Repository navigation
Expand file tree
/
Copy pathreductions.py
More file actions
2562 lines (2142 loc) · 89.2 KB
/
Copy pathreductions.py
File metadata and controls
2562 lines (2142 loc) · 89.2 KB
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
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
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
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
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
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
from __future__ import annotations
import copy
from enum import Enum
from packaging.version import Version
import numpy as np
from datashader.datashape import dshape, isnumeric, Record, Option
from datashader.datashape import coretypes as ct
from toolz import concat, unique
import xarray as xr
from datashader.antialias import AntialiasCombination, AntialiasStage2
from datashader.utils import isminus1, isnull
from numba import cuda as nb_cuda
try:
from datashader.transfer_functions._cuda_utils import (
cuda_atomic_nanmin, cuda_atomic_nanmax, cuda_args, cuda_row_min_in_place,
cuda_nanmax_n_in_place_4d, cuda_nanmax_n_in_place_3d,
cuda_nanmin_n_in_place_4d, cuda_nanmin_n_in_place_3d,
cuda_row_max_n_in_place_4d, cuda_row_max_n_in_place_3d,
cuda_row_min_n_in_place_4d, cuda_row_min_n_in_place_3d, cuda_shift_and_insert,
)
except ImportError:
(cuda_atomic_nanmin, cuda_atomic_nanmax, cuda_args, cuda_row_min_in_place,
cuda_nanmax_n_in_place_4d, cuda_nanmax_n_in_place_3d,
cuda_nanmin_n_in_place_4d, cuda_nanmin_n_in_place_3d,
cuda_row_max_n_in_place_4d, cuda_row_max_n_in_place_3d,
cuda_row_min_n_in_place_4d, cuda_row_min_n_in_place_3d, cuda_shift_and_insert,
) = None, None, None, None, None, None, None, None, None, None, None, None, None
try:
import cudf
import cupy as cp
except Exception:
cudf = cp = None
from .utils import (
Expr, ngjit, nansum_missing, nanmax_in_place, nansum_in_place, row_min_in_place,
nanmax_n_in_place_4d, nanmax_n_in_place_3d, nanmin_n_in_place_4d, nanmin_n_in_place_3d,
row_max_n_in_place_4d, row_max_n_in_place_3d, row_min_n_in_place_4d, row_min_n_in_place_3d,
shift_and_insert,
)
class SpecialColumn(Enum):
"""
Internally datashader identifies the columns required by the user's
Reductions and extracts them from the supplied source (e.g. DataFrame) to
pass through the dynamically-generated append function in compiler.py and
end up as arguments to the Reduction._append* functions. Each column is
a string name or a SpecialColumn. A column of None is used in Reduction
classes to denote that no column is required.
"""
RowIndex = 1
class UsesCudaMutex(Enum):
"""
Enum that encapsulates the need for a Reduction to use a CUDA mutex to
operate correctly on a GPU. Possible values:
No: the Reduction append_cuda function is atomic and no mutex is required.
Local: Reduction append_cuda needs wrapping in a mutex.
Global: the overall compiled append function needs wrapping in a mutex.
"""
No = 0
Local = 1
Global = 2
class Preprocess(Expr):
"""Base class for preprocessing steps."""
def __init__(self, column: str | SpecialColumn | None):
self.column = column
@property
def inputs(self):
return (self.column,)
@property
def nan_check_column(self):
return None
class extract(Preprocess):
"""Extract a column from a dataframe as a numpy array of values."""
def apply(self, df, cuda):
if self.column is SpecialColumn.RowIndex:
attr_name = "_datashader_row_offset"
if isinstance(df, xr.Dataset):
row_offset = df.attrs[attr_name]
row_length = df.attrs["_datashader_row_length"]
else:
attrs = getattr(df, "attrs", None)
row_offset = getattr(attrs or df, attr_name, 0)
row_length = len(df)
if cudf and isinstance(df, cudf.DataFrame):
if self.column is SpecialColumn.RowIndex:
return cp.arange(row_offset, row_offset + row_length, dtype=np.int64)
if df[self.column].dtype.kind == 'f':
nullval = np.nan
else:
nullval = 0
if Version(cudf.__version__) >= Version("22.02"):
return df[self.column].to_cupy(na_value=nullval)
return cp.array(df[self.column].to_gpu_array(fillna=nullval))
elif self.column is SpecialColumn.RowIndex:
if cuda:
return cp.arange(row_offset, row_offset + row_length, dtype=np.int64)
else:
return np.arange(row_offset, row_offset + row_length, dtype=np.int64)
elif isinstance(df, xr.Dataset):
if cuda and not isinstance(df[self.column].data, cp.ndarray):
return cp.asarray(df[self.column])
else:
return df[self.column].data
else:
return df[self.column].values
class CategoryPreprocess(Preprocess):
"""Base class for categorizing preprocessors."""
@property
def cat_column(self):
"""Returns name of categorized column"""
return self.column
def categories(self, input_dshape):
"""Returns list of categories corresponding to input shape"""
raise NotImplementedError("categories not implemented")
def validate(self, in_dshape):
"""Validates input shape"""
raise NotImplementedError("validate not implemented")
def apply(self, df, cuda):
"""Applies preprocessor to DataFrame and returns array"""
raise NotImplementedError("apply not implemented")
class category_codes(CategoryPreprocess):
"""
Extract just the category codes from a categorical column.
To create a new type of categorizer, derive a subclass from this
class or one of its subclasses, implementing ``__init__``,
``_hashable_inputs``, ``categories``, ``validate``, and ``apply``.
See the implementation of ``category_modulo`` in ``reductions.py``
for an example.
"""
def categories(self, input_dshape):
return input_dshape.measure[self.column].categories
def validate(self, in_dshape):
if self.column not in in_dshape.dict:
raise ValueError("specified column not found")
if not isinstance(in_dshape.measure[self.column], ct.Categorical):
raise ValueError("input must be categorical")
def apply(self, df, cuda):
if cudf and isinstance(df, cudf.DataFrame):
if Version(cudf.__version__) >= Version("22.02"):
return df[self.column].cat.codes.to_cupy()
return df[self.column].cat.codes.to_gpu_array()
else:
return df[self.column].cat.codes.values
class category_modulo(category_codes):
"""
A variation on category_codes that assigns categories using an integer column, modulo a base.
Category is computed as (column_value - offset)%modulo.
"""
# couldn't find anything in the datashape docs about how to check if a CType is an integer, so
# just define a big set
IntegerTypes = {ct.bool_, ct.uint8, ct.uint16, ct.uint32, ct.uint64, ct.int8, ct.int16,
ct.int32, ct.int64}
def __init__(self, column, modulo, offset=0):
super().__init__(column)
self.offset = offset
self.modulo = modulo
def _hashable_inputs(self):
return super()._hashable_inputs() + (self.offset, self.modulo)
def categories(self, in_dshape):
return list(range(self.modulo))
def validate(self, in_dshape):
if self.column not in in_dshape.dict:
raise ValueError("specified column not found")
if in_dshape.measure[self.column] not in self.IntegerTypes:
raise ValueError("input must be an integer column")
def apply(self, df, cuda):
result = (df[self.column] - self.offset) % self.modulo
if cudf and isinstance(df, cudf.Series):
if Version(cudf.__version__) >= Version("22.02"):
return result.to_cupy()
return result.to_gpu_array()
else:
return result.values
class category_binning(category_modulo):
"""
A variation on category_codes that assigns categories by binning a continuous-valued column.
The number of categories returned is always nbins+1.
The last category (nbin) is for NaNs in the data column, as well as for values under/over the
binned interval (when include_under or include_over is False).
Parameters
----------
column: column to use
lower: lower bound of first bin
upper: upper bound of last bin
nbins: number of bins
include_under: if True, values below bin 0 are assigned to category 0
include_over: if True, values above the last bin (nbins-1) are assigned to category nbin-1
"""
def __init__(self, column, lower, upper, nbins, include_under=True, include_over=True):
super().__init__(column, nbins + 1) # +1 category for NaNs and clipped values
self.bin0 = lower
self.binsize = (upper - lower) / float(nbins)
self.nbins = nbins
self.bin_under = 0 if include_under else nbins
self.bin_over = nbins-1 if include_over else nbins
def _hashable_inputs(self):
return super()._hashable_inputs() + (self.bin0, self.binsize, self.bin_under, self.bin_over)
def validate(self, in_dshape):
if self.column not in in_dshape.dict:
raise ValueError("specified column not found")
def apply(self, df, cuda):
if cudf and isinstance(df, cudf.DataFrame):
if Version(cudf.__version__) >= Version("22.02"):
values = df[self.column].to_cupy(na_value=cp.nan)
else:
values = cp.array(df[self.column].to_gpu_array(fillna=True))
nan_values = cp.isnan(values)
else:
values = df[self.column].to_numpy()
nan_values = np.isnan(values)
index_float = (values - self.bin0) / self.binsize
# NaN values are corrected below, so set them to zero to avoid warnings when
# converting from float to int.
index_float[nan_values] = 0
index = index_float.astype(int)
index[index < 0] = self.bin_under
index[index >= self.nbins] = self.bin_over
index[nan_values] = self.nbins
return index
class category_values(CategoryPreprocess):
"""Extract a category and a value column from a dataframe as (2,N) numpy array of values."""
def __init__(self, categorizer, value_column):
super().__init__(value_column)
self.categorizer = categorizer
@property
def inputs(self):
return (self.categorizer.column, self.column)
@property
def cat_column(self):
"""Returns name of categorized column"""
return self.categorizer.column
def categories(self, input_dshape):
return self.categorizer.categories
def validate(self, in_dshape):
return self.categorizer.validate(in_dshape)
def apply(self, df, cuda):
a = self.categorizer.apply(df, cuda)
if cudf and isinstance(df, cudf.DataFrame):
import cupy
if self.column == SpecialColumn.RowIndex:
nullval = -1
elif df[self.column].dtype.kind == 'f':
nullval = np.nan
else:
nullval = 0
a = cupy.asarray(a)
if self.column == SpecialColumn.RowIndex:
b = extract(SpecialColumn.RowIndex).apply(df, cuda)
elif Version(cudf.__version__) >= Version("22.02"):
b = df[self.column].to_cupy(na_value=nullval)
else:
b = cupy.asarray(df[self.column].fillna(nullval))
return cupy.stack((a, b), axis=-1)
else:
if self.column == SpecialColumn.RowIndex:
b = extract(SpecialColumn.RowIndex).apply(df, cuda)
else:
b = df[self.column].values
return np.stack((a, b), axis=-1)
class Reduction(Expr):
"""Base class for per-bin reductions."""
def __init__(self, column: str | SpecialColumn | None=None):
self.column = column
self._nan_check_column = None
@property
def nan_check_column(self):
if self._nan_check_column is not None:
return extract(self._nan_check_column)
else:
return None
def uses_cuda_mutex(self) -> UsesCudaMutex:
"""Return ``True`` if this Reduction needs to use a CUDA mutex to
ensure that it is threadsafe across CUDA threads.
If the CUDA append functions are all atomic (i.e. using functions from
the numba.cuda.atomic module) then this is ``False``, otherwise it is
``True``.
"""
return UsesCudaMutex.No
def uses_row_index(self, cuda, partitioned):
"""Return ``True`` if this Reduction uses a row index virtual column.
For some reductions the order of the rows of supplied data is
important. These include ``first`` and ``last`` reductions as well as
``where`` reductions that return a row index. In some situations the
order is intrinsic such as ``first`` reductions that are processed
sequentially (i.e. on a CPU without using Dask) and no extra column is
required. But in situations of parallel processing (using a GPU or
Dask) extra information is needed that is provided by a row index
virtual column.
Returning ``True`` from this function will cause a row index column to
be created and passed to the ``append`` functions in the usual manner.
"""
return False
def validate(self, in_dshape):
if self.column == SpecialColumn.RowIndex:
return
if self.column not in in_dshape.dict:
raise ValueError("specified column not found")
if not isnumeric(in_dshape.measure[self.column]):
raise ValueError("input must be numeric")
@property
def inputs(self):
return (extract(self.column),)
def is_categorical(self):
"""Return ``True`` if this is or contains a categorical reduction."""
return False
def is_where(self):
"""Return ``True`` if this is a ``where`` reduction or directly wraps
a where reduction."""
return False
def _antialias_requires_2_stages(self):
# Return True if this Reduction must be processed with 2 stages,
# False if it doesn't matter.
# Overridden in derived classes as appropriate.
return False
def _antialias_stage_2(self, self_intersect, array_module) -> tuple[AntialiasStage2]:
# Only called if using antialiased lines. Overridden in derived classes.
# Returns a tuple containing an item for each constituent reduction.
# Each item is (AntialiasCombination, zero_value)).
raise NotImplementedError(f"{type(self)}._antialias_stage_2 is not defined")
def _build_bases(self, cuda, partitioned):
return (self,)
def _build_combine_temps(self, cuda, partitioned):
# Temporaries (i.e. not returned to user) that are reductions, the
# aggs of which are passed to the combine() function but not the
# append() functions, as opposed to _build_temps() which are passed
# to both append() and combine().
return ()
def _build_temps(self, cuda=False):
# Temporaries (i.e. not returned to user) that are reductions, the
# aggs of which are passed to both append() and combine() functions.
return ()
def _build_create(self, required_dshape):
fields = getattr(required_dshape.measure, "fields", None)
if fields is not None and len(required_dshape.measure.fields) > 0:
# If more than one field then they all have the same dtype so can just take the first.
first_field = required_dshape.measure.fields[0]
required_dshape = dshape(first_field[1])
if isinstance(required_dshape, Option):
required_dshape = dshape(required_dshape.ty)
if required_dshape == dshape(ct.bool_):
return self._create_bool
elif required_dshape == dshape(ct.float32):
return self._create_float32_nan
elif required_dshape == dshape(ct.float64):
return self._create_float64_nan
elif required_dshape == dshape(ct.int64):
return self._create_int64
elif required_dshape == dshape(ct.uint32):
return self._create_uint32
else:
raise NotImplementedError(f"Unexpected dshape {dshape}")
def _build_append(self, dshape, schema, cuda, antialias, self_intersect):
if cuda:
if antialias and self.column is None:
return self._append_no_field_antialias_cuda
elif antialias:
return self._append_antialias_cuda
elif self.column is None:
return self._append_no_field_cuda
else:
return self._append_cuda
else:
if antialias and self.column is None:
return self._append_no_field_antialias
elif antialias:
return self._append_antialias
elif self.column is None:
return self._append_no_field
else:
return self._append
def _build_combine(self, dshape, antialias, cuda, partitioned, categorical = False):
return self._combine
def _build_finalize(self, dshape):
return self._finalize
@staticmethod
def _create_bool(shape, array_module):
return array_module.zeros(shape, dtype='bool')
@staticmethod
def _create_float32_nan(shape, array_module):
return array_module.full(shape, array_module.nan, dtype='f4')
@staticmethod
def _create_float64_nan(shape, array_module):
return array_module.full(shape, array_module.nan, dtype='f8')
@staticmethod
def _create_float64_empty(shape, array_module):
return array_module.empty(shape, dtype='f8')
@staticmethod
def _create_float64_zero(shape, array_module):
return array_module.zeros(shape, dtype='f8')
@staticmethod
def _create_int64(shape, array_module):
return array_module.full(shape, -1, dtype='i8')
@staticmethod
def _create_uint32(shape, array_module):
return array_module.zeros(shape, dtype='u4')
def __repr__(self):
return f"{type(self).__name__}({self.column!r})"
class OptionalFieldReduction(Reduction):
"""Base class for things like ``count`` or ``any`` for which the field is optional"""
def __init__(self, column=None):
super().__init__(column)
@property
def inputs(self):
return (extract(self.column),) if self.column is not None else ()
def validate(self, in_dshape):
if self.column is not None:
super().validate(in_dshape)
@staticmethod
def _finalize(bases, cuda=False, **kwargs):
return xr.DataArray(bases[0], **kwargs)
class SelfIntersectingOptionalFieldReduction(OptionalFieldReduction):
"""
Base class for optional field reductions for which self-intersecting
geometry may or may not be desirable.
Ignored if not using antialiasing.
"""
def __init__(self, column=None, self_intersect=True):
super().__init__(column)
self.self_intersect = self_intersect
def _antialias_requires_2_stages(self):
return not self.self_intersect
def _build_append(self, dshape, schema, cuda, antialias, self_intersect):
if antialias and not self_intersect:
# append functions specific to antialiased lines without self_intersect
if cuda:
if self.column is None:
return self._append_no_field_antialias_cuda_not_self_intersect
else:
return self._append_antialias_cuda_not_self_intersect
else:
if self.column is None:
return self._append_no_field_antialias_not_self_intersect
else:
return self._append_antialias_not_self_intersect
# Fall back to base class implementation
return super()._build_append(dshape, schema, cuda, antialias, self_intersect)
def _hashable_inputs(self):
# Reductions with different self_intersect attributes much have different hashes otherwise
# toolz.memoize will treat them as the same to give incorrect results.
return super()._hashable_inputs() + (self.self_intersect,)
class count(SelfIntersectingOptionalFieldReduction):
"""Count elements in each bin, returning the result as a uint32, or a
float32 if using antialiasing.
Parameters
----------
column : str, optional
If provided, only counts elements in ``column`` that are not ``NaN``.
Otherwise, counts every element.
"""
def out_dshape(self, in_dshape, antialias, cuda, partitioned):
return dshape(ct.float32) if antialias else dshape(ct.uint32)
def _antialias_stage_2(self, self_intersect, array_module) -> tuple[AntialiasStage2]:
if self_intersect:
return (AntialiasStage2(AntialiasCombination.SUM_1AGG, array_module.nan),)
else:
return (AntialiasStage2(AntialiasCombination.SUM_2AGG, array_module.nan),)
# CPU append functions
@staticmethod
@ngjit
def _append(x, y, agg, field):
if not isnull(field):
agg[y, x] += 1
return 0
return -1
@staticmethod
@ngjit
def _append_antialias(x, y, agg, field, aa_factor, prev_aa_factor):
if not isnull(field):
if isnull(agg[y, x]):
agg[y, x] = aa_factor - prev_aa_factor
else:
agg[y, x] += aa_factor - prev_aa_factor
return 0
return -1
@staticmethod
@ngjit
def _append_antialias_not_self_intersect(x, y, agg, field, aa_factor, prev_aa_factor):
if not isnull(field):
if isnull(agg[y, x]) or aa_factor > agg[y, x]:
agg[y, x] = aa_factor
return 0
return -1
@staticmethod
@ngjit
def _append_no_field(x, y, agg):
agg[y, x] += 1
return 0
@staticmethod
@ngjit
def _append_no_field_antialias(x, y, agg, aa_factor, prev_aa_factor):
if isnull(agg[y, x]):
agg[y, x] = aa_factor - prev_aa_factor
else:
agg[y, x] += aa_factor - prev_aa_factor
return 0
@staticmethod
@ngjit
def _append_no_field_antialias_not_self_intersect(x, y, agg, aa_factor, prev_aa_factor):
if isnull(agg[y, x]) or aa_factor > agg[y, x]:
agg[y, x] = aa_factor
return 0
return -1
# GPU append functions
@staticmethod
@nb_cuda.jit(device=True)
def _append_antialias_cuda(x, y, agg, field, aa_factor, prev_aa_factor):
value = field*aa_factor
if not isnull(value):
old = cuda_atomic_nanmax(agg, (y, x), value)
if isnull(old) or old < value:
return 0
return -1
@staticmethod
@nb_cuda.jit(device=True)
def _append_no_field_antialias_cuda_not_self_intersect(x, y, agg, aa_factor, prev_aa_factor):
if not isnull(aa_factor):
old = cuda_atomic_nanmax(agg, (y, x), aa_factor)
if isnull(old) or old < aa_factor:
return 0
return -1
@staticmethod
@nb_cuda.jit(device=True)
def _append_cuda(x, y, agg, field):
if not isnull(field):
nb_cuda.atomic.add(agg, (y, x), 1)
return 0
return -1
@staticmethod
@nb_cuda.jit(device=True)
def _append_no_field_antialias_cuda(x, y, agg, aa_factor, prev_aa_factor):
if not isnull(aa_factor):
old = cuda_atomic_nanmax(agg, (y, x), aa_factor)
if isnull(old) or old < aa_factor:
return 0
return -1
@staticmethod
@nb_cuda.jit(device=True)
def _append_no_field_cuda(x, y, agg):
nb_cuda.atomic.add(agg, (y, x), 1)
return 0
def _build_combine(self, dshape, antialias, cuda, partitioned, categorical = False):
if antialias:
return self._combine_antialias
else:
return self._combine
@staticmethod
def _combine(aggs):
return aggs.sum(axis=0, dtype='u4')
@staticmethod
def _combine_antialias(aggs):
ret = aggs[0]
for i in range(1, len(aggs)):
nansum_in_place(ret, aggs[i])
return ret
def __repr__(self):
return "count()"
class _count_ignore_antialiasing(count):
"""Count reduction but ignores antialiasing. Used by mean reduction.
"""
def out_dshape(self, in_dshape, antialias, cuda, partitioned):
return dshape(ct.uint32)
def _antialias_stage_2(self, self_intersect, array_module) -> tuple[AntialiasStage2]:
if self_intersect:
return (AntialiasStage2(AntialiasCombination.SUM_1AGG, 0),)
else:
return (AntialiasStage2(AntialiasCombination.SUM_2AGG, 0),)
@staticmethod
@ngjit
def _append_antialias(x, y, agg, field, aa_factor, prev_aa_factor):
if not isnull(field) and prev_aa_factor == 0.0:
agg[y, x] += 1
return 0
return -1
@staticmethod
@ngjit
def _append_antialias_not_self_intersect(x, y, agg, field, aa_factor, prev_aa_factor):
if not isnull(field) and prev_aa_factor == 0.0:
agg[y, x] += 1
return 0
return -1
class by(Reduction):
"""Apply the provided reduction separately per category.
Parameters
----------
cats: str or CategoryPreprocess instance
Name of column to aggregate over, or a categorizer object that returns categories.
Resulting aggregate has an outer dimension axis along the categories present.
reduction : Reduction
Per-category reduction function.
"""
def __init__(self, cat_column, reduction=count()):
super().__init__()
# set basic categorizer
if isinstance(cat_column, CategoryPreprocess):
self.categorizer = cat_column
elif isinstance(cat_column, str):
self.categorizer = category_codes(cat_column)
else:
raise TypeError("first argument must be a column name or a CategoryPreprocess instance")
self.column = self.categorizer.column # for backwards compatibility with count_cat
self.columns = (self.categorizer.column,)
if (columns := getattr(reduction, 'columns', None)) is not None:
# Must reverse columns (from where reduction) so that val_column property
# is the column that is returned to the user.
self.columns += columns[::-1]
else:
self.columns += (getattr(reduction, 'column', None),)
self.reduction = reduction
# if a value column is supplied, set category_values preprocessor
if self.val_column is not None:
self.preprocess = category_values(self.categorizer, self.val_column)
else:
self.preprocess = self.categorizer
def __hash__(self):
return hash((type(self), self._hashable_inputs(), self.categorizer._hashable_inputs(),
self.reduction))
def _build_temps(self, cuda=False):
return tuple(by(self.categorizer, tmp) for tmp in self.reduction._build_temps(cuda))
@property
def cat_column(self):
return self.columns[0]
@property
def val_column(self):
return self.columns[1]
def validate(self, in_dshape):
self.preprocess.validate(in_dshape)
self.reduction.validate(in_dshape)
def out_dshape(self, input_dshape, antialias, cuda, partitioned):
cats = self.categorizer.categories(input_dshape)
red_shape = self.reduction.out_dshape(input_dshape, antialias, cuda, partitioned)
return dshape(Record([(c, red_shape) for c in cats]))
@property
def inputs(self):
return (self.preprocess,)
def is_categorical(self):
return True
def is_where(self):
return self.reduction.is_where()
@property
def nan_check_column(self):
return self.reduction.nan_check_column
def uses_cuda_mutex(self) -> UsesCudaMutex:
return self.reduction.uses_cuda_mutex()
def uses_row_index(self, cuda, partitioned):
return self.reduction.uses_row_index(cuda, partitioned)
def _antialias_requires_2_stages(self):
return self.reduction._antialias_requires_2_stages()
def _antialias_stage_2(self, self_intersect, array_module) -> tuple[AntialiasStage2]:
ret = self.reduction._antialias_stage_2(self_intersect, array_module)
return (AntialiasStage2(combination=ret[0].combination,
zero=ret[0].zero,
n_reduction=ret[0].n_reduction,
categorical=True),)
def _build_create(self, required_dshape):
n_cats = len(required_dshape.measure.fields)
return lambda shape, array_module: self.reduction._build_create(
required_dshape)(shape + (n_cats,), array_module)
def _build_bases(self, cuda, partitioned):
bases = self.reduction._build_bases(cuda, partitioned)
if len(bases) == 1 and bases[0] is self:
return bases
return tuple(by(self.categorizer, base) for base in bases)
def _build_append(self, dshape, schema, cuda, antialias, self_intersect):
return self.reduction._build_append(dshape, schema, cuda, antialias, self_intersect)
def _build_combine(self, dshape, antialias, cuda, partitioned, categorical = False):
return self.reduction._build_combine(dshape, antialias, cuda, partitioned, True)
def _build_combine_temps(self, cuda, partitioned):
return self.reduction._build_combine_temps(cuda, partitioned)
def _build_finalize(self, dshape):
cats = list(self.categorizer.categories(dshape))
def finalize(bases, cuda=False, **kwargs):
# Return a modified copy of kwargs. Cannot modify supplied kwargs as it
# may be used by multiple reductions, e.g. if a summary reduction.
kwargs = copy.deepcopy(kwargs)
kwargs['dims'] += [self.cat_column]
kwargs['coords'][self.cat_column] = cats
return self.reduction._build_finalize(dshape)(bases, cuda=cuda, **kwargs)
return finalize
def __repr__(self):
return f"{type(self).__name__}(column={self.column!r}, reduction={self.reduction!r})"
class any(OptionalFieldReduction):
"""Whether any elements in ``column`` map to each bin.
Parameters
----------
column : str, optional
If provided, any elements in ``column`` that are ``NaN`` are skipped.
"""
def out_dshape(self, in_dshape, antialias, cuda, partitioned):
return dshape(ct.float32) if antialias else dshape(ct.bool_)
def _antialias_stage_2(self, self_intersect, array_module) -> tuple[AntialiasStage2]:
return (AntialiasStage2(AntialiasCombination.MAX, array_module.nan),)
# CPU append functions
@staticmethod
@ngjit
def _append(x, y, agg, field):
if not isnull(field):
agg[y, x] = True
return 0
return -1
@staticmethod
@ngjit
def _append_antialias(x, y, agg, field, aa_factor, prev_aa_factor):
if not isnull(field):
if isnull(agg[y, x]) or aa_factor > agg[y, x]:
agg[y, x] = aa_factor
return 0
return -1
@staticmethod
@ngjit
def _append_no_field(x, y, agg):
agg[y, x] = True
return 0
@staticmethod
@ngjit
def _append_no_field_antialias(x, y, agg, aa_factor, prev_aa_factor):
if isnull(agg[y, x]) or aa_factor > agg[y, x]:
agg[y, x] = aa_factor
return 0
return -1
# GPU append functions
_append_cuda =_append
_append_no_field_cuda = _append_no_field
def _build_combine(self, dshape, antialias, cuda, partitioned, categorical = False):
if antialias:
return self._combine_antialias
else:
return self._combine
@staticmethod
def _combine(aggs):
return aggs.sum(axis=0, dtype='bool')
@staticmethod
def _combine_antialias(aggs):
ret = aggs[0]
for i in range(1, len(aggs)):
nanmax_in_place(ret, aggs[i])
return ret
class _upsample(Reduction):
""""Special internal class used for upsampling"""
def out_dshape(self, in_dshape, antialias, cuda, partitioned):
return dshape(Option(ct.float64))
@staticmethod
def _finalize(bases, cuda=False, **kwargs):
return xr.DataArray(bases[0], **kwargs)
@property
def inputs(self):
return (extract(self.column),)
def _build_create(self, required_dshape):
# Use uninitialized memory, the upsample function must explicitly set unused
# values to nan
return self._create_float64_empty
@staticmethod
@ngjit
def _append(x, y, agg, field):
# not called, the upsample function must set agg directly
pass
@staticmethod
@nb_cuda.jit(device=True)
def _append_cuda(x, y, agg, field):
# not called, the upsample function must set agg directly
pass
@staticmethod
def _combine(aggs):
return np.nanmax(aggs, axis=0)
class FloatingReduction(Reduction):
"""Base classes for reductions that always have floating-point dtype."""
def out_dshape(self, in_dshape, antialias, cuda, partitioned):
return dshape(Option(ct.float64))
@staticmethod
def _finalize(bases, cuda=False, **kwargs):
return xr.DataArray(bases[0], **kwargs)
class _sum_zero(FloatingReduction):
"""Sum of all elements in ``column``.
Parameters
----------
column : str
Name of the column to aggregate over. Column data type must be numeric.
"""
def _antialias_stage_2(self, self_intersect, array_module) -> tuple[AntialiasStage2]:
if self_intersect:
return (AntialiasStage2(AntialiasCombination.SUM_1AGG, 0),)
else:
return (AntialiasStage2(AntialiasCombination.SUM_2AGG, 0),)
def _build_create(self, required_dshape):
return self._create_float64_zero
# CPU append functions.
@staticmethod
@ngjit
def _append(x, y, agg, field):
if not isnull(field):
# agg[y, x] cannot be null as initialised to zero.
agg[y, x] += field
return 0
return -1
@staticmethod
@ngjit
def _append_antialias(x, y, agg, field, aa_factor, prev_aa_factor):
value = field*(aa_factor - prev_aa_factor)
if not isnull(value):
# agg[y, x] cannot be null as initialised to zero.
agg[y, x] += value
return 0
return -1
@staticmethod
@ngjit
def _append_antialias_not_self_intersect(x, y, agg, field, aa_factor, prev_aa_factor):
value = field*aa_factor
if not isnull(value) and value > agg[y, x]:
# agg[y, x] cannot be null as initialised to zero.
agg[y, x] = value
return 0
return -1
# GPU append functions
@staticmethod
@nb_cuda.jit(device=True)
def _append_cuda(x, y, agg, field):
if not isnull(field):
nb_cuda.atomic.add(agg, (y, x), field)
return 0
return -1
@staticmethod
def _combine(aggs):
return aggs.sum(axis=0, dtype='f8')
class SelfIntersectingFloatingReduction(FloatingReduction):
"""