8) + 1; ctr_0 += 8) {
            const __m256 vel0Term = _mm256_add_ps(_mm256_add_ps(_mm256_add_ps(_mm256_add_ps(_mm256_loadu_ps(&_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 + 4 * _stride_pdfs_3 + ctr_0 - 1]), _mm256_loadu_ps(&_data_pdfs[_stride_pdfs_1 * ctr_1 - _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + 8 * _stride_pdfs_3 + ctr_0 - 1])), _mm256_loadu_ps(&_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 - _stride_pdfs_2 + 14 * _stride_pdfs_3 + ctr_0 - 1])), _mm256_loadu_ps(&_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + 10 * _stride_pdfs_3 + ctr_0 - 1])), _mm256_loadu_ps(&_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 + _stride_pdfs_2 + 18 * _stride_pdfs_3 + ctr_0 - 1]));
            const __m256 momdensity_0 = _mm256_add_ps(_mm256_add_ps(_mm256_add_ps(_mm256_add_ps(_mm256_add_ps(_mm256_mul_ps(_mm256_set_ps(-1.0f, -1.0f, -1.0f, -1.0f, -1.0f, -1.0f, -1.0f, -1.0f), _mm256_loadu_ps(&_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 + 3 * _stride_pdfs_3 + ctr_0 + 1])), _mm256_mul_ps(_mm256_set_ps(-1.0f, -1.0f, -1.0f, -1.0f, -1.0f, -1.0f, -1.0f, -1.0f), _mm256_loadu_ps(&_data_pdfs[_stride_pdfs_1 * ctr_1 - _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + 7 * _stride_pdfs_3 + ctr_0 + 1]))), _mm256_mul_ps(_mm256_set_ps(-1.0f, -1.0f, -1.0f, -1.0f, -1.0f, -1.0f, -1.0f, -1.0f), _mm256_loadu_ps(&_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 - _stride_pdfs_2 + 13 * _stride_pdfs_3 + ctr_0 + 1]))), _mm256_mul_ps(_mm256_set_ps(-1.0f, -1.0f, -1.0f, -1.0f, -1.0f, -1.0f, -1.0f, -1.0f), _mm256_loadu_ps(&_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + 9 * _stride_pdfs_3 + ctr_0 + 1]))), _mm256_mul_ps(_mm256_set_ps(-1.0f, -1.0f, -1.0f, -1.0f, -1.0f, -1.0f, -1.0f, -1.0f), _mm256_loadu_ps(&_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 + _stride_pdfs_2 + 17 * _stride_pdfs_3 + ctr_0 + 1]))), vel0Term);
            const __m256 vel1Term = _mm256_add_ps(_mm256_add_ps(_mm256_add_ps(_mm256_loadu_ps(&_data_pdfs[_stride_pdfs_1 * ctr_1 - _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + _stride_pdfs_3 + ctr_0]), _mm256_loadu_ps(&_data_pdfs[_stride_pdfs_1 * ctr_1 - _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + 7 * _stride_pdfs_3 + ctr_0 + 1])), _mm256_loadu_ps(&_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_ps(&_data_pdfs[_stride_pdfs_1 * ctr_1 - _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + _stride_pdfs_2 + 15 * _stride_pdfs_3 + ctr_0]));
            const __m256 momdensity_1 = _mm256_add_ps(_mm256_add_ps(_mm256_add_ps(_mm256_add_ps(_mm256_add_ps(_mm256_add_ps(_mm256_mul_ps(_mm256_set_ps(-1.0f, -1.0f, -1.0f, -1.0f, -1.0f, -1.0f, -1.0f, -1.0f), _mm256_loadu_ps(&_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + 2 * _stride_pdfs_3 + ctr_0])), _mm256_mul_ps(_mm256_set_ps(-1.0f, -1.0f, -1.0f, -1.0f, -1.0f, -1.0f, -1.0f, -1.0f), _mm256_loadu_ps(&_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + 9 * _stride_pdfs_3 + ctr_0 + 1]))), _mm256_mul_ps(_mm256_set_ps(-1.0f, -1.0f, -1.0f, -1.0f, -1.0f, -1.0f, -1.0f, -1.0f), _mm256_loadu_ps(&_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + 10 * _stride_pdfs_3 + ctr_0 - 1]))), _mm256_mul_ps(_mm256_set_ps(-1.0f, -1.0f, -1.0f, -1.0f, -1.0f, -1.0f, -1.0f, -1.0f), _mm256_loadu_ps(&_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_ps(_mm256_set_ps(-1.0f, -1.0f, -1.0f, -1.0f, -1.0f, -1.0f, -1.0f, -1.0f), _mm256_loadu_ps(&_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_ps(&_data_pdfs[_stride_pdfs_1 * ctr_1 - _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + 8 * _stride_pdfs_3 + ctr_0 - 1]));
            const __m256 vel2Term = _mm256_add_ps(_mm256_add_ps(_mm256_loadu_ps(&_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 - _stride_pdfs_2 + 5 * _stride_pdfs_3 + ctr_0]), _mm256_loadu_ps(&_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 - _stride_pdfs_2 + 13 * _stride_pdfs_3 + ctr_0 + 1])), _mm256_loadu_ps(&_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 - _stride_pdfs_2 + 12 * _stride_pdfs_3 + ctr_0]));
            const __m256 delta_rho = _mm256_add_ps(_mm256_add_ps(_mm256_add_ps(_mm256_add_ps(_mm256_add_ps(_mm256_add_ps(_mm256_add_ps(_mm256_add_ps(_mm256_add_ps(vel0Term, vel1Term), vel2Term), _mm256_load_ps(&_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 + ctr_0])), _mm256_loadu_ps(&_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 + 3 * _stride_pdfs_3 + ctr_0 + 1])), _mm256_loadu_ps(&_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + 2 * _stride_pdfs_3 + ctr_0])), _mm256_loadu_ps(&_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 + _stride_pdfs_2 + 6 * _stride_pdfs_3 + ctr_0])), _mm256_loadu_ps(&_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + 9 * _stride_pdfs_3 + ctr_0 + 1])), _mm256_loadu_ps(&_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 + _stride_pdfs_2 + 17 * _stride_pdfs_3 + ctr_0 + 1])), _mm256_loadu_ps(&_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 + _stride_pdfs_2 + 16 * _stride_pdfs_3 + ctr_0]));
            const __m256 momdensity_2 = _mm256_add_ps(_mm256_add_ps(_mm256_add_ps(_mm256_add_ps(_mm256_add_ps(_mm256_add_ps(_mm256_add_ps(_mm256_mul_ps(_mm256_set_ps(-1.0f, -1.0f, -1.0f, -1.0f, -1.0f, -1.0f, -1.0f, -1.0f), _mm256_loadu_ps(&_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 + _stride_pdfs_2 + 6 * _stride_pdfs_3 + ctr_0])), _mm256_mul_ps(_mm256_set_ps(-1.0f, -1.0f, -1.0f, -1.0f, -1.0f, -1.0f, -1.0f, -1.0f), _mm256_loadu_ps(&_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 + _stride_pdfs_2 + 17 * _stride_pdfs_3 + ctr_0 + 1]))), _mm256_mul_ps(_mm256_set_ps(-1.0f, -1.0f, -1.0f, -1.0f, -1.0f, -1.0f, -1.0f, -1.0f), _mm256_loadu_ps(&_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 + _stride_pdfs_2 + 18 * _stride_pdfs_3 + ctr_0 - 1]))), _mm256_mul_ps(_mm256_set_ps(-1.0f, -1.0f, -1.0f, -1.0f, -1.0f, -1.0f, -1.0f, -1.0f), _mm256_loadu_ps(&_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_ps(_mm256_set_ps(-1.0f, -1.0f, -1.0f, -1.0f, -1.0f, -1.0f, -1.0f, -1.0f), _mm256_loadu_ps(&_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_ps(&_data_pdfs[_stride_pdfs_1 * ctr_1 + _stride_pdfs_2 * ctr_2 - _stride_pdfs_2 + 14 * _stride_pdfs_3 + ctr_0 - 1])), _mm256_loadu_ps(&_data_pdfs[_stride_pdfs_1 * ctr_1 - _stride_pdfs_1 + _stride_pdfs_2 * ctr_2 - _stride_pdfs_2 + 11 * _stride_pdfs_3 + ctr_0]));
            const __m256 rho = _mm256_add_ps(_mm256_set_ps(1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f), delta_rho);
            const __m256 u_0 = _mm256_add_ps(_mm256_mul_ps(momdensity_0, _mm256_div_ps(_mm256_set_ps(1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f), rho)), _mm256_mul_ps(_mm256_mul_ps(_mm256_set_ps(0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f), _mm256_div_ps(_mm256_set_ps(1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f), rho)), _mm256_load_ps(&_data_force[_stride_force_1 * ctr_1 + _stride_force_2 * ctr_2 + ctr_0])));
            const __m256 u_1 = _mm256_add_ps(_mm256_mul_ps(momdensity_1, _mm256_div_ps(_mm256_set_ps(1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f), rho)), _mm256_mul_ps(_mm256_mul_ps(_mm256_set_ps(0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f), _mm256_div_ps(_mm256_set_ps(1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f), rho)), _mm256_loadu_ps(&_data_force[_stride_force_1 * ctr_1 + _stride_force_2 * ctr_2 + _stride_force_3 + ctr_0])));
            const __m256 u_2 = _mm256_add_ps(_mm256_mul_ps(momdensity_2, _mm256_div_ps(_mm256_set_ps(1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f), rho)), _mm256_mul_ps(_mm256_mul_ps(_mm256_set_ps(0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f), _mm256_div_ps(_mm256_set_ps(1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f), rho)), _mm256_loadu_ps(&_data_force[_stride_force_1 * ctr_1 + _stride_force_2 * ctr_2 + 2 * _stride_force_3 + ctr_0])));
            _mm256_store_ps(&_data_velocity[_stride_velocity_1 * ctr_1 + _stride_velocity_2 * ctr_2 + ctr_0], u_0);
            _mm256_storeu_ps(&_data_velocity[_stride_velocity_1 * ctr_1 + _stride_velocity_2 * ctr_2 + _stride_velocity_3 + ctr_0], u_1);
            _mm256_storeu_ps(&_data_velocity[_stride_velocity_1 * ctr_1 + _stride_velocity_2 * ctr_2 + 2 * _stride_velocity_3 + ctr_0], u_2);
          }
          for (int64_t ctr_0 = (int64_t)((_size_force_0 - 2) / (8)) * (8) + 1; ctr_0 < _size_force_0 - 1; ctr_0 += 1) {
            const float 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];
            const float 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];
            const float 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];
            const float 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];
            const float 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];
            const float 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];
            const float 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];
            const float rho = delta_rho + 1.0f;
            const float u_0 = momdensity_0 * ((1.0f) / (rho)) + 0.5f * ((1.0f) / (rho)) * _data_force[_stride_force_1 * ctr_1 + _stride_force_2 * ctr_2 + ctr_0];
            const float u_1 = momdensity_1 * ((1.0f) / (rho)) + 0.5f * ((1.0f) / (rho)) * _data_force[_stride_force_1 * ctr_1 + _stride_force_2 * ctr_2 + _stride_force_3 + ctr_0];
            const float u_2 = momdensity_2 * ((1.0f) / (rho)) + 0.5f * ((1.0f) / (rho)) * _data_force[_stride_force_1 * ctr_1 + _stride_force_2 * ctr_2 + 2 * _stride_force_3 + ctr_0];
            _data_velocity[_stride_velocity_1 * ctr_1 + _stride_velocity_2 * ctr_2 + ctr_0] = u_0;
            _data_velocity[_stride_velocity_1 * ctr_1 + _stride_velocity_2 * ctr_2 + _stride_velocity_3 + ctr_0] = u_1;
            _data_velocity[_stride_velocity_1 * ctr_1 + _stride_velocity_2 * ctr_2 + 2 * _stride_velocity_3 + ctr_0] = u_2;
          }
        }
      }
    }
  }
}
} // namespace internal_9f3cf46adc24d4ebfb8e08168d5e7562
 
void UpdateVelFromPDFSinglePrecisionAVX::run(IBlock *block) {
 
  auto force = block->getData<field::GhostLayerField<float, 3>>(forceID);
  auto velocity = block->getData<field::GhostLayerField<float, 3>>(velocityID);
  auto pdfs = block->getData<field::GhostLayerField<float, 19>>(pdfsID);
 
  WALBERLA_ASSERT_GREATER_EQUAL(-1, -int_c(force->nrOfGhostLayers()))
  float *RESTRICT const _data_force = force->dataAt(-1, -1, -1, 0);
  WALBERLA_ASSERT_EQUAL(force->layout(), field::fzyx)
  WALBERLA_ASSERT_EQUAL((uintptr_t)force->dataAt(0, 0, 0, 0) % 32, 0)
  WALBERLA_ASSERT_GREATER_EQUAL(-1, -int_c(pdfs->nrOfGhostLayers()))
  float *RESTRICT const _data_pdfs = pdfs->dataAt(-1, -1, -1, 0);
  WALBERLA_ASSERT_EQUAL(pdfs->layout(), field::fzyx)
  WALBERLA_ASSERT_EQUAL((uintptr_t)pdfs->dataAt(0, 0, 0, 0) % 32, 0)
  WALBERLA_ASSERT_GREATER_EQUAL(-1, -int_c(velocity->nrOfGhostLayers()))
  float *RESTRICT _data_velocity = velocity->dataAt(-1, -1, -1, 0);
  WALBERLA_ASSERT_EQUAL(velocity->layout(), field::fzyx)
  WALBERLA_ASSERT_EQUAL((uintptr_t)velocity->dataAt(0, 0, 0, 0) % 32, 0)
  WALBERLA_ASSERT_GREATER_EQUAL(force->xSizeWithGhostLayer(), int64_t(int64_c(force->xSize()) + 2))
  const int64_t _size_force_0 = int64_t(int64_c(force->xSize()) + 2);
  WALBERLA_ASSERT_EQUAL(force->layout(), field::fzyx)
  WALBERLA_ASSERT_EQUAL((uintptr_t)force->dataAt(0, 0, 0, 0) % 32, 0)
  WALBERLA_ASSERT_GREATER_EQUAL(force->ySizeWithGhostLayer(), int64_t(int64_c(force->ySize()) + 2))
  const int64_t _size_force_1 = int64_t(int64_c(force->ySize()) + 2);
  WALBERLA_ASSERT_EQUAL(force->layout(), field::fzyx)
  WALBERLA_ASSERT_EQUAL((uintptr_t)force->dataAt(0, 0, 0, 0) % 32, 0)
  WALBERLA_ASSERT_GREATER_EQUAL(force->zSizeWithGhostLayer(), int64_t(int64_c(force->zSize()) + 2))
  const int64_t _size_force_2 = int64_t(int64_c(force->zSize()) + 2);
  WALBERLA_ASSERT_EQUAL(force->layout(), field::fzyx)
  WALBERLA_ASSERT_EQUAL((uintptr_t)force->dataAt(0, 0, 0, 0) % 32, 0)
  const int64_t _stride_force_1 = int64_t(force->yStride());
  const int64_t _stride_force_2 = int64_t(force->zStride());
  const int64_t _stride_force_3 = int64_t(1 * int64_t(force->fStride()));
  const int64_t _stride_pdfs_1 = int64_t(pdfs->yStride());
  const int64_t _stride_pdfs_2 = int64_t(pdfs->zStride());
  const int64_t _stride_pdfs_3 = int64_t(1 * int64_t(pdfs->fStride()));
  const int64_t _stride_velocity_1 = int64_t(velocity->yStride());
  const int64_t _stride_velocity_2 = int64_t(velocity->zStride());
  const int64_t _stride_velocity_3 = int64_t(1 * int64_t(velocity->fStride()));
  internal_9f3cf46adc24d4ebfb8e08168d5e7562::updatevelfrompdfsingleprecisionavx_updatevelfrompdfsingleprecisionavx(_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);
}
 
void UpdateVelFromPDFSinglePrecisionAVX::runOnCellInterval(const shared_ptr<StructuredBlockStorage> &blocks, const CellInterval &globalCellInterval, cell_idx_t ghostLayers, IBlock *block) {
 
  CellInterval ci = globalCellInterval;
  CellInterval blockBB = blocks->getBlockCellBB(*block);
  blockBB.expand(ghostLayers);
  ci.intersect(blockBB);
  blocks->transformGlobalToBlockLocalCellInterval(ci, *block);
  if (ci.empty())
    return;
 
  auto force = block->getData<field::GhostLayerField<float, 3>>(forceID);
  auto velocity = block->getData<field::GhostLayerField<float, 3>>(velocityID);
  auto pdfs = block->getData<field::GhostLayerField<float, 19>>(pdfsID);
 
  WALBERLA_ASSERT_GREATER_EQUAL(ci.xMin() - 1, -int_c(force->nrOfGhostLayers()))
  WALBERLA_ASSERT_GREATER_EQUAL(ci.yMin() - 1, -int_c(force->nrOfGhostLayers()))
  WALBERLA_ASSERT_GREATER_EQUAL(ci.zMin() - 1, -int_c(force->nrOfGhostLayers()))
  float *RESTRICT const _data_force = force->dataAt(ci.xMin() - 1, ci.yMin() - 1, ci.zMin() - 1, 0);
  WALBERLA_ASSERT_EQUAL(force->layout(), field::fzyx)
  WALBERLA_ASSERT_EQUAL((uintptr_t)force->dataAt(0, 0, 0, 0) % 32, 0)
  WALBERLA_ASSERT_GREATER_EQUAL(ci.xMin() - 1, -int_c(pdfs->nrOfGhostLayers()))
  WALBERLA_ASSERT_GREATER_EQUAL(ci.yMin() - 1, -int_c(pdfs->nrOfGhostLayers()))
  WALBERLA_ASSERT_GREATER_EQUAL(ci.zMin() - 1, -int_c(pdfs->nrOfGhostLayers()))
  float *RESTRICT const _data_pdfs = pdfs->dataAt(ci.xMin() - 1, ci.yMin() - 1, ci.zMin() - 1, 0);
  WALBERLA_ASSERT_EQUAL(pdfs->layout(), field::fzyx)
  WALBERLA_ASSERT_EQUAL((uintptr_t)pdfs->dataAt(0, 0, 0, 0) % 32, 0)
  WALBERLA_ASSERT_GREATER_EQUAL(ci.xMin() - 1, -int_c(velocity->nrOfGhostLayers()))
  WALBERLA_ASSERT_GREATER_EQUAL(ci.yMin() - 1, -int_c(velocity->nrOfGhostLayers()))
  WALBERLA_ASSERT_GREATER_EQUAL(ci.zMin() - 1, -int_c(velocity->nrOfGhostLayers()))
  float *RESTRICT _data_velocity = velocity->dataAt(ci.xMin() - 1, ci.yMin() - 1, ci.zMin() - 1, 0);
  WALBERLA_ASSERT_EQUAL(velocity->layout(), field::fzyx)
  WALBERLA_ASSERT_EQUAL((uintptr_t)velocity->dataAt(0, 0, 0, 0) % 32, 0)
  WALBERLA_ASSERT_GREATER_EQUAL(force->xSizeWithGhostLayer(), int64_t(int64_c(ci.xSize()) + 2))
  const int64_t _size_force_0 = int64_t(int64_c(ci.xSize()) + 2);
  WALBERLA_ASSERT_EQUAL(force->layout(), field::fzyx)
  WALBERLA_ASSERT_EQUAL((uintptr_t)force->dataAt(0, 0, 0, 0) % 32, 0)
  WALBERLA_ASSERT_GREATER_EQUAL(force->ySizeWithGhostLayer(), int64_t(int64_c(ci.ySize()) + 2))
  const int64_t _size_force_1 = int64_t(int64_c(ci.ySize()) + 2);
  WALBERLA_ASSERT_EQUAL(force->layout(), field::fzyx)
  WALBERLA_ASSERT_EQUAL((uintptr_t)force->dataAt(0, 0, 0, 0) % 32, 0)
  WALBERLA_ASSERT_GREATER_EQUAL(force->zSizeWithGhostLayer(), int64_t(int64_c(ci.zSize()) + 2))
  const int64_t _size_force_2 = int64_t(int64_c(ci.zSize()) + 2);
  WALBERLA_ASSERT_EQUAL(force->layout(), field::fzyx)
  WALBERLA_ASSERT_EQUAL((uintptr_t)force->dataAt(0, 0, 0, 0) % 32, 0)
  const int64_t _stride_force_1 = int64_t(force->yStride());
  const int64_t _stride_force_2 = int64_t(force->zStride());
  const int64_t _stride_force_3 = int64_t(1 * int64_t(force->fStride()));
  const int64_t _stride_pdfs_1 = int64_t(pdfs->yStride());
  const int64_t _stride_pdfs_2 = int64_t(pdfs->zStride());
  const int64_t _stride_pdfs_3 = int64_t(1 * int64_t(pdfs->fStride()));
  const int64_t _stride_velocity_1 = int64_t(velocity->yStride());
  const int64_t _stride_velocity_2 = int64_t(velocity->zStride());
  const int64_t _stride_velocity_3 = int64_t(1 * int64_t(velocity->fStride()));
  internal_9f3cf46adc24d4ebfb8e08168d5e7562::updatevelfrompdfsingleprecisionavx_updatevelfrompdfsingleprecisionavx(_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);
}
 
} // namespace pystencils
} // namespace walberla
 
#if (defined WALBERLA_CXX_COMPILER_IS_GNU) || (defined WALBERLA_CXX_COMPILER_IS_CLANG)
#pragma GCC diagnostic pop
#endif
 
#if (defined WALBERLA_CXX_COMPILER_IS_INTEL)
#pragma warning pop
#endif