gmx::FusedGpuHaloExchange Class Reference

#include <gromacs/domdec/fused_gpuhaloexchange.h>

Description

NVSHMEM-based fused GPU halo exchange.

Aggregates all pulses into a single structure for a fused GPU-initiated exchange path using NVSHMEM. Combining coordinate (X) and force (F) metadata is a by-product of this aggregation. The class:

• Builds per-dimension/pulse metadata and aggregates entries into one struct
• Allocates unified device buffers for send/recv regions and computes per-pulse offsets aligned to c_haloEntryAlignBytes for coalesced access and high-throughput transfers (e.g., NVSHMEM device puts/gets, CUDA vector/TMA async copies)
• Exposes device pointers per entry for pack/unpack kernels
• Manages NVSHMEM signal buffers via a single symmetric allocation and pointer-offset views (see implementation for exact layout)
• Reduces MPI collectives: unified send/recv sizes enable a single MPI_Allreduce for max sizing instead of per-pulse max-reductions
• Unified buffer sizes are max-reduced across ranks to guarantee identical sizes

Public Member Functions
	FusedGpuHaloExchange (const DeviceContext &deviceContext, gmx_wallcycle *wcycle, MPI_Comm mpi_comm_mysim, MPI_Comm mpi_comm_mysim_world)
	Creates NVSHMEM-based fused GPU halo exchange object. More...
	~FusedGpuHaloExchange ()
	Destructor.
GpuEventSynchronizer *	launchAllCoordinateExchanges (const matrix box, GpuEventSynchronizer *dependencyEvent)
	Launch fused coordinate (X) exchanges across all pulses. More...
GpuEventSynchronizer *	launchAllForceExchanges (bool accumulateForces, FixedCapacityVector< GpuEventSynchronizer *, 2 > *dependencyEvents)
	Launch fused force (F) exchanges across all pulses. More...
void	reinitAllHaloExchanges (const t_commrec &cr, DeviceBuffer< RVec > d_coordinatesBuffer, DeviceBuffer< RVec > d_forcesBuffer, DeviceBuffer< uint64_t > d_syncBase, int totalNumPulses)
	(Re-)initialize fused halo exchanges for all dimensions and pulses. Builds per-pulse entries, sets shared buffers, NVSHMEM signals, and prepares metadata. More...
void	destroyAllHaloExchangeBuffers ()
	Destroy unified/symmetric buffers used by fused halo exchange (if any).
void	allocateAndCopyHaloExchangeData ()
	Copy per-pulse metadata to device.
GpuEventSynchronizer *	getForcesReadyOnDeviceEvent ()
	Get the event synchronizer for the fused forces ready on device. More...
void	allocateUnifiedHaloBuffers ()
	Allocate unified send/recv buffers and point entries into them. Uses max-reduced per-pulse extents to size symmetric buffers and assigns aligned per-entry pointers into those buffers. More...

Public Attributes
SharedBuffers	sharedBuffers_
DeviceBuffer< uint64_t >	d_ppHaloExSyncBase_ = nullptr
int	ppHaloExPerSyncBufSize_ = 0

Constructor & Destructor Documentation

gmx::FusedGpuHaloExchange::FusedGpuHaloExchange	(	const DeviceContext &	deviceContext,
		gmx_wallcycle *	wcycle,
		MPI_Comm	mpi_comm_mysim,
		MPI_Comm	mpi_comm_mysim_world
	)

Creates NVSHMEM-based fused GPU halo exchange object.

Parameters

[in]	deviceContext	GPU device context
[in]	wcycle	Wallclock cycle accounting
[in]	mpi_comm_mysim	MPI communicator used for simulation
[in]	mpi_comm_mysim_world	MPI communicator involving PP + PME

Member Function Documentation

void gmx::FusedGpuHaloExchange::allocateUnifiedHaloBuffers

(

)

Allocate unified send/recv buffers and point entries into them. Uses max-reduced per-pulse extents to size symmetric buffers and assigns aligned per-entry pointers into those buffers.

max-reduced per-pulse extent across ranks (elements), unified for send/recv

GpuEventSynchronizer * gmx::FusedGpuHaloExchange::getForcesReadyOnDeviceEvent

(

)

Get the event synchronizer for the fused forces ready on device.

Returns

The event to synchronize the stream that consumes forces on device.

GpuEventSynchronizer * gmx::FusedGpuHaloExchange::launchAllCoordinateExchanges	(	const matrix	box,
		GpuEventSynchronizer *	dependencyEvent
	)

Launch fused coordinate (X) exchanges across all pulses.

Parameters

[in]	box	Coordinate box (for shifts)
[in]	dependencyEvent	Dependency event for this operation

Returns

Event recorded when this operation has been launched

GpuEventSynchronizer * gmx::FusedGpuHaloExchange::launchAllForceExchanges	(	bool	accumulateForces,
		FixedCapacityVector< GpuEventSynchronizer *, 2 > *	dependencyEvents
	)

Launch fused force (F) exchanges across all pulses.

Parameters

[in]	accumulateForces	True if forces should accumulate, otherwise they are set
[in]	dependencyEvents	Dependency events for this operation

Returns

Event recorded when this operation has been launched

void gmx::FusedGpuHaloExchange::reinitAllHaloExchanges	(	const t_commrec &	cr,
		DeviceBuffer< RVec >	d_coordinatesBuffer,
		DeviceBuffer< RVec >	d_forcesBuffer,
		DeviceBuffer< uint64_t >	d_syncBase,
		int	totalNumPulses
	)

(Re-)initialize fused halo exchanges for all dimensions and pulses. Builds per-pulse entries, sets shared buffers, NVSHMEM signals, and prepares metadata.

Parameters

[in]	cr	Communication record
[in]	d_coordinatesBuffer	Pointer to coordinates buffer in GPU memory
[in]	d_forcesBuffer	Pointer to forces buffer in GPU memory
[in]	d_syncBase	Base device pointer for NVSHMEM signal buffer
[in]	totalNumPulses	Total number of pulses across all dimensions

---

The documentation for this class was generated from the following files:

• src/gromacs/domdec/fused_gpuhaloexchange.h
• src/gromacs/domdec/fused_gpuhaloexchange.cpp