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
|
#!/usr/bin/env python3
# encoding: utf-8
"""
_FourierTransformFFTW3MPI.py
Created by Graham Dennis on 2008-06-08.
Copyright (c) 2008-2012, Graham Dennis
This program is free software: you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation, either version 2 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License
along with this program. If not, see <http://www.gnu.org/licenses/>.
"""
from xpdeint.Features.Transforms.FourierTransformFFTW3 import FourierTransformFFTW3
from xpdeint.ParserException import ParserException
from xpdeint.Utilities import permutations, lazy_property
import operator
from itertools import groupby
from functools import reduce
class _FourierTransformFFTW3MPI (FourierTransformFFTW3):
def preflight(self):
super(_FourierTransformFFTW3MPI, self).preflight()
fields = self.getVar('fields')
geometry = self.getVar('geometry')
driver = self._driver
# Check that all vectors that are distributed and need fourier transforms
# contain all the points in the MPI dimensions. Otherwise we can't fourier
# transform them.
for field in filter(driver.isFieldDistributed, fields):
# If all the distributed dimensions are the same in this field as in the geometry, then everything is OK
if all([field.dimensionWithName(name) == geometry.dimensionWithName(name) for name in driver.distributedDimensionNames]):
continue
for vector in [v for v in self.vectorsNeedingThisTransform if v.field == field]:
raise ParserException(vector.xmlElement, "Vector '%s' cannot be fourier transformed because it would be distributed with MPI "
"and it doesn't have the same number of points as the geometry for the distributed dimensions." % vector)
for field in [field for field in fields if not field.isDistributed]:
for dim in [dim for dim in field.dimensions if dim.transform == self]:
for rep in [rep for rep in dim.representations if rep and rep.hasLocalOffset]:
dim.invalidateRepresentation(rep)
def initialiseForMPIWithDimensions(self, dimensions):
# We can only upgrade to MPI support if both the first and second dimensions
# are 'dft' or 'r2r' transforms. In the future, this restriction can be lifted
if len(dimensions) < 2:
raise ParserException(self._driver.xmlElement,
"There must be at least two dimensions to use the 'distributed-mpi' with the '%s' transform." % self.transformName[dimensions[0].name])
if len(dimensions[0].representations) <= 1 or len(dimensions[1].representations) <= 1:
raise ParserException(
self._driver.xmlElement,
"To use the 'distributed-mpi' driver either the first dimension must have no transform or "
"the first two dimensions must both have transforms."
)
self._driver.distributedDimensionNames = [dim.name for dim in dimensions[0:2]]
self.mpiDimensions = dimensions[0:2]
firstMPIDimension = dimensions[0]
secondMPIDimension = dimensions[1]
# Add additional transformed representations for the swapped case.
for rep in firstMPIDimension.representations[:]:
distributedRep = rep.copy(parent = firstMPIDimension)
distributedRep.setHasLocalOffset('unswapped')
firstMPIDimension.addRepresentation(distributedRep)
for rep in secondMPIDimension.representations[:]:
distributedRep = rep.copy(parent = secondMPIDimension)
distributedRep.setHasLocalOffset('swapped')
secondMPIDimension.addRepresentation(distributedRep)
def isFieldDistributed(self, field):
if not field:
return False
return field.hasDimension(self.mpiDimensions[0]) and field.hasDimension(self.mpiDimensions[1])
@lazy_property
def hasFFTWDistributedTransforms(self):
geometry = self.getVar('geometry')
return True if self.fullTransformDimensionsForField(geometry) else False
def fullTransformDimensionsForField(self, field):
keyFunc = lambda x: {'dft': 'complex', 'dct': 'real', 'dst': 'real'}.get(self.transformNameMap.get(x.name))
for transformType, dims in groupby(field.transverseDimensions, keyFunc):
return list(dims) if transformType else []
@property
def vectorsNeedingDistributedTransforms(self):
result = set()
[result.update(transformationDict['vectors'])
for tID, transformationDict in self.transformations if transformationDict.get('distributedTransform', False)]
return result
def availableTransformations(self):
parent_results = super(_FourierTransformFFTW3MPI, self).availableTransformations()
results = []
# Create mpi forward / back operations
geometry = self.getVar('geometry')
sortedDimNames = [(geometry.indexOfDimensionName(dimName), dimName) for dimName in self.transformNameMap]
sortedDimNames.sort()
sortedDimNames = [o[1] for o in sortedDimNames]
untransformedDimReps = dict([(dimName, geometry.dimensionWithName(dimName).firstDimRepWithTagName('coordinate')) for dimName in sortedDimNames])
transformedDimReps = dict([(dimName, geometry.dimensionWithName(dimName).firstDimRepWithTagName('spectral')) for dimName in sortedDimNames])
mpiTransformDimNamesLists = []
fullTransformDims = self.fullTransformDimensionsForField(geometry)
if len(fullTransformDims) > 2:
mpiTransformDimNamesLists.append([dim.name for dim in fullTransformDims])
if len(fullTransformDims) >= 2:
mpiTransformDimNamesLists.append([dim.name for dim in fullTransformDims[0:2]])
for dimNames in mpiTransformDimNamesLists:
untransformedBasis = tuple(untransformedDimReps[dimName].name for dimName in dimNames)
transformedBasis = tuple(transformedDimReps[dimName].name for dimName in dimNames)
transformCost = self.fftCost([dimName for dimName in dimNames])
communicationsCost = reduce(operator.mul, [untransformedDimReps[dimName].latticeEstimate for dimName in dimNames])
results.append(dict(
transformations = [tuple([self.canonicalBasisForBasis(untransformedBasis), self.canonicalBasisForBasis(transformedBasis)])],
communicationsCost = communicationsCost,
cost = transformCost,
distributedTransform = True,
forwardScale = self.scaleFactorForDimReps(untransformedBasis),
backwardScale = self.scaleFactorForDimReps(transformedBasis),
requiresScaling = True,
transformType = 'complex' if self.transformNameMap[self.mpiDimensions[0].name] == 'dft' else 'real',
geometryDependent = True,
transformFunction = self.distributedTransformFunction
))
# Create transpose operations
transposeOperations = []
for firstDimRep, secondDimRep in permutations(*[[rep for rep in dim.representations if not rep.hasLocalOffset] for dim in self.mpiDimensions]):
communicationsCost = firstDimRep.latticeEstimate * secondDimRep.latticeEstimate
basisA = ('distributed ' + firstDimRep.name, secondDimRep.name)
basisB = ('distributed ' + secondDimRep.name, firstDimRep.name)
if not self.hasFFTWDistributedTransforms:
basisA = tuple(reversed(basisA))
basisB = tuple(reversed(basisB))
results.append(dict(
transformations = [tuple([basisA, basisB])],
communicationsCost = communicationsCost,
geometryDependent = True,
transformType = 'real',
distributedTransform = True,
transformFunction = self.transposeTransformFunction,
transposedOrder = not self.hasFFTWDistributedTransforms,
))
final_transforms = []
for transform in results:
final_transforms.append(transform.copy())
transform['outOfPlace'] = True
final_transforms.append(transform)
return parent_results + final_transforms
def canonicalBasisForBasis(self, basis, noTranspose = False):
if all([set(rep.canonicalName for rep in mpiDim.representations).intersection(basis) for mpiDim in self.mpiDimensions]):
# Decide what the order is.
basis = list(basis)
mpiDimRepNames = [rep.canonicalName for mpiDim in self.mpiDimensions for rep in mpiDim.representations if rep.canonicalName in basis]
mpiDimRepIndices = [basis.index(rep.canonicalName) for mpiDim in self.mpiDimensions for rep in mpiDim.representations
if rep.canonicalName in basis]
mpiDimRepIndices.sort()
assert len(mpiDimRepIndices) == 2
assert mpiDimRepIndices[1] - mpiDimRepIndices[0] == 1
basisSlice = slice(mpiDimRepIndices[0], mpiDimRepIndices[1]+1)
nonDistributedMPIDimRepNames = [b.replace('distributed ', '') for b in mpiDimRepNames]
if (not noTranspose) and sum(b.startswith('distributed ') for b in basis[basisSlice]) == 1:
# Transposes are legal, and the basis is already propery distributed.
# Leave it alone.
pass
else:
if (not noTranspose) \
and all([any([rep.canonicalName in mpiDimRepNames
for rep in mpiDim.representations if issubclass(rep.tag, rep.tagForName('spectral'))])
for mpiDim in self.mpiDimensions]):
# Transposes are legal and all MPI dimensions are in spectral representations.
# We decide that this means we are swapped.
basis[basisSlice] = reversed(nonDistributedMPIDimRepNames)
else:
# Either transposes aren't legal or not all MPI dimensions were in spectral representation.
basis[basisSlice] = nonDistributedMPIDimRepNames
distributedIdx = basisSlice.start if self.hasFFTWDistributedTransforms else basisSlice.start + 1
basis[distributedIdx] = 'distributed ' + basis[distributedIdx]
basis = tuple(basis)
else:
# At most one of the mpi dimensions is in this basis. Therefore we must ensure that no part of the basis contains 'distributed '
basis = tuple(b.replace('distributed ','') for b in basis)
assert sum('distributed ' in b for b in basis) <= 1
return basis
|
