Merge pull request #70 from gridap/adding_weights_to_partition

Adding the possibility of prescribing weights before redistributing

Author: amartinhuertas

src/OctreeDistributedDiscreteModels.jl

@@ -1599,9 +1599,23 @@ end
 # Assumptions. Either:
 # A) model.parts MPI tasks are included in parts_redistributed_model MPI tasks; or
 # B) model.parts MPI tasks include parts_redistributed_model MPI tasks
+const WeightsArrayType=Union{Nothing,MPIArray{<:Vector{<:Integer}}}
 function GridapDistributed.redistribute(model::OctreeDistributedDiscreteModel{Dc,Dp}, 
-                                        parts_redistributed_model=model.parts) where {Dc,Dp}
+                                        parts_redistributed_model=model.parts;
+                                        weights::WeightsArrayType=nothing) where {Dc,Dp}
   parts = (parts_redistributed_model === model.parts) ? model.parts : parts_redistributed_model
+  _weights=nothing
+  if (weights !== nothing)
+    Gridap.Helpers.@notimplementedif parts!==model.parts
+    _weights=map(model.dmodel.models,weights) do lmodel,weights
+       # The length of the local weights array has to match the number of 
+       # cells in the model. This includes both owned and ghost cells. 
+       # Only the flags for owned cells are actually taken into account. 
+       @assert num_cells(lmodel)==length(weights)
+       convert(Vector{Cint},weights)
+    end
+  end
+
   comm  = parts.comm
   if (GridapDistributed.i_am_in(model.parts.comm) || GridapDistributed.i_am_in(parts.comm))
     if (parts_redistributed_model !== model.parts)
@@ -1610,7 +1624,7 @@ function GridapDistributed.redistribute(model::OctreeDistributedDiscreteModel{Dc
       @assert A || B
     end
     if (parts_redistributed_model===model.parts || A)
-      _redistribute_parts_subseteq_parts_redistributed(model,parts_redistributed_model)
+      _redistribute_parts_subseteq_parts_redistributed(model,parts_redistributed_model,_weights)
     else
       _redistribute_parts_supset_parts_redistributed(model, parts_redistributed_model)
     end
@@ -1619,7 +1633,9 @@ function GridapDistributed.redistribute(model::OctreeDistributedDiscreteModel{Dc
   end
 end
 
-function _redistribute_parts_subseteq_parts_redistributed(model::OctreeDistributedDiscreteModel{Dc,Dp}, parts_redistributed_model) where {Dc,Dp}
+function _redistribute_parts_subseteq_parts_redistributed(model::OctreeDistributedDiscreteModel{Dc,Dp},
+                                                          parts_redistributed_model, 
+                                                          _weights::WeightsArrayType) where {Dc,Dp}
   parts = (parts_redistributed_model === model.parts) ? model.parts : parts_redistributed_model
   if (parts_redistributed_model === model.parts)
     ptr_pXest_old = model.ptr_pXest
@@ -1631,7 +1647,15 @@ function _redistribute_parts_subseteq_parts_redistributed(model::OctreeDistribut
                                        parts.comm)
   end
   ptr_pXest_new = pXest_copy(model.pXest_type, ptr_pXest_old)
-  pXest_partition!(model.pXest_type, ptr_pXest_new)
+  if (_weights !== nothing)
+    init_fn_callback_c = pXest_reset_callbacks(model.pXest_type)
+    map(_weights) do _weights 
+      pXest_reset_data!(model.pXest_type, ptr_pXest_new, Cint(sizeof(Cint)), init_fn_callback_c, pointer(_weights))
+    end
+    pXest_partition!(model.pXest_type, ptr_pXest_new; weights_set=true)
+  else
+    pXest_partition!(model.pXest_type, ptr_pXest_new; weights_set=false)
+  end 
 
   # Compute RedistributeGlue
   parts_snd, lids_snd, old2new = pXest_compute_migration_control_data(model.pXest_type,ptr_pXest_old,ptr_pXest_new)

src/PXestTypeMethods.jl

@@ -309,16 +309,31 @@ function pXest_balance!(::P8estType, ptr_pXest; k_2_1_balance=0)
   end
 end
 
-function pXest_partition!(::P4estType, ptr_pXest)
-  p4est_partition(ptr_pXest, 0, C_NULL)
+function pXest_partition!(pXest_type::P4estType, ptr_pXest; weights_set=false)
+  if (!weights_set)
+    p4est_partition(ptr_pXest, 0, C_NULL)
+  else
+    wcallback=pXest_weight_callback(pXest_type)
+    p4est_partition(ptr_pXest, 0, wcallback)
+  end
 end 
 
-function pXest_partition!(::P6estType, ptr_pXest)
-  p6est_partition(ptr_pXest, C_NULL)
+function pXest_partition!(pXest_type::P6estType, ptr_pXest; weights_set=false)
+  if (!weights_set)
+    p6est_partition(ptr_pXest, C_NULL)
+  else
+    wcallback=pXest_weight_callback(pXest_type)
+    p6est_partition(ptr_pXest, wcallback)
+  end
 end 
 
-function pXest_partition!(::P8estType, ptr_pXest)
-  p8est_partition(ptr_pXest, 0, C_NULL)
+function pXest_partition!(pXest_type::P8estType, ptr_pXest; weights_set=false)
+  if (!weights_set)
+    p8est_partition(ptr_pXest, 0, C_NULL)
+  else
+    wcallback=pXest_weight_callback(pXest_type)
+    p8est_partition(ptr_pXest, 0, wcallback)
+  end
 end 
 
 
@@ -805,6 +820,30 @@ function pXest_refine_callbacks(::P8estType)
   refine_callback_c, refine_replace_callback_c
 end 
 
+function pXest_weight_callback(::P4estType)
+  function weight_callback(::Ptr{p4est_t},
+    which_tree::p4est_topidx_t,
+    quadrant_ptr::Ptr{p4est_quadrant_t})
+    quadrant = quadrant_ptr[]
+    return unsafe_wrap(Array, Ptr{Cint}(quadrant.p.user_data), 1)[]
+  end
+  @cfunction($weight_callback, Cint, (Ptr{p4est_t}, p4est_topidx_t, Ptr{p4est_quadrant_t}))
+end
+
+function pXest_weight_callback(::P6estType)
+  Gridap.Helpers.@notimplemented
+end 
+
+function pXest_weight_callback(::P8estType)
+  function weight_callback(::Ptr{p8est_t},
+    which_tree::p4est_topidx_t,
+    quadrant_ptr::Ptr{p8est_quadrant_t})
+    quadrant = quadrant_ptr[]
+    return unsafe_wrap(Array, Ptr{Cint}(quadrant.p.user_data), 1)[]
+  end
+  @cfunction($weight_callback, Cint, (Ptr{p8est_t}, p4est_topidx_t, Ptr{p8est_quadrant_t}))
+end 
+
 function _unwrap_ghost_quadrants(::P4estType, pXest_ghost)
   Ptr{p4est_quadrant_t}(pXest_ghost.ghosts.array)
 end

test/PoissonNonConformingOctreeModelsTests.jl

@@ -139,7 +139,14 @@ module PoissonNonConformingOctreeModelsTests
     e       = uH - uhH
     el2     = sqrt(sum( ∫( e⋅e )*dΩH ))
 
-    fmodel_red, red_glue=GridapDistributed.redistribute(fmodel);
+    weights=map(ranks,fmodel.dmodel.models) do rank,lmodel
+      if (rank%2==0)
+        zeros(Cint,num_cells(lmodel))
+      else
+        ones(Cint,num_cells(lmodel))
+      end
+    end 
+    fmodel_red, red_glue=GridapDistributed.redistribute(fmodel,weights=weights);
     Vhred=FESpace(fmodel_red,reffe,conformity=:H1;dirichlet_tags="boundary")
     Uhred=TrialFESpace(Vhred,u)