-
Notifications
You must be signed in to change notification settings - Fork 196
/
SpectralFieldData.cpp
340 lines (302 loc) · 14.2 KB
/
SpectralFieldData.cpp
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
/* Copyright 2019 Maxence Thevenet, Remi Lehe, Revathi Jambunathan
*
*
* This file is part of WarpX.
*
* License: BSD-3-Clause-LBNL
*/
#include "SpectralFieldData.H"
#include "Utils/WarpXAlgorithmSelection.H"
#include "WarpX.H"
#include <AMReX_Array4.H>
#include <AMReX_BLassert.H>
#include <AMReX_Box.H>
#include <AMReX_BoxArray.H>
#include <AMReX_Dim3.H>
#include <AMReX_FArrayBox.H>
#include <AMReX_GpuAtomic.H>
#include <AMReX_GpuComplex.H>
#include <AMReX_GpuDevice.H>
#include <AMReX_GpuLaunch.H>
#include <AMReX_GpuQualifiers.H>
#include <AMReX_IntVect.H>
#include <AMReX_LayoutData.H>
#include <AMReX_MFIter.H>
#include <AMReX_PODVector.H>
#include <AMReX_REAL.H>
#include <AMReX_Utility.H>
#if WARPX_USE_PSATD
using namespace amrex;
/* \brief Initialize fields in spectral space, and FFT plans */
SpectralFieldData::SpectralFieldData( const int lev,
const amrex::BoxArray& realspace_ba,
const SpectralKSpace& k_space,
const amrex::DistributionMapping& dm,
const int n_field_required,
const bool periodic_single_box)
{
amrex::LayoutData<amrex::Real>* cost = WarpX::getCosts(lev);
m_periodic_single_box = periodic_single_box;
const BoxArray& spectralspace_ba = k_space.spectralspace_ba;
// Allocate the arrays that contain the fields in spectral space
// (one component per field)
fields = SpectralField(spectralspace_ba, dm, n_field_required, 0);
// Allocate temporary arrays - in real space and spectral space
// These arrays will store the data just before/after the FFT
tmpRealField = MultiFab(realspace_ba, dm, 1, 0);
tmpSpectralField = SpectralField(spectralspace_ba, dm, 1, 0);
// By default, we assume the FFT is done from/to a nodal grid in real space
// It the FFT is performed from/to a cell-centered grid in real space,
// a correcting "shift" factor must be applied in spectral space.
xshift_FFTfromCell = k_space.getSpectralShiftFactor(dm, 0,
ShiftType::TransformFromCellCentered);
xshift_FFTtoCell = k_space.getSpectralShiftFactor(dm, 0,
ShiftType::TransformToCellCentered);
#if (AMREX_SPACEDIM == 3)
yshift_FFTfromCell = k_space.getSpectralShiftFactor(dm, 1,
ShiftType::TransformFromCellCentered);
yshift_FFTtoCell = k_space.getSpectralShiftFactor(dm, 1,
ShiftType::TransformToCellCentered);
zshift_FFTfromCell = k_space.getSpectralShiftFactor(dm, 2,
ShiftType::TransformFromCellCentered);
zshift_FFTtoCell = k_space.getSpectralShiftFactor(dm, 2,
ShiftType::TransformToCellCentered);
#else
zshift_FFTfromCell = k_space.getSpectralShiftFactor(dm, 1,
ShiftType::TransformFromCellCentered);
zshift_FFTtoCell = k_space.getSpectralShiftFactor(dm, 1,
ShiftType::TransformToCellCentered);
#endif
// Allocate and initialize the FFT plans
forward_plan = AnyFFT::FFTplans(spectralspace_ba, dm);
backward_plan = AnyFFT::FFTplans(spectralspace_ba, dm);
// Loop over boxes and allocate the corresponding plan
// for each box owned by the local MPI proc
for ( MFIter mfi(spectralspace_ba, dm); mfi.isValid(); ++mfi ){
if (cost && WarpX::load_balance_costs_update_algo == LoadBalanceCostsUpdateAlgo::Timers)
{
amrex::Gpu::synchronize();
}
Real wt = amrex::second();
// Note: the size of the real-space box and spectral-space box
// differ when using real-to-complex FFT. When initializing
// the FFT plan, the valid dimensions are those of the real-space box.
IntVect fft_size = realspace_ba[mfi].length();
forward_plan[mfi] = AnyFFT::CreatePlan(
fft_size, tmpRealField[mfi].dataPtr(),
reinterpret_cast<AnyFFT::Complex*>( tmpSpectralField[mfi].dataPtr()),
AnyFFT::direction::R2C, AMREX_SPACEDIM);
backward_plan[mfi] = AnyFFT::CreatePlan(
fft_size, tmpRealField[mfi].dataPtr(),
reinterpret_cast<AnyFFT::Complex*>( tmpSpectralField[mfi].dataPtr()),
AnyFFT::direction::C2R, AMREX_SPACEDIM);
if (cost && WarpX::load_balance_costs_update_algo == LoadBalanceCostsUpdateAlgo::Timers)
{
amrex::Gpu::synchronize();
wt = amrex::second() - wt;
amrex::HostDevice::Atomic::Add( &(*cost)[mfi.index()], wt);
}
}
}
SpectralFieldData::~SpectralFieldData()
{
if (tmpRealField.size() > 0){
for ( MFIter mfi(tmpRealField); mfi.isValid(); ++mfi ){
AnyFFT::DestroyPlan(forward_plan[mfi]);
AnyFFT::DestroyPlan(backward_plan[mfi]);
}
}
}
/* \brief Transform the component `i_comp` of MultiFab `mf`
* to spectral space, and store the corresponding result internally
* (in the spectral field specified by `field_index`) */
void
SpectralFieldData::ForwardTransform (const int lev,
const MultiFab& mf, const int field_index,
const int i_comp, const IntVect& stag)
{
amrex::LayoutData<amrex::Real>* cost = WarpX::getCosts(lev);
// Check field index type, in order to apply proper shift in spectral space
const bool is_nodal_x = (stag[0] == amrex::IndexType::NODE) ? true : false;
#if (AMREX_SPACEDIM == 3)
const bool is_nodal_y = (stag[1] == amrex::IndexType::NODE) ? true : false;
const bool is_nodal_z = (stag[2] == amrex::IndexType::NODE) ? true : false;
#else
const bool is_nodal_z = (stag[1] == amrex::IndexType::NODE) ? true : false;
#endif
// Loop over boxes
for ( MFIter mfi(mf); mfi.isValid(); ++mfi ){
if (cost && WarpX::load_balance_costs_update_algo == LoadBalanceCostsUpdateAlgo::Timers)
{
amrex::Gpu::synchronize();
}
Real wt = amrex::second();
// Copy the real-space field `mf` to the temporary field `tmpRealField`
// This ensures that all fields have the same number of points
// before the Fourier transform.
// As a consequence, the copy discards the *last* point of `mf`
// in any direction that has *nodal* index type.
{
Box realspace_bx;
if (m_periodic_single_box) {
realspace_bx = mfi.validbox(); // Discard guard cells
} else {
realspace_bx = mf[mfi].box(); // Keep guard cells
}
realspace_bx.enclosedCells(); // Discard last point in nodal direction
AMREX_ALWAYS_ASSERT( realspace_bx.contains(tmpRealField[mfi].box()) );
Array4<const Real> mf_arr = mf[mfi].array();
Array4<Real> tmp_arr = tmpRealField[mfi].array();
ParallelFor( tmpRealField[mfi].box(),
[=] AMREX_GPU_DEVICE(int i, int j, int k) noexcept {
tmp_arr(i,j,k) = mf_arr(i,j,k,i_comp);
});
}
// Perform Fourier transform from `tmpRealField` to `tmpSpectralField`
AnyFFT::Execute(forward_plan[mfi]);
// Copy the spectral-space field `tmpSpectralField` to the appropriate
// index of the FabArray `fields` (specified by `field_index`)
// and apply correcting shift factor if the real space data comes
// from a cell-centered grid in real space instead of a nodal grid.
{
Array4<Complex> fields_arr = SpectralFieldData::fields[mfi].array();
Array4<const Complex> tmp_arr = tmpSpectralField[mfi].array();
const Complex* xshift_arr = xshift_FFTfromCell[mfi].dataPtr();
#if (AMREX_SPACEDIM == 3)
const Complex* yshift_arr = yshift_FFTfromCell[mfi].dataPtr();
#endif
const Complex* zshift_arr = zshift_FFTfromCell[mfi].dataPtr();
// Loop over indices within one box
const Box spectralspace_bx = tmpSpectralField[mfi].box();
ParallelFor( spectralspace_bx,
[=] AMREX_GPU_DEVICE(int i, int j, int k) noexcept {
Complex spectral_field_value = tmp_arr(i,j,k);
// Apply proper shift in each dimension
if (is_nodal_x==false) spectral_field_value *= xshift_arr[i];
#if (AMREX_SPACEDIM == 3)
if (is_nodal_y==false) spectral_field_value *= yshift_arr[j];
if (is_nodal_z==false) spectral_field_value *= zshift_arr[k];
#elif (AMREX_SPACEDIM == 2)
if (is_nodal_z==false) spectral_field_value *= zshift_arr[j];
#endif
// Copy field into the right index
fields_arr(i,j,k,field_index) = spectral_field_value;
});
}
if (cost && WarpX::load_balance_costs_update_algo == LoadBalanceCostsUpdateAlgo::Timers)
{
amrex::Gpu::synchronize();
wt = amrex::second() - wt;
amrex::HostDevice::Atomic::Add( &(*cost)[mfi.index()], wt);
}
}
}
/* \brief Transform spectral field specified by `field_index` back to
* real space, and store it in the component `i_comp` of `mf` */
void
SpectralFieldData::BackwardTransform( const int lev,
MultiFab& mf,
const int field_index,
const int i_comp )
{
amrex::LayoutData<amrex::Real>* cost = WarpX::getCosts(lev);
// Check field index type, in order to apply proper shift in spectral space
const bool is_nodal_x = mf.is_nodal(0);
#if (AMREX_SPACEDIM == 3)
const bool is_nodal_y = mf.is_nodal(1);
const bool is_nodal_z = mf.is_nodal(2);
#else
const bool is_nodal_z = mf.is_nodal(1);
#endif
const int si = (is_nodal_x) ? 1 : 0;
#if (AMREX_SPACEDIM == 2)
const int sj = (is_nodal_z) ? 1 : 0;
const int sk = 0;
#elif (AMREX_SPACEDIM == 3)
const int sj = (is_nodal_y) ? 1 : 0;
const int sk = (is_nodal_z) ? 1 : 0;
#endif
// Loop over boxes
for ( MFIter mfi(mf); mfi.isValid(); ++mfi ){
if (cost && WarpX::load_balance_costs_update_algo == LoadBalanceCostsUpdateAlgo::Timers)
{
amrex::Gpu::synchronize();
}
Real wt = amrex::second();
// Copy the spectral-space field `tmpSpectralField` to the appropriate
// field (specified by the input argument field_index)
// and apply correcting shift factor if the field is to be transformed
// to a cell-centered grid in real space instead of a nodal grid.
{
Array4<const Complex> field_arr = SpectralFieldData::fields[mfi].array();
Array4<Complex> tmp_arr = tmpSpectralField[mfi].array();
const Complex* xshift_arr = xshift_FFTtoCell[mfi].dataPtr();
#if (AMREX_SPACEDIM == 3)
const Complex* yshift_arr = yshift_FFTtoCell[mfi].dataPtr();
#endif
const Complex* zshift_arr = zshift_FFTtoCell[mfi].dataPtr();
// Loop over indices within one box
const Box spectralspace_bx = tmpSpectralField[mfi].box();
ParallelFor( spectralspace_bx,
[=] AMREX_GPU_DEVICE(int i, int j, int k) noexcept {
Complex spectral_field_value = field_arr(i,j,k,field_index);
// Apply proper shift in each dimension
if (is_nodal_x==false) spectral_field_value *= xshift_arr[i];
#if (AMREX_SPACEDIM == 3)
if (is_nodal_y==false) spectral_field_value *= yshift_arr[j];
if (is_nodal_z==false) spectral_field_value *= zshift_arr[k];
#elif (AMREX_SPACEDIM == 2)
if (is_nodal_z==false) spectral_field_value *= zshift_arr[j];
#endif
// Copy field into temporary array
tmp_arr(i,j,k) = spectral_field_value;
});
}
// Perform Fourier transform from `tmpSpectralField` to `tmpRealField`
AnyFFT::Execute(backward_plan[mfi]);
// Copy the temporary field tmpRealField to the real-space field mf and
// normalize, dividing by N, since (FFT + inverse FFT) results in a factor N
{
amrex::Box const& mf_box = (m_periodic_single_box) ? mfi.validbox() : mfi.fabbox();
amrex::Array4<amrex::Real> mf_arr = mf[mfi].array();
amrex::Array4<const amrex::Real> tmp_arr = tmpRealField[mfi].array();
const amrex::Real inv_N = 1._rt / tmpRealField[mfi].box().numPts();
// Total number of cells, including ghost cells (nj represents ny in 3D and nz in 2D)
const int ni = mf_box.length(0);
const int nj = mf_box.length(1);
#if (AMREX_SPACEDIM == 2)
constexpr int nk = 1;
#elif (AMREX_SPACEDIM == 3)
const int nk = mf_box.length(2);
#endif
// Lower bound of the box (lo_j represents lo_y in 3D and lo_z in 2D)
const int lo_i = amrex::lbound(mf_box).x;
const int lo_j = amrex::lbound(mf_box).y;
#if (AMREX_SPACEDIM == 2)
constexpr int lo_k = 0;
#elif (AMREX_SPACEDIM == 3)
const int lo_k = amrex::lbound(mf_box).z;
#endif
// Loop over cells within full box, including ghost cells
ParallelFor(mf_box, [=] AMREX_GPU_DEVICE(int i, int j, int k) noexcept
{
// Assume periodicity and set the last outer guard cell equal to the first one:
// this is necessary in order to get the correct value along a nodal direction,
// because the last point along a nodal direction is always discarded when FFTs
// are computed, as the real-space box is always cell-centered.
const int ii = (i == lo_i + ni - si) ? lo_i : i;
const int jj = (j == lo_j + nj - sj) ? lo_j : j;
const int kk = (k == lo_k + nk - sk) ? lo_k : k;
// Copy and normalize field
mf_arr(i,j,k,i_comp) = inv_N * tmp_arr(ii,jj,kk);
});
}
if (cost && WarpX::load_balance_costs_update_algo == LoadBalanceCostsUpdateAlgo::Timers)
{
amrex::Gpu::synchronize();
wt = amrex::second() - wt;
amrex::HostDevice::Atomic::Add( &(*cost)[mfi.index()], wt);
}
}
}
#endif // WARPX_USE_PSATD