Method AddTransSIMDIR

comp/l2hofespace.cpp:3194–3227

    using DiffOp<DiffOpIdVectorL2Piola<DIM_SPC,VB>>::AddTransSIMDIR;
    static void AddTransSIMDIR (const FiniteElement & bfel, const SIMD_BaseMappedIntegrationRule & bmir,
                                BareSliceMatrix<SIMD<double>> y, BareSliceVector<double> x)
    {
      //static Timer t("DiffpIdVectorL2Piola::AddTransSIMD");
      // static Timer tc("DiffpIdVectorL2Piola::AddTransSIMD calc");
      // RegionTracer rt(TaskManager::GetThreadId(), t);

      auto & mir = static_cast<const SIMD_MappedIntegrationRule<DIM_ELEMENT,DIM_SPC>&> (bmir);
      auto & fel = static_cast<const VectorFiniteElement&> (bfel);
      auto & feli = static_cast<const BaseScalarFiniteElement&> (fel[0]);
      size_t ndofi = feli.GetNDof();

      STACK_ARRAY(SIMD<double>, mempt, mir.Size()*DIM_SPACE);
      FlatMatrix<SIMD<double>> hy(DIM_SPACE, mir.Size(), &mempt[0]);

      for (size_t i = 0; i < mir.Size(); i++)
        {
          auto jac = mir[i].GetJacobian();
          Vec<DIM_SPACE,SIMD<double>> val = y.Col(i);
          val *= 1/mir[i].GetJacobiDet();
          hy.Col(i).Range(0,DIM_SPACE) = Trans(jac) * val;
        }

      STACK_ARRAY(double, memx, DIM_ELEMENT*ndofi);
      FlatMatrixFixWidth<DIM_ELEMENT, double> matx(ndofi, &memx[0]);

      for (size_t i = 0; i < ndofi; i++)
        matx.Row(i) = x.Slice(i, ndofi);

      feli.AddTrans(mir.IR(), hy, matx);

      for (size_t k = 0; k < DIM_ELEMENT; k++)
        x.Range(k*ndofi, (k+1)*ndofi) = matx.Col(k);
    }


/*