Improve Datatype handling (#471)

* Improve Datatype handling

- Uses generated functions to avoid creating new instances `Datatype`s on every communication operation (#462). These are lazily initialized to allow for use in precompiled modules.
- Attach the corresponding Julia to an MPI Datatype as an attribute
- Nicer printing of Datatypes
- Allow specifying bits type as return type of MPI.Scatter
- Minor optimisation when constructing buffers of a single NTuple

* duplicate Datatypes if same size as existing type

Author: simonbyrne

deps/consts_microsoftmpi.jl

@@ -83,6 +83,7 @@ const MPI_LOCK_EXCLUSIVE = Cint(234)
 const MPI_LOCK_SHARED = Cint(235)
 const MPI_MAX_INFO_KEY = Cint(255)
 const MPI_MAX_INFO_VAL = Cint(1024)
+const MPI_MAX_OBJECT_NAME = Cint(128)
 const MPI_TAG_UB = reinterpret(Cint, 0x64400001)
 const MPI_COMM_TYPE_SHARED = Cint(1)
 const MPI_ORDER_C = Cint(56)

deps/consts_mpich.jl

@@ -88,6 +88,7 @@ const MPI_LOCK_EXCLUSIVE = Cint(234)
 const MPI_LOCK_SHARED = Cint(235)
 const MPI_MAX_INFO_KEY = Cint(255)
 const MPI_MAX_INFO_VAL = Cint(1024)
+const MPI_MAX_OBJECT_NAME = Cint(128)
 const MPI_TAG_UB = Cint(1681915905)
 const MPI_COMM_TYPE_SHARED = Cint(1)
 const MPI_ORDER_C = Cint(56)

deps/consts_openmpi.jl

@@ -95,6 +95,7 @@ const MPI_LOCK_EXCLUSIVE = Cint(1)
 const MPI_LOCK_SHARED = Cint(2)
 const MPI_MAX_INFO_KEY = Cint(36)
 const MPI_MAX_INFO_VAL = Cint(256)
+const MPI_MAX_OBJECT_NAME = Cint(64)
 const MPI_TAG_UB = Cint(0)
 const MPI_COMM_TYPE_SHARED = Cint(0)
 const MPI_ORDER_C = Cint(0)

deps/gen_consts.jl

@@ -126,6 +126,7 @@ MPI_Cints = [
     :MPI_LOCK_SHARED,
     :MPI_MAX_INFO_KEY,
     :MPI_MAX_INFO_VAL,
+    :MPI_MAX_OBJECT_NAME,
     :MPI_TAG_UB,
     :MPI_COMM_TYPE_SHARED,
     :MPI_ORDER_C,

docs/examples/06-scatterv.jl

@@ -36,6 +36,7 @@ if rank == root
 
     # store sizes in 2 * comm_size Array
     sizes = vcat(M_counts', N_counts')
+    size_ubuf = UBuffer(sizes, 2)
 
     # store number of values to send to each rank in comm_size length Vector
     counts = vec(prod(sizes, dims=1))
@@ -44,7 +45,7 @@ if rank == root
     output_vbuf = VBuffer(output, counts) # VBuffer for gather
 else
     # these variables can be set to `nothing` on non-root processes
-    sizes = nothing
+    size_ubuf = UBuffer(nothing)
     output_vbuf = test_vbuf = VBuffer(nothing)
 end
 
@@ -58,8 +59,8 @@ if rank == root
 end 
 MPI.Barrier(comm)
 
-local_M, local_N = MPI.Scatter!(sizes, zeros(Int, 2), root, comm)
-local_test = MPI.Scatterv!(test_vbuf, zeros(Float64, local_M, local_N), root, comm)
+local_size = MPI.Scatter(size_ubuf, NTuple{2,Int}, root, comm)
+local_test = MPI.Scatterv!(test_vbuf, zeros(Float64, local_size), root, comm)
 
 for i = 0:comm_size-1
     if rank == i

docs/src/advanced.md

@@ -17,6 +17,7 @@ MPI.Types.create_subarray
 MPI.Types.create_struct
 MPI.Types.create_resized
 MPI.Types.commit!
+MPI.Types.duplicate
 ```
 
 ## Operator objects

src/buffers.jl

@@ -147,22 +147,29 @@ function Buffer(sub::Base.FastContiguousSubArray)
 end
 function Buffer(sub::Base.FastSubArray)
     datatype = Types.create_vector(length(sub), 1, sub.stride1,
-                                   Datatype(eltype(sub); commit=false))
+                                   Datatype(eltype(sub)))
     Types.commit!(datatype)
     Buffer(sub, Cint(1), datatype)
 end
 function Buffer(sub::SubArray{T,N,P,I,false}) where {T,N,P,I<:Tuple{Vararg{Union{Base.ScalarIndex, Base.Slice, AbstractUnitRange}}}}
     datatype = Types.create_subarray(size(parent(sub)),
                                      map(length, sub.indices),
                                      map(i -> first(i)-1, sub.indices),
-                                     Datatype(eltype(sub), commit=false))
+                                     Datatype(eltype(sub)))
     Types.commit!(datatype)
     Buffer(parent(sub), Cint(1), datatype)
 end
 
+# NTuple: avoid creating a new datatype if possible
+function Buffer(data::Ref{NTuple{N,T}}) where {N,T}
+    Buffer(data, Cint(N), Datatype(T))
+end
+
+
 Buffer(::InPlace) = Buffer(IN_PLACE, 0, DATATYPE_NULL)
 Buffer(::Nothing) = Buffer(nothing, 0, DATATYPE_NULL)
 
+
 """
     Buffer_send(data)
 

src/collective.jl

@@ -113,11 +113,25 @@ Scatter!(sendbuf::Nothing, recvbuf, root::Integer, comm::Comm) =
     Scatter!(UBuffer(nothing), recvbuf, root, comm)
 
 # determine UBuffer count from recvbuf
-Scatter!(sendbuf::AbstractArray, recvbuf::Union{Ref,AbstractArray}, root::Integer, comm::Comm) =
+Scatter!(sendbuf::AbstractArray{T}, recvbuf::Union{Ref{T},AbstractArray{T}}, root::Integer, comm::Comm) where {T} =
     Scatter!(UBuffer(sendbuf,length(recvbuf)), recvbuf, root, comm)    
 
 """
-    Scatterv!(sendbuf::Union{VBuffer,Nothing}, recvbuf, root, comm)
+    Scatter(sendbuf, T, root::Integer, comm::Comm)
+
+Splits the buffer `sendbuf` in the `root` process into `Comm_size(comm)` chunks,
+sending the `j`-th chunk to the process of rank `j-1` as an object of type `T`.
+
+# See also
+- [`Scatter!`](@ref)
+"""
+function Scatter(sendbuf, ::Type{T}, root::Integer, comm::Comm) where {T}
+    Scatter!(sendbuf, Ref{T}(), root, comm)[]
+end
+
+
+"""
+    Scatterv!(sendbuf, T, root, comm)
 
 Splits the buffer `sendbuf` in the `root` process into `Comm_size(comm)` chunks and sends
 the `j`th chunk to the process of rank `j-1` into the `recvbuf` buffer.