rapidsmpf::coll::detail Namespace Reference

Classes
struct	HostOp
	Host-side range-based reduction operator. More...
struct	DeviceOp
	Device-side range-based reduction operator. More...
class	Chunk
	Represents a data chunk in the allgather operation. More...
class	PostBox
	A thread-safe container for managing chunks in collectives. More...

Typedefs
using	ChunkID = std::uint64_t
	Type alias for chunk identifiers.

Functions
template<typename T , typename Op >
requires std::invocable< Op, T const &, T const & > ReduceOperator	make_host_reduce_operator (Op op)
	Create a host-based reduction operator from a typed binary operation. More...
template<typename T , typename Op >
requires std::invocable< Op, T const &, T const & > ReduceOperator	make_device_reduce_operator (Op op)
	Create a device-based reduction operator from a typed binary operation. More...
std::vector< std::unique_ptr< Chunk > >	test_some (std::vector< std::unique_ptr< Chunk >> &chunks, std::vector< std::unique_ptr< Communicator::Future >> &futures, Communicator *comm)
	Complete posted chunk receives. More...

Detailed Description

SPDX-FileCopyrightText: Copyright (c) 2025-2026, NVIDIA CORPORATION & AFFILIATES. All rights reserved. SPDX-License-Identifier: Apache-2.0

Function Documentation

make_device_reduce_operator()

template<typename T , typename Op >

requires std::invocable<Op, T const&, T const&> ReduceOperator rapidsmpf::coll::detail::make_device_reduce_operator

(

Op

op

)

Create a device-based reduction operator from a typed binary operation.

Template Parameters

T	The element type.
Op	The binary operation type.

Parameters

op	The binary operation (e.g., cuda::std::plus<T>{}).

Returns

A ReduceOperator wrapping the DeviceOp.

Note

This function requires CUDA compilation (CUDACC defined) to be used. Attempting to use it without CUDA will result in a compilation error.

Definition at line 355 of file allreduce.hpp.

make_host_reduce_operator()

template<typename T , typename Op >

requires std::invocable<Op, T const&, T const&> ReduceOperator rapidsmpf::coll::detail::make_host_reduce_operator

(

Op

op

)

Create a host-based reduction operator from a typed binary operation.

Template Parameters

T	The element type.
Op	The binary operation type.

Parameters

op	The binary operation (e.g., std::plus<T>{}).

Returns

A ReduceOperator wrapping the HostOp.

Definition at line 338 of file allreduce.hpp.

test_some()

std::vector<std::unique_ptr<Chunk> > rapidsmpf::coll::detail::test_some	(	std::vector< std::unique_ptr< Chunk >> &	chunks,
		std::vector< std::unique_ptr< Communicator::Future >> &	futures,
		Communicator *	comm
	)

Complete posted chunk receives.

Tests a set of outstanding communicator futures, and for every completed receive, reattaches the received data buffer to the corresponding chunk before returning it.

Parameters

chunks	Chunks with receives currently in flight.
futures	Receive futures corresponding 1:1 with chunks.
comm	Communicator used to test and release completed receives.

Returns

The subset of chunks whose receives completed during this call.