Add Metis extension (#1)

Project.toml

@@ -6,14 +6,23 @@ version = "0.0.1-DEV"
 [deps]
 Ferrite = "c061ca5d-56c9-439f-9c0e-210fe06d3992"
 MPI = "da04e1cc-30fd-572f-bb4f-1f8673147195"
+Reexport = "189a3867-3050-52da-a836-e630ba90ab69"
 SparseArrays = "2f01184e-e22b-5df5-ae63-d93ebab69eaf"
 Tensors = "48a634ad-e948-5137-8d70-aa71f2a747f4"
 
+[weakdeps]
+Metis = "2679e427-3c69-5b7f-982b-ece356f1e94b"
+
+[extensions]
+FerriteDistributedMetisPartitioning = "Metis"
+
 [compat]
+Ferrite = "0.3"
+Metis = "1.3"
 julia = "1"
 
 [extras]
 Test = "8dfed614-e22c-5e08-85e1-65c5234f0b40"
 
 [targets]
-test = ["Test"]
+test = ["Metis", "Test"]

ext/FerriteDistributedMetisPartitioning.jl

@@ -0,0 +1,52 @@
+
+module FerriteDistributedMetisPartitioning
+
+# This extension requires modules as type parameters
+# https://github.com/JuliaLang/julia/pull/47749
+if VERSION >= v"1.10.0-DEV.90"
+
+using FerriteDistributed, Metis
+import FerriteDistributed: CoverTopology, PartitioningAlgorithm
+
+struct MetisPartitioning <: PartitioningAlgorithm.Ext{Metis}
+    alg::Symbol
+end
+
+PartitioningAlgorithm.Ext{Metis}(alg::Symbol) = MetisPartitioning(alg)
+
+function FerriteDistributed.create_partitioning(grid::Grid{dim,C,T}, grid_topology::CoverTopology, n_partitions::Int, partition_alg::PartitioningAlgorithm.Ext{Metis}) where {dim,C,T}
+    n_cells_global = getncells(grid)
+    @assert n_cells_global > 0
+
+    if n_partitions == 1
+        return ones(Metis.idx_t, n_cells_global)
+    end
+
+    # Set up the element connectivity graph
+    xadj = Vector{Metis.idx_t}(undef, n_cells_global+1)
+    xadj[1] = 1
+    adjncy = Vector{Metis.idx_t}(undef, 0)
+    @inbounds for i in 1:n_cells_global
+        n_neighbors = 0
+        for neighbor ∈ getneighborhood(grid_topology, grid, CellIndex(i))
+            push!(adjncy, neighbor)
+            n_neighbors += 1
+        end
+        xadj[i+1] = xadj[i] + n_neighbors
+    end
+
+    # Generate a partitioning
+    return Metis.partition(
+        Metis.Graph(
+            Metis.idx_t(n_cells_global),
+            xadj,
+            adjncy
+        ),
+        n_partitions;
+        alg=partition_alg.alg
+    )
+end
+
+end # VERSION check
+
+end # module FerriteDistributedMetisPartitioning

src/FerriteDistributed.jl

@@ -1,6 +1,7 @@
 module FerriteDistributed
 
-using Ferrite, MPI
+using Reexport
+@reexport using Ferrite, MPI
 
 include("CoverTopology.jl")
 

src/NODGrid.jl

@@ -58,7 +58,7 @@ Get the number of ranks on the global communicator of the distributed grid.
 Construct a non-overlapping distributed grid from a grid with the SFC induced by the element ordering on a specified MPI communicator.
 It is assumed that this function is called with exactly the same grid on each MPI process in the communicator.
 """
-function NODGrid(grid_comm::MPI.Comm, grid_to_distribute::Grid{dim,C,T}) where {dim,C,T}
+function NODGrid(grid_comm::MPI.Comm, grid_to_distribute::Grid{dim,C,T}, alg = PartitioningAlgorithm.SFC()) where {dim,C,T}
     grid_topology = CoverTopology(grid_to_distribute)
     nparts = MPI.Comm_size(grid_comm)
     partitioning = create_partitioning(grid_to_distribute, grid_topology, nparts, PartitioningAlgorithm.SFC())

test/runtests.jl

@@ -29,4 +29,25 @@ end
     @test FerriteDistributed.global_nranks(dgrid) == 1
 end
 
+if VERSION >= v"1.10.0-DEV.90"
+    using Metis
+    @testset "Metis integration" begin
+        grid = generate_grid(Quadrilateral, (4,4))
+        dgrid = NODGrid(MPI.COMM_WORLD, grid, FerriteDistributed.PartitioningAlgorithm.Ext{Metis}(:RECURSIVE))
+
+        @test isempty(get_shared_vertices(dgrid))
+        @test isempty(get_shared_edges(dgrid))
+        @test isempty(get_shared_faces(dgrid))
+        @test getvertexsets(getlocalgrid(dgrid)) == getvertexsets(grid)
+        @test getfacesets(getlocalgrid(dgrid)) == getfacesets(grid)
+        @test getcellsets(getlocalgrid(dgrid)) == getcellsets(grid)
+        @test getnnodes(getlocalgrid(dgrid) ) == getnnodes(grid)
+        @test getncells(getlocalgrid(dgrid) ) == getncells(grid)
+
+        @test FerriteDistributed.global_rank(dgrid) == 1
+
+        @test FerriteDistributed.global_nranks(dgrid) == 1
+    end
+end # VERSION >= v"1.10.0-DEV.90"
+
 #TODO MPI based test sets with more than 1 proc