HYPRE extension

Project.toml

@@ -11,9 +11,11 @@ SparseArrays = "2f01184e-e22b-5df5-ae63-d93ebab69eaf"
 Tensors = "48a634ad-e948-5137-8d70-aa71f2a747f4"
 
 [weakdeps]
+HYPRE = "b5ffcf37-a2bd-41ab-a3da-4bd9bc8ad771"
 Metis = "2679e427-3c69-5b7f-982b-ece356f1e94b"
 
 [extensions]
+FerriteDistributedHYPREAssembly = "HYPRE"
 FerriteDistributedMetisPartitioning = "Metis"
 
 [compat]

ext/FerriteDistributedHYPREAssembly.jl

@@ -0,0 +1,20 @@
+"""
+Module containing the code for distributed assembly via HYPRE.jl
+"""
+module FerriteDistributedHYPREAssembly
+
+using FerriteDistributed
+# TODO remove me. These are merely hotfixes to split the extensions trasiently via an internal API.
+import FerriteDistributed: getglobalgrid, num_local_true_dofs, num_local_dofs, global_comm, interface_comm, global_rank, compute_owner, remote_entities
+using MPI
+using HYPRE
+using Base: @propagate_inbounds
+
+include("FerriteDistributedHYPREAssembly/assembler.jl")
+include("FerriteDistributedHYPREAssembly/conversion.jl")
+
+function __init__()
+    @info "FerriteHYPRE extension loaded."
+end
+
+end # module FerriteHYPRE

ext/FerriteDistributedHYPREAssembly/assembler.jl

@@ -0,0 +1,48 @@
+function Ferrite.create_sparsity_pattern(::Type{<:HYPREMatrix}, dh::Ferrite.AbstractDofHandler, ch::Union{ConstraintHandler,Nothing}=nothing; kwargs...)
+    K = create_sparsity_pattern(dh, ch; kwargs...)
+    fill!(K.nzval, 1)
+    return HYPREMatrix(K)
+end
+
+###########################################
+## HYPREAssembler and associated methods ##
+###########################################
+
+struct HYPREAssembler <: Ferrite.AbstractSparseAssembler
+    A::HYPRE.HYPREAssembler
+end
+
+Ferrite.matrix_handle(a::HYPREAssembler) = a.A.A.A # :)
+Ferrite.vector_handle(a::HYPREAssembler) = a.A.b.b # :)
+
+function Ferrite.start_assemble(K::HYPREMatrix, f::HYPREVector)
+    return HYPREAssembler(HYPRE.start_assemble!(K, f))
+end
+
+function Ferrite.assemble!(a::HYPREAssembler, dofs::AbstractVector{<:Integer}, ke::AbstractMatrix, fe::AbstractVector)
+    HYPRE.assemble!(a.A, dofs, ke, fe)
+end
+
+function Ferrite.end_assemble(a::HYPREAssembler)
+    HYPRE.finish_assemble!(a.A)
+end
+
+## Methods for arrayutils.jl ##
+
+function Ferrite.addindex!(A::HYPREMatrix, v, i::Int, j::Int)
+    nrows = HYPRE_Int(1)
+    ncols = Ref{HYPRE_Int}(1)
+    rows = Ref{HYPRE_BigInt}(i)
+    cols = Ref{HYPRE_BigInt}(j)
+    values = Ref{HYPRE_Complex}(v)
+    HYPRE.@check HYPRE_IJMatrixAddToValues(A.ijmatrix, nrows, ncols, rows, cols, values)
+    return A
+end
+
+function Ferrite.addindex!(b::HYPREVector, v, i::Int)
+    nvalues = HYPRE_Int(1)
+    indices = Ref{HYPRE_BigInt}(i)
+    values = Ref{HYPRE_Complex}(v)
+    HYPRE.@check HYPRE_IJVectorAddToValues(b.ijvector, nvalues, indices, values)
+    return b
+end

ext/FerriteDistributedHYPREAssembly/conversion.jl

@@ -0,0 +1,119 @@
+# Hypre to Ferrite vector
+function hypre_to_ferrite!(u::Vector{T}, uh::HYPREVector, dh::Ferrite.AbstractDofHandler) where {T}
+    # Copy solution from HYPRE to Julia
+    uj = Vector{Float64}(undef, num_local_true_dofs(dh))
+    copy!(uj, uh)
+
+    my_rank = global_rank(getglobalgrid(dh))
+
+    # Helper to gather which global dof and values have to be send to which process
+    gdof_value_send = [Dict{Int,Float64}() for i ∈ 1:MPI.Comm_size(MPI.COMM_WORLD)]
+    # Helper to get the global dof to local dof mapping
+    rank_recv_count = [0 for i∈1:MPI.Comm_size(MPI.COMM_WORLD)]
+    gdof_to_ldof = Dict{Int,Int}()
+
+    next_dof = 1
+    for (ldof,rank) ∈ enumerate(dh.ldof_to_rank)
+        if rank == my_rank
+            u[ldof] = uj[next_dof]
+            next_dof += 1
+        else 
+            # We have to sync these later.
+            gdof_to_ldof[dh.ldof_to_gdof[ldof]] = ldof
+            rank_recv_count[rank] += 1
+        end
+    end
+
+    # TODO speed this up and better API
+    dgrid = getglobalgrid(dh)
+    for sv ∈ get_shared_vertices(dgrid)
+        lvi = sv.local_idx
+        my_rank != compute_owner(dgrid, sv) && continue
+        for field_idx in 1:num_fields(dh)
+            if Ferrite.has_vertex_dofs(dh, field_idx, lvi)
+                local_dofs = Ferrite.vertex_dofs(dh, field_idx, lvi)
+                global_dofs = dh.ldof_to_gdof[local_dofs]
+                for receiver_rank ∈ keys(remote_entities(sv))
+                    for i ∈ 1:length(global_dofs)
+                        # Note that u already has the correct values for all locally owned dofs due to the loop above!
+                        gdof_value_send[receiver_rank][global_dofs[i]] = u[local_dofs[i]]
+                    end
+                end
+            end
+        end
+    end
+
+    for se ∈ get_shared_edges(dgrid)
+        lei = se.local_idx
+        my_rank != compute_owner(dgrid, se) && continue
+        for field_idx in 1:num_fields(dh)
+            if Ferrite.has_edge_dofs(dh, field_idx, lei)
+                local_dofs = Ferrite.edge_dofs(dh, field_idx, lei)
+                global_dofs = dh.ldof_to_gdof[local_dofs]
+                for receiver_rank ∈ keys(remote_entities(se))
+                    for i ∈ 1:length(global_dofs)
+                        # Note that u already has the correct values for all locally owned dofs due to the loop above!
+                        gdof_value_send[receiver_rank][global_dofs[i]] = u[local_dofs[i]]
+                    end
+                end
+            end
+        end
+    end
+
+    for sf ∈ get_shared_faces(dgrid)
+        lfi = sf.local_idx
+        my_rank != compute_owner(dgrid, sf) && continue
+        for field_idx in 1:num_fields(dh)
+            if Ferrite.has_face_dofs(dh, field_idx, lfi)
+                local_dofs = Ferrite.face_dofs(dh, field_idx, lfi)
+                global_dofs = dh.ldof_to_gdof[local_dofs]
+                for receiver_rank ∈ keys(remote_entities(sf))
+                    for i ∈ 1:length(global_dofs)
+                        # Note that u already has the correct values for all locally owned dofs due to the loop above!
+                        gdof_value_send[receiver_rank][global_dofs[i]] = u[local_dofs[i]]
+                    end
+                end
+            end
+        end
+    end
+
+    Ferrite.@debug println("preparing to distribute $gdof_value_send (R$my_rank)")
+
+    # TODO precompute graph at it is static
+    graph_source   = Cint[my_rank-1]
+    graph_dest   = Cint[]
+    for r ∈ 1:MPI.Comm_size(MPI.COMM_WORLD)
+        !isempty(gdof_value_send[r]) && push!(graph_dest, r-1)
+    end
+
+    graph_degree = Cint[length(graph_dest)]
+    graph_comm = MPI.Dist_graph_create(MPI.COMM_WORLD, graph_source, graph_degree, graph_dest)
+    indegree, outdegree, _ = MPI.Dist_graph_neighbors_count(graph_comm)
+
+    inranks = Vector{Cint}(undef, indegree)
+    outranks = Vector{Cint}(undef, outdegree)
+    MPI.Dist_graph_neighbors!(graph_comm, inranks, outranks)
+
+    send_count = [length(gdof_value_send[outrank+1]) for outrank ∈ outranks]
+    recv_count = [rank_recv_count[inrank+1] for inrank ∈ inranks]
+
+    send_gdof = Cint[]
+    for outrank ∈ outranks
+        append!(send_gdof, Cint.(keys(gdof_value_send[outrank+1])))
+    end
+    recv_gdof = Vector{Cint}(undef, sum(recv_count))
+    MPI.Neighbor_alltoallv!(VBuffer(send_gdof,send_count), VBuffer(recv_gdof,recv_count), graph_comm)
+
+    send_val = Cdouble[]
+    for outrank ∈ outranks
+        append!(send_val, Cdouble.(values(gdof_value_send[outrank+1])))
+    end
+    recv_val = Vector{Cdouble}(undef, sum(recv_count))
+    MPI.Neighbor_alltoallv!(VBuffer(send_val,send_count), VBuffer(recv_val,recv_count), graph_comm)
+
+    for (gdof, val) ∈ zip(recv_gdof, recv_val)
+        u[gdof_to_ldof[gdof]] = val
+    end
+
+    return u
+end