Reduce particle shape when a particle approaches the EB #5209
Conversation
…lgorithms with PML BC
This reverts commit 0051a17.
![github-advanced-security[bot]](https://avatars.githubusercontent.com/in/57789?s=60&v=4){kind=small}
….com:oshapoval/WarpX into add_analysis_embedded_boundary_removal_depth
RemiLehe
changed the title
| } | ||
|
|
||
| } | ||
|
|
||
| ComputeDistanceToEB(); | ||
| MarkReducedShapeCells( m_eb_reduce_particle_shape[lev], eb_fact, WarpX::nox ); |
There was a problem hiding this comment.
I think you don't need to pass m_eb_reduce_particle_shape (or even all) arguments here and could call it like ComputeDistanceToEB(). (Because both are member functions of the same class.)
There was a problem hiding this comment.
Sounds good.
Part of the motivation here is to maybe do a follow-up PR where some of the Mark...Cells functions become free functions.
|
|
||
| // Every cell in box is regular: do not reduce particle shape in any cell | ||
| amrex::ParallelFor(box, [=] AMREX_GPU_DEVICE (int i, int j, int k) { | ||
| eb_reduce_particle_shape_arr(i, j, k) = 0; |
There was a problem hiding this comment.
Default from setVal, but ok to be explicit.
| @@ -656,7 +659,10 @@ void doEsirkepovDepositionShapeN (const GetParticlePosition<PIdx>& GetPosition, | |||
| const amrex::XDim3 & xyzmin, | |||
| amrex::Dim3 lo, | |||
| amrex::Real q, | |||
| [[maybe_unused]]int n_rz_azimuthal_modes) | |||
| [[maybe_unused]]int n_rz_azimuthal_modes, | |||
There was a problem hiding this comment.
Not important formatting of existing code:
| [[maybe_unused]]int n_rz_azimuthal_modes, | |
| [[maybe_unused]] int n_rz_azimuthal_modes, |
| // If particle is close to the embedded boundary, recompute deposition with order 1 shape | ||
| if constexpr (reduce_shape_control == has_reduced_shape) { | ||
| if (reduce_shape_new) { | ||
| for (int i=0; i<depos_order+3; i++) sx_new[i] = 0.; // Erase previous deposition |
There was a problem hiding this comment.
I took a look on godbolt and the compiler seems to unroll this constant-size loop, so no need to write it more obscurely.
| for (int i=0; i<depos_order+3; i++) sx_old[i] = 0.; // Erase previous deposition | ||
| compute_shifted_shape_factor_order1( sx_old+depos_order/2, x_old, i_new+depos_order/2 ); // Redeposit with order 1 | ||
| } | ||
| // Note: depos_order/2 in the above code corresponds to the shift between the index of the lowest point |
There was a problem hiding this comment.
Thanks for the comment, very helpful to parse this.
If it helps to name this, you could also write this in an const int before the function call.
|
The performance of CurrentDeposition was tested by running a uniform plasma test (input attached below) case three times, on Perlmutter. Results: feature branch does not change performance EB was set via implicit function: |
| amrex::Array4<int> const & eb_reduce_particle_shape_arr = eb_reduce_particle_shape->array(mfi); | ||
|
|
||
| // Check if the box (including one layer of guard cells) contains a mix of covered and regular cells | ||
| const amrex::Box& eb_info_box = mfi.tilebox(amrex::IntVect::TheCellVector()).grow(1); |
Follow-up to #5209: My compiler says those locations would reference temporary objects that were destroyed at the end of the line. That seems to be the case indeed. Copy instead to make the temporary a named and thus persistent variable. 
Description edited by @RemiLehe
Overview
This PR reduces the particle shape to order 1, when the particle gets closer to the embedded boundary:
This ensures that the particle does not deposit charge in valid cells, at the time when it is removed, which in turn ensures proper charge conservation with the electromagnetic solver.
Implementation
eb_reduce_particle_shape(andiMultiFab) that indicates in which cells to reduce the particle shape.Tests
This PR adds tests similar to the ones introduced in #5562 to check for charge conservation near the embedded boundary, but with higher-order shape factors:
developmentfor shape 2 and 3 but pass on this PR.developmentfor shape 3 ; they do pass for this PR. It is not clear why the tests do not fail ondevelopmentwith shape 2.Note: For now, this PR only modifies the current deposition (and only the Esirkepov kernel). A follow-up PR will also modify the charge deposition.