2015-10-29 17:08:16 +08:00
|
|
|
!
|
2015-11-21 18:37:48 +08:00
|
|
|
! Copyright (C) Quantum ESPRESSO group
|
|
|
|
|
!
|
2015-10-29 17:08:16 +08:00
|
|
|
! This file is distributed under the terms of the
|
|
|
|
|
! GNU General Public License. See the file `License'
|
|
|
|
|
! in the root directory of the present distribution,
|
|
|
|
|
! or http://www.gnu.org/copyleft/gpl.txt .
|
|
|
|
|
!
|
|
|
|
|
!----------------------------------------------------------------------
|
|
|
|
|
! FFT base Module.
|
|
|
|
|
! Written by Carlo Cavazzoni, modified by Paolo Giannozzi
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
! Rewritten by Stefano de Gironcoli
|
2015-10-29 17:08:16 +08:00
|
|
|
!----------------------------------------------------------------------
|
|
|
|
|
!
|
|
|
|
|
!=----------------------------------------------------------------------=!
|
|
|
|
|
MODULE scatter_mod
|
|
|
|
|
!=----------------------------------------------------------------------=!
|
|
|
|
|
|
2016-08-10 05:38:57 +08:00
|
|
|
USE fft_types, ONLY: fft_type_descriptor
|
2016-11-25 22:11:28 +08:00
|
|
|
USE fft_param
|
2015-10-29 17:08:16 +08:00
|
|
|
|
|
|
|
|
IMPLICIT NONE
|
|
|
|
|
|
|
|
|
|
INTERFACE gather_grid
|
|
|
|
|
MODULE PROCEDURE gather_real_grid, gather_complex_grid
|
|
|
|
|
END INTERFACE
|
2015-11-21 18:37:48 +08:00
|
|
|
|
2015-10-29 17:08:16 +08:00
|
|
|
INTERFACE scatter_grid
|
|
|
|
|
MODULE PROCEDURE scatter_real_grid, scatter_complex_grid
|
|
|
|
|
END INTERFACE
|
|
|
|
|
|
|
|
|
|
SAVE
|
|
|
|
|
|
|
|
|
|
PRIVATE
|
|
|
|
|
|
2016-08-10 05:38:57 +08:00
|
|
|
PUBLIC :: fft_type_descriptor
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
PUBLIC :: gather_grid, scatter_grid
|
2017-08-05 05:04:39 +08:00
|
|
|
PUBLIC :: fft_scatter_xy, fft_scatter_yz, fft_scatter_tg, fft_scatter_tg_opt
|
2015-10-29 17:08:16 +08:00
|
|
|
PUBLIC :: cgather_sym, cgather_sym_many, cscatter_sym_many
|
|
|
|
|
|
|
|
|
|
!=----------------------------------------------------------------------=!
|
|
|
|
|
CONTAINS
|
|
|
|
|
!=----------------------------------------------------------------------=!
|
|
|
|
|
!
|
|
|
|
|
!
|
|
|
|
|
!-----------------------------------------------------------------------
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
SUBROUTINE fft_scatter_xy ( desc, f_in, f_aux, nxx_, isgn )
|
2015-10-29 17:08:16 +08:00
|
|
|
!-----------------------------------------------------------------------
|
|
|
|
|
!
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
! transpose of the fft xy planes across the desc%comm2 communicator
|
|
|
|
|
!
|
|
|
|
|
! a) From Y-oriented columns to X-oriented partial slices (isgn > 0)
|
|
|
|
|
! Active columns along the Y direction corresponding to a subset of the
|
|
|
|
|
! active X values and a range of Z values (in this order) are stored
|
|
|
|
|
! consecutively for each processor and are such that the subgroup owns
|
|
|
|
|
! all data for a range of Z values.
|
|
|
|
|
!
|
|
|
|
|
! The Y pencil -> X-oriented partial slices transposition is performed
|
|
|
|
|
! in the subgroup of processors (desc%comm2) owning this range of Z values.
|
|
|
|
|
!
|
|
|
|
|
! The transpose takes place in two steps:
|
|
|
|
|
! 1) on each processor the columns are sliced into sections along Y
|
|
|
|
|
! that are stored one after the other. On each processor, slices for
|
|
|
|
|
! processor "iproc2" are desc%nr2p(iproc2)*desc%nr1p(me2)*desc%my_nr3p big.
|
|
|
|
|
! 2) all processors communicate to exchange slices (all sectin of columns with
|
|
|
|
|
! Y in the slice belonging to "me" must be received, all the others
|
|
|
|
|
! must be sent to "iproc2")
|
|
|
|
|
!
|
|
|
|
|
! Finally one gets the "partial slice" representation: each processor has
|
|
|
|
|
! all the X values of desc%my_nr2p Y and desc%my_nr3p Z values.
|
|
|
|
|
! Data are organized with the X index running fastest, then Y, then Z.
|
|
|
|
|
!
|
|
|
|
|
! f_in contains the input Y columns, is destroyed on output
|
|
|
|
|
! f_aux contains the output X-oriented partial slices.
|
2015-10-29 17:08:16 +08:00
|
|
|
!
|
|
|
|
|
! b) From planes to columns (isgn < 0)
|
|
|
|
|
!
|
|
|
|
|
! Quite the same in the opposite direction
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
! f_aux contains the input X-oriented partial slices, is destroyed on output
|
|
|
|
|
! f_in contains the output Y columns.
|
2015-10-29 17:08:16 +08:00
|
|
|
!
|
|
|
|
|
IMPLICIT NONE
|
|
|
|
|
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
TYPE (fft_type_descriptor), INTENT(in) :: desc
|
|
|
|
|
INTEGER, INTENT(in) :: nxx_, isgn
|
2015-10-29 17:08:16 +08:00
|
|
|
COMPLEX (DP), INTENT(inout) :: f_in (nxx_), f_aux (nxx_)
|
|
|
|
|
|
2015-11-21 18:37:48 +08:00
|
|
|
#if defined(__MPI)
|
2015-10-29 17:08:16 +08:00
|
|
|
!
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
INTEGER :: ierr, me2, nproc2, iproc2, ncpx, my_nr2p, nr2px, ip, ip0
|
2017-08-05 18:05:26 +08:00
|
|
|
INTEGER :: i, it, j, k, kfrom, kdest, offset, ioff, mc, m1, m3, i1, icompact, sendsize
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
INTEGER, ALLOCATABLE :: ncp_(:), nr1p_(:), indx(:,:)
|
|
|
|
|
!
|
2017-09-14 21:24:20 +08:00
|
|
|
#if defined(__NON_BLOCKING_SCATTER)
|
|
|
|
|
INTEGER :: sh(desc%nproc2), rh(desc%nproc2)
|
|
|
|
|
#endif
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
me2 = desc%mype2 + 1
|
|
|
|
|
nproc2 = desc%nproc2 ; if ( abs(isgn) == 3 ) nproc2 = 1
|
|
|
|
|
|
|
|
|
|
! allocate auxiliary array for columns distribution
|
|
|
|
|
ALLOCATE ( ncp_(nproc2), nr1p_(nproc2), indx(desc%nr1x,nproc2) )
|
|
|
|
|
if ( abs (isgn) == 1 ) then ! It's a potential FFT
|
|
|
|
|
ncp_ = desc%nr1p * desc%my_nr3p
|
|
|
|
|
nr1p_= desc%nr1p
|
|
|
|
|
indx = desc%indp
|
|
|
|
|
my_nr2p=desc%my_nr2p
|
|
|
|
|
else if ( abs (isgn) == 2 ) then ! It's a wavefunction FFT
|
|
|
|
|
ncp_ = desc%nr1w * desc%my_nr3p
|
|
|
|
|
nr1p_= desc%nr1w
|
|
|
|
|
indx = desc%indw
|
|
|
|
|
my_nr2p=desc%my_nr2p
|
|
|
|
|
else if ( abs (isgn) == 3 ) then ! It's a wavefunction FFT with task group
|
|
|
|
|
ncp_ = desc%nr1w_tg * desc%my_nr3p!
|
|
|
|
|
nr1p_= desc%nr1w_tg !
|
|
|
|
|
indx(:,1) = desc%indw_tg !
|
|
|
|
|
my_nr2p=desc%nr2x ! in task group FFTs whole Y colums are distributed
|
|
|
|
|
end if
|
|
|
|
|
!
|
|
|
|
|
CALL start_clock ('fft_scatt_xy')
|
|
|
|
|
!
|
|
|
|
|
! calculate the message size
|
|
|
|
|
!
|
2017-08-05 18:05:26 +08:00
|
|
|
if ( abs (isgn) == 3 ) then
|
|
|
|
|
nr2px = desc%nr2x ! if it's a task group FFT whole planes are distributed
|
|
|
|
|
else
|
|
|
|
|
nr2px = MAXVAL ( desc%nr2p ) ! maximum number of Y values to be disributed
|
|
|
|
|
end if
|
|
|
|
|
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
ncpx = MAXVAL ( ncp_ ) ! maximum number of Y columns to be disributed
|
|
|
|
|
|
|
|
|
|
sendsize = ncpx * nr2px ! dimension of the scattered chunks (safe value)
|
2015-10-29 17:08:16 +08:00
|
|
|
|
|
|
|
|
ierr = 0
|
|
|
|
|
IF (isgn.gt.0) THEN
|
|
|
|
|
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
IF (nproc2==1) GO TO 10
|
2015-10-29 17:08:16 +08:00
|
|
|
!
|
|
|
|
|
! "forward" scatter from columns to planes
|
|
|
|
|
!
|
|
|
|
|
! step one: store contiguously the slices
|
|
|
|
|
!
|
2016-09-15 15:41:30 +08:00
|
|
|
offset = 0
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
DO iproc2 = 1, nproc2
|
|
|
|
|
kdest = ( iproc2 - 1 ) * sendsize
|
2016-09-15 15:41:30 +08:00
|
|
|
kfrom = offset
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
DO k = 1, ncp_(me2)
|
|
|
|
|
DO i = 1, desc%nr2p( iproc2 )
|
2015-10-29 17:08:16 +08:00
|
|
|
f_aux ( kdest + i ) = f_in ( kfrom + i )
|
|
|
|
|
ENDDO
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
kdest = kdest + nr2px
|
|
|
|
|
kfrom = kfrom + desc%nr2x
|
2015-10-29 17:08:16 +08:00
|
|
|
ENDDO
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
offset = offset + desc%nr2p( iproc2 )
|
2017-09-14 21:24:20 +08:00
|
|
|
#if defined(__NON_BLOCKING_SCATTER)
|
2017-09-14 23:13:08 +08:00
|
|
|
CALL mpi_isend( f_aux( (iproc2-1)*sendsize + 1 ), sendsize, &
|
|
|
|
|
MPI_DOUBLE_COMPLEX, iproc2-1, me2, desc%comm2, &
|
|
|
|
|
sh( iproc2 ), ierr )
|
2017-09-14 21:24:20 +08:00
|
|
|
#endif
|
2015-10-29 17:08:16 +08:00
|
|
|
ENDDO
|
|
|
|
|
!
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
! step two: communication across the nproc3 group
|
2015-10-29 17:08:16 +08:00
|
|
|
!
|
2017-09-14 21:24:20 +08:00
|
|
|
#if defined(__NON_BLOCKING_SCATTER)
|
|
|
|
|
DO iproc2 = 1, nproc2
|
|
|
|
|
!
|
|
|
|
|
! now post the receive
|
|
|
|
|
!
|
2017-09-14 23:13:08 +08:00
|
|
|
CALL mpi_irecv( f_in( (iproc2-1)*sendsize + 1 ), sendsize, &
|
|
|
|
|
MPI_DOUBLE_COMPLEX, iproc2-1, MPI_ANY_TAG, &
|
|
|
|
|
desc%comm2, rh( iproc2 ), ierr )
|
2017-09-14 21:24:20 +08:00
|
|
|
!IF( abs(ierr) /= 0 ) CALL fftx_error__ ('fft_scatter', ' forward receive info<>0', abs(ierr) )
|
|
|
|
|
ENDDO
|
|
|
|
|
|
|
|
|
|
call mpi_waitall( nproc2, sh, MPI_STATUSES_IGNORE, ierr )
|
|
|
|
|
!
|
|
|
|
|
#else
|
2017-09-14 23:13:08 +08:00
|
|
|
CALL mpi_alltoall (f_aux(1), sendsize, MPI_DOUBLE_COMPLEX, f_in(1), &
|
|
|
|
|
sendsize, MPI_DOUBLE_COMPLEX, desc%comm2, ierr)
|
2017-09-14 21:24:20 +08:00
|
|
|
!
|
2015-11-21 18:37:48 +08:00
|
|
|
IF( abs(ierr) /= 0 ) CALL fftx_error__ ('fft_scatter', 'info<>0', abs(ierr) )
|
2017-09-14 21:24:20 +08:00
|
|
|
#endif
|
2015-10-29 17:08:16 +08:00
|
|
|
!
|
|
|
|
|
10 CONTINUE
|
|
|
|
|
!
|
2017-08-20 00:43:31 +08:00
|
|
|
f_aux = (0.0_DP, 0.0_DP)
|
2015-10-29 17:08:16 +08:00
|
|
|
!
|
2017-09-14 21:24:20 +08:00
|
|
|
#if defined(__NON_BLOCKING_SCATTER)
|
|
|
|
|
if (nproc2 > 1) call mpi_waitall( nproc2, rh, MPI_STATUSES_IGNORE, ierr )
|
|
|
|
|
#endif
|
|
|
|
|
!
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
DO iproc2 = 1, nproc2
|
|
|
|
|
it = ( iproc2 - 1 ) * sendsize
|
|
|
|
|
DO i = 1, ncp_( iproc2 )
|
|
|
|
|
m3 = (i-1)/nr1p_(iproc2)+1 ; i1 = mod(i-1,nr1p_(iproc2))+1 ; m1 = indx(i1,iproc2)
|
2017-08-05 18:05:26 +08:00
|
|
|
icompact = m1 + (m3-1)*desc%nr1x*my_nr2p
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
DO j = 1, my_nr2p
|
2017-08-05 18:05:26 +08:00
|
|
|
!f_aux( m1 + (j-1)*desc%nr1x + (m3-1)*desc%nr1x*my_nr2p ) = f_in( j + it )
|
|
|
|
|
f_aux( icompact ) = f_in( j + it )
|
|
|
|
|
icompact = icompact + desc%nr1x
|
2015-10-29 17:08:16 +08:00
|
|
|
ENDDO
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
it = it + nr2px
|
2015-10-29 17:08:16 +08:00
|
|
|
ENDDO
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
ENDDO
|
2015-10-29 17:08:16 +08:00
|
|
|
|
|
|
|
|
ELSE
|
|
|
|
|
!
|
|
|
|
|
! "backward" scatter from planes to columns
|
|
|
|
|
!
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
DO iproc2 = 1, nproc2
|
|
|
|
|
it = ( iproc2 - 1 ) * sendsize
|
|
|
|
|
DO i = 1, ncp_( iproc2 )
|
|
|
|
|
m3 = (i-1)/nr1p_(iproc2)+1 ; i1 = mod(i-1,nr1p_(iproc2))+1 ; m1 = indx(i1,iproc2)
|
2017-08-05 18:05:26 +08:00
|
|
|
icompact = m1 + (m3-1)*desc%nr1x*my_nr2p
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
DO j = 1, my_nr2p
|
2017-08-05 18:05:26 +08:00
|
|
|
!f_in( j + it ) = f_aux( m1 + (j-1)*desc%nr1x + (m3-1)*desc%nr1x*my_nr2p )
|
|
|
|
|
f_in( j + it ) = f_aux( icompact )
|
|
|
|
|
icompact = icompact + desc%nr1x
|
2015-10-29 17:08:16 +08:00
|
|
|
ENDDO
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
it = it + nr2px
|
2015-10-29 17:08:16 +08:00
|
|
|
ENDDO
|
2017-09-14 21:24:20 +08:00
|
|
|
#if defined(__NON_BLOCKING_SCATTER)
|
2017-09-14 23:13:08 +08:00
|
|
|
IF( nproc2 > 1 ) CALL mpi_isend( f_in( ( iproc2 - 1 ) * sendsize + 1 ), &
|
|
|
|
|
sendsize, MPI_DOUBLE_COMPLEX, iproc2-1, me2, &
|
|
|
|
|
desc%comm2, sh( iproc2 ), ierr )
|
2017-09-14 21:24:20 +08:00
|
|
|
#endif
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
ENDDO
|
|
|
|
|
IF (nproc2==1) GO TO 20
|
2015-10-29 17:08:16 +08:00
|
|
|
|
|
|
|
|
!
|
|
|
|
|
! step two: communication
|
|
|
|
|
!
|
2017-09-14 21:24:20 +08:00
|
|
|
#if ! defined(__NON_BLOCKING_SCATTER)
|
2017-09-14 23:13:08 +08:00
|
|
|
CALL mpi_alltoall (f_in(1), sendsize, MPI_DOUBLE_COMPLEX, f_aux(1), &
|
|
|
|
|
sendsize, MPI_DOUBLE_COMPLEX, desc%comm2, ierr)
|
2015-10-29 17:08:16 +08:00
|
|
|
|
2015-11-21 18:37:48 +08:00
|
|
|
IF( abs(ierr) /= 0 ) CALL fftx_error__ ('fft_scatter', 'info<>0', abs(ierr) )
|
2017-09-14 21:24:20 +08:00
|
|
|
#else
|
|
|
|
|
DO iproc2 = 1, nproc2
|
2017-09-14 23:13:08 +08:00
|
|
|
CALL mpi_irecv( f_aux( (iproc2-1)*sendsize + 1 ), sendsize, &
|
|
|
|
|
MPI_DOUBLE_COMPLEX, iproc2-1, MPI_ANY_TAG, &
|
|
|
|
|
desc%comm2, rh(iproc2), ierr )
|
2017-09-14 21:24:20 +08:00
|
|
|
! IF( abs(ierr) /= 0 ) CALL fftx_error__ ('fft_scatter', ' backward receive info<>0', abs(ierr) )
|
|
|
|
|
ENDDO
|
|
|
|
|
|
|
|
|
|
call mpi_waitall( nproc2, sh, MPI_STATUSES_IGNORE, ierr )
|
|
|
|
|
call mpi_waitall( nproc2, rh, MPI_STATUSES_IGNORE, ierr )
|
|
|
|
|
#endif
|
2015-10-29 17:08:16 +08:00
|
|
|
!
|
|
|
|
|
! step one: store contiguously the columns
|
|
|
|
|
!
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
! ensures that no garbage is present in the output
|
|
|
|
|
! not useless ... clean the array to be returned from the garbage of previous A2A step
|
|
|
|
|
f_in = (0.0_DP, 0.0_DP) !
|
2015-10-29 17:08:16 +08:00
|
|
|
!
|
2016-09-15 15:41:30 +08:00
|
|
|
offset = 0
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
DO iproc2 = 1, nproc2
|
|
|
|
|
kdest = ( iproc2 - 1 ) * sendsize
|
2016-09-15 15:41:30 +08:00
|
|
|
kfrom = offset
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
DO k = 1, ncp_(me2)
|
|
|
|
|
DO i = 1, desc%nr2p( iproc2 )
|
2015-10-29 17:08:16 +08:00
|
|
|
f_in ( kfrom + i ) = f_aux ( kdest + i )
|
|
|
|
|
ENDDO
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
kdest = kdest + nr2px
|
|
|
|
|
kfrom = kfrom + desc%nr2x
|
2015-10-29 17:08:16 +08:00
|
|
|
ENDDO
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
offset = offset + desc%nr2p( iproc2 )
|
2015-10-29 17:08:16 +08:00
|
|
|
ENDDO
|
|
|
|
|
|
|
|
|
|
20 CONTINUE
|
|
|
|
|
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
|
2015-10-29 17:08:16 +08:00
|
|
|
ENDIF
|
|
|
|
|
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
DEALLOCATE ( ncp_ , nr1p_, indx )
|
|
|
|
|
CALL stop_clock ('fft_scatt_xy')
|
2015-10-29 17:08:16 +08:00
|
|
|
|
|
|
|
|
#endif
|
|
|
|
|
|
2016-03-10 19:05:20 +08:00
|
|
|
RETURN
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
99 format ( 20 ('(',2f12.9,')') )
|
2016-03-10 19:05:20 +08:00
|
|
|
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
END SUBROUTINE fft_scatter_xy
|
2016-03-10 19:05:20 +08:00
|
|
|
!
|
|
|
|
|
!-----------------------------------------------------------------------
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
SUBROUTINE fft_scatter_yz ( desc, f_in, f_aux, nxx_, isgn )
|
2016-03-10 19:05:20 +08:00
|
|
|
!-----------------------------------------------------------------------
|
|
|
|
|
!
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
! transpose of the fft yz planes across the desc%comm3 communicator
|
2016-03-10 19:05:20 +08:00
|
|
|
!
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
! a) From Z-oriented columns to Y-oriented colums (isgn > 0)
|
|
|
|
|
! Active columns (or sticks or pencils) along the Z direction for each
|
|
|
|
|
! processor are stored consecutively and are such that they correspond
|
|
|
|
|
! to a subset of the active X values.
|
2016-03-10 19:05:20 +08:00
|
|
|
!
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
! The pencil -> slices transposition is performed in the subgroup
|
|
|
|
|
! of processors (desc%comm3) owning these X values.
|
2016-03-10 19:05:20 +08:00
|
|
|
!
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
! The transpose takes place in two steps:
|
|
|
|
|
! 1) on each processor the columns are sliced into sections along Z
|
|
|
|
|
! that are stored one after the other. On each processor, slices for
|
|
|
|
|
! processor "iproc3" are desc%nr3p(iproc3)*desc%nsw/nsp(me) big.
|
|
|
|
|
! 2) all processors communicate to exchange slices (all columns with
|
|
|
|
|
! Z in the slice belonging to "me" must be received, all the others
|
|
|
|
|
! must be sent to "iproc3")
|
2016-03-10 19:05:20 +08:00
|
|
|
!
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
! Finally one gets the "slice" representation: each processor has
|
|
|
|
|
! desc%nr3p(mype3) Z values of all the active pencils along Y for the
|
|
|
|
|
! X values of the current group. Data are organized with the Y index
|
|
|
|
|
! running fastest, then the reordered X values, then Z.
|
2016-03-10 19:05:20 +08:00
|
|
|
!
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
! f_in contains the input Z columns, is destroyed on output
|
|
|
|
|
! f_aux contains the output Y colums.
|
2016-03-10 19:05:20 +08:00
|
|
|
!
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
! b) From planes to columns (isgn < 0)
|
|
|
|
|
!
|
|
|
|
|
! Quite the same in the opposite direction
|
|
|
|
|
! f_aux contains the input Y columns, is destroyed on output
|
|
|
|
|
! f_in contains the output Z columns.
|
2016-03-10 19:05:20 +08:00
|
|
|
!
|
|
|
|
|
IMPLICIT NONE
|
|
|
|
|
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
TYPE (fft_type_descriptor), INTENT(in) :: desc
|
|
|
|
|
INTEGER, INTENT(in) :: nxx_, isgn
|
2016-03-10 19:05:20 +08:00
|
|
|
COMPLEX (DP), INTENT(inout) :: f_in (nxx_), f_aux (nxx_)
|
|
|
|
|
|
|
|
|
|
#if defined(__MPI)
|
|
|
|
|
!
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
INTEGER :: ierr, me, me2, me2_start, me2_end, me3, nproc3, iproc3, ncpx, nr3px, ip, ip0
|
|
|
|
|
INTEGER :: i, it, k, kfrom, kdest, offset, ioff, mc, m1, m2, i1, sendsize
|
|
|
|
|
INTEGER, ALLOCATABLE :: ncp_(:), ir1p_(:)
|
|
|
|
|
INTEGER :: my_nr1p_
|
|
|
|
|
!
|
2017-09-14 21:24:20 +08:00
|
|
|
#if defined(__NON_BLOCKING_SCATTER)
|
|
|
|
|
INTEGER :: sh(desc%nproc3), rh(desc%nproc3)
|
|
|
|
|
#endif
|
|
|
|
|
!
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
me = desc%mype + 1
|
|
|
|
|
me2 = desc%mype2 + 1
|
|
|
|
|
me3 = desc%mype3 + 1
|
|
|
|
|
nproc3 = desc%nproc3
|
|
|
|
|
|
|
|
|
|
! allocate auxiliary array for columns distribution
|
|
|
|
|
ALLOCATE ( ncp_( desc%nproc), ir1p_(desc%nr1x) )
|
|
|
|
|
|
|
|
|
|
me2_start = me2 ; me2_end = me2
|
|
|
|
|
if ( abs (isgn) == 1 ) then ! It's a potential FFT
|
|
|
|
|
ncp_ = desc%nsp
|
|
|
|
|
ir1p_= desc%ir1p
|
|
|
|
|
else if ( abs (isgn) == 2 ) then ! It's a wavefunction FFT
|
|
|
|
|
ncp_ = desc%nsw
|
|
|
|
|
ir1p_= desc%ir1w
|
|
|
|
|
else if ( abs (isgn) == 3 ) then ! It's a wavefunction FFT with task group
|
|
|
|
|
ncp_ = desc%nsw
|
|
|
|
|
ir1p_= desc%ir1w_tg
|
|
|
|
|
me2_start = 1 ; me2_end = desc%nproc2
|
|
|
|
|
end if
|
|
|
|
|
my_nr1p_ = count (ir1p_ > 0)
|
|
|
|
|
!
|
|
|
|
|
CALL start_clock ('fft_scatt_yz')
|
|
|
|
|
!
|
|
|
|
|
! calculate the message size
|
|
|
|
|
!
|
|
|
|
|
nr3px = MAXVAL ( desc%nr3p ) ! maximum number of Z values to be exchanged
|
|
|
|
|
ncpx = MAXVAL ( ncp_ ) ! maximum number of Z columns to be exchanged
|
2017-08-05 18:05:26 +08:00
|
|
|
if (abs(isgn)==3) then
|
|
|
|
|
ncpx = ncpx * desc%nproc2 ! if it's a task group FFT groups of columns are exchanged
|
|
|
|
|
end if
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
|
|
|
|
|
sendsize = ncpx * nr3px ! dimension of the scattered chunks
|
2016-03-10 19:05:20 +08:00
|
|
|
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
ierr = 0
|
|
|
|
|
IF (isgn.gt.0) THEN
|
2016-03-10 19:05:20 +08:00
|
|
|
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
IF (nproc3==1) GO TO 10
|
2016-03-10 19:05:20 +08:00
|
|
|
!
|
|
|
|
|
! "forward" scatter from columns to planes
|
|
|
|
|
!
|
|
|
|
|
! step one: store contiguously the slices
|
|
|
|
|
!
|
2016-09-15 15:41:30 +08:00
|
|
|
offset = 0
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
DO iproc3 = 1, nproc3
|
|
|
|
|
kdest = ( iproc3 - 1 ) * sendsize
|
|
|
|
|
kfrom = offset
|
|
|
|
|
DO me2 = me2_start, me2_end
|
|
|
|
|
ip = desc%iproc(me2,me3)
|
|
|
|
|
DO k = 1, ncp_ (ip) ! was ncp_(me3)
|
|
|
|
|
DO i = 1, desc%nr3p( iproc3 )
|
2016-03-31 17:29:50 +08:00
|
|
|
f_aux ( kdest + i ) = f_in ( kfrom + i )
|
|
|
|
|
ENDDO
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
kdest = kdest + nr3px
|
|
|
|
|
kfrom = kfrom + desc%nr3x
|
2016-03-31 17:29:50 +08:00
|
|
|
ENDDO
|
|
|
|
|
ENDDO
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
offset = offset + desc%nr3p( iproc3 )
|
2017-09-14 21:24:20 +08:00
|
|
|
#if defined(__NON_BLOCKING_SCATTER)
|
2017-09-14 23:13:08 +08:00
|
|
|
CALL mpi_isend( f_aux( ( iproc3 - 1 ) * sendsize + 1 ), sendsize, &
|
|
|
|
|
MPI_DOUBLE_COMPLEX, iproc3-1, me3, desc%comm3, &
|
|
|
|
|
sh( iproc3 ), ierr )
|
2017-09-14 21:24:20 +08:00
|
|
|
#endif
|
2016-03-10 19:05:20 +08:00
|
|
|
ENDDO
|
|
|
|
|
!
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
! ensures that no garbage is present in the output
|
|
|
|
|
! useless; the later accessed elements are overwritten by the A2A step
|
2016-03-10 19:05:20 +08:00
|
|
|
!
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
! step two: communication across the nproc3 group
|
2016-03-10 19:05:20 +08:00
|
|
|
!
|
2017-09-14 21:24:20 +08:00
|
|
|
#if defined(__NON_BLOCKING_SCATTER)
|
|
|
|
|
DO iproc3 = 1, nproc3
|
|
|
|
|
!
|
|
|
|
|
! now post the receive
|
|
|
|
|
!
|
2017-09-14 23:13:08 +08:00
|
|
|
CALL mpi_irecv( f_in( (iproc3-1)*sendsize + 1 ), sendsize, &
|
|
|
|
|
MPI_DOUBLE_COMPLEX, iproc3-1, MPI_ANY_TAG, &
|
|
|
|
|
desc%comm3, rh( iproc3 ), ierr )
|
2017-09-14 21:24:20 +08:00
|
|
|
!IF( abs(ierr) /= 0 ) CALL fftx_error__ ('fft_scatter', ' forward receive info<>0', abs(ierr) )
|
|
|
|
|
!
|
|
|
|
|
!
|
|
|
|
|
ENDDO
|
|
|
|
|
|
|
|
|
|
call mpi_waitall( nproc3, sh, MPI_STATUSES_IGNORE, ierr )
|
|
|
|
|
!
|
|
|
|
|
#else
|
2017-09-14 23:13:08 +08:00
|
|
|
CALL mpi_alltoall (f_aux(1), sendsize, MPI_DOUBLE_COMPLEX, f_in(1), &
|
|
|
|
|
sendsize, MPI_DOUBLE_COMPLEX, desc%comm3, ierr)
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
|
|
|
|
|
IF( abs(ierr) /= 0 ) CALL fftx_error__ ('fft_scatter', 'info<>0', abs(ierr) )
|
2017-09-14 21:24:20 +08:00
|
|
|
#endif
|
2016-03-10 19:05:20 +08:00
|
|
|
!
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
10 CONTINUE
|
2016-03-10 19:05:20 +08:00
|
|
|
!
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
f_aux = (0.0_DP, 0.0_DP) !
|
2017-09-14 21:24:20 +08:00
|
|
|
#if defined(__NON_BLOCKING_SCATTER)
|
|
|
|
|
if (nproc3 > 1) call mpi_waitall( nproc3, rh, MPI_STATUSES_IGNORE, ierr )
|
|
|
|
|
#endif
|
2016-03-11 17:24:23 +08:00
|
|
|
!
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
DO iproc3 = 1, desc%nproc3
|
|
|
|
|
it = ( iproc3 - 1 ) * sendsize
|
|
|
|
|
DO me2 = me2_start, me2_end
|
|
|
|
|
ip = desc%iproc( me2, iproc3)
|
|
|
|
|
ioff = desc%iss(ip)
|
|
|
|
|
DO i = 1, ncp_( ip ) ! was ncp_(iproc3)
|
|
|
|
|
mc = desc%ismap( i + ioff ) ! this is m1+(m2-1)*nr1x of the current pencil
|
|
|
|
|
m1 = mod (mc-1,desc%nr1x) + 1 ; m2 = (mc-1)/desc%nr1x + 1
|
2017-08-05 05:04:39 +08:00
|
|
|
i1 = m2 + ( ir1p_(m1) - 1 ) * desc%nr2x
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
DO k = 1, desc%my_nr3p
|
2017-08-05 05:04:39 +08:00
|
|
|
f_aux( i1 ) = f_in( k + it )
|
|
|
|
|
i1 = i1 + desc%nr2x*my_nr1p_
|
2016-03-10 19:05:20 +08:00
|
|
|
ENDDO
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
it = it + nr3px
|
2016-03-10 19:05:20 +08:00
|
|
|
ENDDO
|
|
|
|
|
ENDDO
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
ENDDO
|
2016-03-10 19:05:20 +08:00
|
|
|
|
|
|
|
|
ELSE
|
|
|
|
|
!
|
|
|
|
|
! "backward" scatter from planes to columns
|
|
|
|
|
!
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
DO iproc3 = 1, desc%nproc3
|
|
|
|
|
it = ( iproc3 - 1 ) * sendsize
|
|
|
|
|
DO me2 = me2_start, me2_end
|
|
|
|
|
ip = desc%iproc(me2, iproc3)
|
|
|
|
|
ioff = desc%iss(ip)
|
|
|
|
|
DO i = 1, ncp_( ip )
|
|
|
|
|
mc = desc%ismap( i + ioff ) ! this is m1+(m2-1)*nr1x of the current pencil
|
|
|
|
|
m1 = mod (mc-1,desc%nr1x) + 1 ; m2 = (mc-1)/desc%nr1x + 1
|
2017-08-05 05:04:39 +08:00
|
|
|
i1 = m2 + ( ir1p_(m1) - 1 ) * desc%nr2x
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
DO k = 1, desc%my_nr3p
|
2017-08-05 05:04:39 +08:00
|
|
|
f_in( k + it ) = f_aux( i1 )
|
|
|
|
|
i1 = i1 + desc%nr2x * my_nr1p_
|
2016-03-31 17:29:50 +08:00
|
|
|
ENDDO
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
it = it + nr3px
|
2016-03-31 17:29:50 +08:00
|
|
|
ENDDO
|
2016-03-10 19:05:20 +08:00
|
|
|
ENDDO
|
2017-09-14 21:24:20 +08:00
|
|
|
#if defined(__NON_BLOCKING_SCATTER)
|
2017-09-14 23:13:08 +08:00
|
|
|
IF( nproc3 > 1 ) CALL mpi_isend( f_in( ( iproc3 - 1 ) * sendsize + 1 ), &
|
|
|
|
|
sendsize, MPI_DOUBLE_COMPLEX, iproc3-1, &
|
|
|
|
|
me3, desc%comm3, sh( iproc3 ), ierr )
|
2017-09-14 21:24:20 +08:00
|
|
|
#endif
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
ENDDO
|
|
|
|
|
|
|
|
|
|
IF( nproc3 == 1 ) GO TO 20
|
|
|
|
|
!
|
|
|
|
|
! step two: communication
|
|
|
|
|
!
|
2017-09-14 21:24:20 +08:00
|
|
|
#if ! defined(__NON_BLOCKING_SCATTER)
|
2017-09-14 23:13:08 +08:00
|
|
|
CALL mpi_alltoall (f_in(1), sendsize, MPI_DOUBLE_COMPLEX, f_aux(1), &
|
|
|
|
|
sendsize, MPI_DOUBLE_COMPLEX, desc%comm3, ierr)
|
2017-08-05 05:04:39 +08:00
|
|
|
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
IF( abs(ierr) /= 0 ) CALL fftx_error__ ('fft_scatter', 'info<>0', abs(ierr) )
|
2017-09-14 21:24:20 +08:00
|
|
|
#else
|
|
|
|
|
DO iproc3 = 1, desc%nproc3
|
2017-09-14 23:13:08 +08:00
|
|
|
CALL mpi_irecv( f_aux( (iproc3-1)*sendsize + 1 ), sendsize, &
|
|
|
|
|
MPI_DOUBLE_COMPLEX, iproc3-1, MPI_ANY_TAG, &
|
|
|
|
|
desc%comm3, rh(iproc3), ierr )
|
2017-09-14 21:24:20 +08:00
|
|
|
! IF( abs(ierr) /= 0 ) CALL fftx_error__ ('fft_scatter', ' backward receive info<>0', abs(ierr) )
|
|
|
|
|
ENDDO
|
|
|
|
|
|
|
|
|
|
call mpi_waitall( desc%nproc3, sh, MPI_STATUSES_IGNORE, ierr )
|
|
|
|
|
call mpi_waitall( desc%nproc3, rh, MPI_STATUSES_IGNORE, ierr )
|
|
|
|
|
#endif
|
2016-03-10 19:05:20 +08:00
|
|
|
!
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
! step one: store contiguously the columns
|
|
|
|
|
!
|
2016-09-15 15:41:30 +08:00
|
|
|
offset = 0
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
DO iproc3 = 1, nproc3
|
|
|
|
|
kdest = ( iproc3 - 1 ) * sendsize
|
|
|
|
|
kfrom = offset
|
|
|
|
|
DO me2 = me2_start, me2_end
|
|
|
|
|
ip = desc%iproc(me2,me3)
|
|
|
|
|
DO k = 1, ncp_ (ip)
|
|
|
|
|
DO i = 1, desc%nr3p( iproc3 )
|
2016-03-31 17:29:50 +08:00
|
|
|
f_in ( kfrom + i ) = f_aux ( kdest + i )
|
|
|
|
|
ENDDO
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
kdest = kdest + nr3px
|
|
|
|
|
kfrom = kfrom + desc%nr3x
|
2016-03-31 17:29:50 +08:00
|
|
|
ENDDO
|
|
|
|
|
ENDDO
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
offset = offset + desc%nr3p( iproc3 )
|
|
|
|
|
ENDDO
|
2017-08-18 01:25:28 +08:00
|
|
|
|
|
|
|
|
! clean extra array elements in each stick
|
|
|
|
|
|
|
|
|
|
IF( desc%nr3x /= desc%nr3 ) THEN
|
|
|
|
|
DO k = 1, ncp_ ( desc%mype+1 )
|
|
|
|
|
DO i = desc%nr3, desc%nr3x
|
|
|
|
|
f_in( (k-1)*desc%nr3x + i ) = 0.0d0
|
|
|
|
|
END DO
|
|
|
|
|
END DO
|
|
|
|
|
END IF
|
|
|
|
|
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
|
|
|
|
|
20 CONTINUE
|
2016-03-10 19:05:20 +08:00
|
|
|
|
|
|
|
|
ENDIF
|
|
|
|
|
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
DEALLOCATE ( ncp_ , ir1p_ )
|
|
|
|
|
CALL stop_clock ('fft_scatt_yz')
|
2016-03-10 19:05:20 +08:00
|
|
|
|
2015-10-29 17:08:16 +08:00
|
|
|
#endif
|
|
|
|
|
|
|
|
|
|
RETURN
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
98 format ( 10 ('(',2f12.9,')') )
|
|
|
|
|
99 format ( 20 ('(',2f12.9,')') )
|
2015-10-29 17:08:16 +08:00
|
|
|
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
END SUBROUTINE fft_scatter_yz
|
2015-10-29 17:08:16 +08:00
|
|
|
!
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
!-----------------------------------------------------------------------
|
|
|
|
|
SUBROUTINE fft_scatter_tg ( desc, f_in, f_aux, nxx_, isgn )
|
|
|
|
|
!-----------------------------------------------------------------------
|
|
|
|
|
!
|
|
|
|
|
! task group wavefunction redistribution
|
|
|
|
|
!
|
|
|
|
|
! a) (isgn >0 ) From many-wfc partial-plane arrangement to single-wfc whole-plane one
|
|
|
|
|
!
|
|
|
|
|
! b) (isgn <0 ) From single-wfc whole-plane arrangement to many-wfc partial-plane one
|
2016-04-08 17:21:11 +08:00
|
|
|
!
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
! in both cases:
|
|
|
|
|
! f_in contains the input data, is overwritten with the desired output
|
|
|
|
|
! f_aux is used as working array, may contain garbage in output
|
2016-04-08 17:21:11 +08:00
|
|
|
!
|
|
|
|
|
IMPLICIT NONE
|
|
|
|
|
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
TYPE (fft_type_descriptor), INTENT(in) :: desc
|
2016-04-08 17:21:11 +08:00
|
|
|
INTEGER, INTENT(in) :: nxx_, isgn
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
COMPLEX (DP), INTENT(inout) :: f_in (nxx_), f_aux (nxx_)
|
|
|
|
|
|
|
|
|
|
INTEGER :: ierr
|
|
|
|
|
|
|
|
|
|
CALL start_clock ('fft_scatt_tg')
|
|
|
|
|
|
|
|
|
|
if ( abs (isgn) /= 3 ) call fftx_error__ ('fft_scatter_tg', 'wrong call', 1 )
|
|
|
|
|
|
|
|
|
|
#if defined(__MPI)
|
|
|
|
|
!
|
2017-08-03 22:41:36 +08:00
|
|
|
f_aux = f_in
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
if ( isgn > 0 ) then
|
|
|
|
|
|
2017-08-03 22:41:36 +08:00
|
|
|
CALL MPI_ALLTOALLV( f_aux, desc%tg_snd, desc%tg_sdsp, MPI_DOUBLE_COMPLEX, &
|
|
|
|
|
f_in, desc%tg_rcv, desc%tg_rdsp, MPI_DOUBLE_COMPLEX, desc%comm2, ierr)
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
IF( ierr /= 0 ) CALL fftx_error__( 'fft_scatter_tg', ' alltoall error 1 ', abs(ierr) )
|
|
|
|
|
|
|
|
|
|
else
|
|
|
|
|
|
2017-08-03 22:41:36 +08:00
|
|
|
CALL MPI_ALLTOALLV( f_aux, desc%tg_rcv, desc%tg_rdsp, MPI_DOUBLE_COMPLEX, &
|
|
|
|
|
f_in, desc%tg_snd, desc%tg_sdsp, MPI_DOUBLE_COMPLEX, desc%comm2, ierr)
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
IF( ierr /= 0 ) CALL fftx_error__( 'fft_scatter_tg', ' alltoall error 2 ', abs(ierr) )
|
|
|
|
|
end if
|
|
|
|
|
|
|
|
|
|
#endif
|
|
|
|
|
CALL stop_clock ('fft_scatt_tg')
|
|
|
|
|
|
|
|
|
|
RETURN
|
|
|
|
|
99 format ( 20 ('(',2f12.9,')') )
|
|
|
|
|
|
|
|
|
|
END SUBROUTINE fft_scatter_tg
|
2017-08-05 05:04:39 +08:00
|
|
|
!
|
|
|
|
|
!-----------------------------------------------------------------------
|
|
|
|
|
SUBROUTINE fft_scatter_tg_opt ( desc, f_in, f_out, nxx_, isgn )
|
|
|
|
|
!-----------------------------------------------------------------------
|
|
|
|
|
!
|
|
|
|
|
! task group wavefunction redistribution
|
|
|
|
|
!
|
|
|
|
|
! a) (isgn >0 ) From many-wfc partial-plane arrangement to single-wfc whole-plane one
|
|
|
|
|
!
|
|
|
|
|
! b) (isgn <0 ) From single-wfc whole-plane arrangement to many-wfc partial-plane one
|
|
|
|
|
!
|
|
|
|
|
! in both cases:
|
|
|
|
|
! f_in contains the input data
|
|
|
|
|
! f_out contains the output data
|
|
|
|
|
!
|
|
|
|
|
IMPLICIT NONE
|
|
|
|
|
|
|
|
|
|
TYPE (fft_type_descriptor), INTENT(in) :: desc
|
|
|
|
|
INTEGER, INTENT(in) :: nxx_, isgn
|
|
|
|
|
COMPLEX (DP), INTENT(inout) :: f_in (nxx_), f_out (nxx_)
|
|
|
|
|
|
|
|
|
|
INTEGER :: ierr
|
|
|
|
|
|
|
|
|
|
CALL start_clock ('fft_scatt_tg')
|
|
|
|
|
|
|
|
|
|
if ( abs (isgn) /= 3 ) call fftx_error__ ('fft_scatter_tg', 'wrong call', 1 )
|
|
|
|
|
|
|
|
|
|
#if defined(__MPI)
|
|
|
|
|
!
|
|
|
|
|
if ( isgn > 0 ) then
|
|
|
|
|
|
|
|
|
|
CALL MPI_ALLTOALLV( f_in, desc%tg_snd, desc%tg_sdsp, MPI_DOUBLE_COMPLEX, &
|
|
|
|
|
f_out, desc%tg_rcv, desc%tg_rdsp, MPI_DOUBLE_COMPLEX, desc%comm2, ierr)
|
|
|
|
|
IF( ierr /= 0 ) CALL fftx_error__( 'fft_scatter_tg', ' alltoall error 1 ', abs(ierr) )
|
|
|
|
|
|
|
|
|
|
else
|
|
|
|
|
|
|
|
|
|
CALL MPI_ALLTOALLV( f_in, desc%tg_rcv, desc%tg_rdsp, MPI_DOUBLE_COMPLEX, &
|
|
|
|
|
f_out, desc%tg_snd, desc%tg_sdsp, MPI_DOUBLE_COMPLEX, desc%comm2, ierr)
|
|
|
|
|
IF( ierr /= 0 ) CALL fftx_error__( 'fft_scatter_tg', ' alltoall error 2 ', abs(ierr) )
|
|
|
|
|
end if
|
|
|
|
|
|
|
|
|
|
#endif
|
|
|
|
|
CALL stop_clock ('fft_scatt_tg')
|
|
|
|
|
|
|
|
|
|
RETURN
|
|
|
|
|
99 format ( 20 ('(',2f12.9,')') )
|
|
|
|
|
|
|
|
|
|
END SUBROUTINE fft_scatter_tg_opt
|
2016-04-08 17:21:11 +08:00
|
|
|
!
|
2015-10-29 17:08:16 +08:00
|
|
|
!----------------------------------------------------------------------------
|
|
|
|
|
SUBROUTINE gather_real_grid ( dfft, f_in, f_out )
|
|
|
|
|
!----------------------------------------------------------------------------
|
|
|
|
|
!
|
|
|
|
|
! ... gathers a distributed real-space FFT grid to dfft%root, that is,
|
|
|
|
|
! ... the first processor of input descriptor dfft - version for real arrays
|
|
|
|
|
!
|
|
|
|
|
! ... REAL*8 f_in = distributed variable (dfft%nnr)
|
|
|
|
|
! ... REAL*8 f_out = gathered variable (dfft%nr1x*dfft%nr2x*dfft%nr3x)
|
|
|
|
|
!
|
|
|
|
|
IMPLICIT NONE
|
|
|
|
|
!
|
|
|
|
|
REAL(DP), INTENT(in) :: f_in (:)
|
|
|
|
|
REAL(DP), INTENT(inout):: f_out(:)
|
2016-08-10 05:38:57 +08:00
|
|
|
TYPE ( fft_type_descriptor ), INTENT(IN) :: dfft
|
2015-10-29 17:08:16 +08:00
|
|
|
!
|
2015-11-21 18:37:48 +08:00
|
|
|
#if defined(__MPI)
|
2015-10-29 17:08:16 +08:00
|
|
|
!
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
INTEGER :: proc, info, offset_in, offset_aux, ir3
|
2015-10-29 17:08:16 +08:00
|
|
|
! ... the following are automatic arrays
|
|
|
|
|
INTEGER :: displs(0:dfft%nproc-1), recvcount(0:dfft%nproc-1)
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
REAL(DP), ALLOCATABLE :: f_aux(:)
|
2015-10-29 17:08:16 +08:00
|
|
|
!
|
|
|
|
|
IF( size( f_in ) < dfft%nnr ) &
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
CALL fftx_error__( ' gather_real_grid ', ' f_in too small ', dfft%nnr-size( f_in ) )
|
|
|
|
|
!
|
|
|
|
|
CALL start_clock( 'rgather_grid' )
|
|
|
|
|
|
|
|
|
|
!write (6,*) 'gather grid ok 0 ', dfft%nproc, dfft%nproc2, dfft%nproc3
|
|
|
|
|
ALLOCATE ( f_aux(dfft%nr1x * dfft%nr2x * dfft%my_nr3p) )
|
|
|
|
|
! 1) gather within the comm2 communicator
|
|
|
|
|
displs = 0
|
|
|
|
|
DO proc =0, ( dfft%nproc2 -1 )
|
|
|
|
|
recvcount(proc) = dfft%nr1x * dfft%nr2p(proc+1)
|
|
|
|
|
if (proc > 0) displs(proc) = displs(proc-1) + recvcount(proc-1)
|
|
|
|
|
end do
|
|
|
|
|
offset_in = 1; offset_aux = 1
|
|
|
|
|
do ir3 = 1, dfft%my_nr3p
|
|
|
|
|
!write (6,*) 'gather grid ok 1 ir3=', ir3
|
|
|
|
|
info = 0
|
|
|
|
|
CALL MPI_GATHERV( f_in(offset_in) , recvcount(dfft%mype2), MPI_DOUBLE_PRECISION, &
|
|
|
|
|
f_aux(offset_aux), recvcount, displs, MPI_DOUBLE_PRECISION, dfft%root, &
|
|
|
|
|
dfft%comm2, info )
|
|
|
|
|
CALL fftx_error__( 'gather_real_grid', 'info<>0', info )
|
|
|
|
|
!write (6,*) 'gather grid ok 2 ir3=', ir3
|
|
|
|
|
offset_in = offset_in + dfft%nr1x * dfft%my_nr2p
|
|
|
|
|
offset_aux = offset_aux + dfft%nr1x * dfft%nr2
|
|
|
|
|
end do
|
|
|
|
|
|
|
|
|
|
! 2) gather within the comm3 communicator
|
|
|
|
|
displs = 0
|
|
|
|
|
DO proc = 0, ( dfft%nproc3 - 1 )
|
|
|
|
|
recvcount(proc) = dfft%nr1x * dfft%nr2x * dfft%nr3p(proc+1)
|
|
|
|
|
if (proc > 0) displs(proc) = displs(proc-1) + recvcount(proc-1)
|
2015-10-29 17:08:16 +08:00
|
|
|
ENDDO
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
info = 0
|
|
|
|
|
!write (6,*) 'gather grid ok 3'
|
|
|
|
|
CALL MPI_GATHERV( f_aux, recvcount(dfft%mype3), MPI_DOUBLE_PRECISION, &
|
|
|
|
|
f_out, recvcount, displs, MPI_DOUBLE_PRECISION, dfft%root, &
|
|
|
|
|
dfft%comm3, info )
|
|
|
|
|
!write (6,*) 'gather grid ok 4'
|
|
|
|
|
CALL fftx_error__( ' gather_real_grid', 'info<>0', info )
|
2015-10-29 17:08:16 +08:00
|
|
|
!
|
|
|
|
|
! ... the following check should be performed only on processor dfft%root
|
|
|
|
|
! ... otherwise f_out must be allocated on all processors even if not used
|
|
|
|
|
!
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
info = size( f_out ) - displs( dfft%nproc3-1 ) - recvcount( dfft%nproc3-1 )
|
2015-10-29 17:08:16 +08:00
|
|
|
IF( info < 0 ) &
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
CALL fftx_error__( ' gather_real_grid ', ' f_out too small ', -info )
|
2015-10-29 17:08:16 +08:00
|
|
|
!
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
DEALLOCATE ( f_aux )
|
2015-10-29 17:08:16 +08:00
|
|
|
!
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
CALL stop_clock( 'rgather_grid' )
|
2015-10-29 17:08:16 +08:00
|
|
|
!
|
|
|
|
|
#else
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
CALL fftx_error__(' gather_real_grid', 'do not use in serial execution', 1)
|
2015-10-29 17:08:16 +08:00
|
|
|
#endif
|
|
|
|
|
!
|
|
|
|
|
RETURN
|
|
|
|
|
!
|
|
|
|
|
END SUBROUTINE gather_real_grid
|
|
|
|
|
|
|
|
|
|
!----------------------------------------------------------------------------
|
|
|
|
|
SUBROUTINE gather_complex_grid ( dfft, f_in, f_out )
|
|
|
|
|
!----------------------------------------------------------------------------
|
|
|
|
|
!
|
|
|
|
|
! ... gathers a distributed real-space FFT grid to dfft%root, that is,
|
|
|
|
|
! ... the first processor of input descriptor dfft - complex arrays
|
|
|
|
|
!
|
|
|
|
|
! ... COMPLEX*16 f_in = distributed variable (dfft%nnr)
|
|
|
|
|
! ... COMPLEX*16 f_out = gathered variable (dfft%nr1x*dfft%nr2x*dfft%nr3x)
|
|
|
|
|
!
|
|
|
|
|
IMPLICIT NONE
|
|
|
|
|
!
|
|
|
|
|
COMPLEX(DP), INTENT(in) :: f_in (:)
|
|
|
|
|
COMPLEX(DP), INTENT(inout):: f_out(:)
|
2016-08-10 05:38:57 +08:00
|
|
|
TYPE ( fft_type_descriptor ), INTENT(IN) :: dfft
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
COMPLEX(DP), ALLOCATABLE :: f_aux(:)
|
2015-10-29 17:08:16 +08:00
|
|
|
!
|
2015-11-21 18:37:48 +08:00
|
|
|
#if defined(__MPI)
|
2015-10-29 17:08:16 +08:00
|
|
|
!
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
INTEGER :: proc, info, offset_in, offset_aux, ir3
|
2015-10-29 17:08:16 +08:00
|
|
|
! ... the following are automatic arrays
|
|
|
|
|
INTEGER :: displs(0:dfft%nproc-1), recvcount(0:dfft%nproc-1)
|
|
|
|
|
!
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
|
|
|
|
|
CALL start_clock( 'cgather_grid' )
|
|
|
|
|
!write (*,*) 'gcgather_grid size(f_in),dfft%nnr',size(f_in), dfft%nnr ; FLUSH(6)
|
2015-10-29 17:08:16 +08:00
|
|
|
IF( 2*size( f_in ) < dfft%nnr ) &
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
CALL fftx_error__( ' gather_complex_grid ', ' f_in too small ', dfft%nnr-size( f_in ) )
|
|
|
|
|
!
|
|
|
|
|
!write (6,*) 'gather grid ok 0 ', dfft%nproc, dfft%nproc2, dfft%nproc3
|
|
|
|
|
ALLOCATE ( f_aux(dfft%nr1x * dfft%nr2x * dfft%my_nr3p ) )
|
|
|
|
|
! 1) gather within the comm2 communicator
|
|
|
|
|
displs = 0
|
|
|
|
|
DO proc =0, ( dfft%nproc2 -1 )
|
|
|
|
|
recvcount(proc) = 2 * dfft%nr1x * dfft%nr2p(proc+1)
|
|
|
|
|
if (proc > 0) displs(proc) = displs(proc-1) + recvcount(proc-1)
|
|
|
|
|
end do
|
|
|
|
|
offset_in = 1; offset_aux = 1
|
|
|
|
|
do ir3 = 1, dfft%my_nr3p
|
|
|
|
|
info = 0
|
|
|
|
|
!write (6,*) 'gather grid ok 1 ir3=', ir3
|
|
|
|
|
CALL MPI_GATHERV( f_in(offset_in) , recvcount(dfft%mype2), MPI_DOUBLE_PRECISION, &
|
|
|
|
|
f_aux(offset_aux), recvcount, displs, MPI_DOUBLE_PRECISION, dfft%root, &
|
|
|
|
|
dfft%comm2, info )
|
|
|
|
|
CALL fftx_error__( 'gather_complex_grid', 'info<>0', info )
|
|
|
|
|
!write (6,*) 'gather grid ok 2 ir3=', ir3
|
|
|
|
|
offset_in = offset_in + dfft%nr1x * dfft%my_nr2p
|
|
|
|
|
offset_aux = offset_aux + dfft%nr1x * dfft%nr2
|
|
|
|
|
end do
|
|
|
|
|
|
|
|
|
|
! 2) gather within the comm3 communicator
|
|
|
|
|
displs = 0
|
|
|
|
|
DO proc = 0, ( dfft%nproc3 - 1 )
|
|
|
|
|
recvcount(proc) = 2 * dfft%nr1x * dfft%nr2x * dfft%nr3p(proc+1)
|
|
|
|
|
if (proc > 0) displs(proc) = displs(proc-1) + recvcount(proc-1)
|
2015-10-29 17:08:16 +08:00
|
|
|
ENDDO
|
|
|
|
|
!
|
|
|
|
|
! ... the following check should be performed only on processor dfft%root
|
|
|
|
|
! ... otherwise f_out must be allocated on all processors even if not used
|
|
|
|
|
!
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
!write (*,*) 'gcgather_grid 2*size(f_out)',2*size(f_out) ; FLUSH(6)
|
|
|
|
|
!write (*,*) 'gcgather_grid displ+recv',dfft%nproc3, displs(dfft%nproc3-1) + recvcount(dfft%nproc3-1); FLUSH(6)
|
|
|
|
|
info = 2*size( f_out ) - displs( dfft%nproc3 - 1 ) - recvcount( dfft%nproc3-1 ) ; FLUSH(6)
|
|
|
|
|
IF( info < 0 ) CALL fftx_error__( ' gather_complex_grid ', ' f_out too small ', -info )
|
|
|
|
|
|
2015-10-29 17:08:16 +08:00
|
|
|
info = 0
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
!write (6,*) 'gather grid ok 3'
|
|
|
|
|
CALL MPI_GATHERV( f_aux, recvcount(dfft%mype3), MPI_DOUBLE_PRECISION, &
|
|
|
|
|
f_out, recvcount, displs, MPI_DOUBLE_PRECISION, dfft%root, &
|
|
|
|
|
dfft%comm3, info )
|
|
|
|
|
!write (6,*) 'gather grid ok 4'
|
|
|
|
|
CALL fftx_error__( 'gather_complex_grid', 'info<>0', info )
|
2015-10-29 17:08:16 +08:00
|
|
|
!
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
DEALLOCATE ( f_aux )
|
2015-10-29 17:08:16 +08:00
|
|
|
!
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
CALL stop_clock( 'cgather_grid' )
|
2015-10-29 17:08:16 +08:00
|
|
|
!
|
|
|
|
|
#else
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
CALL fftx_error__('gather_complex_grid', 'do not use in serial execution', 1)
|
2015-10-29 17:08:16 +08:00
|
|
|
#endif
|
|
|
|
|
!
|
|
|
|
|
RETURN
|
|
|
|
|
!
|
|
|
|
|
END SUBROUTINE gather_complex_grid
|
|
|
|
|
|
|
|
|
|
!----------------------------------------------------------------------------
|
|
|
|
|
SUBROUTINE scatter_real_grid ( dfft, f_in, f_out )
|
|
|
|
|
!----------------------------------------------------------------------------
|
|
|
|
|
!
|
|
|
|
|
! ... scatters a real-space FFT grid from dfft%root, first processor of
|
|
|
|
|
! ... input descriptor dfft, to all others - opposite of "gather_grid"
|
|
|
|
|
!
|
|
|
|
|
! ... REAL*8 f_in = gathered variable (dfft%nr1x*dfft%nr2x*dfft%nr3x)
|
|
|
|
|
! ... REAL*8 f_out = distributed variable (dfft%nnr)
|
|
|
|
|
!
|
|
|
|
|
IMPLICIT NONE
|
|
|
|
|
!
|
|
|
|
|
REAL(DP), INTENT(in) :: f_in (:)
|
|
|
|
|
REAL(DP), INTENT(inout):: f_out(:)
|
2016-08-10 05:38:57 +08:00
|
|
|
TYPE ( fft_type_descriptor ), INTENT(IN) :: dfft
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
REAL(DP), ALLOCATABLE :: f_aux(:)
|
2015-10-29 17:08:16 +08:00
|
|
|
!
|
2015-11-21 18:37:48 +08:00
|
|
|
#if defined(__MPI)
|
2015-10-29 17:08:16 +08:00
|
|
|
!
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
INTEGER :: proc, info, offset_in, offset_aux, ir3
|
2015-10-29 17:08:16 +08:00
|
|
|
! ... the following are automatic arrays
|
|
|
|
|
INTEGER :: displs(0:dfft%nproc-1), sendcount(0:dfft%nproc-1)
|
|
|
|
|
!
|
|
|
|
|
!
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
CALL start_clock( 'rscatter_grid' )
|
2015-10-29 17:08:16 +08:00
|
|
|
!
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
!write (6,*) 'scatter grid ok 0'
|
|
|
|
|
ALLOCATE ( f_aux(dfft%nr1x * dfft%nr2x * dfft%my_nr3p) )
|
|
|
|
|
! 1) scatter within the comm3 communicator
|
|
|
|
|
displs = 0
|
|
|
|
|
DO proc = 0, ( dfft%nproc3 - 1 )
|
|
|
|
|
sendcount(proc) = dfft%nr1x * dfft%nr2x * dfft%nr3p(proc+1)
|
|
|
|
|
if (proc > 0) displs(proc) = displs(proc-1) + sendcount(proc-1)
|
2015-10-29 17:08:16 +08:00
|
|
|
ENDDO
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
info = size( f_in ) - displs( dfft%nproc3 - 1 ) - sendcount( dfft%nproc3 - 1 )
|
|
|
|
|
IF( info < 0 ) CALL fftx_error__( ' scatter_real_grid ', ' f_in too small ', -info )
|
2015-10-29 17:08:16 +08:00
|
|
|
info = 0
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
!write (6,*) 'scatter grid ok 1'
|
2015-10-29 17:08:16 +08:00
|
|
|
CALL MPI_SCATTERV( f_in, sendcount, displs, MPI_DOUBLE_PRECISION, &
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
f_aux, sendcount(dfft%mype3), MPI_DOUBLE_PRECISION, &
|
|
|
|
|
dfft%root, dfft%comm3, info )
|
|
|
|
|
!write (6,*) 'scatter grid ok 2'
|
|
|
|
|
CALL fftx_error__( 'scatter_real_grid', 'info<>0', info )
|
|
|
|
|
|
|
|
|
|
! 2) scatter within the comm2 communicator
|
|
|
|
|
IF( size( f_out ) < dfft%nnr ) &
|
|
|
|
|
CALL fftx_error__( ' scatter_real_grid ', ' f_out too small ', dfft%nnr-size( f_out ) )
|
|
|
|
|
!
|
|
|
|
|
displs = 0 ; f_out = 0.0D0
|
|
|
|
|
DO proc =0, ( dfft%nproc2 -1 )
|
|
|
|
|
sendcount(proc) = dfft%nr1x * dfft%nr2p(proc+1)
|
|
|
|
|
if (proc > 0) displs(proc) = displs(proc-1) + sendcount(proc-1)
|
|
|
|
|
end do
|
|
|
|
|
offset_in = 1 ; offset_aux = 1
|
|
|
|
|
do ir3 = 1, dfft%my_nr3p
|
|
|
|
|
info = 0
|
|
|
|
|
!write (6,*) 'scatter grid ok 3, ir3=', ir3
|
|
|
|
|
CALL MPI_SCATTERV( f_aux(offset_aux), sendcount, displs, MPI_DOUBLE_PRECISION, &
|
|
|
|
|
f_out(offset_in), sendcount(dfft%mype2), MPI_DOUBLE_PRECISION, &
|
|
|
|
|
dfft%root, dfft%comm2, info )
|
|
|
|
|
!write (6,*) 'scatter grid ok 4, ir3=', ir3
|
|
|
|
|
CALL fftx_error__( 'scatter_real_grid', 'info<>0', info )
|
|
|
|
|
offset_in = offset_in + dfft%nr1x * dfft%my_nr2p
|
|
|
|
|
offset_aux = offset_aux + dfft%nr1x * dfft%nr2
|
|
|
|
|
end do
|
|
|
|
|
!
|
|
|
|
|
DEALLOCATE ( f_aux )
|
|
|
|
|
!
|
|
|
|
|
CALL stop_clock( 'rscatter_grid' )
|
2015-10-29 17:08:16 +08:00
|
|
|
!
|
|
|
|
|
#else
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
CALL fftx_error__('scatter_real_grid', 'do not use in serial execution', 1)
|
2015-10-29 17:08:16 +08:00
|
|
|
#endif
|
|
|
|
|
!
|
|
|
|
|
RETURN
|
|
|
|
|
!
|
|
|
|
|
END SUBROUTINE scatter_real_grid
|
|
|
|
|
!----------------------------------------------------------------------------
|
|
|
|
|
SUBROUTINE scatter_complex_grid ( dfft, f_in, f_out )
|
|
|
|
|
!----------------------------------------------------------------------------
|
|
|
|
|
!
|
|
|
|
|
! ... scatters a real-space FFT grid from dfft%root, first processor of
|
|
|
|
|
! ... input descriptor dfft, to all others - opposite of "gather_grid"
|
|
|
|
|
!
|
|
|
|
|
! ... COMPLEX*16 f_in = gathered variable (dfft%nr1x*dfft%nr2x*dfft%nr3x)
|
|
|
|
|
! ... COMPLEX*16 f_out = distributed variable (dfft%nnr)
|
|
|
|
|
!
|
|
|
|
|
IMPLICIT NONE
|
|
|
|
|
!
|
|
|
|
|
COMPLEX(DP), INTENT(in) :: f_in (:)
|
|
|
|
|
COMPLEX(DP), INTENT(inout):: f_out(:)
|
2016-08-10 05:38:57 +08:00
|
|
|
TYPE ( fft_type_descriptor ), INTENT(IN) :: dfft
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
COMPLEX(DP), ALLOCATABLE :: f_aux(:)
|
2015-10-29 17:08:16 +08:00
|
|
|
!
|
2015-11-21 18:37:48 +08:00
|
|
|
#if defined(__MPI)
|
2015-10-29 17:08:16 +08:00
|
|
|
!
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
INTEGER :: proc, info, offset_in, offset_aux, ir3
|
2015-10-29 17:08:16 +08:00
|
|
|
! ... the following are automatic arrays
|
|
|
|
|
INTEGER :: displs(0:dfft%nproc-1), sendcount(0:dfft%nproc-1)
|
|
|
|
|
!
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
CALL start_clock( 'cscatter_grid' )
|
2015-10-29 17:08:16 +08:00
|
|
|
!
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
!write (6,*) 'scatter grid ok 0'
|
|
|
|
|
ALLOCATE ( f_aux(dfft%nr1x * dfft%nr2x * dfft%my_nr3p ) )
|
|
|
|
|
! 1) scatter within the comm3 communicator
|
|
|
|
|
displs = 0
|
|
|
|
|
DO proc = 0, ( dfft%nproc3 - 1 )
|
|
|
|
|
sendcount(proc) = 2 * dfft%nr1x * dfft%nr2x * dfft%nr3p(proc+1)
|
|
|
|
|
if (proc > 0) displs(proc) = displs(proc-1) + sendcount(proc-1)
|
2015-10-29 17:08:16 +08:00
|
|
|
ENDDO
|
|
|
|
|
!
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
!write(*,*) 'cscatter_grid 2*size(f_in) ', 2*size(f_in); FLUSH(6)
|
|
|
|
|
!write(*,*) 'cscatter_grid displ+send ', dfft%nproc3, displs(dfft%nproc3-1) + sendcount(dfft%nproc3-1); FLUSH(6)
|
|
|
|
|
info = 2*size( f_in ) - displs( dfft%nproc3 - 1 ) - sendcount( dfft%nproc3 - 1 )
|
2015-10-29 17:08:16 +08:00
|
|
|
IF( info < 0 ) &
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
CALL fftx_error__( ' scatter_complex_grid ', ' f_in too small ', -info )
|
2015-10-29 17:08:16 +08:00
|
|
|
!
|
|
|
|
|
info = 0
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
!write (6,*) 'scatter grid ok 1'
|
2015-10-29 17:08:16 +08:00
|
|
|
CALL MPI_SCATTERV( f_in, sendcount, displs, MPI_DOUBLE_PRECISION, &
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
f_aux, sendcount(dfft%mype3), MPI_DOUBLE_PRECISION, &
|
|
|
|
|
dfft%root, dfft%comm3, info )
|
|
|
|
|
!write (6,*) 'scatter grid ok 2'
|
|
|
|
|
CALL fftx_error__( ' scatter_complex_grid', 'info<>0', info )
|
|
|
|
|
|
|
|
|
|
! 2) scatter within the comm2 communicator
|
|
|
|
|
!write(*,*) 'cscatter_grid size(f_out), dfft%nnr ', size(f_out),dfft%nnr; FLUSH(6)
|
|
|
|
|
IF( size( f_out ) < dfft%nnr ) &
|
|
|
|
|
CALL fftx_error__( ' scatter_complex_grid ', ' f_out too small ', dfft%nnr-size( f_out ) )
|
|
|
|
|
!
|
|
|
|
|
displs = 0 ; f_out = 0.0D0
|
|
|
|
|
DO proc =0, ( dfft%nproc2 -1 )
|
|
|
|
|
sendcount(proc) = 2 * dfft%nr1x * dfft%nr2p(proc+1)
|
|
|
|
|
if (proc > 0) displs(proc) = displs(proc-1) + sendcount(proc-1)
|
|
|
|
|
end do
|
|
|
|
|
offset_in = 1 ; offset_aux = 1
|
|
|
|
|
do ir3 = 1, dfft%my_nr3p
|
|
|
|
|
info = 0
|
|
|
|
|
!write (6,*) 'scatter grid ok 3, ir3=', ir3
|
|
|
|
|
CALL MPI_SCATTERV( f_aux(offset_aux), sendcount, displs, MPI_DOUBLE_PRECISION, &
|
|
|
|
|
f_out(offset_in), sendcount(dfft%mype2), MPI_DOUBLE_PRECISION, &
|
|
|
|
|
dfft%root, dfft%comm2, info )
|
|
|
|
|
!write (6,*) 'scatter grid ok 4, ir3=', ir3
|
|
|
|
|
CALL fftx_error__( 'scatter_complex_grid', 'info<>0', info )
|
|
|
|
|
offset_in = offset_in + dfft%nr1x * dfft%my_nr2p
|
|
|
|
|
offset_aux = offset_aux + dfft%nr1x * dfft%nr2x
|
|
|
|
|
end do
|
2015-10-29 17:08:16 +08:00
|
|
|
!
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
! ... the following check should be performed only on processor dfft%root
|
|
|
|
|
! ... otherwise f_in must be allocated on all processors even if not used
|
2015-10-29 17:08:16 +08:00
|
|
|
!
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
DEALLOCATE ( f_aux )
|
2015-10-29 17:08:16 +08:00
|
|
|
!
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
CALL stop_clock( 'cscatter_grid' )
|
2015-10-29 17:08:16 +08:00
|
|
|
!
|
|
|
|
|
#else
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
CALL fftx_error__('scatter_complex_grid', 'do not use in serial execution', 1)
|
2015-10-29 17:08:16 +08:00
|
|
|
#endif
|
|
|
|
|
!
|
|
|
|
|
RETURN
|
|
|
|
|
!
|
|
|
|
|
END SUBROUTINE scatter_complex_grid
|
|
|
|
|
!
|
|
|
|
|
! ... "gather"-like subroutines
|
|
|
|
|
!
|
|
|
|
|
!-----------------------------------------------------------------------
|
|
|
|
|
SUBROUTINE cgather_sym( dfftp, f_in, f_out )
|
|
|
|
|
!-----------------------------------------------------------------------
|
|
|
|
|
!
|
|
|
|
|
! ... gather complex data for symmetrization (used in phonon code)
|
|
|
|
|
! ... Differs from gather_grid because mpi_allgatherv is used instead
|
|
|
|
|
! ... of mpi_gatherv - all data is gathered on ALL processors
|
|
|
|
|
! ... COMPLEX*16 f_in = distributed variable (nrxx)
|
|
|
|
|
! ... COMPLEX*16 f_out = gathered variable (nr1x*nr2x*nr3x)
|
|
|
|
|
!
|
|
|
|
|
IMPLICIT NONE
|
|
|
|
|
!
|
2016-08-10 05:38:57 +08:00
|
|
|
TYPE (fft_type_descriptor), INTENT(in) :: dfftp
|
2015-10-29 17:08:16 +08:00
|
|
|
COMPLEX(DP) :: f_in( : ), f_out(:)
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
COMPLEX(DP), ALLOCATABLE :: f_aux(:)
|
2015-10-29 17:08:16 +08:00
|
|
|
!
|
2015-11-21 18:37:48 +08:00
|
|
|
#if defined(__MPI)
|
2015-10-29 17:08:16 +08:00
|
|
|
!
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
INTEGER :: proc, info, offset_in, offset_aux, ir3
|
|
|
|
|
! ... the following are automatic arrays
|
2015-10-29 17:08:16 +08:00
|
|
|
INTEGER :: displs(0:dfftp%nproc-1), recvcount(0:dfftp%nproc-1)
|
|
|
|
|
!
|
|
|
|
|
CALL start_clock( 'cgather' )
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
|
|
|
|
|
ALLOCATE ( f_aux(dfftp%nr1x * dfftp%nr2x * dfftp%my_nr3p ) )
|
2015-10-29 17:08:16 +08:00
|
|
|
!
|
2015-11-21 18:37:48 +08:00
|
|
|
CALL MPI_BARRIER( dfftp%comm, info )
|
2015-10-29 17:08:16 +08:00
|
|
|
!
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
! 1) gather within the comm2 communicator
|
|
|
|
|
displs = 0
|
|
|
|
|
DO proc =0, ( dfftp%nproc2 -1 )
|
|
|
|
|
recvcount(proc) = 2 * dfftp%nr1x * dfftp%nr2p(proc+1)
|
|
|
|
|
if (proc > 0) displs(proc) = displs(proc-1) + recvcount(proc-1)
|
|
|
|
|
end do
|
|
|
|
|
offset_in = 1; offset_aux = 1
|
|
|
|
|
do ir3 = 1, dfftp%my_nr3p
|
|
|
|
|
info = 0
|
|
|
|
|
CALL MPI_ALLGATHERV( f_in(offset_in), recvcount(dfftp%mype2), MPI_DOUBLE_PRECISION, &
|
|
|
|
|
f_aux(offset_aux),recvcount, displs, MPI_DOUBLE_PRECISION, &
|
|
|
|
|
dfftp%comm2, info )
|
|
|
|
|
CALL fftx_error__( 'cgather_sym', 'info<>0', info )
|
|
|
|
|
offset_in = offset_in + dfftp%nr1x * dfftp%my_nr2p
|
|
|
|
|
offset_aux = offset_aux + dfftp%nr1x * dfftp%nr2x
|
|
|
|
|
end do
|
|
|
|
|
|
|
|
|
|
! 2) gather within the comm3 communicator
|
|
|
|
|
displs = 0
|
|
|
|
|
DO proc = 0, ( dfftp%nproc3 - 1 )
|
|
|
|
|
recvcount(proc) = 2 * dfftp%nr1x * dfftp%nr2x * dfftp%nr3p(proc+1)
|
|
|
|
|
if (proc > 0) displs(proc) = displs(proc-1) + recvcount(proc-1)
|
|
|
|
|
ENDDO
|
|
|
|
|
info = 0
|
|
|
|
|
CALL MPI_ALLGATHERV( f_aux, recvcount(dfftp%mype3), MPI_DOUBLE_PRECISION, &
|
2015-10-29 17:08:16 +08:00
|
|
|
f_out, recvcount, displs, MPI_DOUBLE_PRECISION, &
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
dfftp%comm3, info )
|
2015-11-21 18:37:48 +08:00
|
|
|
CALL fftx_error__( 'cgather_sym', 'info<>0', info )
|
2015-10-29 17:08:16 +08:00
|
|
|
!
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
! ... the following check should be performed only on processor dfft%root
|
|
|
|
|
! ... otherwise f_out must be allocated on all processors even if not used
|
|
|
|
|
!
|
|
|
|
|
DEALLOCATE ( f_aux )
|
|
|
|
|
!
|
2015-10-29 17:08:16 +08:00
|
|
|
CALL stop_clock( 'cgather' )
|
|
|
|
|
!
|
|
|
|
|
#else
|
2015-11-21 18:37:48 +08:00
|
|
|
CALL fftx_error__('cgather_sym', 'do not use in serial execution', 1)
|
2015-10-29 17:08:16 +08:00
|
|
|
#endif
|
|
|
|
|
!
|
|
|
|
|
RETURN
|
|
|
|
|
!
|
|
|
|
|
END SUBROUTINE cgather_sym
|
|
|
|
|
!
|
|
|
|
|
!
|
|
|
|
|
!-----------------------------------------------------------------------
|
|
|
|
|
SUBROUTINE cgather_sym_many( dfftp, f_in, f_out, nbnd, nbnd_proc, start_nbnd_proc )
|
|
|
|
|
!-----------------------------------------------------------------------
|
|
|
|
|
!
|
|
|
|
|
! ... Written by A. Dal Corso
|
|
|
|
|
!
|
|
|
|
|
! ... This routine generalizes cgather_sym, receiveng nbnd complex
|
|
|
|
|
! ... distributed functions and collecting nbnd_proc(dfftp%mype+1)
|
|
|
|
|
! ... functions in each processor.
|
|
|
|
|
! ... start_nbnd_proc(dfftp%mype+1), says where the data for each processor
|
|
|
|
|
! ... start in the distributed variable
|
|
|
|
|
! ... COMPLEX*16 f_in = distributed variable (nrxx,nbnd)
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
! ... COMPLEX*16 f_out = gathered variable (nr1x*nr2x*nr3x,nbnd_proc(dfftp%mype+1))
|
2015-10-29 17:08:16 +08:00
|
|
|
!
|
|
|
|
|
IMPLICIT NONE
|
|
|
|
|
!
|
2016-08-10 05:38:57 +08:00
|
|
|
TYPE (fft_type_descriptor), INTENT(in) :: dfftp
|
2015-10-29 17:08:16 +08:00
|
|
|
INTEGER :: nbnd, nbnd_proc(dfftp%nproc), start_nbnd_proc(dfftp%nproc)
|
|
|
|
|
COMPLEX(DP) :: f_in(dfftp%nnr,nbnd)
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
COMPLEX(DP) :: f_out(dfftp%nr1x*dfftp%nr2x*dfftp%nr3x,nbnd_proc(dfftp%mype+1))
|
|
|
|
|
COMPLEX(DP), ALLOCATABLE :: f_aux(:)
|
2015-10-29 17:08:16 +08:00
|
|
|
!
|
2015-11-21 18:37:48 +08:00
|
|
|
#if defined(__MPI)
|
2015-10-29 17:08:16 +08:00
|
|
|
!
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
INTEGER :: proc_, proc2_, proc3_, proc, info, offset_in, offset_aux, ir3, nr2px
|
|
|
|
|
INTEGER :: ibnd, jbnd, iii
|
2015-10-29 17:08:16 +08:00
|
|
|
INTEGER :: displs(0:dfftp%nproc-1), recvcount(0:dfftp%nproc-1)
|
|
|
|
|
!
|
|
|
|
|
!
|
|
|
|
|
CALL start_clock( 'cgather' )
|
|
|
|
|
!
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
ALLOCATE ( f_aux(dfftp%nr1x * dfftp%nr2x * dfftp%my_nr3p ) )
|
2015-10-29 17:08:16 +08:00
|
|
|
!
|
2015-11-21 18:37:48 +08:00
|
|
|
CALL MPI_BARRIER( dfftp%comm, info )
|
2015-10-29 17:08:16 +08:00
|
|
|
!
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
f_out = (0.d0,0.d0)
|
|
|
|
|
|
|
|
|
|
!write (*,*) 'enter cgather ', nbnd
|
|
|
|
|
!write (*,*) nbnd_proc
|
|
|
|
|
!write (*,*) start_nbnd_proc
|
|
|
|
|
!write (*,*) 'dfftp%nproc',dfftp%nproc
|
|
|
|
|
!do ibnd =1,3
|
|
|
|
|
!write (*,*) 'evc = ',ibnd
|
|
|
|
|
!write (*,*) f_in(1:3,ibnd)
|
|
|
|
|
!end do
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
nr2px = MAXVAL ( dfftp%nr2p ) ! maximum number of Y values to be disributed
|
|
|
|
|
|
|
|
|
|
DO proc_ = 0, dfftp%nproc - 1
|
|
|
|
|
! define the processor index of the two sub-communicators
|
|
|
|
|
proc2_ = dfftp%iproc2(proc_ + 1) -1 ; proc3_ = dfftp%iproc3(proc_ + 1) -1
|
|
|
|
|
!write (*,*) ' proc_, proc2_, proc3_, dfftp%mype2', proc_, proc2_, proc3_, dfftp%mype2
|
|
|
|
|
DO ibnd = 1, nbnd_proc(proc_ +1)
|
|
|
|
|
jbnd = start_nbnd_proc(proc_ +1) + ibnd - 1
|
|
|
|
|
!write (*,*) ' proc_, ibnd, jbnd ', proc_, ibnd, jbnd
|
|
|
|
|
f_aux(:) = (0.d0,0.d0)
|
|
|
|
|
! 1) gather within the comm2 communicator
|
|
|
|
|
displs = 0
|
|
|
|
|
DO proc =0, ( dfftp%nproc2 -1 )
|
|
|
|
|
recvcount(proc) = 2 * dfftp%nr1x * dfftp%nr2p(proc+1)
|
|
|
|
|
if (proc > 0) displs(proc) = displs(proc-1) + recvcount(proc-1)
|
|
|
|
|
end do
|
|
|
|
|
!write (*,*) 'dfftp%nr1x ', dfftp%nr1x
|
|
|
|
|
!write (*,*) 'dfftp%nr2x ', dfftp%nr2x
|
|
|
|
|
!write (*,*) 'dfftp%nr3x ', dfftp%nr3x
|
|
|
|
|
!write (*,*) 'dfftp%nr2p ', dfftp%nr2p, ' nnr3px ', nr2px
|
|
|
|
|
!write (*,*) 'recvcount ', recvcount(0:dfftp%nproc2 -1)
|
|
|
|
|
!write (*,*) 'displs ', displs(0:dfftp%nproc2 -1)
|
|
|
|
|
offset_in = 1; offset_aux = 1
|
|
|
|
|
do ir3 = 1, dfftp%my_nr3p
|
|
|
|
|
info = 0
|
|
|
|
|
CALL MPI_GATHERV( f_in(offset_in,jbnd), recvcount(dfftp%mype2), MPI_DOUBLE_PRECISION, &
|
|
|
|
|
f_aux(offset_aux), recvcount, displs, MPI_DOUBLE_PRECISION, &
|
|
|
|
|
proc2_, dfftp%comm2, info )
|
|
|
|
|
CALL fftx_error__( 'cgather_sym_many', 'info<>0', info )
|
|
|
|
|
offset_in = offset_in + dfftp%nr1x * dfftp%my_nr2p
|
|
|
|
|
offset_aux = offset_aux + dfftp%nr1x * dfftp%nr2x
|
|
|
|
|
end do
|
|
|
|
|
!write(*,*) ' -> f_in(...+1:...+3,jbnd) ',jbnd
|
|
|
|
|
!offset_in = 0
|
|
|
|
|
!do iii =1,2*dfftp%nr2x
|
|
|
|
|
! write(*,'(i4,3("(",2f10.7,") "))') iii, f_in(offset_in+1:offset_in+3, jbnd)
|
|
|
|
|
! offset_in = offset_in + dfftp%nr1x
|
|
|
|
|
!enddo
|
|
|
|
|
IF (dfftp%mype2==proc2_) THEN
|
|
|
|
|
! write(*,*) ' -> f_aux(...+1:...+3)'
|
|
|
|
|
! offset_aux = 0
|
|
|
|
|
! do iii =1,4*dfftp%nr2x
|
|
|
|
|
! write(*,'(i4,3("(",2f10.7,") "))') iii, f_aux(offset_aux+1:offset_aux+3)
|
|
|
|
|
! offset_aux = offset_aux + dfftp%nr1x
|
|
|
|
|
! enddo
|
|
|
|
|
! write(*,*) ' -> F_AUX(...+1:...+3)'
|
|
|
|
|
! offset_aux = 0
|
|
|
|
|
! do iii =1,dfftp%my_nr3p
|
|
|
|
|
! write(*,'(i4,3("(",2f10.7,") "))') iii, f_aux(offset_aux+1:offset_aux+3)
|
|
|
|
|
! offset_aux = offset_aux + dfftp%nr1x*dfftp%nr2x
|
|
|
|
|
! enddo
|
|
|
|
|
! 2) gather within the comm3 communicator
|
|
|
|
|
displs = 0
|
|
|
|
|
DO proc = 0, ( dfftp%nproc3 - 1 )
|
|
|
|
|
recvcount(proc) = 2 * dfftp%nr1x * dfftp%nr2x * dfftp%nr3p(proc+1)
|
|
|
|
|
if (proc > 0) displs(proc) = displs(proc-1) + recvcount(proc-1)
|
|
|
|
|
ENDDO
|
|
|
|
|
info = 0
|
|
|
|
|
CALL MPI_GATHERV( f_aux, recvcount(dfftp%mype3), MPI_DOUBLE_PRECISION, &
|
|
|
|
|
f_out(1,ibnd), recvcount, displs, MPI_DOUBLE_PRECISION, &
|
|
|
|
|
proc3_, dfftp%comm3, info )
|
|
|
|
|
CALL fftx_error__( 'cgather_sym_many', 'info<>0', info )
|
|
|
|
|
|
|
|
|
|
! IF (dfftp%mype3==proc3_) THEN
|
|
|
|
|
! write(*,*) ' -> f_out(...+1:...+3,ibnd) ',ibnd
|
|
|
|
|
! offset_in = 0
|
|
|
|
|
! do iii =1,dfftp%nr3x
|
|
|
|
|
! write(*,'(i4,3("(",2f10.7,") "))') iii, f_out(offset_in+1:offset_in+3, ibnd)
|
|
|
|
|
! offset_in = offset_in + dfftp%nr1x*dfftp%nr2x
|
|
|
|
|
! enddo
|
|
|
|
|
! END IF
|
|
|
|
|
END IF
|
|
|
|
|
|
2015-10-29 17:08:16 +08:00
|
|
|
END DO
|
|
|
|
|
END DO
|
|
|
|
|
!
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
DEALLOCATE ( f_aux )
|
2015-10-29 17:08:16 +08:00
|
|
|
!
|
|
|
|
|
CALL stop_clock( 'cgather' )
|
|
|
|
|
!
|
|
|
|
|
#else
|
2015-11-21 18:37:48 +08:00
|
|
|
CALL fftx_error__('cgather_sym_many', 'do not use in serial execution', 1)
|
2015-10-29 17:08:16 +08:00
|
|
|
#endif
|
|
|
|
|
!
|
|
|
|
|
RETURN
|
|
|
|
|
!
|
|
|
|
|
END SUBROUTINE cgather_sym_many
|
|
|
|
|
!
|
|
|
|
|
!----------------------------------------------------------------------------
|
|
|
|
|
SUBROUTINE cscatter_sym_many( dfftp, f_in, f_out, target_ibnd, nbnd, nbnd_proc, &
|
|
|
|
|
start_nbnd_proc )
|
|
|
|
|
!----------------------------------------------------------------------------
|
|
|
|
|
!
|
|
|
|
|
! ... Written by A. Dal Corso
|
|
|
|
|
!
|
|
|
|
|
! ... generalizes cscatter_sym. It assumes that each processor has
|
|
|
|
|
! ... a certain number of bands (nbnd_proc(dfftp%mype+1)). The processor
|
|
|
|
|
! ... that has target_ibnd scatters it to all the other processors
|
|
|
|
|
! ... that receive a distributed part of the target function.
|
|
|
|
|
! ... start_nbnd_proc(dfftp%mype+1) is used to identify the processor
|
|
|
|
|
! ... that has the required band
|
|
|
|
|
!
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
! ... COMPLEX*16 f_in = gathered variable (nr1x*nr2x*nr3x, nbnd_proc(dfftp%mype+1) )
|
2015-10-29 17:08:16 +08:00
|
|
|
! ... COMPLEX*16 f_out = distributed variable (nrxx)
|
|
|
|
|
!
|
|
|
|
|
IMPLICIT NONE
|
|
|
|
|
!
|
2016-08-10 05:38:57 +08:00
|
|
|
TYPE (fft_type_descriptor), INTENT(in) :: dfftp
|
2015-10-29 17:08:16 +08:00
|
|
|
INTEGER :: nbnd, nbnd_proc(dfftp%nproc), start_nbnd_proc(dfftp%nproc)
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
COMPLEX(DP) :: f_in(dfftp%nr1x*dfftp%nr2x*dfftp%nr3x,nbnd_proc(dfftp%mype+1))
|
2015-10-29 17:08:16 +08:00
|
|
|
COMPLEX(DP) :: f_out(dfftp%nnr)
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
COMPLEX(DP), ALLOCATABLE :: f_aux(:)
|
2015-10-29 17:08:16 +08:00
|
|
|
INTEGER :: target_ibnd
|
|
|
|
|
!
|
2015-11-21 18:37:48 +08:00
|
|
|
#if defined(__MPI)
|
2015-10-29 17:08:16 +08:00
|
|
|
!
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
INTEGER :: proc_, proc2_, proc3_, proc, info, offset_out, offset_aux, ir3
|
2015-10-29 17:08:16 +08:00
|
|
|
INTEGER :: displs(0:dfftp%nproc-1), sendcount(0:dfftp%nproc-1)
|
|
|
|
|
INTEGER :: ibnd, jbnd
|
|
|
|
|
!
|
|
|
|
|
CALL start_clock( 'cscatter_sym' )
|
|
|
|
|
!
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
ALLOCATE ( f_aux(dfftp%nr1x * dfftp%nr2x * dfftp%my_nr3p ) )
|
2015-10-29 17:08:16 +08:00
|
|
|
!
|
|
|
|
|
f_out = (0.0_DP, 0.0_DP)
|
|
|
|
|
!
|
2015-11-21 18:37:48 +08:00
|
|
|
CALL MPI_BARRIER( dfftp%comm, info )
|
2015-10-29 17:08:16 +08:00
|
|
|
!
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
DO proc_ = 0, dfftp%nproc - 1
|
|
|
|
|
! define the processor index of the two sub-communicators
|
|
|
|
|
proc2_ = dfftp%iproc2(proc_ + 1) -1 ; proc3_ = dfftp%iproc3(proc_ + 1) -1
|
|
|
|
|
DO ibnd = 1, nbnd_proc(proc_+1)
|
|
|
|
|
jbnd = start_nbnd_proc(proc_+1) + ibnd - 1
|
|
|
|
|
IF (jbnd/=target_ibnd) CYCLE
|
|
|
|
|
IF (dfftp%mype2==proc2_) THEN
|
|
|
|
|
! 1) scatter within the comm3 communicator
|
|
|
|
|
displs = 0
|
|
|
|
|
DO proc = 0, ( dfftp%nproc3 - 1 )
|
|
|
|
|
sendcount(proc) = 2 * dfftp%nr1x * dfftp%nr2x * dfftp%nr3p(proc+1)
|
|
|
|
|
if (proc > 0) displs(proc) = displs(proc-1) + sendcount(proc-1)
|
|
|
|
|
ENDDO
|
|
|
|
|
info = 0
|
|
|
|
|
CALL MPI_SCATTERV( f_in(1,ibnd), sendcount, displs, MPI_DOUBLE_PRECISION, &
|
|
|
|
|
f_aux, sendcount(dfftp%mype3), MPI_DOUBLE_PRECISION, &
|
|
|
|
|
proc3_, dfftp%comm3, info )
|
|
|
|
|
ELSE
|
|
|
|
|
f_aux=(0.d0,0.d0)
|
|
|
|
|
END IF
|
|
|
|
|
! 2) scatter within the comm2 communicator
|
|
|
|
|
displs = 0 ; f_out = 0.0D0
|
|
|
|
|
DO proc =0, ( dfftp%nproc2 -1 )
|
|
|
|
|
sendcount(proc) = 2 * dfftp%nr1x * dfftp%nr2p(proc+1)
|
|
|
|
|
if (proc > 0) displs(proc) = displs(proc-1) + sendcount(proc-1)
|
|
|
|
|
end do
|
|
|
|
|
offset_out = 1 ; offset_aux = 1
|
|
|
|
|
do ir3 = 1, dfftp%my_nr3p
|
|
|
|
|
info = 0
|
|
|
|
|
CALL MPI_SCATTERV( f_aux(offset_aux), sendcount, displs, MPI_DOUBLE_PRECISION, &
|
|
|
|
|
f_out(offset_out), sendcount(dfftp%mype2), MPI_DOUBLE_PRECISION, &
|
|
|
|
|
proc2_, dfftp%comm2, info )
|
|
|
|
|
CALL fftx_error__( 'gather_grid', 'info<>0', info )
|
|
|
|
|
offset_out = offset_out + dfftp%nr1x * dfftp%my_nr2p
|
|
|
|
|
offset_aux = offset_aux + dfftp%nr1x * dfftp%nr2x
|
|
|
|
|
end do
|
2015-10-29 17:08:16 +08:00
|
|
|
ENDDO
|
|
|
|
|
ENDDO
|
|
|
|
|
!
|
MAJOR restructuring of the FFTXlib library
In real space processors are organized in a 2D pattern.
Each processor owns data from a sub-set of Z-planes and a sub-set of Y-planes.
In reciprocal space each processor owns Z-columns that belong to a sub set of
X-values. This allows to split the processors in two sets for communication
in the YZ and XY planes.
In alternative, if the situation allows for it, a task group paralelization is used
(with ntg=nyfft) where complete XY planes of ntg wavefunctions are collected and Fourier
trasnformed in G space by different task-groups. This is preferable to the Z-proc + Y-proc
paralleization if task group can be used because a smaller number of larger ammounts of
data are transferred. Hence three types of fft are implemented:
!
!! ... isgn = +-1 : parallel 3d fft for rho and for the potential
!
!! ... isgn = +-2 : parallel 3d fft for wavefunctions
!
!! ... isgn = +-3 : parallel 3d fft for wavefunctions with task group
!
!! ... isgn = + : G-space to R-space, output = \sum_G f(G)exp(+iG*R)
!! ... fft along z using pencils (cft_1z)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along x using pencils (cft_1x)
!
!! ... isgn = - : R-space to G-space, output = \int_R f(R)exp(-iG*R)/Omega
!! ... fft along x using pencils (cft_1x)
!! ... transpose across nodes (fft_scatter_xy)
!! ... fft along y using pencils (cft_1y)
!! ... transpose across nodes (fft_scatter_yz)
!! ... fft along z using pencils (cft_1z)
!
! If task_group_fft_is_active the FFT acts on a number of wfcs equal to
! dfft%nproc2, the number of Y-sections in which a plane is divided.
! Data are reshuffled by the fft_scatter_tg routine so that each of the
! dfft%nproc2 subgroups (made by dfft%nproc3 procs) deals with whole planes
! of a single wavefunciton.
!
fft_type module heavily modified, a number of variables renamed with more intuitive names
(at least to me), a number of more variables introduced for the Y-proc parallelization.
Task_group module made void. task_group management is now reduced to the logical component
fft_desc%have_task_groups of fft_type_descriptor type variable fft_desc.
In term of interfaces, the 'easy' calling sequences are
SUBROUTINE invfft/fwfft( grid_type, f, dfft, howmany )
!! where:
!!
!! **grid_type = 'Dense'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the dense grid (dfftp). On output, f is overwritten
!!
!! **grid_type = 'Smooth'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Wave'** :
!! inverse/direct fourier transform of wave functions f
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'tgWave'** :
!! inverse/direct fourier transform of wave functions f with task group
!! on the smooth grid (dffts). On output, f is overwritten
!!
!! **grid_type = 'Custom'** :
!! inverse/direct fourier transform of potentials and charge density f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **grid_type = 'CustomWave'** :
!! inverse/direct fourier transform of wave functions f
!! on a custom grid (dfft_exx). On output, f is overwritten
!!
!! **dfft = FFT descriptor**, IMPORTANT NOTICE: grid is specified only by dfft.
!! No check is performed on the correspondence between dfft and grid_type.
!! grid_type is now used only to distinguish cases 'Wave' / 'CustomWave'
!! from all other cases
Many more files modified.
git-svn-id: http://qeforge.qe-forge.org/svn/q-e/trunk/espresso@13676 c92efa57-630b-4861-b058-cf58834340f0
2017-08-02 04:31:02 +08:00
|
|
|
DEALLOCATE ( f_aux )
|
2015-10-29 17:08:16 +08:00
|
|
|
!
|
|
|
|
|
CALL stop_clock( 'cscatter_sym' )
|
|
|
|
|
!
|
|
|
|
|
#else
|
2015-11-21 18:37:48 +08:00
|
|
|
CALL fftx_error__('cscatter_sym_many', 'do not use in serial execution', 1)
|
2015-10-29 17:08:16 +08:00
|
|
|
#endif
|
|
|
|
|
!
|
|
|
|
|
RETURN
|
|
|
|
|
!
|
|
|
|
|
END SUBROUTINE cscatter_sym_many
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
!=----------------------------------------------------------------------=!
|
|
|
|
|
END MODULE scatter_mod
|
|
|
|
|
!=----------------------------------------------------------------------=!
|
2016-03-11 17:24:23 +08:00
|
|
|
!
|
|
|
|
|
!
|
|
|
|
|
!---------------------------------------------------------------------
|
|
|
|
|
subroutine fftsort (n, ia)
|
|
|
|
|
!---------------------------------------------------------------------
|
|
|
|
|
! sort an integer array ia(1:n) into ascending order using heapsort algorithm.
|
|
|
|
|
! n is input, ia is replaced on output by its sorted rearrangement.
|
|
|
|
|
! create an index table (ind) by making an exchange in the index array
|
|
|
|
|
! whenever an exchange is made on the sorted data array (ia).
|
|
|
|
|
! in case of equal values in the data array (ia) the values in the
|
|
|
|
|
! index array (ind) are used to order the entries.
|
|
|
|
|
! if on input ind(1) = 0 then indices are initialized in the routine,
|
|
|
|
|
! if on input ind(1) != 0 then indices are assumed to have been
|
|
|
|
|
! initialized before entering the routine and these
|
|
|
|
|
! indices are carried around during the sorting process
|
|
|
|
|
!
|
|
|
|
|
! no work space needed !
|
|
|
|
|
! free us from machine-dependent sorting-routines !
|
|
|
|
|
!
|
|
|
|
|
! adapted from Numerical Recipes pg. 329 (new edition)
|
|
|
|
|
!
|
|
|
|
|
implicit none
|
|
|
|
|
!-input/output variables
|
|
|
|
|
integer :: n
|
2016-04-09 23:01:59 +08:00
|
|
|
integer :: ia (2,n)
|
2016-03-11 17:24:23 +08:00
|
|
|
!-local variables
|
|
|
|
|
integer :: i, ir, j, l
|
|
|
|
|
integer :: iia(2)
|
|
|
|
|
! nothing to order
|
|
|
|
|
if (n.lt.2) return
|
|
|
|
|
! initialize indices for hiring and retirement-promotion phase
|
|
|
|
|
l = n / 2 + 1
|
|
|
|
|
ir = n
|
|
|
|
|
10 continue
|
|
|
|
|
! still in hiring phase
|
|
|
|
|
if (l.gt.1) then
|
|
|
|
|
l = l - 1
|
|
|
|
|
iia(:) = ia (:,l)
|
|
|
|
|
! in retirement-promotion phase.
|
|
|
|
|
else
|
|
|
|
|
! clear a space at the end of the array
|
|
|
|
|
iia(:) = ia (:,ir)
|
|
|
|
|
!
|
|
|
|
|
! retire the top of the heap into it
|
|
|
|
|
ia (:,ir) = ia (:,1)
|
|
|
|
|
!
|
|
|
|
|
! decrease the size of the corporation
|
|
|
|
|
ir = ir - 1
|
|
|
|
|
! done with the last promotion
|
|
|
|
|
if (ir.eq.1) then
|
|
|
|
|
! the least competent worker at all !
|
|
|
|
|
ia (:,1) = iia(:)
|
|
|
|
|
!
|
|
|
|
|
return
|
|
|
|
|
endif
|
|
|
|
|
endif
|
|
|
|
|
! wheter in hiring or promotion phase, we
|
|
|
|
|
i = l
|
|
|
|
|
! set up to place iia in its proper level
|
|
|
|
|
j = l + l
|
|
|
|
|
!
|
|
|
|
|
do while (j.le.ir)
|
|
|
|
|
if (j.lt.ir) then
|
|
|
|
|
if (ia (1,j) .lt. ia (1,j + 1) ) then
|
|
|
|
|
j = j + 1
|
|
|
|
|
endif
|
|
|
|
|
endif
|
|
|
|
|
! demote iia
|
|
|
|
|
if (iia(1).lt.ia (1,j) ) then
|
|
|
|
|
ia (:,i) = ia (:,j)
|
|
|
|
|
i = j
|
|
|
|
|
j = j + j
|
|
|
|
|
else
|
|
|
|
|
! set j to terminate do-while loop
|
|
|
|
|
j = ir + 1
|
|
|
|
|
endif
|
|
|
|
|
enddo
|
|
|
|
|
ia (:,i) = iia(:)
|
|
|
|
|
goto 10
|
|
|
|
|
!
|
|
|
|
|
end subroutine fftsort
|
|
|
|
|
|