AlexanderCicchino
commented
1 year ago For the overintegration curvilinear tests, it runs out of memory on the p=20 solve after doing the scaling. The last major memory issue for curvilinear overintegration comes from creating the MappingQGeneric object and fe value for get_position_vector in high_order_mesh. An issue after this PR should be raised to fix this by having an option to extract the volume_nodes vector from the mesh->grids->warp function. This will prevent high_order_mesh from growing in memory proportional to quadrature nodes, but instead being solely dependent on the mesh nodes. Lastly, another large memory usage comes from allocation the dim volume_quadrature_collection and fe_collection_lagrange. For DG strong, neither of these are needed so an option will need to be added where the volume quadrature weights are obtained through the tensor product of the oneD quadrature weights-- similarly for the surface quadrature weights.
jbrillon
This PR addresses: 1) Remaining memory issues for high polynomial orders for linear and curvilinear meshes. 2) Provides scaling for DG strong conservative, overintegrated and entropy conserving on a straight and curvilinear mesh. Showing the entire strong DG scales at O(n^{d+1}), up to 136^3 dofs ran locally on 4GB of ram. 3) Restructuring the two-point flux and creating a vector sparsity structure as to only compute O(n^{d+1}) terms on the fly and apply them. 4) Overintegrated entropy conserving schemes now conserve entropy at machine precision on general integration strength by introducing a new solution basis entropy projected variable. Below, please find results for conservative strong DG, FR c+ with and without overintegration. Also, DG and c+ entropy conserving "EC" two-point flux form using Ranocha's two-point flux for both Cartesian and nonsymmetric curvilinear meshes. 3D_TGV_scaling_curv_legend_right.pdf 3D_scaling_cartesian_TGV_legend_right.pdf