voxel_assembly_common.hpp

// FEAT3: Finite Element Analysis Toolbox, Version 3
// Copyright (C) 2010 by Stefan Turek & the FEAT group
// FEAT3 is released under the GNU General Public License version 3,
// see the file 'copyright.txt' in the top level directory for details.
 
#pragma once
 
#include <kernel/base_header.hpp>
#include <kernel/shape.hpp>
#include <kernel/geometry/conformal_mesh.hpp>
#include <kernel/trafo/standard/mapping.hpp>
#include <kernel/space/lagrange2/element.hpp>
 
#ifdef FEAT_HAVE_OMP
#include "omp.h"
#endif
 
#ifdef __CUDACC__
#include <cooperative_groups.h>
#include <cooperative_groups/memcpy_async.h>
#include <cooperative_groups/reduce.h>
 
namespace cg = cooperative_groups;
#endif
 
namespace FEAT
{
  namespace VoxelAssembly
  {
    typedef Space::Lagrange2::Element<Trafo::Standard::Mapping<Geometry::ConformalMesh<Shape::Hypercube<2>>>> Q2StandardQuad;
    typedef Space::Lagrange2::Element<Trafo::Standard::Mapping<Geometry::ConformalMesh<Shape::Hypercube<3>>>> Q2StandardHexa;
 
    template<int dim_>
    using Q2StandardHyperCube = Space::Lagrange2::Element<Trafo::Standard::Mapping<Geometry::ConformalMesh<Shape::Hypercube<dim_>>>>;
 
 
    template<typename DT_, typename IT_>
    struct AssemblyMappingData
    {
      const IT_* cell_to_dof;
      const IT_* cell_to_dof_sorter;
      Index cell_num;
      const void* nodes;
      Index node_size;
    };
 
    template<typename DT_>
    struct AssemblyCubatureData
    {
      const void* cub_pt;
      const DT_* cub_wg;
      int num_cubs;
    };
 
    template<typename DT_, typename IT_>
    struct CSRMatrixData
    {
      DT_* data;
      const IT_* row_ptr;
      const IT_* col_idx;
      Index num_rows;
      Index num_cols;
    };
 
    template<typename DT_>
    struct AssemblyBurgersData
    {
      DT_ nu;
      DT_ theta;
      DT_ beta;
      DT_ frechet_beta;
      DT_ sd_delta;
      DT_ sd_nu;
      DT_ sd_v_norm;
      bool deformation;
    };
 
    template<typename SpaceHelp_>
    struct BurgersSharedDataGlobalWrapper
    {
      typedef SpaceHelp_ SpaceHelp;
      typedef typename SpaceHelp::SpaceType SpaceType;
      typedef typename SpaceHelp::DataType DataType;
      typedef typename SpaceHelp::IndexType IndexType;
      static constexpr int dim = SpaceHelp::dim;
      static constexpr int num_loc_dofs = SpaceType::DofMappingType::dof_count;
      static constexpr int num_loc_verts = SpaceHelp::num_verts;
      DataType local_coeffs[dim*num_loc_verts];
      DataType local_conv_dofs[dim*num_loc_dofs];
      IndexType local_dofs[num_loc_dofs];
      // these can be defined in shared memory
      DataType tol_eps;
      bool need_streamline;
      bool need_convection;
    };
 
    template<typename SpaceHelp_>
    struct BurgersDefectSharedDataGlobalWrapper
    {
      typedef SpaceHelp_ SpaceHelp;
      typedef typename SpaceHelp::SpaceType SpaceType;
      typedef typename SpaceHelp::DataType DataType;
      typedef typename SpaceHelp::IndexType IndexType;
      static constexpr int dim = SpaceHelp::dim;
      static constexpr int num_loc_dofs = SpaceType::DofMappingType::dof_count;
      static constexpr int num_loc_verts = SpaceHelp::num_verts;
      DataType local_coeffs[dim*num_loc_verts];
      DataType local_conv_dofs[dim*num_loc_dofs];
      DataType local_prim_dofs[dim*num_loc_dofs];
      IndexType local_dofs[num_loc_dofs];
      // these can be defined in shared memory
      bool need_convection;
    };
 
    #ifdef __CUDACC__
    template<typename DT_, int dim_>
    CUDA_DEVICE inline DT_ dot(const Tiny::Vector<DT_, dim_>& grad, const DT_* conv, DT_ alpha = DT_(1))
    {
      DT_ lv(DT_(0));
      for(int k = 0; k < dim_; ++k)
      {
        lv += alpha * grad[k] * conv[k];
      }
      return lv;
    }
 
    template<typename DT_, typename TG_>
    CUDA_DEVICE inline void coalesced_format(const TG_& thread_group, DT_* target, int size)
    {
      for(int i = thread_group.thread_rank(); i < size; i += thread_group.num_threads())
      {
        target[i] = DT_(0);
      }
    }
 
    template<typename DT_, typename TG_, int dim_>
    CUDA_DEVICE inline void grouped_special_dot(const TG_& tg, DT_& loc_val, const DT_ phi, const DT_* conv, DT_ alpha = DT_(1))
    {
      auto coal_g = cg::coalesced_threads();
      DT_ lv(DT_(0));
      for(int k = 0; k < dim_; ++k)
      {
        lv += alpha * phi * conv[k];
      }
      lv = cg::reduce(coal_g, lv, cg::plus<DT_>());
      cg::invoke_one(coal_g, [&](){atomicAdd(&loc_val, lv);});
    }
 
    template<typename DT_, typename TG_, int dim_>
    CUDA_DEVICE inline void grouped_dot(const TG_& tg, DT_& loc_val, const Tiny::Vector<DT_, dim_>& grad, const DT_* conv, DT_ alpha = DT_(1))
    {
      auto coal_g = cg::coalesced_threads();
      DT_ lv(DT_(0));
      for(int k = 0; k < dim_; ++k)
      {
        lv += alpha * grad[k] * conv[k];
      }
      lv = cg::reduce(coal_g, lv, cg::plus<DT_>());
      cg::invoke_one(coal_g, [&](){atomicAdd(&loc_val, lv);});
    }
 
    template<typename DT_, typename TG_, int dim_>
    CUDA_DEVICE inline void grouped_axpy(const TG_& tg, DT_* loc_v, const DT_ phi, const DT_* conv, DT_ alpha = DT_(1))
    {
      auto coal_g = cg::coalesced_threads();
      for(int k = 0; k < dim_; ++k)
      {
        //thos could also be done a bit nicer by using async reduce versions...
        DT_ lv(DT_(0));
        lv = cg::reduce(coal_g, alpha * phi * conv[k], cg::plus<DT_>());
        cg::invoke_one(coal_g, [&](){atomicAdd(loc_v+k, lv);});
        // coal_g.snyc();
      }
    }
 
    template<typename DT_, typename TG_, int dim_>
    CUDA_DEVICE inline void grouped_add_outer_product(const TG_& tg, DT_* loc_grad_v, const DT_* conv, const Tiny::Vector<DT_, dim_>& grad, DT_ alpha = DT_(1))
    {
      auto coal_g = cg::coalesced_threads();
      for(int i(0); i < dim_; ++i)
      {
        for(int j(0); j < dim_; ++j)
        {
          DT_ lv = cg::reduce(coal_g, alpha * conv[i] * grad[j], cg::plus<DT_>());
          cg::invoke_one(coal_g, [&](){atomicAdd(&loc_grad_v[i*dim_ + j], lv);});
        }
      }
    }
 
    template<typename DT_>
    CUDA_DEVICE DT_* get_aligned_address(void* ptr)
    {
      // alignment is always a power of 2, so can use the following bit magic
      // first add the (alignment-1) of the datatype, which leads to the next aligned
      // address modulo bits that fall into the alignment.
      // The simply cut off everything from the address larger then our aligned address by using
      // AND with the negative alignment, which is simply all bits flipped of (alignment-1)
      constexpr std::uint64_t align = alignof(DT_);
      std::uint64_t ptr_a = reinterpret_cast<std::uint64_t>(ptr);
      return (DT_*)((ptr_a + align - 1u)&(-align));
    }
 
    #endif
 
    template<typename DT_>
    struct AssemblyMaterialData
    {
      DT_ frechet_material;
      DT_ reg_eps;
      DT_ mu_0;
      DT_ exp;
      DT_ lambda;
      DT_ a_T;
      DT_ yasuda_a;
      DT_ mu_inf;
      bool need_frechet_material;
    };
 
    enum MaterialType
    {
      carreau = 0,
      carreauYasuda = 1
    };
 
    namespace Intern
    {
      template<typename DT_, MaterialType mat_type>
      struct ViscoFunctor;
 
      template<typename DT_, MaterialType mat_type>
      struct ViscoFunctorNew;
 
      template<typename DT_>
      struct ViscoFunctorNew<DT_, MaterialType::carreau>
      {
        DT_ mu_0, exp, lambda;
 
        CUDA_HOST_DEVICE DT_ operator()(DT_ gamma_dot, DT_ a_T) const
        {
        #ifdef __CUDACC__
          return a_T * mu_0 * CudaMath::cuda_pow(DT_(1) + lambda * a_T * gamma_dot, - exp);
        #else
          return a_T * mu_0 * Math::pow(DT_(1) + lambda * a_T * gamma_dot, - exp);
        #endif
        }
      };
 
      template<typename DT_>
      struct ViscoFunctor<DT_, MaterialType::carreau>
      {
        DT_ mu_0, exp, lambda, a_T;
 
        CUDA_HOST_DEVICE DT_ operator()(DT_ gamma_dot) const
        {
        #ifdef __CUDACC__
          return a_T * mu_0 * CudaMath::cuda_pow(DT_(1) + lambda * a_T * gamma_dot, - exp);
        #else
          return a_T * mu_0 * Math::pow(DT_(1) + lambda * a_T * gamma_dot, - exp);
        #endif
        }
      };
 
      template<typename DT_>
      struct ViscoFunctorNew<DT_, MaterialType::carreauYasuda>
      {
        DT_ mu_0, exp, lambda, yasuda_a, mu_inf;
 
        CUDA_HOST_DEVICE DT_ operator()(DT_ gamma_dot, DT_ a_T) const
        {
        #ifdef __CUDACC__
          return a_T * mu_inf + a_T * (mu_0 - mu_inf) *
                  CudaMath::cuda_pow(DT_(1) + CudaMath::cuda_pow(lambda * a_T * gamma_dot, yasuda_a), (exp - 1) / yasuda_a);
        #else
          return a_T * mu_inf + a_T * (mu_0 - mu_inf) *
                  Math::pow(DT_(1) + Math::pow(lambda * a_T * gamma_dot, yasuda_a), (exp - 1) / yasuda_a);
        #endif
        }
      };
 
      template<typename DT_>
      struct ViscoFunctor<DT_, MaterialType::carreauYasuda>
      {
        DT_ mu_0, exp, lambda, a_T, yasuda_a, mu_inf;
 
        CUDA_HOST_DEVICE DT_ operator()(DT_ gamma_dot) const
        {
        #ifdef __CUDACC__
          return a_T * mu_inf + a_T * (mu_0 - mu_inf) *
                  CudaMath::cuda_pow(DT_(1) + CudaMath::cuda_pow(lambda * a_T * gamma_dot, yasuda_a), (exp - 1) / yasuda_a);
        #else
          return a_T * mu_inf + a_T * (mu_0 - mu_inf) *
                  Math::pow(DT_(1) + Math::pow(lambda * a_T * gamma_dot, yasuda_a), (exp - 1) / yasuda_a);
        #endif
        }
      };
 
      template<typename DT_, MaterialType mat_type>
      struct ViscoDFunctor;
 
      template<typename DT_, MaterialType mat_type>
      struct ViscoDFunctorNew;
 
      template<typename DT_>
      struct ViscoDFunctorNew<DT_, MaterialType::carreau>
      {
        DT_ mu_0, exp, lambda;
 
        CUDA_HOST_DEVICE DT_ operator()(DT_ gamma_dot, DT_ a_T) const
        {
        #ifdef __CUDACC__
          return - a_T * a_T * mu_0 * lambda * exp *
                CudaMath::cuda_pow(DT_(1) + lambda * a_T * gamma_dot, - exp - DT_(1));
        #else
          return - a_T * a_T * mu_0 * lambda * exp *
                Math::pow(DT_(1) + lambda * a_T * gamma_dot, - exp - DT_(1));
        #endif
        }
      };
 
      template<typename DT_>
      struct ViscoDFunctor<DT_, MaterialType::carreau>
      {
        DT_ mu_0, exp, lambda, a_T;
 
        CUDA_HOST_DEVICE DT_ operator()(DT_ gamma_dot) const
        {
        #ifdef __CUDACC__
          return - a_T * a_T * mu_0 * lambda * exp *
                CudaMath::cuda_pow(DT_(1) + lambda * a_T * gamma_dot, - exp - DT_(1));
        #else
          return - a_T * a_T * mu_0 * lambda * exp *
                Math::pow(DT_(1) + lambda * a_T * gamma_dot, - exp - DT_(1));
        #endif
        }
      };
 
      template<typename DT_>
      struct ViscoDFunctorNew<DT_, MaterialType::carreauYasuda>
      {
        DT_ mu_0, exp, lambda, yasuda_a, mu_inf;
 
        CUDA_HOST_DEVICE DT_ operator()(DT_ gamma_dot, DT_ a_T) const
        {
        #ifdef __CUDACC__
        return a_T * (mu_0 - mu_inf) * ((exp-DT_(1.0)) / yasuda_a) *
                CudaMath::cuda_pow( DT_(1.0) + CudaMath::cuda_pow(lambda * gamma_dot, yasuda_a), ((exp - DT_(1.0) - yasuda_a) / yasuda_a)) *
                yasuda_a * lambda * a_T * CudaMath::cuda_pow(lambda * gamma_dot * a_T, yasuda_a-DT_(1.0));
        #else
        return a_T * (mu_0 - mu_inf) * ((exp-DT_(1.0)) / yasuda_a) *
                Math::pow( DT_(1.0) + Math::pow(lambda * gamma_dot, yasuda_a), ((exp - DT_(1.0) - yasuda_a) / yasuda_a)) *
                yasuda_a * lambda * a_T * Math::pow(lambda * gamma_dot * a_T, yasuda_a-DT_(1.0));
        #endif
        }
      };
 
      template<typename DT_>
      struct ViscoDFunctor<DT_, MaterialType::carreauYasuda>
      {
        DT_ mu_0, exp, lambda, a_T, yasuda_a, mu_inf;
 
        CUDA_HOST_DEVICE DT_ operator()(DT_ gamma_dot) const
        {
        #ifdef __CUDACC__
        return a_T * (mu_0 - mu_inf) * ((exp-DT_(1.0)) / yasuda_a) *
                CudaMath::cuda_pow( DT_(1.0) + CudaMath::cuda_pow(lambda * gamma_dot, yasuda_a), ((exp - DT_(1.0) - yasuda_a) / yasuda_a)) *
                yasuda_a * lambda * a_T * CudaMath::cuda_pow(lambda * gamma_dot * a_T, yasuda_a-DT_(1.0));
        #else
        return a_T * (mu_0 - mu_inf) * ((exp-DT_(1.0)) / yasuda_a) *
                Math::pow( DT_(1.0) + Math::pow(lambda * gamma_dot, yasuda_a), ((exp - DT_(1.0) - yasuda_a) / yasuda_a)) *
                yasuda_a * lambda * a_T * Math::pow(lambda * gamma_dot * a_T, yasuda_a-DT_(1.0));
        #endif
        }
      };
    }
  }
}