Better mesh generation + missing kernels#206
Conversation
| // 2D Q3Q2 - ok. | ||
| // 2D Q4Q3. | ||
| TENS::Specialization<2,QVectorLayout::byNODES,1,4,8>::Opt<1>::Add(); | ||
| TENS::Specialization<2,QVectorLayout::byVDIM,2,5,8>::Opt<1>::Add(); |
| mesh = PartitionMPI(dim, Mpi::WorldSize(), elem_per_mpi, myid == 0, | ||
| rp_levels, mpi_partitioning); |
There was a problem hiding this comment.
Do we want to extend this function to allow generating domains which are not the unit square/cube (similar to the Sx/Sy/Sz options I added to #205) so you can generate meshes for the triple point problem? I think this would also require specific multiples of elements in each dimension to ensure that interface boundaries match element boundaries.
This function also seems to generate some interesting elongated (in terms of zone count per dimension) meshes and only allows cartesian partitioning.
For 512 ranks in 3D I get 16x4x8, 1024 ranks gives 32x4x8. I think I prefer the existing mesh generation method as you can specify the exact domain shape and test metis partitioned mesh communication patterns/balance.
There was a problem hiding this comment.
As you said, that won't solve the problem for the 3point mesh. I prefer to keep it simple with [0,1]^dim - the intention is not to test physics, just a generic mesh that can be used for performance tests.
I will debug the weird elongations.
There was a problem hiding this comment.
We can keep the Sx/Sy/Sz option + metis. I wanted to replace the complex -partitionning stuff.
There was a problem hiding this comment.
fixed "default" Co-authored-by: Andrew Ho <ho37@llnl.gov>
camierjs
left a comment
camierjs
left a comment
There was a problem hiding this comment.
Tested with mfem/mfem#5227 - looks good!
3 participants
No description provided.