accfft_gpu.cpp File Reference

#include "accfft_gpu.h"
#include <mpi.h>
#include <omp.h>
#include <iostream>
#include <cmath>
#include <math.h>
#include "transpose_cuda.h"
#include <cuda_runtime_api.h>
#include <string.h>
#include <cuda.h>
#include <cufft.h>
#include "accfft_common.h"

Functions
void	accfft_cleanup_gpu ()
int	accfft_local_size_dft_r2c_gpu (int *n, int *isize, int *istart, int *osize, int *ostart, MPI_Comm c_comm, bool inplace)
accfft_plan_gpu *	accfft_plan_dft_3d_r2c_gpu (int *n, double *data_d, double *data_out_d, MPI_Comm c_comm, unsigned flags)
void	accfft_execute_r2c_gpu (accfft_plan_gpu *plan, double *data, Complex *data_out, double *timer, std::bitset< 3 > xyz)
void	accfft_execute_c2r_gpu (accfft_plan_gpu *plan, Complex *data, double *data_out, double *timer, std::bitset< 3 > xyz)
int	accfft_local_size_dft_c2c_gpu (int *n, int *isize, int *istart, int *osize, int *ostart, MPI_Comm c_comm)
accfft_plan_gpu *	accfft_plan_dft_3d_c2c_gpu (int *n, Complex *data_d, Complex *data_out_d, MPI_Comm c_comm, unsigned flags)
void	accfft_execute_c2c_gpu (accfft_plan_gpu *plan, int direction, Complex *data_d, Complex *data_out_d, double *timer, std::bitset< 3 > xyz)
void	accfft_destroy_plan (accfft_plan_gpu *plan)
void	accfft_destroy_plan_gpu (accfft_plan_gpu *plan)

Detailed Description

GPU functions of AccFFT

Function Documentation

void accfft_cleanup_gpu

(

)

Cleanup all CPU resources

void accfft_destroy_plan

(

accfft_plan_gpu *

plan

)

Destroy AccFFT CPU plan. This function calls accfft_destroy_plan_gpu.

Parameters

plan	Input plan to be destroyed.

void accfft_destroy_plan_gpu

(

accfft_plan_gpu *

plan

)

Destroy AccFFT GPU plan.

Parameters

plan	Input plan to be destroyed.

void accfft_execute_c2c_gpu	(	accfft_plan_gpu *	plan,
		int	direction,
		Complex *	data_d,
		Complex *	data_out_d,
		double *	timer,
		std::bitset< 3 >	xyz
	)

Execute C2C plan. This function is blocking and only returns after the transform is completed.

Note

For inplace transforms, data_out should point to the same memory address as data, AND the plan must have been created as inplace.

Parameters

plan	FFT plan created by accfft_plan_dft_3d_r2c.
data	Input data in frequency domain.
data_out	Output data in frequency domain.
timer	See Timing AccFFT for more details.
XYZ	a bit set field that determines which directions FFT should be executed

void accfft_execute_c2r_gpu	(	accfft_plan_gpu *	plan,
		Complex *	data,
		double *	data_out,
		double *	timer,
		std::bitset< 3 >	xyz
	)

Execute C2R plan. This function is blocking and only returns after the transform is completed.

Note

For inplace transforms, data_out should point to the same memory address as data, AND the plan must have been created as inplace.

Parameters

plan	FFT plan created by accfft_plan_dft_3d_r2c.
data	Input data in frequency domain.
data_out	Output data in frequency domain.
timer	See Timing AccFFT for more details.
XYZ	a bit set field that determines which directions FFT should be executed

void accfft_execute_r2c_gpu	(	accfft_plan_gpu *	plan,
		double *	data,
		Complex *	data_out,
		double *	timer,
		std::bitset< 3 >	xyz
	)

Execute R2C plan. This function is blocking and only returns after the transform is completed.

Note

For inplace transforms, data_out should point to the same memory address as data, AND the plan must have been created as inplace.

Parameters

plan	FFT plan created by accfft_plan_dft_3d_r2c.
data	Input data in spatial domain.
data_out	Output data in frequency domain.
timer	See Timing AccFFT for more details.
XYZ	a bit set field that determines which directions FFT should be executed

int accfft_local_size_dft_c2c_gpu	(	int *	n,
		int *	isize,
		int *	istart,
		int *	osize,
		int *	ostart,
		MPI_Comm	c_comm
	)

Get the local sizes of the distributed global data for a GPU C2C transform

Parameters

n	Integer array of size 3, corresponding to the global data size
isize	The size of the data that is locally distributed to the calling process
istart	The starting index of the data that locally resides on the calling process
osize	The output size of the data that locally resides on the calling process, after the C2C transform is finished
ostart	The output starting index of the data that locally resides on the calling process, after the R2C transform is finished
c_comm	Cartesian communicator returned by accfft_create_comm

Returns

int accfft_local_size_dft_r2c_gpu	(	int *	n,
		int *	isize,
		int *	istart,
		int *	osize,
		int *	ostart,
		MPI_Comm	c_comm,
		bool	inplace
	)

Get the local sizes of the distributed global data for a GPU R2C transform

Parameters

n	Integer array of size 3, corresponding to the global data size
isize	The size of the data that is locally distributed to the calling process
istart	The starting index of the data that locally resides on the calling process
osize	The output size of the data that locally resides on the calling process, after the R2C transform is finished
ostart	The output starting index of the data that locally resides on the calling process, after the R2C transform is finished
c_comm	Cartesian communicator returned by accfft_create_comm

Returns

accfft_plan_gpu* accfft_plan_dft_3d_c2c_gpu	(	int *	n,
		Complex *	data_d,
		Complex *	data_out_d,
		MPI_Comm	c_comm,
		unsigned	flags
	)

Creates a 3D C2C parallel FFT plan. If data_out point to the same location as the input data, then an inplace plan will be created. Otherwise the plan would be outplace.

Parameters

n	Integer array of size 3, corresponding to the global data size
data	Input data in spatial domain
data_out	Output data in frequency domain
c_comm	Cartesian communicator returned by accfft_create_comm
flags	AccFFT flags, See AccFFT Flags for more details.

Returns

accfft_plan_gpu* accfft_plan_dft_3d_r2c_gpu	(	int *	n,
		double *	data_d,
		double *	data_out_d,
		MPI_Comm	c_comm,
		unsigned	flags
	)

Creates a 3D R2C parallel FFT plan.If data_out point to the same location as the input data, then an inplace plan will be created. Otherwise the plan would be outplace.

Parameters

n	Integer array of size 3, corresponding to the global data size
data	Input data in spatial domain
data_out	Output data in frequency domain
c_comm	Cartesian communicator returned by accfft_create_comm
flags	AccFFT flags, See AccFFT Flags for more details.

Returns