pme_gpu.cpp File Reference

#include "gmxpre.h"
#include "config.h"
#include <list>
#include "gromacs/ewald/ewald_utils.h"
#include "gromacs/ewald/pme.h"
#include "gromacs/fft/parallel_3dfft.h"
#include "gromacs/math/invertmatrix.h"
#include "gromacs/mdlib/gmx_omp_nthreads.h"
#include "gromacs/mdtypes/enerdata.h"
#include "gromacs/mdtypes/forceoutput.h"
#include "gromacs/mdtypes/inputrec.h"
#include "gromacs/utility/exceptions.h"
#include "gromacs/utility/fatalerror.h"
#include "gromacs/utility/gmxassert.h"
#include "gromacs/utility/stringutil.h"
#include "pme_gpu_internal.h"
#include "pme_grid.h"
#include "pme_internal.h"
#include "pme_solve.h"

Include dependency graph for pme_gpu.cpp:

Description

Implements high-level PME GPU functions which do not require GPU framework-specific code.

Author

Aleksei Iupinov a.yup.nosp@m.inov.nosp@m.@gmai.nosp@m.l.co.nosp@m.m

Functions
void	pme_gpu_reset_timings (const gmx_pme_t *pme)
	Resets the PME GPU timings. To be called at the reset step. More...
void	pme_gpu_get_timings (const gmx_pme_t *pme, gmx_wallclock_gpu_pme_t *timings)
	Copies the PME GPU timings to the gmx_wallclock_gpu_pme_t structure (for log output). To be called at the run end. More...
int	pme_gpu_get_padding_size (const gmx_pme_t *pme)
	Returns the size of the padding needed by GPU version of PME in the coordinates array. More...
void	parallel_3dfft_execute_gpu_wrapper (gmx_pme_t *pme, const int gridIndex, enum gmx_fft_direction dir, gmx_wallcycle_t wcycle)
	A convenience wrapper for launching either the GPU or CPU FFT. More...
void	pme_gpu_prepare_computation (gmx_pme_t *pme, bool needToUpdateBox, const matrix box, gmx_wallcycle *wcycle, int flags, bool useGpuForceReduction)
	Prepares PME on GPU computation (updating the box if needed) More...
void	pme_gpu_launch_spread (gmx_pme_t *pme, GpuEventSynchronizer *xReadyOnDevice, gmx_wallcycle *wcycle)
	Launches first stage of PME on GPU - spreading kernel. More...
void	pme_gpu_launch_complex_transforms (gmx_pme_t *pme, gmx_wallcycle *wcycle)
	Launches middle stages of PME (FFT R2C, solving, FFT C2R) either on GPU or on CPU, depending on the run mode. More...
void	pme_gpu_launch_gather (const gmx_pme_t *pme, gmx_wallcycle *wcycle, PmeForceOutputHandling forceTreatment)
	Launches last stage of PME on GPU - force gathering and D2H force transfer. More...
static void	sum_forces (gmx::ArrayRef< gmx::RVec > f, gmx::ArrayRef< const gmx::RVec > forceToAdd)
	Accumulate the forcesToAdd to f, using the available threads.
static void	pme_gpu_reduce_outputs (const int flags, const PmeOutput &output, gmx_wallcycle *wcycle, gmx::ForceWithVirial *forceWithVirial, gmx_enerdata_t *enerd)
	Reduce quantities from output to forceWithVirial and enerd.
bool	pme_gpu_try_finish_task (gmx_pme_t *pme, const int flags, gmx_wallcycle *wcycle, gmx::ForceWithVirial *forceWithVirial, gmx_enerdata_t *enerd, GpuTaskCompletion completionKind)
	Attempts to complete PME GPU tasks. More...
PmeOutput	pme_gpu_wait_finish_task (gmx_pme_t *pme, const int flags, gmx_wallcycle *wcycle)
	Blocks until PME GPU tasks are completed, and gets the output forces and virial/energy (if they were to be computed). More...
void	pme_gpu_wait_and_reduce (gmx_pme_t *pme, const int flags, gmx_wallcycle *wcycle, gmx::ForceWithVirial *forceWithVirial, gmx_enerdata_t *enerd)
	Blocks until PME GPU tasks are completed, and gets the output forces and virial/energy (if they were to be computed). More...
void	pme_gpu_reinit_computation (const gmx_pme_t *pme, gmx_wallcycle *wcycle)
	The PME GPU reinitialization function that is called both at the end of any PME computation and on any load balancing. More...
DeviceBuffer< float >	pme_gpu_get_device_x (const gmx_pme_t *pme)
	Get pointer to device copy of coordinate data. More...
void *	pme_gpu_get_device_f (const gmx_pme_t *pme)
	Get pointer to device copy of force data. More...
void	pme_gpu_set_device_x (const gmx_pme_t *pme, DeviceBuffer< float > d_x)
	Set pointer to device copy of coordinate data. More...
void *	pme_gpu_get_device_stream (const gmx_pme_t *pme)
	Returns the pointer to the GPU stream. More...
void *	pme_gpu_get_device_context (const gmx_pme_t *pme)
	Returns the pointer to the GPU context. More...
GpuEventSynchronizer *	pme_gpu_get_f_ready_synchronizer (const gmx_pme_t *pme)
	Get pointer to the device synchronizer object that allows syncing on PME force calculation completion. More...

Function Documentation

void parallel_3dfft_execute_gpu_wrapper ( gmx_pme_t *  pme, const int  gridIndex, enum gmx_fft_direction  dir, gmx_wallcycle_t  wcycle  )

inline

A convenience wrapper for launching either the GPU or CPU FFT.

Parameters

[in]	pme	The PME structure.
[in]	gridIndex	The grid index - should currently always be 0.
[in]	dir	The FFT direction enum.
[in]	wcycle	The wallclock counter.

void* pme_gpu_get_device_context

(

const gmx_pme_t *

pme

)

Returns the pointer to the GPU context.

Parameters

[in]

pme

The PME data structure.

Returns

Pointer to GPU context object.

void* pme_gpu_get_device_f

(

const gmx_pme_t *

pme

)

Get pointer to device copy of force data.

Parameters

[in]

pme

The PME data structure.

Returns

Pointer to force data

void* pme_gpu_get_device_stream

(

const gmx_pme_t *

pme

)

Returns the pointer to the GPU stream.

Parameters

[in]

pme

The PME data structure.

Returns

Pointer to GPU stream object.

DeviceBuffer<float> pme_gpu_get_device_x

(

const gmx_pme_t *

pme

)

Get pointer to device copy of coordinate data.

Parameters

[in]

pme

The PME data structure.

Returns

Pointer to coordinate data

GpuEventSynchronizer* pme_gpu_get_f_ready_synchronizer

(

const gmx_pme_t *

pme

)

Get pointer to the device synchronizer object that allows syncing on PME force calculation completion.

Parameters

[in]

pme

The PME data structure.

Returns

Pointer to sychronizer

int pme_gpu_get_padding_size

(

const gmx_pme_t *

pme

)

Returns the size of the padding needed by GPU version of PME in the coordinates array.

Parameters

[in]

pme

The PME data structure.

void pme_gpu_get_timings	(	const gmx_pme_t *	pme,
		gmx_wallclock_gpu_pme_t *	timings
	)

Copies the PME GPU timings to the gmx_wallclock_gpu_pme_t structure (for log output). To be called at the run end.

Parameters

[in]	pme	The PME structure.
[in]	timings	The gmx_wallclock_gpu_pme_t structure.

void pme_gpu_launch_complex_transforms	(	gmx_pme_t *	pme,
		gmx_wallcycle *	wcycle
	)

Launches middle stages of PME (FFT R2C, solving, FFT C2R) either on GPU or on CPU, depending on the run mode.

Parameters

[in]	pme	The PME data structure.
[in]	wcycle	The wallclock counter.

void pme_gpu_launch_gather	(	const gmx_pme_t *	pme,
		gmx_wallcycle *	wcycle,
		PmeForceOutputHandling	forceTreatment
	)

Launches last stage of PME on GPU - force gathering and D2H force transfer.

Parameters

[in]	pme	The PME data structure.
[in]	wcycle	The wallclock counter.
[in]	forceTreatment	Tells how data should be treated. The gathering kernel either stores the output reciprocal forces into the host array, or copies its contents to the GPU first and accumulates. The reduction is non-atomic.

void pme_gpu_launch_spread	(	gmx_pme_t *	pme,
		GpuEventSynchronizer *	xReadyOnDevice,
		gmx_wallcycle *	wcycle
	)

Launches first stage of PME on GPU - spreading kernel.

Parameters

[in]	pme	The PME data structure.
[in]	xReadyOnDevice	Event synchronizer indicating that the coordinates are ready in the device memory; nullptr allowed only on separate PME ranks.
[in]	wcycle	The wallclock counter.

void pme_gpu_prepare_computation	(	gmx_pme_t *	pme,
		bool	needToUpdateBox,
		const matrix	box,
		gmx_wallcycle *	wcycle,
		int	flags,
		bool	useGpuForceReduction
	)

Prepares PME on GPU computation (updating the box if needed)

Parameters

[in]	pme	The PME data structure.
[in]	needToUpdateBox	Tells if the stored unit cell parameters should be updated from box.
[in]	box	The unit cell box.
[in]	wcycle	The wallclock counter.
[in]	flags	The combination of flags to affect this PME computation. The flags are the GMX_PME_ flags from pme.h.
[in]	useGpuForceReduction	Whether PME forces are reduced on GPU this step or should be downloaded for CPU reduction

void pme_gpu_reinit_computation	(	const gmx_pme_t *	pme,
		gmx_wallcycle *	wcycle
	)

The PME GPU reinitialization function that is called both at the end of any PME computation and on any load balancing.

Clears the internal grid and energy/virial buffers; it is not safe to start the PME computation without calling this. Note that unlike in the nbnxn module, the force buffer does not need clearing.

Todo:

Rename this function to clear – it clearly only does output resetting and we should be clear about what the function does..

Parameters

[in]	pme	The PME data structure.
[in]	wcycle	The wallclock counter.

void pme_gpu_reset_timings

(

const gmx_pme_t *

pme

)

Resets the PME GPU timings. To be called at the reset step.

Parameters

[in]

pme

The PME structure.

void pme_gpu_set_device_x	(	const gmx_pme_t *	pme,
		DeviceBuffer< float >	d_x
	)

Set pointer to device copy of coordinate data.

Parameters

[in]	pme	The PME data structure.
[in]	d_x	The pointer to the positions buffer to be set

bool pme_gpu_try_finish_task	(	gmx_pme_t *	pme,
		int	flags,
		gmx_wallcycle *	wcycle,
		gmx::ForceWithVirial *	forceWithVirial,
		gmx_enerdata_t *	enerd,
		GpuTaskCompletion	completionKind
	)

Attempts to complete PME GPU tasks.

The completionKind argument controls whether the function blocks until all PME GPU tasks enqueued completed (as pme_gpu_wait_finish_task() does) or only checks and returns immediately if they did not. When blocking or the tasks have completed it also gets the output forces by assigning the ArrayRef to the forces pointer passed in. Virial/energy are also outputs if they were to be computed.

Parameters

[in]	pme	The PME data structure.
[in]	flags	The combination of flags to affect this PME computation. The flags are the GMX_PME_ flags from pme.h.
[in]	wcycle	The wallclock counter.
[out]	forceWithVirial	The output force and virial
[out]	enerd	The output energies
[in]	flags	The combination of flags to affect this PME computation. The flags are the GMX_PME_ flags from pme.h.
[in]	completionKind	Indicates whether PME task completion should only be checked rather than waited for

Returns

True if the PME GPU tasks have completed

void pme_gpu_wait_and_reduce	(	gmx_pme_t *	pme,
		int	flags,
		gmx_wallcycle *	wcycle,
		gmx::ForceWithVirial *	forceWithVirial,
		gmx_enerdata_t *	enerd
	)

Blocks until PME GPU tasks are completed, and gets the output forces and virial/energy (if they were to be computed).

Parameters

[in]	pme	The PME data structure.
[in]	flags	The combination of flags to affect this PME computation. The flags are the GMX_PME_ flags from pme.h.
[in]	wcycle	The wallclock counter.
[out]	forceWithVirial	The output force and virial
[out]	enerd	The output energies

PmeOutput pme_gpu_wait_finish_task	(	gmx_pme_t *	pme,
		int	flags,
		gmx_wallcycle *	wcycle
	)

Blocks until PME GPU tasks are completed, and gets the output forces and virial/energy (if they were to be computed).

Parameters

[in]	pme	The PME data structure.
[in]	flags	The combination of flags to affect this PME computation. The flags are the GMX_PME_ flags from pme.h.
[out]	wcycle	The wallclock counter.

Returns

The output forces, energy and virial