Buffers

MPI.IN_PLACE

A sentinel value that can be passed as a buffer argument for certain collective operations to use the same buffer for send and receive operations.

• Scatter! and Scatterv!: can be used as the recvbuf argument on the root process.

• Gather! and Gatherv!: can be used as the sendbuf argument on the root process.

• Allgather!, Allgatherv!, Alltoall! and Alltoallv!: can be used as the sendbuf argument on all processes.

• Reduce! (root only), Allreduce!, Scan! and Exscan!: can be used as sendbuf argument.

MPI.Buffer

An MPI buffer for communication with a single rank. It is used for point-to-point and one-sided operations, as well as some collective operations. Operations will implicitly construct a Buffer when required via the generic constructor, but it can be advantageous to manually construct Buffers when doing so incurs additional overhead, for example when using a non-predefined MPI.Datatype.

Fields

• data: a Julia object referencing a region of memory to be used for communication. It is required that the object can be cconverted to an MPIPtr.

• count: the number of elements of datatype in the buffer. Note that this may not correspond to the number of elements in the array if derived types are used.

• datatype: the MPI.Datatype stored in the buffer.

Usage

Buffer(data, count::Integer, datatype::Datatype)

Generic constructor.

Buffer(data)

Construct a Buffer backed by data, automatically determining the appropriate count and datatype. Methods are provided for

• Ref
• Array
• CUDA.CuArray if CUDA.jl is loaded.
• AMDGPU.ROCArray if AMDGPU.jl is loaded.
• SubArrays of an Array, CUDA.CuArray or AMDGPU.ROCArray where the layout is contiguous, sequential or blocked.

See also

• Buffer_send

Buffer_send(data)

Construct a Buffer object for a send operation from data, allowing cases where isbits(data).

MPI.UBuffer

An MPI buffer for chunked collective communication, where all chunks are of uniform size.

Fields

• data: A Julia object referencing a region of memory to be used for communication. It is required that the object can be cconverted to an MPIPtr.

• count: The number of elements of datatype in each chunk.

• nchunks: The maximum number of chunks stored in the buffer. This is used only for validation, and can be set to nothing to disable checks.

• datatype: The MPI.Datatype stored in the buffer.

Usage

UBuffer(data, count::Integer, nchunks::Union{Nothing, Integer}, datatype::Datatype)

Generic constructor.

UBuffer(data, count::Integer)

Construct a UBuffer backed by data, where count is the number of elements in each chunk.

See also

• VBuffer: similar, but supports chunks of non-uniform sizes.

MPI.VBuffer

An MPI buffer for chunked collective communication, where chunks can be of different sizes and at different offsets.

Fields

• data: A Julia object referencing a region of memory to be used for communication. It is required that the object can be cconverted to an MPIPtr.

• counts: An array containing the length of each chunk.

• displs: An array containing the (0-based) displacements of each chunk.

• datatype: The MPI.Datatype stored in the buffer.

Usage

VBuffer(data, counts[, displs[, datatype]])

Construct a VBuffer backed by data, where counts[j] is the number of elements in the jth chunk, and displs[j] is the 0-based displacement. In other words, the jth chunk occurs in indices displs[j]+1:displs[j]+counts[j].

The default value for displs[j] = sum(counts[1:j-1]).

See also

• UBuffer when chunks are all of the same size.

MPI.RBuffer

An MPI buffer for reduction operations (MPI.Reduce!, MPI.Allreduce!, MPI.Scan!, MPI.Exscan!).

Fields

• senddata: A Julia object referencing a region of memory to be used for the send buffer. It is required that the object can be cconverted to an MPIPtr.

• recvdata: A Julia object referencing a region of memory to be used for the receive buffer. It is required that the object can be cconverted to an MPIPtr.

• count: the number of elements of datatype in the buffer. Note that this may not correspond to the number of elements in the array if derived types are used.

• datatype: the MPI.Datatype stored in the buffer.

Usage

RBuffer(senddata, recvdata[, count, datatype])

Generic constructor.

RBuffer(senddata, recvdata)

Construct a Buffer backed by senddata and recvdata, automatically determining the appropriate count and datatype.

• senddata can be MPI.IN_PLACE
• recvdata can be nothing on a non-root node with MPI.Reduce!

MPI.MPIPtr

A pointer to an MPI buffer. This type is used only as part of the implicit conversion in ccall: a Julia object can be passed to MPI by defining methods for Base.cconvert(::Type{MPIPtr}, ...)/Base.unsafe_convert(::Type{MPIPtr}, ...).

Currently supported are:

• Ptr
• Ref
• Array
• SubArray
• CUDA.CuArray if CUDA.jl is loaded.
• AMDGPU.ROCArray if AMDGPU.jl is loaded.

Additionally, certain sentinel values can be used, e.g. MPI_IN_PLACE or MPI_BOTTOM.