From: Jean-Paul Pelteret Date: Sat, 24 Oct 2020 19:51:46 +0000 (+0200) Subject: Update step-44: Further optimise assembly loop X-Git-Tag: v9.3.0-rc1~960^2 X-Git-Url: https://gitweb.dealii.org/cgi-bin/gitweb.cgi?a=commitdiff_plain;h=refs%2Fpull%2F11099%2Fhead;p=dealii.git Update step-44: Further optimise assembly loop --- diff --git a/examples/step-44/step-44.cc b/examples/step-44/step-44.cc index 33943bd98e..dfeeb33205 100644 --- a/examples/step-44/step-44.cc +++ b/examples/step-44/step-44.cc @@ -841,8 +841,8 @@ namespace Step44 void determine_component_extractors(); - // Apply Dirichlet boundary conditions on the displacement field - void make_constraints(const int &it_nr); + // Create Dirichlet constraints for the incremental displacement field: + void make_constraints(const int it_nr); // Several functions to assemble the system and right hand side matrices // using multithreading. Each of them comes as a wrapper function, one @@ -856,8 +856,6 @@ namespace Step44 ScratchData_ASM & scratch, PerTaskData_ASM & data) const; - void copy_local_to_global_system(const PerTaskData_ASM &data); - // And similar to perform global static condensation: void assemble_sc(); @@ -1699,7 +1697,7 @@ namespace Step44 // Although for this particular problem we could potentially construct the // RHS vector before assembling the system matrix, for the sake of // extensibility we choose not to do so. The benefit to assembling the RHS - // vector and system matrix seperately is that latter is an expensive + // vector and system matrix separately is that the latter is an expensive // operation and we can potentially avoid an extra assembly process by not // assembling the tangent matrix when convergence is attained. However, this // makes parallelizing the code using MPI more difficult. Furthermore, when @@ -1961,12 +1959,11 @@ namespace Step44 // Since we use TBB for assembly, we simply setup a copy of the // data structures required for the process and pass them, along - // with the memory addresses of the assembly functions to the - // WorkStream object for processing. Note that we must ensure that - // the matrix and RHS vector are reset before any assembly operations can - // occur. Furthermore, since we are describing a problem with Neumann BCs, we - // will need the face normals and so must specify this in the face update - // flags. + // with the assembly functions to the WorkStream object for processing. Note + // that we must ensure that the matrix and RHS vector are reset before any + // assembly operations can occur. Furthermore, since we are describing a + // problem with Neumann BCs, we will need the face normals and so must specify + // this in the face update flags. template void Solid::assemble_system() { @@ -1994,7 +1991,11 @@ namespace Step44 this->assemble_system_one_cell(cell, scratch, data); }, [this](const PerTaskData_ASM &data) { - this->copy_local_to_global_system(data); + this->constraints.distribute_local_to_global(data.cell_matrix, + data.cell_rhs, + data.local_dof_indices, + tangent_matrix, + system_rhs); }, scratch_data, per_task_data); @@ -2002,20 +2003,6 @@ namespace Step44 timer.leave_subsection(); } - // This function adds the local contribution to the system matrix. - // Note that we choose not to use the constraint matrix to do the - // job for us because the tangent matrix and residual processes have - // been split up into two separate functions. - template - void Solid::copy_local_to_global_system(const PerTaskData_ASM &data) - { - constraints.distribute_local_to_global(data.cell_matrix, - data.cell_rhs, - data.local_dof_indices, - tangent_matrix, - system_rhs); - } - // Of course, we still have to define how we assemble the tangent matrix // contribution for a single cell. We first need to reset and initialize some // of the scratch data structures and retrieve some basic information @@ -2092,8 +2079,13 @@ namespace Step44 const SymmetricTensor<2, dim> &I = Physics::Elasticity::StandardTensors::I; + // These two tensors store some precomputed data. Their use will + // explained shortly. + SymmetricTensor<2, dim> symm_grad_Nx_i_x_Jc; + Tensor<1, dim> grad_Nx_i_comp_i_x_tau; + // Next we define some aliases to make the assembly process easier to - // follow + // follow. const std::vector & N = scratch.Nx[q_point]; const std::vector> &symm_grad_Nx = scratch.symm_grad_Nx[q_point]; @@ -2120,7 +2112,35 @@ namespace Step44 else Assert(i_group <= J_dof, ExcInternalError()); - // Next comes the tangent matrix contributions: + // Before we go into the inner loop, we have one final chance to + // introduce some optimizations. We've already taken into account + // the symmetry of the system, and we can now precompute some + // common terms that are repeatedly applied in the inner loop. + // We won't be excessive here, but will rather focus on expensive + // operations, namely those involving the rank-4 material stiffness + // tensor and the rank-2 stress tensor. + // + // What we may observe is that both of these tensors are contracted + // with shape function gradients indexed on the "i" DoF. This + // implies that this particular operation remains constant as we + // loop over the "j" DoF. For that reason, we can extract this from + // the inner loop and save the many operations that, for each + // quadrature point and DoF index "i" and repeated over index "j" + // are required to double contract a rank-2 symmetric tensor with a + // rank-4 symmetric tensor, and a rank-1 tensor with a rank-2 + // tensor. + // + // At the loss of some readability, this small change will reduce + // the assembly time of the symmetrized system by about half when + // using the simulation default parameters, and becomes more + // significant as the h-refinement level increases. + if (i_group == u_dof) + { + symm_grad_Nx_i_x_Jc = symm_grad_Nx[i] * Jc; + grad_Nx_i_comp_i_x_tau = grad_Nx[i][component_i] * tau_ns; + } + + // Now we're prepared to compute the tangent matrix contributions: for (const unsigned int j : scratch.fe_values.dof_indices_ending_at(i)) { @@ -2136,14 +2156,13 @@ namespace Step44 if ((i_group == j_group) && (i_group == u_dof)) { // The material contribution: - data.cell_matrix(i, j) += symm_grad_Nx[i] * Jc * // + data.cell_matrix(i, j) += symm_grad_Nx_i_x_Jc * // symm_grad_Nx[j] * JxW; // // The geometrical stress contribution: if (component_i == component_j) - data.cell_matrix(i, j) += grad_Nx[i][component_i] * - tau_ns * - grad_Nx[j][component_j] * JxW; + data.cell_matrix(i, j) += + grad_Nx_i_comp_i_x_tau * grad_Nx[j][component_j] * JxW; } // Next is the $\mathsf{\mathbf{k}}_{ \widetilde{p} u}$ // contribution @@ -2236,7 +2255,7 @@ namespace Step44 // additional contributions are to be made since the constraints // are already exactly satisfied. template - void Solid::make_constraints(const int &it_nr) + void Solid::make_constraints(const int it_nr) { std::cout << " CST " << std::flush;