ngsolve 6.2.2506.post74.dev0__cp314-cp314-macosx_10_15_universal2.whl

netgen/include/hcurlhdiv_dshape.hpp

@@ -0,0 +1,857 @@
+#ifndef FILE_HCURLHDIV_DSHAPE
+#define FILE_HCURLHDIV_DSHAPE
+
+
+#include "finiteelement.hpp"
+#include "diffop.hpp"
+
+namespace ngfem
+{
+
+  /** calculates [du1/dx1 du2/dx1 (du3/dx1) du1/dx2 du2/dx2 (du3/dx2) (du1/dx3 du2/dx3 du3/dx3)] */
+  template<typename FEL, int DIMSPACE, int DIM, int DIM_STRESS>
+  void CalcDShapeFE(const FEL & fel, const MappedIntegrationPoint<DIM,DIMSPACE>& mip,
+                    BareSliceMatrix<> bmatu, LocalHeap& lh, double eps = 1e-4)
+  {
+    HeapReset hr(lh);
+    // bmatu = 0;
+    // evaluate dshape by numerical diff
+    //fel, eltrans, mip, returnval, lh
+    int nd_u = fel.GetNDof();
+    const IntegrationPoint& ip = mip.IP();//volume_ir[i];
+    const ElementTransformation & eltrans = mip.GetTransformation();
+    FlatMatrixFixWidth<DIM_STRESS> shape_ul(nd_u, lh);
+    FlatMatrixFixWidth<DIM_STRESS> shape_ur(nd_u, lh);
+    FlatMatrixFixWidth<DIM_STRESS> shape_ull(nd_u, lh);
+    FlatMatrixFixWidth<DIM_STRESS> shape_urr(nd_u, lh);
+    FlatMatrixFixWidth<DIM_STRESS> dshape_u_ref(nd_u, lh);//(shape_ur); ///saves "reserved lh-memory"
+
+    FlatMatrixFixWidth<DIM> dshape_u_ref_comp(nd_u, lh);
+    FlatMatrixFixWidth<DIMSPACE> dshape_u(nd_u, lh);//(shape_ul);///saves "reserved lh-memory"
+    
+    for (int j = 0; j < DIM; j++)   // d / dxj
+      {
+        IntegrationPoint ipl(ip);
+        ipl(j) -= eps;
+        IntegrationPoint ipr(ip);
+        ipr(j) += eps;
+        IntegrationPoint ipll(ip);
+        ipll(j) -= 2*eps;
+        IntegrationPoint iprr(ip);
+        iprr(j) += 2*eps;
+        
+        MappedIntegrationPoint<DIM,DIMSPACE> mipl(ipl, eltrans);
+        MappedIntegrationPoint<DIM,DIMSPACE> mipr(ipr, eltrans);
+        MappedIntegrationPoint<DIM,DIMSPACE> mipll(ipll, eltrans);
+        MappedIntegrationPoint<DIM,DIMSPACE> miprr(iprr, eltrans);
+        
+        fel.CalcMappedShape (mipl, shape_ul);
+        fel.CalcMappedShape (mipr, shape_ur);
+        fel.CalcMappedShape (mipll, shape_ull);
+        fel.CalcMappedShape (miprr, shape_urr);
+        
+        dshape_u_ref = (1.0/(12.0*eps)) * (8.0*shape_ur-8.0*shape_ul-shape_urr+shape_ull);
+        for (int l = 0; l < DIM_STRESS; l++)
+          bmatu.Col(j*DIM_STRESS+l) = dshape_u_ref.Col(l);
+      }
+    
+    for (int j = 0; j < DIM_STRESS; j++)
+      {
+        for (int k = 0; k < nd_u; k++)
+          for (int l = 0; l < DIM; l++)
+            dshape_u_ref_comp(k,l) = bmatu(k, l*DIM_STRESS+j);
+        
+        dshape_u = dshape_u_ref_comp * mip.GetJacobianInverse();
+        
+        for (int k = 0; k < nd_u; k++)
+          for (int l = 0; l < DIMSPACE; l++)
+            bmatu(k, l*DIM_STRESS+j) = dshape_u(k,l);
+      }
+  }
+
+
+  template<typename FEL, int DIMSPACE, int DIM, int DIM_STRESS, class TVX, class TVY>
+  void ApplyDShapeFE(const FEL & fel, const MappedIntegrationPoint<DIM,DIMSPACE>& mip, const TVX & x, TVY & y, LocalHeap& lh, double eps = 1e-4)
+  {
+    const IntegrationPoint& ip = mip.IP();
+    const ElementTransformation & eltrans = mip.GetTransformation();
+    Mat<DIM_STRESS,1> shape_ul;
+    Mat<DIM_STRESS,1> shape_ur;
+    Mat<DIM_STRESS,1> shape_ull;
+    Mat<DIM_STRESS,1> shape_urr;
+    Mat<DIM_STRESS,1> dshape_u_ref;
+
+    Vec<DIM> dshape_u_ref_comp;
+    Vec<DIMSPACE> dshape_u;
+    
+    for (int j = 0; j < DIM; j++)   // d / dxj
+      {
+        IntegrationPoint ipl(ip);
+        ipl(j) -= eps;
+        IntegrationPoint ipr(ip);
+        ipr(j) += eps;
+        IntegrationPoint ipll(ip);
+        ipll(j) -= 2*eps;
+        IntegrationPoint iprr(ip);
+        iprr(j) += 2*eps;
+        
+        MappedIntegrationPoint<DIM,DIMSPACE> mipl(ipl, eltrans);
+        MappedIntegrationPoint<DIM,DIMSPACE> mipr(ipr, eltrans);
+        MappedIntegrationPoint<DIM,DIMSPACE> mipll(ipll, eltrans);
+        MappedIntegrationPoint<DIM,DIMSPACE> miprr(iprr, eltrans);
+        
+        fel.EvaluateMappedShape (mipl,  x, shape_ul);
+        fel.EvaluateMappedShape (mipr,  x, shape_ur);
+        fel.EvaluateMappedShape (mipll, x, shape_ull);
+        fel.EvaluateMappedShape (miprr, x, shape_urr);
+        
+        dshape_u_ref = (1.0/(12.0*eps)) * (8.0*shape_ur-8.0*shape_ul-shape_urr+shape_ull);
+        
+        for (int l = 0; l < DIM_STRESS; l++)
+          y(j*DIM_STRESS+l) = dshape_u_ref(l);
+      }
+    
+    for (int j = 0; j < DIM_STRESS; j++)
+      {
+        for (int l = 0; l < DIM; l++)
+          dshape_u_ref_comp(l) = y(l*DIM_STRESS+j);
+        
+        dshape_u = Trans(mip.GetJacobianInverse()) * dshape_u_ref_comp;
+        
+        for (int l = 0; l < DIMSPACE; l++)
+          y(l*DIM_STRESS+j) = dshape_u(l);
+      }
+  }
+
+  template<typename FEL, int DIMSPACE, int DIM, int DIM_STRESS, class TVX, class TVY>
+  void ApplyTransDShapeFE(const FEL & fel_u, const MappedIntegrationPoint<DIM,DIMSPACE>& mip, const TVX & x, TVY & by, LocalHeap & lh, double eps = 1e-4)
+  {
+    typedef typename TVX::TSCAL TSCALX;
+
+    HeapReset hr(lh);
+    int nd_u = fel_u.GetNDof();
+    FlatMatrixFixWidth<DIM_STRESS*DIMSPACE> bmatu(nd_u,lh);
+    
+    auto y = by.Range(0, nd_u);
+    const IntegrationPoint& ip = mip.IP();
+    const ElementTransformation & eltrans = mip.GetTransformation();
+    FlatMatrixFixWidth<DIM_STRESS> shape_ul(nd_u, lh);
+    FlatMatrixFixWidth<DIM_STRESS> shape_ur(nd_u, lh);
+    FlatMatrixFixWidth<DIM_STRESS> shape_ull(nd_u, lh);
+    FlatMatrixFixWidth<DIM_STRESS> shape_urr(nd_u, lh);
+    FlatMatrixFixWidth<DIM_STRESS> dshape_u_ref(nd_u, lh);
+    FlatMatrixFixWidth<DIM_STRESS> dshape_u(nd_u, lh);
+    
+    FlatMatrix<TSCALX> hx(DIMSPACE,DIM_STRESS,&x(0));
+    Mat<DIM,DIM_STRESS,TSCALX> tx = mip.GetJacobianInverse() * hx;
+    
+    y = 0;
+    for (int j = 0; j < DIM; j++)   // d / dxj
+      {
+        IntegrationPoint ipts[4];
+        
+        ipts[0] = ip;
+        ipts[0](j) -= eps;
+        ipts[1] = ip;
+        ipts[1](j) += eps;              
+        ipts[2] = ip;
+        ipts[2](j) -= 2*eps;
+        ipts[3] = ip;
+        ipts[3](j) += 2*eps;
+        
+        IntegrationRule ir(4, ipts);
+        MappedIntegrationRule<DIM,DIMSPACE> mirl(ir, eltrans, lh);
+        
+        fel_u.CalcMappedShape (mirl[0], shape_ul);
+        fel_u.CalcMappedShape (mirl[1], shape_ur);
+        fel_u.CalcMappedShape (mirl[2], shape_ull);
+        fel_u.CalcMappedShape (mirl[3], shape_urr);
+        
+        dshape_u_ref = (1.0/(12.0*eps)) * (8.0*shape_ur-8.0*shape_ul-shape_urr+shape_ull);
+        y += dshape_u_ref * tx.Row(j);
+      }
+  }
+
+
+
+  template<typename FEL, int DIMSPACE, int DIM, int DIM_STRESS>
+  void CalcSIMDDShapeFE(const FEL & fel, const SIMD_MappedIntegrationRule<DIM,DIMSPACE>& mir, BareSliceMatrix<SIMD<double>> mat, double eps = 1e-4)
+    {
+      size_t nd_u = fel.GetNDof();
+
+      STACK_ARRAY(SIMD<double>, mem1, 2*DIM_STRESS*nd_u);
+      FlatMatrix<SIMD<double>> shape_u_tmp(nd_u*DIM_STRESS, 1, &mem1[0]);
+
+      FlatMatrix<SIMD<double>> dshape_u_ref(nd_u*DIM_STRESS, 1, &mem1[DIM_STRESS*nd_u]);
+
+      LocalHeapMem<10000> lh("diffopgrad-lh");
+
+      auto & ir = mir.IR();
+      for (size_t i = 0; i < mir.Size(); i++)
+        {
+          const SIMD<IntegrationPoint> & ip = ir[i];
+          const ElementTransformation & eltrans = mir[i].GetTransformation();
+
+          for (int j = 0; j < DIM; j++)   // d / dxj
+            {
+              HeapReset hr(lh);
+              SIMD<IntegrationPoint> ipts[4];
+              ipts[0] = ip;
+              ipts[0](j) -= eps;
+              ipts[1] = ip;
+              ipts[1](j) += eps;              
+              ipts[2] = ip;
+              ipts[2](j) -= 2*eps;
+              ipts[3] = ip;
+              ipts[3](j) += 2*eps;
+
+              SIMD_IntegrationRule ir(4, ipts);
+              SIMD_MappedIntegrationRule<DIM,DIMSPACE> mirl(ir, eltrans, lh);
+
+              fel.CalcMappedShape (mirl[2], shape_u_tmp);
+              dshape_u_ref = 1.0/(12.0*eps) * shape_u_tmp;
+              fel.CalcMappedShape (mirl[3], shape_u_tmp);
+              dshape_u_ref -= 1.0/(12.0*eps) * shape_u_tmp;
+              fel.CalcMappedShape (mirl[0], shape_u_tmp);
+              dshape_u_ref -= 8.0/(12.0*eps) * shape_u_tmp;
+              fel.CalcMappedShape (mirl[1], shape_u_tmp);
+              dshape_u_ref += 8.0/(12.0*eps) * shape_u_tmp;
+
+              // dshape_u_ref =  (8.0*shape_ur-8.0*shape_ul-shape_urr+shape_ull);
+              for (size_t l = 0; l < DIM_STRESS; l++)
+                for (size_t k = 0; k < nd_u; k++)
+                  mat(k*DIM_STRESS*DIM+j*DIM_STRESS+l, i) = dshape_u_ref(k*DIM_STRESS+l, 0);
+            }
+          
+          for (size_t j = 0; j < DIM_STRESS; j++)
+            for (size_t k = 0; k < nd_u; k++)
+              {
+                Vec<DIM,SIMD<double>> dshape_u_ref, dshape_u;
+                for (size_t l = 0; l < DIM; l++)
+                  dshape_u_ref(l) = mat(k*DIM_STRESS*DIM+l*DIM+j, i);
+                
+                dshape_u = Trans(mir[i].GetJacobianInverse()) * dshape_u_ref;
+                
+                for (size_t l = 0; l < DIMSPACE; l++)
+                  mat(k*DIM_STRESS*DIMSPACE+l*DIM_STRESS+j, i) = dshape_u(l);
+              }
+        }
+    }
+  
+  template<typename FEL, int DIMSPACE, int DIM, int DIM_STRESS>
+  void ApplySIMDDShapeFE (const FEL & fel_u, const SIMD_BaseMappedIntegrationRule & bmir,
+                          BareSliceVector<double> x, BareSliceMatrix<SIMD<double>> y, double eps = 1e-4)
+  {
+    constexpr size_t BS = 64; // number of simd-points
+    size_t maxnp = min2(BS, bmir.Size());
+    size_t size = (maxnp+1)*SIMD<double>::Size()*500  +  5*DIM_STRESS*BS*sizeof(SIMD<double>);
+    STACK_ARRAY(char, data, size);
+    LocalHeap lh(data, size);
+    
+    auto & mir = static_cast<const SIMD_MappedIntegrationRule<DIM,DIMSPACE>&> (bmir);
+    auto & ir = mir.IR();
+    const ElementTransformation & trafo = mir.GetTransformation();
+
+    for (int k = 0; k < mir.Size(); k++)
+      for (int m = 0; m < DIM_STRESS*DIMSPACE; m++)
+        y(m, k) = SIMD<double> (0.0);
+    
+    for (size_t base = 0; base < ir.Size(); base += BS)
+      {
+        HeapReset hr(lh);
+        size_t num = min2(BS, ir.Size()-base);
+        
+        FlatMatrix<SIMD<double>> hxl(DIM_STRESS, num, lh);
+        FlatMatrix<SIMD<double>> hxr(DIM_STRESS, num, lh);
+        FlatMatrix<SIMD<double>> hxll(DIM_STRESS, num, lh);
+        FlatMatrix<SIMD<double>> hxrr(DIM_STRESS, num, lh);
+        FlatMatrix<SIMD<double>> hx(DIM_STRESS, num, lh);
+        
+        for (int j = 0; j < DIM; j++)
+          {
+            // hx = (F^-1 * x).Row(j)
+            {
+              HeapReset hr(lh);
+              SIMD_IntegrationRule irl(num*SIMD<double>::Size(), lh);
+              for (int k = 0; k < irl.Size(); k++)
+                {
+                  irl[k] = ir[base+k];
+                  irl[k](j) -= eps;
+                }
+              SIMD_MappedIntegrationRule<DIM,DIMSPACE> mirl(irl, trafo, lh);
+              fel_u.Evaluate (mirl, x, hxl);
+            }
+            {
+              HeapReset hr(lh);
+              SIMD_IntegrationRule irr(num*SIMD<double>::Size(), lh);
+              for (int k = 0; k < irr.Size(); k++)
+                {
+                  irr[k] = ir[base+k];              
+                  irr[k](j) += eps;
+                }
+              SIMD_MappedIntegrationRule<DIM,DIMSPACE> mirr(irr, trafo, lh);
+              fel_u.Evaluate (mirr, x, hxr);
+            }
+            {
+              HeapReset hr(lh);
+              SIMD_IntegrationRule irll(num*SIMD<double>::Size(), lh);
+              for (int k = 0; k < irll.Size(); k++)
+                {
+                  irll[k] = ir[base+k];
+                  irll[k](j) -= 2*eps;
+                }
+              SIMD_MappedIntegrationRule<DIM,DIMSPACE> mirll(irll, trafo, lh);
+              fel_u.Evaluate (mirll, x, hxll);
+            }
+            {
+              HeapReset hr(lh);
+              SIMD_IntegrationRule irrr(num*SIMD<double>::Size(), lh);
+              for (int k = 0; k < irrr.Size(); k++)
+                {
+                  irrr[k] = ir[base+k];              
+                  irrr[k](j) += 2*eps;
+                }
+              SIMD_MappedIntegrationRule<DIM,DIMSPACE> mirrr(irrr, trafo, lh);
+              fel_u.Evaluate (mirrr, x, hxrr);
+            }
+            // hx = 1.0/(2*eps) * (hxr-hxl);
+            hx = 1.0/(12*eps) * (8*hxr-8*hxl-hxrr+hxll);
+            for (int k = 0; k < num; k++)
+              {
+                auto jacinv = mir[base+k].GetJacobianInverse();
+                for (int l = 0; l < DIM_STRESS; l++)
+                  {
+                    for (int m = 0; m < DIMSPACE; m++)
+                      y(m*DIM_STRESS+l, base+k) += jacinv(j,m) * hx(l, k);
+                  }
+              }
+          }
+      }
+  }
+
+
+
+
+  template<typename FEL, int DIMSPACE, int DIM, int DIM_STRESS>
+  void AddTransSIMDDShapeFE (const FEL & fel_u, const SIMD_BaseMappedIntegrationRule & bmir,
+                             BareSliceMatrix<SIMD<double>> x, BareSliceVector<double> y, double eps = 1e-4)
+  {
+    constexpr size_t BS = 64; // number of simd-points
+    size_t maxnp = min2(BS, bmir.Size());
+    size_t size = (maxnp+1)*SIMD<double>::Size()*500;
+    
+    STACK_ARRAY(char, data, size);
+    LocalHeap lh(data, size);
+    
+    auto & mir = static_cast<const SIMD_MappedIntegrationRule<DIM,DIMSPACE>&> (bmir);
+    auto & ir = mir.IR();
+    const ElementTransformation & trafo = mir.GetTransformation();
+    
+    for (size_t base = 0; base < ir.Size(); base += BS)
+      {
+        HeapReset hr(lh);
+        size_t num = min2(BS, ir.Size()-base);
+        
+        FlatMatrix<SIMD<double>> hx1(DIM_STRESS, num, lh);
+        FlatMatrix<SIMD<double>> hx2(DIM_STRESS, num, lh);
+        
+        for (size_t j = 0; j < DIM; j++)
+          {
+            // hx = (F^-1 * x).Row(j)
+            for (size_t k = 0; k < num; k++)
+              {
+                auto jacinv = mir[base+k].GetJacobianInverse();
+                for (int l = 0; l < DIM_STRESS; l++)
+                  {
+                    SIMD<double> sum = 0;
+                    for (int m = 0; m < DIMSPACE; m++)
+                      sum += jacinv(j,m) * x(m*DIM_STRESS+l, base+k);
+                    
+                    hx1(l,k) = (-(8/(12*eps)) * sum).Data();
+                    hx2(l,k) = ( (1/(12*eps)) * sum).Data();
+                  }
+              }
+            
+            {
+              HeapReset hr(lh);
+              SIMD_IntegrationRule irl(num*SIMD<double>::Size(), lh);
+              for (size_t k = 0; k < irl.Size(); k++)
+                {
+                  irl[k] = ir[base+k];
+                  irl[k](j) -= eps;
+                }
+              SIMD_MappedIntegrationRule<DIM,DIMSPACE> mirl(irl, trafo, lh);
+              fel_u.AddTrans (mirl, hx1, y);
+              irl.NothingToDelete();
+            }
+            {
+              HeapReset hr(lh);
+              hx1 *= -1;
+              SIMD_IntegrationRule irr(num*SIMD<double>::Size(), lh);
+              for (int k = 0; k < irr.Size(); k++)
+                {
+                  irr[k] = ir[base+k];              
+                  irr[k](j) += eps;
+                }
+              SIMD_MappedIntegrationRule<DIM,DIMSPACE> mirr(irr, trafo, lh);
+              fel_u.AddTrans (mirr, hx1, y);
+            }
+            {
+              HeapReset hr(lh);
+              SIMD_IntegrationRule irl(num*SIMD<double>::Size(), lh);
+              for (int k = 0; k < irl.Size(); k++)
+                {
+                  irl[k] = ir[base+k];
+                  irl[k](j) -= 2*eps;
+                }
+              SIMD_MappedIntegrationRule<DIM,DIMSPACE> mirl(irl, trafo, lh);
+              fel_u.AddTrans (mirl, hx2, y);
+            }
+            {
+              HeapReset hr(lh);
+              hx2 *= -1;
+              SIMD_IntegrationRule irr(num*SIMD<double>::Size(), lh);
+              for (int k = 0; k < irr.Size(); k++)
+                {
+                  irr[k] = ir[base+k];              
+                  irr[k](j) += 2*eps;
+                }
+              SIMD_MappedIntegrationRule<DIM,DIMSPACE> mirr(irr, trafo, lh);
+              fel_u.AddTrans (mirr, hx2, y);
+            }
+          }
+      }
+  }
+
+
+
+  template <typename ME, typename ORIG, typename FEL>
+  class NumDiffGradient : public DiffOp<ME>
+  {
+    typedef DiffOp<ME> BASE;
+  public:
+    static constexpr double eps() { return 1e-4; }
+
+    static constexpr int DIM = ORIG::DIM;    
+    static constexpr int DIM_SPACE = ORIG::DIM_SPACE;
+    static constexpr int DIM_ELEMENT = ORIG::DIM_ELEMENT;
+    static constexpr int DIM_DMAT = ORIG::DIM_DMAT*ORIG::DIM_SPACE;
+    static constexpr int DIFFORDER = ORIG::DIFFORDER+1;
+
+
+    template <typename AFEL, typename SIP, typename MAT,
+              typename std::enable_if<!std::is_convertible<MAT,BareSliceMatrix<double,ColMajor>>::value, int>::type = 0>
+    static void GenerateMatrix (const AFEL & fel, const SIP & sip,
+                                MAT & mat, LocalHeap & lh)
+    {
+      cout << "nicht gut" << endl;
+      cout << "type(fel) = " << typeid(fel).name() << ", sip = " << typeid(sip).name()
+           << ", mat = " << typeid(mat).name() << endl;
+    }
+
+
+    static int DimRef() { return DIM_DMAT + ORIG::DIM_DMAT; }   // grad(u), u
+    
+    template <typename IP, typename MAT>
+    static void GenerateMatrixRef (const FiniteElement & fel, const IP & ip,
+                                   MAT && mat, LocalHeap & lh)
+    {
+      int nd_u = fel.GetNDof();
+
+      FlatMatrixFixWidth<ORIG::DIM_DMAT> shape_ul(nd_u, lh);
+      FlatMatrixFixWidth<ORIG::DIM_DMAT> shape_ur(nd_u, lh);
+      FlatMatrixFixWidth<ORIG::DIM_DMAT> shape_ull(nd_u, lh);
+      FlatMatrixFixWidth<ORIG::DIM_DMAT> shape_urr(nd_u, lh);
+      FlatMatrixFixWidth<ORIG::DIM_DMAT> dshape_u_ref(nd_u, lh);
+      
+      FlatMatrixFixWidth<DIM_ELEMENT> dshape_u_ref_comp(nd_u, lh);
+      FlatMatrixFixWidth<DIM_SPACE> dshape_u(nd_u, lh);
+      
+      for (int j = 0; j < DIM_ELEMENT; j++)   // d / dxj
+        {
+          IntegrationPoint ipl(ip);
+          ipl(j) -= eps();
+          IntegrationPoint ipr(ip);
+          ipr(j) += eps();
+          IntegrationPoint ipll(ip);
+          ipll(j) -= 2*eps();
+          IntegrationPoint iprr(ip);
+          iprr(j) += 2*eps();
+          
+
+          ORIG::GenerateMatrixRef (fel, ipl, Trans(shape_ul), lh);
+          ORIG::GenerateMatrixRef (fel, ipr, Trans(shape_ur), lh);
+          ORIG::GenerateMatrixRef (fel, ipll, Trans(shape_ull), lh);
+          ORIG::GenerateMatrixRef (fel, iprr, Trans(shape_urr), lh);
+          
+          dshape_u_ref = (1.0/(12.0*eps())) * (8.0*shape_ur-8.0*shape_ul-shape_urr+shape_ull);
+          for (int l = 0; l < ORIG::DIM_DMAT; l++)
+            mat.Row(j*ORIG::DIM_DMAT+l) = dshape_u_ref.Col(l);
+        }
+
+      ORIG::GenerateMatrixRef (fel, ip, mat.Rows(DIM_DMAT, DIM_DMAT+ORIG::DIM_DMAT), lh);
+    }
+
+    
+
+    template <typename MIP, typename MAT>
+    static void CalcTransformationMatrix (const MIP & mip,
+                                          MAT & mat, LocalHeap & lh)
+    {
+      auto matgrad = Trans(static_cast<const MappedIntegrationPoint<DIM_SPACE,DIM_SPACE>&>(mip).GetJacobianInverse());
+      Mat<DIM_SPACE,DIM_SPACE> matshape;
+      ORIG::CalcTransformationMatrix (mip, matshape, lh);
+
+      // input: du0/dx0, du1/dx0, du2/dx0,  d0/dx1, ....
+      // output du0/dx0, du1/dx0, du2/dx0 ... 
+      for (int i = 0; i < DIM_SPACE; i++)
+        for (int j = 0; j < DIM_SPACE; j++)
+          for (int k = 0; k < DIM_SPACE; k++)
+            for (int l = 0; l < DIM_SPACE; l++)
+              mat(k*DIM_SPACE+i, l*DIM_SPACE+j) = matshape(i,j) * matgrad(k,l);
+
+
+      // numerical diff of ORIG::CalcTransformationMatrix (mip, matshape, lh);
+
+      for (int i = 0; i < DIM_SPACE; i++)
+        for (int j = 0; j < DIM_SPACE; j++)
+          for (int k = 0; k < DIM_SPACE; k++)
+            mat(k*DIM_SPACE+i, DIM_SPACE*DIM_SPACE+j) = 0.0;
+      
+      Mat<DIM_SPACE,DIM_SPACE> dmatshape_dxj;
+      const IntegrationPoint& ip = mip.IP();
+      const ElementTransformation & eltrans = mip.GetTransformation();      
+      for (int j = 0; j < DIM_ELEMENT; j++)   // d / dxj
+        {
+          IntegrationPoint ipl(ip);
+          ipl(j) -= eps();
+          IntegrationPoint ipr(ip);
+          ipr(j) += eps();
+          IntegrationPoint ipll(ip);
+          ipll(j) -= 2*eps();
+          IntegrationPoint iprr(ip);
+          iprr(j) += 2*eps();
+          
+          MappedIntegrationPoint<DIM_ELEMENT,DIM_SPACE> mipl(ipl, eltrans);
+          MappedIntegrationPoint<DIM_ELEMENT,DIM_SPACE> mipr(ipr, eltrans);
+          MappedIntegrationPoint<DIM_ELEMENT,DIM_SPACE> mipll(ipll, eltrans);
+          MappedIntegrationPoint<DIM_ELEMENT,DIM_SPACE> miprr(iprr, eltrans);
+
+          Mat<DIM_SPACE,DIM_SPACE> dml, dmll, dmr, dmrr;
+          ORIG::CalcTransformationMatrix (mipl, dml, lh);          
+          ORIG::CalcTransformationMatrix (mipr, dmr, lh);          
+          ORIG::CalcTransformationMatrix (mipll, dmll, lh);          
+          ORIG::CalcTransformationMatrix (miprr, dmrr, lh);
+
+          dmatshape_dxj = (1.0/(12.0*eps())) * (8.0*dmr-8.0*dml-dmrr+dmll);
+
+          for (int i = 0; i < DIM_SPACE; i++)
+            for (int l = 0; l < DIM_SPACE; l++)
+              for (int k = 0; k < DIM_SPACE; k++)
+                mat(k*DIM_SPACE+i, DIM_SPACE*DIM_SPACE+l) += matgrad(k,j) * dmatshape_dxj(i,l);
+        }
+
+      
+    }
+
+
+    
+
+    template <typename AFEL, typename MIP, typename MAT,
+              typename std::enable_if<std::is_convertible<MAT,BareSliceMatrix<double,ColMajor>>::value, int>::type = 0>
+    static void GenerateMatrix (const AFEL & bfel, const MIP & mip,
+                                MAT mat, LocalHeap & lh)
+    {
+      // CalcDShapeFE<FEL,DIM_SPACE,DIM_ELEMENT,ORIG::DIM_DMAT>
+      // (static_cast<const FEL&>(fel), mip, Trans(mat), lh, eps());
+
+      HeapReset hr(lh);
+      
+      int nd_u = bfel.GetNDof();
+      const IntegrationPoint& ip = mip.IP(); 
+      const ElementTransformation & eltrans = mip.GetTransformation();
+      FlatMatrixFixWidth<ORIG::DIM_DMAT> shape_ul(nd_u, lh);
+      FlatMatrixFixWidth<ORIG::DIM_DMAT> shape_ur(nd_u, lh);
+      FlatMatrixFixWidth<ORIG::DIM_DMAT> shape_ull(nd_u, lh);
+      FlatMatrixFixWidth<ORIG::DIM_DMAT> shape_urr(nd_u, lh);
+      FlatMatrixFixWidth<ORIG::DIM_DMAT> dshape_u_ref(nd_u, lh);
+      
+      FlatMatrixFixWidth<DIM_ELEMENT> dshape_u_ref_comp(nd_u, lh);
+      FlatMatrixFixWidth<DIM_SPACE> dshape_u(nd_u, lh);
+      
+      for (int j = 0; j < DIM_ELEMENT; j++)   // d / dxj
+        {
+          IntegrationPoint ipl(ip);
+          ipl(j) -= eps();
+          IntegrationPoint ipr(ip);
+          ipr(j) += eps();
+          IntegrationPoint ipll(ip);
+          ipll(j) -= 2*eps();
+          IntegrationPoint iprr(ip);
+          iprr(j) += 2*eps();
+          
+          MappedIntegrationPoint<DIM_ELEMENT,DIM_SPACE> mipl(ipl, eltrans);
+          MappedIntegrationPoint<DIM_ELEMENT,DIM_SPACE> mipr(ipr, eltrans);
+          MappedIntegrationPoint<DIM_ELEMENT,DIM_SPACE> mipll(ipll, eltrans);
+          MappedIntegrationPoint<DIM_ELEMENT,DIM_SPACE> miprr(iprr, eltrans);
+
+          ORIG::GenerateMatrix  (bfel, mipl, Trans(shape_ul), lh);
+          ORIG::GenerateMatrix  (bfel, mipr, Trans(shape_ur), lh);
+          ORIG::GenerateMatrix  (bfel, mipll, Trans(shape_ull), lh);
+          ORIG::GenerateMatrix  (bfel, miprr, Trans(shape_urr), lh);
+          
+          dshape_u_ref = (1.0/(12.0*eps())) * (8.0*shape_ur-8.0*shape_ul-shape_urr+shape_ull);
+          for (int l = 0; l < ORIG::DIM_DMAT; l++)
+            mat.Row(j*ORIG::DIM_DMAT+l) = dshape_u_ref.Col(l);
+        }
+      
+      for (int j = 0; j < ORIG::DIM_DMAT; j++)
+        {
+          for (int k = 0; k < nd_u; k++)
+            for (int l = 0; l < DIM_ELEMENT; l++)
+              dshape_u_ref_comp(k,l) = mat(l*ORIG::DIM_DMAT+j, k);
+          
+          dshape_u = dshape_u_ref_comp * mip.GetJacobianInverse();
+          
+          for (int k = 0; k < nd_u; k++)
+            for (int l = 0; l < DIM_SPACE; l++)
+              mat(l*ORIG::DIM_DMAT+j, k) = dshape_u(k,l);
+        }
+    }
+
+
+    
+    /*
+    template <typename AFEL, typename MIP, typename MAT,
+              typename std::enable_if<std::is_convertible<MAT,SliceMatrix<double,ColMajor>>::value, int>::type = 0>
+    static void GenerateMatrix (const AFEL & fel, const MIP & mip,
+                                MAT mat, LocalHeap & lh)
+    {
+      CalcDShapeFE<FEL,DIM_SPACE,DIM_ELEMENT,ORIG::DIM_DMAT>
+        (static_cast<const FEL&>(fel), mip, Trans(mat), lh, eps());
+    }
+
+    template <typename AFEL, typename MIP, class TVX, class TVY>
+    static void Apply (const AFEL & fel, const MIP & mip,
+                       const TVX & x, TVY && y,
+                       LocalHeap & lh) 
+    {
+      HeapReset hr(lh);
+      FlatMatrixFixWidth<DIM_SPACE*ORIG::DIM_DMAT> hm(fel.GetNDof(),lh);
+      CalcDShapeFE<FEL,DIM_SPACE,DIM_ELEMENT,ORIG::DIM_DMAT>
+        (static_cast<const FEL&>(fel), mip, hm, lh, eps());
+      y = Trans(hm)*x;
+    }
+
+    template <typename AFEL, typename MIP, class TVX, class TVY>
+    static void ApplyTrans (const AFEL & fel, const MIP & mip,
+			    const TVX & x, TVY & by,
+			    LocalHeap & lh) 
+    {
+      ApplyTransDShapeFE<FEL,DIM_SPACE,DIM_ELEMENT,ORIG::DIM_DMAT>
+        (static_cast<const FEL&>(fel), mip, x, by, lh, eps());
+    }
+    */
+
+    
+    static void GenerateMatrixSIMDIR (const FiniteElement & bfel,
+                                      const SIMD_BaseMappedIntegrationRule & bmir, BareSliceMatrix<SIMD<double>> mat)
+    {
+      /*
+      CalcSIMDDShapeFE<FEL,DIM_SPACE,DIM_ELEMENT,ORIG::DIM_DMAT>
+        (static_cast<const FEL&>(bfel), static_cast<const SIMD_MappedIntegrationRule<DIM_ELEMENT,DIM_SPACE> &>(bmir), mat, eps());
+      return;
+      */
+      
+      auto & mir = static_cast<const SIMD_MappedIntegrationRule<DIM_ELEMENT,DIM_SPACE>&> (bmir);      
+      // auto & fel_u = static_cast<const FEL&>(bfel);
+      size_t nd_u = bfel.GetNDof();
+
+      STACK_ARRAY(SIMD<double>, mem1, 5*ORIG::DIM_DMAT*nd_u);
+      FlatMatrix<SIMD<double>> shape_u_tmp(nd_u*ORIG::DIM_DMAT, 4, &mem1[0]);
+      FlatMatrix<SIMD<double>> dshape_u_ref(nd_u*ORIG::DIM_DMAT, 1, &mem1[4*ORIG::DIM_DMAT*nd_u]);
+
+      LocalHeapMem<10000> lh("diffopgrad-lh");
+
+      auto & ir = mir.IR();
+      for (size_t i = 0; i < mir.Size(); i++)
+        {
+          const SIMD<IntegrationPoint> & ip = ir[i];
+          const ElementTransformation & eltrans = mir[i].GetTransformation();
+
+          double dist[] = { 1, -1, 2, -2 };
+          double weight[] = { 8/12., -8/12., -1/12., 1/12. };
+          
+          for (int j = 0; j < DIM_ELEMENT; j++)   // d / dxj
+            {
+              HeapReset hr(lh);
+              SIMD<IntegrationPoint> ipts[4];
+              for (int i = 0; i < 4; i++)
+                {
+                  ipts[i] = ip;
+                  ipts[i](j) += dist[i]*eps();                  
+                }
+              SIMD_IntegrationRule ir(4, ipts);
+              SIMD_MappedIntegrationRule<DIM_ELEMENT,DIM_SPACE> mirl(ir, eltrans, lh);
+              ORIG::GenerateMatrixSIMDIR (bfel, mirl, shape_u_tmp);
+
+              dshape_u_ref.Col(0) =
+                weight[0]/eps() * shape_u_tmp.Col(0) +
+                weight[1]/eps() * shape_u_tmp.Col(1) +
+                weight[2]/eps() * shape_u_tmp.Col(2) +
+                weight[3]/eps() * shape_u_tmp.Col(3);
+              
+              for (size_t l = 0; l < ORIG::DIM_DMAT; l++)
+                for (size_t k = 0; k < nd_u; k++)
+                  mat(k*ORIG::DIM_DMAT*DIM_ELEMENT+j*ORIG::DIM_DMAT+l, i) = dshape_u_ref(k*ORIG::DIM_DMAT+l, 0);
+            }
+          
+          for (size_t j = 0; j < ORIG::DIM_DMAT; j++)
+            for (size_t k = 0; k < nd_u; k++)
+              {
+                Vec<DIM_ELEMENT,SIMD<double>> dshape_u_ref;
+                Vec<DIM_SPACE,SIMD<double>> dshape_u;
+                
+                for (size_t l = 0; l < DIM_ELEMENT; l++)
+                  dshape_u_ref(l) = mat(k*ORIG::DIM_DMAT*DIM_ELEMENT+l*DIM_ELEMENT+j, i);
+                
+                dshape_u = Trans(mir[i].GetJacobianInverse()) * dshape_u_ref;
+                
+                for (size_t l = 0; l < DIM_SPACE; l++)
+                  mat(k*ORIG::DIM_DMAT*DIM_SPACE+l*ORIG::DIM_DMAT+j, i) = dshape_u(l);
+              }
+        }
+    }
+
+
+
+    using BASE::ApplySIMDIR;
+    static void ApplySIMDIR (const FiniteElement & fel, const SIMD_BaseMappedIntegrationRule & bmir,
+                             BareSliceVector<double> x, BareSliceMatrix<SIMD<double>> y)
+    {
+      // ApplySIMDDShapeFE<FEL,ORIG::DIM_SPACE,ORIG::DIM_ELEMENT,ORIG::DIM_DMAT>(static_cast<const FEL&>(fel), bmir, x, y, eps());
+      // return;
+      
+      constexpr size_t BS = 16; // number of simd-points
+      size_t maxnp = min2(BS, bmir.Size());
+      size_t size = (maxnp+1)*SIMD<double>::Size()*500  +  5*ORIG::DIM_DMAT*BS*sizeof(SIMD<double>);
+      size *= 4;
+      STACK_ARRAY(char, data, size);
+      LocalHeap lh(data, size);
+      
+      auto & mir = static_cast<const SIMD_MappedIntegrationRule<DIM_ELEMENT,DIM_SPACE>&> (bmir);
+      auto & ir = mir.IR();
+      const ElementTransformation & trafo = mir.GetTransformation();
+      
+      y.AddSize(ORIG::DIM_DMAT*DIM_SPACE, mir.Size()) = SIMD<double>(0.);
+      for (size_t base = 0; base < ir.Size(); base += BS)
+        {
+          HeapReset hr(lh);
+          size_t num = std::min(BS, ir.Size()-base);
+          FlatMatrix<SIMD<double>> hxi(ORIG::DIM_DMAT, 4*num, lh);          
+
+          static constexpr double dist[] = { 1, -1, 2, -2 };
+          static constexpr double weight[] = { 8/12., -8/12., -1/12., 1/12. };
+          
+          for (int j = 0; j < DIM_ELEMENT; j++)
+            {
+              HeapReset hr(lh);
+
+              SIMD_IntegrationRule iri(4*num*SIMD<double>::Size(), lh);
+              
+              for (int k = 0; k < num; k++)
+                for (int i = 0; i < 4; i++)
+                  {
+                    iri[4*k+i] = ir[base+k];
+                    iri[4*k+i](j) += dist[i]*eps();
+                  }
+
+              SIMD_MappedIntegrationRule<DIM_ELEMENT,DIM_SPACE> mir2(iri, trafo, lh);
+              ORIG::ApplySIMDIR (fel, mir2, x, hxi);
+              
+              for (int k = 0; k < num; k++)
+                {
+                  Vec<ORIG::DIM_DMAT, SIMD<double>> hxk =
+                    weight[0]/eps() * hxi.Col(4*k) +
+                    weight[1]/eps() * hxi.Col(4*k+1) +
+                    weight[2]/eps() * hxi.Col(4*k+2) +
+                    weight[3]/eps() * hxi.Col(4*k+3);
+                    
+                  auto jacinv = mir[base+k].GetJacobianInverse();
+                  for (int l = 0; l < ORIG::DIM_DMAT; l++)
+                    for (int m = 0; m < DIM_SPACE; m++)
+                      y(m*ORIG::DIM_DMAT+l, base+k) += jacinv(j,m) * hxk(l); // , k);
+                }
+            }
+        }
+    }
+
+
+
+    
+    using BASE::AddTransSIMDIR;    
+    static void AddTransSIMDIR (const FiniteElement & fel, const SIMD_BaseMappedIntegrationRule & bmir,
+                                BareSliceMatrix<SIMD<double>> x, BareSliceVector<double> y)
+    {
+      // AddTransSIMDDShapeFE<FEL,DIM_SPACE,DIM_ELEMENT,ORIG::DIM_DMAT>(static_cast<const FEL&>(fel), bmir, x, y, eps());
+      // return;
+      
+      // auto & fel_u = static_cast<const FEL&>(fel);
+      constexpr size_t BS = 64; // number of simd-points
+      size_t maxnp = std::min(BS, bmir.Size());
+      size_t size = (maxnp+1)*SIMD<double>::Size()*500;
+    
+      STACK_ARRAY(char, data, size);
+      LocalHeap lh(data, size);
+      
+      auto & mir = static_cast<const SIMD_MappedIntegrationRule<DIM_ELEMENT,DIM_SPACE>&> (bmir);
+      auto & ir = mir.IR();
+      const ElementTransformation & trafo = mir.GetTransformation();
+      
+      for (size_t base = 0; base < ir.Size(); base += BS)
+        {
+          HeapReset hr(lh);
+          size_t num = min2(BS, ir.Size()-base);
+          
+          FlatMatrix<SIMD<double>> hx1(ORIG::DIM_DMAT, num, lh);
+          FlatMatrix<SIMD<double>> hx(ORIG::DIM_DMAT, num, lh);
+          
+          for (size_t j = 0; j < DIM_ELEMENT; j++)
+            {
+              // hx = (F^-1 * x).Row(j)
+              for (size_t k = 0; k < num; k++)
+                {
+                  auto jacinv = mir[base+k].GetJacobianInverse();
+                  for (int l = 0; l < ORIG::DIM_DMAT; l++)
+                    {
+                      SIMD<double> sum = 0;
+                      for (int m = 0; m < DIM_SPACE; m++)
+                        sum += jacinv(j,m) * x(m*ORIG::DIM_DMAT+l, base+k);
+                      hx(l,k) = sum;
+                    }
+                }
+              
+              double dist[] = { 1, -1, 2, -2 };
+              double weight[] = { 8/12., -8/12., -1/12., 1/12. };
+              for (int i = 0; i < 4; i++)
+                {
+                  HeapReset hr(lh);
+                  SIMD_IntegrationRule irl(num*SIMD<double>::Size(), lh);
+                  for (size_t k = 0; k < irl.Size(); k++)
+                    {
+                      irl[k] = ir[base+k];
+                      irl[k](j) += dist[i]*eps();
+                    }
+                  SIMD_MappedIntegrationRule<DIM_ELEMENT,DIM_SPACE> mirl(irl, trafo, lh);
+                  hx1 = weight[i]/eps() * hx;
+                  // fel_u.AddTrans (mirl, hx1, y);
+                  ORIG::AddTransSIMDIR(fel, mirl, hx1, y);
+                  irl.NothingToDelete();
+                }
+            }
+        }
+    }
+  };
+}
+
+#endif