field.py 20 KB
Newer Older
1
2
3
4
5
from itertools import chain
import numpy as np
import sympy as sp
from sympy.core.cache import cacheit
from sympy.tensor import IndexedBase
6
from pystencils.types import TypedSymbol, createType
7
8


Michael Kuron's avatar
Michael Kuron committed
9
class Field(object):
10
11
12
13
    """
    With fields one can formulate stencil-like update rules on structured grids.
    This Field class knows about the dimension, memory layout (strides) and optionally about the size of an array.

Martin Bauer's avatar
Martin Bauer committed
14
15
16
17
    Creating Fields:

        To create a field use one of the static create* members. There are two options:

18
        1. create a kernel with fixed loop sizes i.e. the shape of the array is already known. This is usually the
Martin Bauer's avatar
Martin Bauer committed
19
           case if just-in-time compilation directly from Python is done. (see :func:`Field.createFromNumpyArray`)
20
        2. create a more general kernel that works for variable array sizes. This can be used to create kernels
Martin Bauer's avatar
Martin Bauer committed
21
           beforehand for a library. (see :func:`Field.createGeneric`)
22
23
24
25
26
27

    Dimensions:
        A field has spatial and index dimensions, where the spatial dimensions come first.
        The interpretation is that the field has multiple cells in (usually) two or three dimensional space which are
        looped over. Additionally  N values are stored per cell. In this case spatialDimensions is two or three,
        and indexDimensions equals N. If you want to store a matrix on each point in a two dimensional grid, there
Martin Bauer's avatar
Martin Bauer committed
28
        are four dimensions, two spatial and two index dimensions: ``len(arr.shape) == spatialDims + indexDims``
29
30
31
32

    Indexing:
        When accessing (indexing) a field the result is a FieldAccess which is derived from sympy Symbol.
        First specify the spatial offsets in [], then in case indexDimension>0 the indices in ()
Martin Bauer's avatar
Martin Bauer committed
33
        e.g. ``f[-1,0,0](7)``
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49

    Example without index dimensions:
        >>> a = np.zeros([10, 10])
        >>> f = Field.createFromNumpyArray("f", a, indexDimensions=0)
        >>> jacobi = ( f[-1,0] + f[1,0] + f[0,-1] + f[0,1] ) / 4

    Example with index dimensions: LBM D2Q9 stream pull
        >>> stencil = np.array([[0,0], [0,1], [0,-1]])
        >>> src = Field.createGeneric("src", spatialDimensions=2, indexDimensions=1)
        >>> dst = Field.createGeneric("dst", spatialDimensions=2, indexDimensions=1)
        >>> for i, offset in enumerate(stencil):
        ...     sp.Eq(dst[0,0](i), src[-offset](i))
        Eq(dst_C^0, src_C^0)
        Eq(dst_C^1, src_S^1)
        Eq(dst_C^2, src_N^2)
    """
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

    @staticmethod
    def createGeneric(fieldName, spatialDimensions, dtype=np.float64, indexDimensions=0, layout='numpy'):
        """
        Creates a generic field where the field size is not fixed i.e. can be called with arrays of different sizes

        :param fieldName: symbolic name for the field
        :param dtype: numpy data type of the array the kernel is called with later
        :param spatialDimensions: see documentation of Field
        :param indexDimensions: see documentation of Field
        :param layout: tuple specifying the loop ordering of the spatial dimensions e.g. (2, 1, 0 ) means that
                       the outer loop loops over dimension 2, the second outer over dimension 1, and the inner loop
                       over dimension 0. Also allowed: the strings 'numpy' (0,1,..d) or 'reverseNumpy' (d, ..., 1, 0)
        """
        if isinstance(layout, str) and (layout == 'numpy' or layout.lower() == 'c'):
            layout = tuple(range(spatialDimensions))
        elif isinstance(layout, str) and (layout == 'reverseNumpy' or layout.lower() == 'f'):
            layout = tuple(reversed(range(spatialDimensions)))
        if len(layout) != spatialDimensions:
            raise ValueError("Layout")
        shapeSymbol = IndexedBase(TypedSymbol(Field.SHAPE_PREFIX + fieldName, Field.SHAPE_DTYPE), shape=(1,))
        strideSymbol = IndexedBase(TypedSymbol(Field.STRIDE_PREFIX + fieldName, Field.STRIDE_DTYPE), shape=(1,))
        totalDimensions = spatialDimensions + indexDimensions
        shape = tuple([shapeSymbol[i] for i in range(totalDimensions)])
        strides = tuple([strideSymbol[i] for i in range(totalDimensions)])
75
76
77
78
79
80
81
82

        npDataType = np.dtype(dtype)
        if npDataType.fields is not None:
            if indexDimensions != 0:
                raise ValueError("Structured arrays/fields are not allowed to have an index dimension")
            shape += (1,)
            strides += (1,)

83
84
        return Field(fieldName, dtype, layout, shape, strides)

85
86
87
88
89
90
91
92
93
94
95
96
97
    @staticmethod
    def createFromNumpyArray(fieldName, npArray, indexDimensions=0):
        """
        Creates a field based on the layout, data type, and shape of a given numpy array.
        Kernels created for these kind of fields can only be called with arrays of the same layout, shape and type.
        :param fieldName: symbolic name for the field
        :param npArray: numpy array
        :param indexDimensions: see documentation of Field
        """
        spatialDimensions = len(npArray.shape) - indexDimensions
        if spatialDimensions < 1:
            raise ValueError("Too many index dimensions. At least one spatial dimension required")

98
        fullLayout = getLayoutOfArray(npArray)
99
100
101
102
        spatialLayout = tuple([i for i in fullLayout if i < spatialDimensions])
        assert len(spatialLayout) == spatialDimensions

        strides = tuple([s // np.dtype(npArray.dtype).itemsize for s in npArray.strides])
103
        shape = tuple(int(s) for s in npArray.shape)
104

105
106
107
108
109
110
111
        npDataType = np.dtype(npArray.dtype)
        if npDataType.fields is not None:
            if indexDimensions != 0:
                raise ValueError("Structured arrays/fields are not allowed to have an index dimension")
            shape += (1,)
            strides += (1,)

112
113
114
        return Field(fieldName, npArray.dtype, spatialLayout, shape, strides)

    @staticmethod
115
    def createFixedSize(fieldName, shape, indexDimensions=0, dtype=np.float64, layout='numpy'):
116
        """
117
        Creates a field with fixed sizes i.e. can be called only with arrays of the same size and layout
118

119
        :param fieldName: symbolic name for the field
120
121
        :param shape: overall shape of the array
        :param indexDimensions: how many of the trailing dimensions are interpreted as index (as opposed to spatial)
122
        :param dtype: numpy data type of the array the kernel is called with later
123
        :param layout: full layout of array, not only spatial dimensions
124
        """
125
126
127
        spatialDimensions = len(shape) - indexDimensions
        assert spatialDimensions >= 1

128
        if isinstance(layout, str) and (layout == 'numpy' or layout.lower() == 'c'):
129
            layout = tuple(range(spatialDimensions))
130
        elif isinstance(layout, str) and (layout == 'reverseNumpy' or layout.lower() == 'f'):
Martin Bauer's avatar
Martin Bauer committed
131
            layout = tuple(reversed(range(spatialDimensions)))
132
133
134

        shape = tuple(int(s) for s in shape)
        strides = computeStrides(shape, layout)
135
136
137
138
139
140
141
142

        npDataType = np.dtype(dtype)
        if npDataType.fields is not None:
            if indexDimensions != 0:
                raise ValueError("Structured arrays/fields are not allowed to have an index dimension")
            shape += (1,)
            strides += (1,)

143
144
145
146
        spatialLayout = list(layout)
        for i in range(spatialDimensions, len(layout)):
            spatialLayout.remove(i)
        return Field(fieldName, dtype, tuple(spatialLayout), shape, strides)
147
148
149
150

    def __init__(self, fieldName, dtype, layout, shape, strides):
        """Do not use directly. Use static create* methods"""
        self._fieldName = fieldName
151
        self._dtype = createType(dtype)
152
        self._layout = normalizeLayout(layout)
153
154
        self.shape = shape
        self.strides = strides
155
156
157
        # index fields are currently only used for boundary handling
        # the coordinates are not the loop counters in that case, but are read from this index field
        self.isIndexField = False
158
159
160
161
162
163
164

    @property
    def spatialDimensions(self):
        return len(self._layout)

    @property
    def indexDimensions(self):
165
        return len(self.shape) - len(self._layout)
166
167
168
169
170
171
172
173
174
175
176

    @property
    def layout(self):
        return self._layout

    @property
    def name(self):
        return self._fieldName

    @property
    def spatialShape(self):
177
        return self.shape[:self.spatialDimensions]
178

179
180
181
182
183
184
185
186
    @property
    def hasFixedShape(self):
        try:
            [int(i) for i in self.shape]
            return True
        except TypeError:
            return False

187
188
    @property
    def indexShape(self):
189
        return self.shape[self.spatialDimensions:]
190
191
192

    @property
    def spatialStrides(self):
193
        return self.strides[:self.spatialDimensions]
194
195
196

    @property
    def indexStrides(self):
197
        return self.strides[self.spatialDimensions:]
198
199
200
201
202
203
204
205

    @property
    def dtype(self):
        return self._dtype

    def __repr__(self):
        return self._fieldName

206
207
208
209
210
    def neighbor(self, coordId, offset):
        offsetList = [0] * self.spatialDimensions
        offsetList[coordId] = offset
        return Field.Access(self, tuple(offsetList))

211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
    def __getitem__(self, offset):
        if type(offset) is np.ndarray:
            offset = tuple(offset)
        if type(offset) is str:
            offset = tuple(directionStringToOffset(offset, self.spatialDimensions))
        if type(offset) is not tuple:
            offset = (offset,)
        if len(offset) != self.spatialDimensions:
            raise ValueError("Wrong number of spatial indices: "
                             "Got %d, expected %d" % (len(offset), self.spatialDimensions))
        return Field.Access(self, offset)

    def __call__(self, *args, **kwargs):
        center = tuple([0]*self.spatialDimensions)
        return Field.Access(self, center)(*args, **kwargs)

    def __hash__(self):
228
        return hash((self._layout, self.shape, self.strides, self._dtype, self._fieldName))
229
230
231
232
233
234
235
236
237
238
239

    def __eq__(self, other):
        selfTuple = (self.shape, self.strides, self.name, self.dtype)
        otherTuple = (other.shape, other.strides, other.name, other.dtype)
        return selfTuple == otherTuple

    PREFIX = "f"
    STRIDE_PREFIX = PREFIX + "stride_"
    SHAPE_PREFIX = PREFIX + "shape_"
    STRIDE_DTYPE = "const int *"
    SHAPE_DTYPE = "const int *"
240
    DATA_PREFIX = PREFIX + "d_"
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

    class Access(sp.Symbol):
        def __new__(cls, name, *args, **kwargs):
            obj = Field.Access.__xnew_cached_(cls, name, *args, **kwargs)
            return obj

        def __new_stage2__(self, field, offsets=(0, 0, 0), idx=None):
            fieldName = field.name
            offsetsAndIndex = chain(offsets, idx) if idx is not None else offsets
            constantOffsets = not any([isinstance(o, sp.Basic) for o in offsetsAndIndex])

            if not idx:
                idx = tuple([0] * field.indexDimensions)

            if constantOffsets:
                offsetName = offsetToDirectionString(offsets)

                if field.indexDimensions == 0:
                    obj = super(Field.Access, self).__xnew__(self, fieldName + "_" + offsetName)
                elif field.indexDimensions == 1:
                    obj = super(Field.Access, self).__xnew__(self, fieldName + "_" + offsetName + "^" + str(idx[0]))
                else:
                    idxStr = ",".join([str(e) for e in idx])
                    obj = super(Field.Access, self).__xnew__(self, fieldName + "_" + offsetName + "^" + idxStr)

            else:
                offsetName = "%0.10X" % (abs(hash(tuple(offsetsAndIndex))))
                obj = super(Field.Access, self).__xnew__(self, fieldName + "_" + offsetName)

            obj._field = field
            obj._offsets = []
            for o in offsets:
                if isinstance(o, sp.Basic):
                    obj._offsets.append(o)
                else:
                    obj._offsets.append(int(o))
            obj._offsetName = offsetName
            obj._index = idx

            return obj

282
        def __getnewargs__(self):
283
            return self.field, self.offsets, self.index
284

285
286
287
288
289
290
291
292
293
294
295
296
297
298
        __xnew__ = staticmethod(__new_stage2__)
        __xnew_cached_ = staticmethod(cacheit(__new_stage2__))

        def __call__(self, *idx):
            if self._index != tuple([0]*self.field.indexDimensions):
                print(self._index, tuple([0]*self.field.indexDimensions))
                raise ValueError("Indexing an already indexed Field.Access")

            idx = tuple(idx)
            if len(idx) != self.field.indexDimensions and idx != (0,):
                raise ValueError("Wrong number of indices: "
                                 "Got %d, expected %d" % (len(idx), self.field.indexDimensions))
            return Field.Access(self.field, self._offsets, idx)

Martin Bauer's avatar
Martin Bauer committed
299
300
301
        def __getitem__(self, *idx):
            return self.__call__(*idx)

Martin Bauer's avatar
Martin Bauer committed
302
303
304
305
306
        def __iter__(self):
            """This is necessary to work with parts of sympy that test if an object is iterable (e.g. simplify).
            The __getitem__ would make it iterable"""
            raise TypeError("Field access is not iterable")

307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
        @property
        def field(self):
            return self._field

        @property
        def offsets(self):
            return self._offsets

        @property
        def requiredGhostLayers(self):
            return int(np.max(np.abs(self._offsets)))

        @property
        def nrOfCoordinates(self):
            return len(self._offsets)

        @property
        def offsetName(self):
            return self._offsetName

        @property
        def index(self):
            return self._index

331
332
333
334
335
336
        def getNeighbor(self, *offsets):
            return Field.Access(self.field, offsets, self.index)

        def getShifted(self, *shift):
            return Field.Access(self.field, tuple(a + b for a, b in zip(shift, self.offsets)), self.index)

337
338
        def _hashable_content(self):
            superClassContents = list(super(Field.Access, self)._hashable_content())
Martin Bauer's avatar
Martin Bauer committed
339
            t = tuple(superClassContents + [hash(self._field), self._index] + self._offsets)
340
            return t
Martin Bauer's avatar
Martin Bauer committed
341
342
343
344
345
346
347
348
349


def extractCommonSubexpressions(equations):
    """
    Uses sympy to find common subexpressions in equations and returns
    them in a topologically sorted order, ready for evaluation.
    Usually called before list of equations is passed to :func:`createKernel`
    """
    replacements, newEq = sp.cse(equations)
Martin Bauer's avatar
Martin Bauer committed
350
351
352
353
354
    # Workaround for older sympy versions: here subexpressions (temporary = True) are extracted
    # which leads to problems in Piecewise functions which have to a default case indicated by True
    symbolsEqualToTrue = {r[0]: True for r in replacements if r[1] is sp.true}

    replacementEqs = [sp.Eq(*r) for r in replacements if r[1] is not sp.true]
Martin Bauer's avatar
Martin Bauer committed
355
356
    equations = replacementEqs + newEq
    topologicallySortedPairs = sp.cse_main.reps_toposort([[e.lhs, e.rhs] for e in equations])
Martin Bauer's avatar
Martin Bauer committed
357
    equations = [sp.Eq(a[0], a[1].subs(symbolsEqualToTrue)) for a in topologicallySortedPairs]
Martin Bauer's avatar
Martin Bauer committed
358
359
360
    return equations


361
def getLayoutOfArray(arr, indexDimensionIds=[]):
Martin Bauer's avatar
Martin Bauer committed
362
363
364
    """
    Returns a list indicating the memory layout (linearization order) of the numpy array.
    Example:
365
    >>> getLayoutOfArray(np.zeros([3,3,3]))
Martin Bauer's avatar
Martin Bauer committed
366
    (0, 1, 2)
Martin Bauer's avatar
Martin Bauer committed
367
368
369
370
371

    In this example the loop over the zeroth coordinate should be the outermost loop,
    followed by the first and second. Elements arr[x,y,0] and arr[x,y,1] are adjacent in memory.
    Normally constructed numpy arrays have this order, however by stride tricks or other frameworks, arrays
    with different memory layout can be created.
Martin Bauer's avatar
Martin Bauer committed
372
373

    The indexDimensionIds parameter leaves specifies which coordinates should not be
Martin Bauer's avatar
Martin Bauer committed
374
375
    """
    coordinates = list(range(len(arr.shape)))
Martin Bauer's avatar
Martin Bauer committed
376
    relevantStrides = [stride for i, stride in enumerate(arr.strides) if i not in indexDimensionIds]
377
378
379
380
    result = [x for (y, x) in sorted(zip(relevantStrides, coordinates), key=lambda pair: pair[0], reverse=True)]
    return normalizeLayout(result)


381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
def createNumpyArrayWithLayout(shape, layout):
    """
    Creates a numpy array with
    :param shape: shape of the resulting array
    :param layout: layout as tuple, where the coordinates are ordered from slow to fast
    >>> res = createNumpyArrayWithLayout(shape=(2, 3, 4, 5), layout=(3, 2, 0, 1))
    >>> res.shape
    (2, 3, 4, 5)
    >>> getLayoutOfArray(res)
    (3, 2, 0, 1)
    """
    assert set(layout) == set(range(len(shape))), "Wrong layout descriptor"
    currentLayout = list(range(len(shape)))
    swaps = []
    for i in range(len(layout)):
        if currentLayout[i] != layout[i]:
            indexToSwapWith = currentLayout.index(layout[i])
            swaps.append((i, indexToSwapWith))
            currentLayout[i], currentLayout[indexToSwapWith] = currentLayout[indexToSwapWith], currentLayout[i]
    assert tuple(currentLayout) == tuple(layout)

    shape = list(shape)
    for a, b in swaps:
        shape[a], shape[b] = shape[b], shape[a]

    res = np.empty(shape, order='c')
    for a, b in reversed(swaps):
        res = res.swapaxes(a, b)
    return res


412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
def normalizeLayout(layout):
    """Takes a layout tuple and subtracts the minimum from all entries"""
    minEntry = min(layout)
    return tuple(i - minEntry for i in layout)


def computeStrides(shape, layout):
    """
    Computes strides assuming no padding exists
    :param shape: shape (size) of array
    :param layout: layout specification as tuple
    :return: strides in elements, not in bytes
    """
    N = len(shape)
    assert len(layout) == N
    assert len(set(layout)) == N
    strides = [0] * N
    product = 1
430
    for j in reversed(layout):
431
432
433
        strides[j] = product
        product *= shape[j]
    return tuple(strides)
Martin Bauer's avatar
Martin Bauer committed
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526


def offsetComponentToDirectionString(coordinateId, value):
    """
    Translates numerical offset to string notation.
    x offsets are labeled with east 'E' and 'W',
    y offsets with north 'N' and 'S' and
    z offsets with top 'T' and bottom 'B'
    If the absolute value of the offset is bigger than 1, this number is prefixed.
    :param coordinateId: integer 0, 1 or 2 standing for x,y and z
    :param value: integer offset

    Example:
    >>> offsetComponentToDirectionString(0, 1)
    'E'
    >>> offsetComponentToDirectionString(1, 2)
    '2N'
    """
    nameComponents = (('W', 'E'),  # west, east
                      ('S', 'N'),  # south, north
                      ('B', 'T'),  # bottom, top
                      )
    if value == 0:
        result = ""
    elif value < 0:
        result = nameComponents[coordinateId][0]
    else:
        result = nameComponents[coordinateId][1]
    if abs(value) > 1:
        result = "%d%s" % (abs(value), result)
    return result


def offsetToDirectionString(offsetTuple):
    """
    Translates numerical offset to string notation.
    For details see :func:`offsetComponentToDirectionString`
    :param offsetTuple: 3-tuple with x,y,z offset

    Example:
    >>> offsetToDirectionString([1, -1, 0])
    'SE'
    >>> offsetToDirectionString(([-3, 0, -2]))
    '2B3W'
    """
    names = ["", "", ""]
    for i in range(len(offsetTuple)):
        names[i] = offsetComponentToDirectionString(i, offsetTuple[i])
    name = "".join(reversed(names))
    if name == "":
        name = "C"
    return name


def directionStringToOffset(directionStr, dim=3):
    """
    Reverse mapping of :func:`offsetToDirectionString`
    :param directionStr: string representation of offset
    :param dim: dimension of offset, i.e the length of the returned list

    >>> directionStringToOffset('NW', dim=3)
    array([-1,  1,  0])
    >>> directionStringToOffset('NW', dim=2)
    array([-1,  1])
    >>> directionStringToOffset(offsetToDirectionString([3,-2,1]))
    array([ 3, -2,  1])
    """
    offsetMap = {
        'C': np.array([0, 0, 0]),

        'W': np.array([-1, 0, 0]),
        'E': np.array([1, 0, 0]),

        'S': np.array([0, -1, 0]),
        'N': np.array([0, 1, 0]),

        'B': np.array([0, 0, -1]),
        'T': np.array([0, 0, 1]),
    }
    offset = np.array([0, 0, 0])

    while len(directionStr) > 0:
        factor = 1
        firstNonDigit = 0
        while directionStr[firstNonDigit].isdigit():
            firstNonDigit += 1
        if firstNonDigit > 0:
            factor = int(directionStr[:firstNonDigit])
            directionStr = directionStr[firstNonDigit:]
        curOffset = offsetMap[directionStr[0]]
        offset += factor * curOffset
        directionStr = directionStr[1:]
    return offset[:dim]