This repository was archived by the owner on Feb 2, 2024. It is now read-only.
Repository navigation
Expand file tree
/
Copy pathdistributed.py
More file actions
2129 lines (1881 loc) · 95.9 KB
/
Copy pathdistributed.py
File metadata and controls
2129 lines (1881 loc) · 95.9 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
# *****************************************************************************
# Copyright (c) 2020, Intel Corporation All rights reserved.
#
# Redistribution and use in source and binary forms, with or without
# modification, are permitted provided that the following conditions are met:
#
# Redistributions of source code must retain the above copyright notice,
# this list of conditions and the following disclaimer.
#
# Redistributions in binary form must reproduce the above copyright notice,
# this list of conditions and the following disclaimer in the documentation
# and/or other materials provided with the distribution.
#
# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
# AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO,
# THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
# PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR
# CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
# EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
# PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
# OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY,
# WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR
# OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE,
# EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
# *****************************************************************************
"""
.. module:: distributed.py
The description of the entire module will be here.
Supported and unsupported list can also be added here
"""
from __future__ import print_function, division, absolute_import
import operator
import types as pytypes # avoid confusion with numba.types
import copy
import warnings
from collections import defaultdict
import numpy as np
import numba
from numba import types
from numba.core import ir, ir_utils
from numba.core import postproc
from numba.core.ir_utils import (
mk_unique_var,
replace_vars_inner,
find_topo_order,
dprint_func_ir,
remove_dead,
mk_alloc,
get_global_func_typ,
get_name_var_table,
get_call_table,
get_tuple_table,
remove_dels,
compile_to_numba_ir,
replace_arg_nodes,
guard,
get_definition,
require,
GuardException,
find_callname,
build_definitions,
find_build_sequence,
find_const,
is_get_setitem)
from numba.core.inline_closurecall import inline_closure_call
from numba.parfors.parfor import (
Parfor,
lower_parfor_sequential,
get_parfor_reductions,
get_parfor_params,
wrap_parfor_blocks,
unwrap_parfor_blocks)
from numba.core.compiler_machinery import FunctionPass, register_pass
import sdc
import sdc.utilities.utils
from sdc import distributed_api, distributed_lower
from sdc.str_ext import string_type
from sdc.str_arr_type import string_array_type
from sdc.distributed_api import Reduce_Type
from sdc.distributed_analysis import Distribution, DistributedAnalysis
from sdc.utilities.utils import (
is_alloc_callname,
is_whole_slice,
is_array_container,
get_slice_step,
is_array,
is_np_array,
find_build_tuple,
debug_prints,
ReplaceFunc,
gen_getitem,
is_call,
is_const_slice,
update_globals)
from sdc.hiframes.pd_dataframe_ext import DataFrameType
distributed_run_extensions = {}
# analysis data for debugging
dist_analysis = None
fir_text = None
@register_pass(mutates_CFG=True, analysis_only=False)
class DistributedPass(FunctionPass):
"""The summary of the class should be here for example below is the summary line for this class
This is an adapter for a new numba passes interface. Numba pass must be stateless. This class wraps statefull DistributedPassImpl
"""
_name = "sdc_extention_distributed_pass"
def __init__(self):
pass
def run_pass(self, state):
return DistributedPassImpl(state).run_pass()
class DistributedPassImpl(object):
"""The summary of the class should be here for example below is the summary line for this class
This class analyzes program and transforms to distributed
"""
def __init__(self, state):
self._dist_analysis = None
self._T_arrs = None # set of transposed arrays (taken from analysis)
self._rank_var = None # will be set in run
self._size_var = None
self._g_dist_var = None
self._set1_var = None # variable set to 1
self._set0_var = None # variable set to 0
self._array_starts = {}
self._array_counts = {}
# keep shape attr calls on parallel arrays like X.shape
self._shape_attrs = {}
# keep array sizes of parallel arrays to handle shape attrs
self._array_sizes = {}
# save output of converted 1DVar array len() variables
# which are global sizes, in order to recover local
# size for 1DVar allocs and parfors
self.oneDVar_len_vars = {}
self.state = state
def run_pass(self):
remove_dels(self.state.func_ir.blocks)
dprint_func_ir(self.state.func_ir, "starting distributed pass")
self.state.func_ir._definitions = build_definitions(self.state.func_ir.blocks)
dist_analysis_pass = DistributedAnalysis(
self.state.func_ir, self.state.typemap, self.state.calltypes, self.state.typingctx,
self.state.metadata)
self._dist_analysis = dist_analysis_pass.run()
# dprint_func_ir(self.state.func_ir, "after analysis distributed")
self._T_arrs = dist_analysis_pass._T_arrs
self._parallel_accesses = dist_analysis_pass._parallel_accesses
if debug_prints(): # pragma: no cover
print("distributions: ", self._dist_analysis)
self._gen_dist_inits()
self.state.func_ir._definitions = build_definitions(self.state.func_ir.blocks)
self.state.func_ir.blocks = self._run_dist_pass(self.state.func_ir.blocks)
self.state.func_ir.blocks = self._dist_prints(self.state.func_ir.blocks)
remove_dead(self.state.func_ir.blocks, self.state.func_ir.arg_names, self.state.func_ir, self.state.typemap)
dprint_func_ir(self.state.func_ir, "after distributed pass")
lower_parfor_sequential(
self.state.typingctx, self.state.func_ir, self.state.typemap, self.state.calltypes)
if sdc.multithread_mode:
# parfor params need to be updated for multithread_mode since some
# new variables like alloc_start are introduced by distributed pass
# and are used in later parfors
parfor_ids = get_parfor_params(
self.state.func_ir.blocks, True, defaultdict(list))
post_proc = postproc.PostProcessor(self.state.func_ir)
post_proc.run()
# save data for debug and test
global dist_analysis, fir_text
dist_analysis = self._dist_analysis
import io
str_io = io.StringIO()
self.state.func_ir.dump(str_io)
fir_text = str_io.getvalue()
str_io.close()
return True
def _run_dist_pass(self, blocks):
"""This function does something"""
topo_order = find_topo_order(blocks)
namevar_table = get_name_var_table(blocks)
work_list = list((l, blocks[l]) for l in reversed(topo_order))
while work_list:
label, block = work_list.pop()
new_body = []
replaced = False
for i, inst in enumerate(block.body):
out_nodes = None
if type(inst) in distributed_run_extensions:
f = distributed_run_extensions[type(inst)]
out_nodes = f(inst, self._dist_analysis.array_dists,
self.state.typemap, self.state.calltypes, self.state.typingctx,
self.state.targetctx, self)
elif isinstance(inst, Parfor):
out_nodes = self._run_parfor(inst, namevar_table)
# run dist pass recursively
p_blocks = wrap_parfor_blocks(inst)
# build_definitions(p_blocks, self.state.func_ir._definitions)
self._run_dist_pass(p_blocks)
unwrap_parfor_blocks(inst)
elif isinstance(inst, ir.Assign):
lhs = inst.target.name
rhs = inst.value
if isinstance(rhs, ir.Expr):
out_nodes = self._run_expr(inst, namevar_table)
elif isinstance(rhs, ir.Var) and (self._is_1D_arr(rhs.name)
and not is_array_container(self.state.typemap, rhs.name)):
self._array_starts[lhs] = self._array_starts[rhs.name]
self._array_counts[lhs] = self._array_counts[rhs.name]
self._array_sizes[lhs] = self._array_sizes[rhs.name]
elif isinstance(rhs, ir.Arg):
out_nodes = self._run_arg(inst)
elif isinstance(inst, (ir.StaticSetItem, ir.SetItem)):
if isinstance(inst, ir.SetItem):
index = inst.index
else:
index = inst.index_var
out_nodes = self._run_getsetitem(inst.target,
index, inst, inst)
elif isinstance(inst, ir.Return):
out_nodes = self._gen_barrier() + [inst]
if out_nodes is None:
new_body.append(inst)
elif isinstance(out_nodes, list):
new_body += out_nodes
elif isinstance(out_nodes, ReplaceFunc):
rp_func = out_nodes
if rp_func.pre_nodes is not None:
new_body.extend(rp_func.pre_nodes)
# inline_closure_call expects a call assignment
dummy_call = ir.Expr.call(
ir.Var(block.scope, "dummy", inst.loc),
rp_func.args, (), inst.loc)
if isinstance(inst, ir.Assign):
# replace inst.value to a call with target args
# as expected by inline_closure_call
inst.value = dummy_call
else:
# replace inst with dummy assignment
# for cases like SetItem
loc = block.loc
dummy_var = ir.Var(
block.scope, mk_unique_var("r_dummy"), loc)
block.body[i] = ir.Assign(dummy_call, dummy_var, loc)
block.body = new_body + block.body[i:]
# TODO: use Parfor loop blocks when replacing funcs in
# parfor loop body
update_globals(rp_func.func, rp_func.glbls)
inline_closure_call(self.state.func_ir, rp_func.glbls,
block, len(new_body), rp_func.func, self.state.typingctx,
rp_func.arg_types,
self.state.typemap, self.state.calltypes, work_list)
replaced = True
break
else:
assert False, "invalid dist pass out nodes"
if not replaced:
blocks[label].body = new_body
return blocks
def _run_expr(self, inst, namevar_table):
lhs = inst.target.name
rhs = inst.value
nodes = [inst]
if rhs.op == 'call':
return self._run_call(inst)
# we save array start/count for data pointer to enable
# file read
if (rhs.op == 'getattr' and rhs.attr == 'ctypes'
and (self._is_1D_arr(rhs.value.name))):
arr_name = rhs.value.name
self._array_starts[lhs] = self._array_starts[arr_name]
self._array_counts[lhs] = self._array_counts[arr_name]
self._array_sizes[lhs] = self._array_sizes[arr_name]
if (rhs.op == 'getattr'
and (self._is_1D_arr(rhs.value.name)
or self._is_1D_Var_arr(rhs.value.name))
and rhs.attr == 'size'):
return self._run_array_size(inst.target, rhs.value)
if (rhs.op == 'static_getitem'
and rhs.value.name in self._shape_attrs):
arr = self._shape_attrs[rhs.value.name]
ndims = self.state.typemap[arr].ndim
if arr not in self._T_arrs and rhs.index == 0:
# return parallel size
if self._is_1D_arr(arr):
# XXX hack for array container case, TODO: handle properly
if arr not in self._array_sizes:
arr_var = namevar_table[arr]
nodes = self._gen_1D_Var_len(arr_var)
nodes[-1].target = inst.target
return nodes
inst.value = self._array_sizes[arr][rhs.index]
else:
assert self._is_1D_Var_arr(arr)
arr_var = namevar_table[arr]
nodes = self._gen_1D_Var_len(arr_var)
nodes[-1].target = inst.target
# save output of converted 1DVar array len() variables
# which are global sizes, in order to recover local
# size for 1DVar allocs and parfors
self.oneDVar_len_vars[inst.target.name] = arr_var
return nodes
# last dimension of transposed arrays is partitioned
if arr in self._T_arrs and rhs.index == ndims - 1:
assert not self._is_1D_Var_arr(
arr), "1D_Var arrays cannot transpose"
inst.value = self._array_sizes[arr][rhs.index]
if rhs.op in ['getitem', 'static_getitem']:
if rhs.op == 'getitem':
index = rhs.index
else:
index = rhs.index_var
return self._run_getsetitem(rhs.value, index, rhs, inst)
if (rhs.op == 'getattr'
and (self._is_1D_arr(rhs.value.name)
or self._is_1D_Var_arr(rhs.value.name))
and rhs.attr == 'shape'):
# XXX: return a new tuple using sizes here?
self._shape_attrs[lhs] = rhs.value.name
if (rhs.op == 'getattr'
and self._is_1D_arr(rhs.value.name)
and rhs.attr == 'T'):
assert lhs in self._T_arrs
orig_arr = rhs.value.name
self._array_starts[lhs] = copy.copy(
self._array_starts[orig_arr]).reverse()
self._array_counts[lhs] = copy.copy(
self._array_counts[orig_arr]).reverse()
self._array_sizes[lhs] = copy.copy(
self._array_sizes[orig_arr]).reverse()
if (rhs.op == 'exhaust_iter'
and rhs.value.name in self._shape_attrs):
self._shape_attrs[lhs] = self._shape_attrs[rhs.value.name]
if rhs.op == 'inplace_binop' and self._is_1D_arr(rhs.lhs.name):
self._array_starts[lhs] = self._array_starts[rhs.lhs.name]
self._array_counts[lhs] = self._array_counts[rhs.lhs.name]
self._array_sizes[lhs] = self._array_sizes[rhs.lhs.name]
return nodes
def _gen_1D_Var_len(self, arr):
def f(A, op): # pragma: no cover
c = len(A)
res = sdc.distributed_api.dist_reduce(c, op)
f_block = compile_to_numba_ir(f, {'sdc': sdc}, self.state.typingctx,
(self.state.typemap[arr.name], types.int32),
self.state.typemap, self.state.calltypes).blocks.popitem()[1]
replace_arg_nodes(
f_block, [arr, ir.Const(Reduce_Type.Sum.value, arr.loc)])
nodes = f_block.body[:-3] # remove none return
return nodes
def _gen_dist_inits(self):
# add initializations
topo_order = find_topo_order(self.state.func_ir.blocks)
first_block = self.state.func_ir.blocks[topo_order[0]]
# set scope and loc of generated code to the first variable in block
scope = first_block.scope
loc = first_block.loc
out = []
self._set1_var = ir.Var(scope, mk_unique_var("$const_parallel"), loc)
self.state.typemap[self._set1_var.name] = types.int64
set1_assign = ir.Assign(ir.Const(1, loc), self._set1_var, loc)
out.append(set1_assign)
self._set0_var = ir.Var(scope, mk_unique_var("$const_parallel"), loc)
self.state.typemap[self._set0_var.name] = types.int64
set0_assign = ir.Assign(ir.Const(0, loc), self._set0_var, loc)
out.append(set0_assign)
# g_dist_var = Global(sdc.distributed_api)
g_dist_var = ir.Var(scope, mk_unique_var("$distributed_g_var"), loc)
self._g_dist_var = g_dist_var
self.state.typemap[g_dist_var.name] = types.misc.Module(sdc.distributed_api)
g_dist = ir.Global('distributed_api', sdc.distributed_api, loc)
g_dist_assign = ir.Assign(g_dist, g_dist_var, loc)
# attr call: rank_attr = getattr(g_dist_var, get_rank)
rank_attr_call = ir.Expr.getattr(g_dist_var, "get_rank", loc)
rank_attr_var = ir.Var(scope, mk_unique_var("$get_rank_attr"), loc)
self.state.typemap[rank_attr_var.name] = get_global_func_typ(
distributed_api.get_rank)
rank_attr_assign = ir.Assign(rank_attr_call, rank_attr_var, loc)
# rank_var = sdc.distributed_api.get_rank()
rank_var = ir.Var(scope, mk_unique_var("$rank"), loc)
self.state.typemap[rank_var.name] = types.int32
rank_call = ir.Expr.call(rank_attr_var, [], (), loc)
self.state.calltypes[rank_call] = self.state.typemap[rank_attr_var.name].get_call_type(
self.state.typingctx, [], {})
rank_assign = ir.Assign(rank_call, rank_var, loc)
self._rank_var = rank_var
out += [g_dist_assign, rank_attr_assign, rank_assign]
# attr call: size_attr = getattr(g_dist_var, get_size)
size_attr_call = ir.Expr.getattr(g_dist_var, "get_size", loc)
size_attr_var = ir.Var(scope, mk_unique_var("$get_size_attr"), loc)
self.state.typemap[size_attr_var.name] = get_global_func_typ(
distributed_api.get_size)
size_attr_assign = ir.Assign(size_attr_call, size_attr_var, loc)
# size_var = sdc.distributed_api.get_size()
size_var = ir.Var(scope, mk_unique_var("$dist_size"), loc)
self.state.typemap[size_var.name] = types.int32
size_call = ir.Expr.call(size_attr_var, [], (), loc)
self.state.calltypes[size_call] = self.state.typemap[size_attr_var.name].get_call_type(
self.state.typingctx, [], {})
size_assign = ir.Assign(size_call, size_var, loc)
self._size_var = size_var
out += [size_attr_assign, size_assign]
first_block.body = out + first_block.body
def _run_call(self, assign):
lhs = assign.target.name
rhs = assign.value
func_var = rhs.func.name
scope = assign.target.scope
loc = assign.target.loc
out = [assign]
func_name = ""
func_mod = ""
fdef = guard(find_callname, self.state.func_ir, rhs, self.state.typemap)
if fdef is None:
# FIXME: since parfors are transformed and then processed
# recursively, some funcs don't have definitions. The generated
# arrays should be assigned REP and the var definitions added.
# warnings.warn(
# "function call couldn't be found for distributed pass")
return out
else:
func_name, func_mod = fdef
# divide 1D alloc
# XXX allocs should be matched before going to _run_call_np
if self._is_1D_arr(lhs) and is_alloc_callname(func_name, func_mod):
# XXX for pre_alloc_string_array(n, nc), we assume nc is local
# value (updated only in parfor like _str_replace_regex_impl)
size_var = rhs.args[0]
out, new_size_var = self._run_alloc(size_var, lhs, scope, loc)
# empty_inferred is tuple for some reason
rhs.args = list(rhs.args)
rhs.args[0] = new_size_var
out.append(assign)
return out
# fix 1D_Var allocs in case global len of another 1DVar is used
if self._is_1D_Var_arr(lhs) and is_alloc_callname(func_name, func_mod):
size_var = rhs.args[0]
out, new_size_var = self._fix_1D_Var_alloc(
size_var, lhs, scope, loc)
# empty_inferred is tuple for some reason
rhs.args = list(rhs.args)
rhs.args[0] = new_size_var
out.append(assign)
return out
# numpy direct functions
if isinstance(func_mod, str) and func_mod == 'numpy':
return self._run_call_np(lhs, func_name, assign, rhs.args)
# array.func calls
if isinstance(func_mod, ir.Var) and is_np_array(self.state.typemap, func_mod.name):
return self._run_call_array(lhs, func_mod, func_name, assign, rhs.args)
# df.func calls
if isinstance(func_mod, ir.Var) and isinstance(self.state.typemap[func_mod.name], DataFrameType):
return self._run_call_df(lhs, func_mod, func_name, assign, rhs.args)
# string_array.func_calls
if (self._is_1D_arr(lhs) and isinstance(func_mod, ir.Var)
and self.state.typemap[func_mod.name] == string_array_type):
if func_name == 'copy':
self._array_starts[lhs] = self._array_starts[func_mod.name]
self._array_counts[lhs] = self._array_counts[func_mod.name]
self._array_sizes[lhs] = self._array_sizes[func_mod.name]
if fdef == ('permutation', 'numpy.random'):
if self.state.typemap[rhs.args[0].name] == types.int64:
self._array_sizes[lhs] = [rhs.args[0]]
return self._run_permutation_int(assign, rhs.args)
# len(A) if A is 1D
if fdef == ('len', 'builtins') and rhs.args and self._is_1D_arr(rhs.args[0].name):
arr = rhs.args[0].name
assign.value = self._array_sizes[arr][0]
# len(A) if A is 1D_Var
if fdef == ('len', 'builtins') and rhs.args and self._is_1D_Var_arr(rhs.args[0].name):
arr_var = rhs.args[0]
out = self._gen_1D_Var_len(arr_var)
out[-1].target = assign.target
self.oneDVar_len_vars[assign.target.name] = arr_var
if (sdc.config._has_pyarrow
and fdef == ('read_parquet', 'sdc.io.parquet_pio')
and self._is_1D_arr(rhs.args[2].name)):
arr = rhs.args[2].name
assert len(self._array_starts[arr]) == 1, "only 1D arrs in parquet"
start_var = self._array_starts[arr][0]
count_var = self._array_counts[arr][0]
rhs.args += [start_var, count_var]
def f(fname, cindex, arr, out_dtype, start, count): # pragma: no cover
return sdc.io.parquet_pio.read_parquet_parallel(fname, cindex,
arr, out_dtype, start, count)
return self._replace_func(f, rhs.args)
if (sdc.config._has_pyarrow
and fdef == ('read_parquet_str', 'sdc.io.parquet_pio')
and self._is_1D_arr(lhs)):
arr = lhs
size_var = rhs.args[2]
assert self.state.typemap[size_var.name] == types.intp
self._array_sizes[arr] = [size_var]
out, start_var, count_var = self._gen_1D_div(size_var, scope, loc,
"$alloc", "get_node_portion", distributed_api.get_node_portion)
self._array_starts[lhs] = [start_var]
self._array_counts[lhs] = [count_var]
rhs.args[2] = start_var
rhs.args.append(count_var)
def f(fname, cindex, start, count): # pragma: no cover
return sdc.io.parquet_pio.read_parquet_str_parallel(fname, cindex,
start, count)
f_block = compile_to_numba_ir(f, {'sdc': sdc}, self.state.typingctx,
(self.state.typemap[rhs.args[0].name], types.intp,
types.intp, types.intp),
self.state.typemap, self.state.calltypes).blocks.popitem()[1]
replace_arg_nodes(f_block, rhs.args)
out += f_block.body[:-2]
out[-1].target = assign.target
if (func_mod == 'sdc.hiframes.api' and func_name in (
'to_arr_from_series', 'ts_series_to_arr_typ',
'to_date_series_type', 'init_series')
and self._is_1D_arr(rhs.args[0].name)):
# TODO: handle index
in_arr = rhs.args[0].name
self._array_starts[lhs] = self._array_starts[in_arr]
self._array_counts[lhs] = self._array_counts[in_arr]
self._array_sizes[lhs] = self._array_sizes[in_arr]
if (fdef == ('init_dataframe', 'sdc.hiframes.pd_dataframe_ext')
and self._is_1D_arr(rhs.args[0].name)):
in_arr = rhs.args[0].name
self._array_starts[lhs] = self._array_starts[in_arr]
self._array_counts[lhs] = self._array_counts[in_arr]
self._array_sizes[lhs] = self._array_sizes[in_arr]
if (fdef == ('compute_split_view', 'sdc.hiframes.split_impl')
and self._is_1D_arr(rhs.args[0].name)):
in_arr = rhs.args[0].name
self._array_starts[lhs] = self._array_starts[in_arr]
self._array_counts[lhs] = self._array_counts[in_arr]
self._array_sizes[lhs] = self._array_sizes[in_arr]
if (fdef == ('get_split_view_index', 'sdc.hiframes.split_impl')
and self._is_1D_arr(rhs.args[0].name)):
arr = rhs.args[0]
index_var = rhs.args[1]
sub_nodes = self._get_ind_sub(
index_var, self._array_starts[arr.name][0])
out = sub_nodes
rhs.args[1] = sub_nodes[-1].target
out.append(assign)
return out
if (fdef == ('setitem_str_arr_ptr', 'sdc.str_arr_ext')
and self._is_1D_arr(rhs.args[0].name)):
arr = rhs.args[0]
index_var = rhs.args[1]
sub_nodes = self._get_ind_sub(
index_var, self._array_starts[arr.name][0])
out = sub_nodes
rhs.args[1] = sub_nodes[-1].target
out.append(assign)
return out
if (fdef == ('str_arr_item_to_numeric', 'sdc.str_arr_ext')
and self._is_1D_arr(rhs.args[0].name)):
# TODO: test parallel
arr = rhs.args[0]
index_var = rhs.args[1]
sub_nodes = self._get_ind_sub(
index_var, self._array_starts[arr.name][0])
out = sub_nodes
rhs.args[1] = sub_nodes[-1].target
# input string array
arr = rhs.args[2]
index_var = rhs.args[3]
sub_nodes = self._get_ind_sub(
index_var, self._array_starts[arr.name][0])
out += sub_nodes
rhs.args[3] = sub_nodes[-1].target
out.append(assign)
return out
if fdef == ('isna', 'sdc.hiframes.api') and self._is_1D_arr(rhs.args[0].name):
# fix index in call to isna
arr = rhs.args[0]
ind = rhs.args[1]
out = self._get_ind_sub(ind, self._array_starts[arr.name][0])
rhs.args[1] = out[-1].target
out.append(assign)
if fdef == ('rolling_fixed', 'sdc.hiframes.rolling') and (
self._is_1D_arr(rhs.args[0].name)
or self._is_1D_Var_arr(rhs.args[0].name)):
in_arr = rhs.args[0].name
if self._is_1D_arr(in_arr):
self._array_starts[lhs] = self._array_starts[in_arr]
self._array_counts[lhs] = self._array_counts[in_arr]
self._array_sizes[lhs] = self._array_sizes[in_arr]
# set parallel flag to true
true_var = ir.Var(scope, mk_unique_var("true_var"), loc)
self.state.typemap[true_var.name] = types.boolean
rhs.args[3] = true_var
out = [ir.Assign(ir.Const(True, loc), true_var, loc), assign]
if fdef == ('rolling_variable', 'sdc.hiframes.rolling') and (
self._is_1D_arr(rhs.args[0].name)
or self._is_1D_Var_arr(rhs.args[0].name)):
in_arr = rhs.args[0].name
if self._is_1D_arr(in_arr):
self._array_starts[lhs] = self._array_starts[in_arr]
self._array_counts[lhs] = self._array_counts[in_arr]
self._array_sizes[lhs] = self._array_sizes[in_arr]
# set parallel flag to true
true_var = ir.Var(scope, mk_unique_var("true_var"), loc)
self.state.typemap[true_var.name] = types.boolean
rhs.args[4] = true_var
out = [ir.Assign(ir.Const(True, loc), true_var, loc), assign]
if (func_mod == 'sdc.hiframes.rolling'
and func_name in ('shift', 'pct_change')
and (self._is_1D_arr(rhs.args[0].name)
or self._is_1D_Var_arr(rhs.args[0].name))):
in_arr = rhs.args[0].name
if self._is_1D_arr(in_arr):
self._array_starts[lhs] = self._array_starts[in_arr]
self._array_counts[lhs] = self._array_counts[in_arr]
self._array_sizes[lhs] = self._array_sizes[in_arr]
# set parallel flag to true
true_var = ir.Var(scope, mk_unique_var("true_var"), loc)
self.state.typemap[true_var.name] = types.boolean
rhs.args[2] = true_var
out = [ir.Assign(ir.Const(True, loc), true_var, loc), assign]
if fdef == ('quantile', 'sdc.hiframes.api') and (self._is_1D_arr(rhs.args[0].name)
or self._is_1D_Var_arr(rhs.args[0].name)):
arr = rhs.args[0].name
if arr in self._array_sizes:
assert len(self._array_sizes[arr]
) == 1, "only 1D arrs in quantile"
size_var = self._array_sizes[arr][0]
else:
size_var = self._set0_var
rhs.args += [size_var]
def f(arr, q, size):
return sdc.hiframes.api.quantile_parallel(arr, q, size)
return self._replace_func(f, rhs.args)
if fdef == (
'nunique', 'sdc.hiframes.api') and (
self._is_1D_arr(
rhs.args[0].name) or self._is_1D_Var_arr(
rhs.args[0].name)):
def f(arr):
return sdc.hiframes.api.nunique_parallel(arr)
return self._replace_func(f, rhs.args)
if fdef == (
'unique', 'sdc.hiframes.api') and (
self._is_1D_arr(
rhs.args[0].name) or self._is_1D_Var_arr(
rhs.args[0].name)):
def f(arr):
return sdc.hiframes.api.unique_parallel(arr)
return self._replace_func(f, rhs.args)
if fdef == (
'nlargest', 'sdc.hiframes.api') and (
self._is_1D_arr(
rhs.args[0].name) or self._is_1D_Var_arr(
rhs.args[0].name)):
def f(arr, k, i, f):
return sdc.hiframes.api.nlargest_parallel(arr, k, i, f)
return self._replace_func(f, rhs.args)
if fdef == (
'median', 'sdc.hiframes.api') and (
self._is_1D_arr(
rhs.args[0].name) or self._is_1D_Var_arr(
rhs.args[0].name)):
def f(arr):
return sdc.hiframes.api.median(arr, True)
return self._replace_func(f, rhs.args)
if fdef == ('convert_rec_to_tup', 'sdc.hiframes.api'):
# optimize Series back to back map pattern with tuples
# TODO: create another optimization pass?
arg_def = guard(get_definition, self.state.func_ir, rhs.args[0])
if (is_call(arg_def) and guard(find_callname, self.state.func_ir, arg_def)
== ('convert_tup_to_rec', 'sdc.hiframes.api')):
assign.value = arg_def.args[0]
return out
if fdef == ('dist_return', 'sdc.distributed_api'):
# always rebalance returned distributed arrays
# TODO: need different flag for 1D_Var return (distributed_var)?
# TODO: rebalance strings?
# return [assign] # self._run_call_rebalance_array(lhs, assign, rhs.args)
assign.value = rhs.args[0]
return [assign]
if ((fdef == ('get_series_data', 'sdc.hiframes.api')
or fdef == ('get_series_index', 'sdc.hiframes.api')
or fdef == ('get_dataframe_data', 'sdc.hiframes.pd_dataframe_ext'))):
out = [assign]
arr = assign.target
# gen len() using 1D_Var reduce approach.
# TODO: refactor to avoid reduction for 1D
# arr_typ = self.state.typemap[arr.name]
ndim = 1
out += self._gen_1D_Var_len(arr)
total_length = out[-1].target
div_nodes, start_var, count_var = self._gen_1D_div(
total_length, arr.scope, arr.loc, "$input", "get_node_portion", distributed_api.get_node_portion)
out += div_nodes
# XXX: get sizes in lower dimensions
self._array_starts[lhs] = [-1] * ndim
self._array_counts[lhs] = [-1] * ndim
self._array_sizes[lhs] = [-1] * ndim
self._array_starts[lhs][0] = start_var
self._array_counts[lhs][0] = count_var
self._array_sizes[lhs][0] = total_length
return out
if fdef == ('threaded_return', 'sdc.distributed_api'):
assign.value = rhs.args[0]
return [assign]
if fdef == ('rebalance_array', 'sdc.distributed_api'):
return self._run_call_rebalance_array(lhs, assign, rhs.args)
if fdef == ('file_read', 'sdc.io.np_io') and rhs.args[1].name in self._array_starts:
_fname = rhs.args[0]
_data_ptr = rhs.args[1]
_start = self._array_starts[_data_ptr.name][0]
_count = self._array_counts[_data_ptr.name][0]
def f(fname, data_ptr, start, count): # pragma: no cover
return sdc.io.np_io.file_read_parallel(fname, data_ptr, start, count)
return self._replace_func(f, [_fname, _data_ptr, _start, _count])
return out
def _run_call_np(self, lhs, func_name, assign, args):
"""transform np.func() calls
"""
# allocs are handled separately
is_1D_bool = (self._is_1D_Var_arr(lhs) or self._is_1D_arr(lhs))
err_str = "allocation calls handled separately 'empty', 'zeros', 'ones', 'full' etc."
assert not (is_1D_bool and func_name in sdc.utilities.utils.np_alloc_callnames), err_str
out = [assign]
scope = assign.target.scope
loc = assign.loc
# numba doesn't support np.reshape() form yet
# if func_name == 'reshape':
# size_var = args[1]
# # handle reshape like new allocation
# out, new_size_var = self._run_alloc(size_var, lhs)
# args[1] = new_size_var
# out.append(assign)
if (func_name == 'array' and is_array(self.state.typemap, args[0].name) and self._is_1D_arr(args[0].name)):
in_arr = args[0].name
self._array_starts[lhs] = self._array_starts[in_arr]
self._array_counts[lhs] = self._array_counts[in_arr]
self._array_sizes[lhs] = self._array_sizes[in_arr]
# output array has same properties (starts etc.) as input array
if (func_name in ['cumsum', 'cumprod', 'empty_like', 'zeros_like', 'ones_like',
'full_like', 'copy', 'ravel', 'ascontiguousarray'] and self._is_1D_arr(args[0].name)):
if func_name == 'ravel':
assert self.state.typemap[args[0].name].ndim == 1, "only 1D ravel supported"
in_arr = args[0].name
self._array_starts[lhs] = self._array_starts[in_arr]
self._array_counts[lhs] = self._array_counts[in_arr]
self._array_sizes[lhs] = self._array_sizes[in_arr]
if (func_name in ['cumsum', 'cumprod'] and self._is_1D_arr(args[0].name)):
in_arr = args[0].name
in_arr_var = args[0]
lhs_var = assign.target
# allocate output array
# TODO: compute inplace if input array is dead
out = mk_alloc(
self.state.typemap,
self.state.calltypes,
lhs_var,
tuple(
self._array_sizes[in_arr]),
self.state.typemap[in_arr].dtype,
scope,
loc)
# generate distributed call
dist_attr_var = ir.Var(scope, mk_unique_var("$dist_attr"), loc)
dist_func_name = "dist_" + func_name
dist_func = getattr(distributed_api, dist_func_name)
dist_attr_call = ir.Expr.getattr(self._g_dist_var, dist_func_name, loc)
self.state.typemap[dist_attr_var.name] = get_global_func_typ(dist_func)
dist_func_assign = ir.Assign(dist_attr_call, dist_attr_var, loc)
err_var = ir.Var(scope, mk_unique_var("$dist_err_var"), loc)
self.state.typemap[err_var.name] = types.int32
dist_call = ir.Expr.call(dist_attr_var, [in_arr_var, lhs_var], (), loc)
self.state.calltypes[dist_call] = self.state.typemap[dist_attr_var.name].get_call_type(
self.state.typingctx, [self.state.typemap[in_arr], self.state.typemap[lhs]], {})
dist_assign = ir.Assign(dist_call, err_var, loc)
return out + [dist_func_assign, dist_assign]
# sum over the first axis is distributed, A.sum(0)
if func_name == 'sum' and len(args) == 2:
axis_def = guard(get_definition, self.state.func_ir, args[1])
if isinstance(axis_def, ir.Const) and axis_def.value == 0:
reduce_op = Reduce_Type.Sum
reduce_var = assign.target
return out + self._gen_reduce(reduce_var, reduce_op, scope, loc)
if func_name == 'dot':
return self._run_call_np_dot(lhs, assign, args)
if func_name == 'stack' and self._is_1D_arr(lhs):
# TODO: generalize
in_arrs, _ = guard(find_build_sequence, self.state.func_ir, args[0])
arr0 = in_arrs[0].name
self._array_starts[lhs] = [self._array_starts[arr0][0], None]
self._array_counts[lhs] = [self._array_counts[arr0][0], None]
self._array_sizes[lhs] = [self._array_sizes[arr0][0], None]
return out
def _run_call_array(self, lhs, arr, func_name, assign, args):
#
out = [assign]
if func_name in ('astype', 'copy') and self._is_1D_arr(lhs):
self._array_starts[lhs] = self._array_starts[arr.name]
self._array_counts[lhs] = self._array_counts[arr.name]
self._array_sizes[lhs] = self._array_sizes[arr.name]
# HACK support A.reshape(n, 1) for 1D_Var
if func_name == 'reshape' and self._is_1D_Var_arr(arr.name):
assert len(args) == 2 and guard(find_const, self.state.func_ir, args[1]) == 1
assert args[0].name in self.oneDVar_len_vars
size_var = args[0]
out, new_size_var = self._fix_1D_Var_alloc(size_var, lhs, assign.target.scope, assign.loc)
# empty_inferred is tuple for some reason
assign.value.args = list(args)
assign.value.args[0] = new_size_var
out.append(assign)
return out
if func_name == 'reshape' and self._is_1D_arr(arr.name):
return self._run_reshape(assign, arr, args)
if func_name == 'transpose' and self._is_1D_arr(lhs):
# Currently only 1D arrays are supported
assert self._is_1D_arr(arr.name)
ndim = self.state.typemap[arr.name].ndim
self._array_starts[lhs] = [-1] * ndim
self._array_counts[lhs] = [-1] * ndim
self._array_sizes[lhs] = [-1] * ndim
self._array_starts[lhs][0] = self._array_starts[arr.name][0]
self._array_counts[lhs][0] = self._array_counts[arr.name][0]
self._array_sizes[lhs][0] = self._array_sizes[arr.name][0]
# TODO: refactor
# TODO: add unittest
if func_name == 'tofile':
if self._is_1D_arr(arr.name):
_fname = args[0]
_start = self._array_starts[arr.name][0]
_count = self._array_counts[arr.name][0]
def f(fname, arr, start, count): # pragma: no cover
return sdc.io.np_io.file_write_parallel(fname, arr, start, count)
return self._replace_func(f, [_fname, arr, _start, _count])
if self._is_1D_Var_arr(arr.name):
_fname = args[0]
def f(fname, arr): # pragma: no cover
count = len(arr)
start = sdc.distributed_api.dist_exscan(count)
return sdc.io.np_io.file_write_parallel(fname, arr, start, count)
return self._replace_func(f, [_fname, arr])
return out
def _run_call_df(self, lhs, df, func_name, assign, args):
if func_name == 'to_csv' and self._is_1D_arr(df.name):
# set index to proper range if None
# avoid header for non-zero ranks
# write to string then parallel file write
# df.to_csv(fname) ->
# l = len(df)
# index_start = dist_exscan(l)
# df2 = df(index=range(index_start, index_start+l))
# header = header and is_root # only first line has header
# str_out = df2.to_csv(None, header=header)
# sdc.io.np_io._file_write_parallel(fname, str_out)
df_typ = self.state.typemap[df.name]
rhs = assign.value
fname = args[0]
# update df index and get to_csv from new df
nodes = self._fix_parallel_df_index(df)
new_df = nodes[-1].target
new_df_typ = self.state.typemap[new_df.name]
new_to_csv = ir.Expr.getattr(new_df, 'to_csv', new_df.loc)
new_func = ir.Var(new_df.scope, mk_unique_var('func'), new_df.loc)
self.state.typemap[new_func.name] = self.state.typingctx.resolve_getattr(
new_df_typ, 'to_csv')
nodes.append(ir.Assign(new_to_csv, new_func, new_df.loc))
rhs.func = new_func
# # header = header and is_root
kws = dict(rhs.kws)
true_var = ir.Var(assign.target.scope, mk_unique_var('true'), rhs.loc)
self.state.typemap[true_var.name] = types.bool_
nodes.append(
ir.Assign(ir.Const(True, new_df.loc), true_var, new_df.loc))
header_var = self._get_arg(
'to_csv', rhs.args, kws, 5, 'header', true_var)
nodes += self._gen_is_root_and_cond(header_var)
header_var = nodes[-1].target
if len(rhs.args) > 5:
rhs.args[5] = header_var
else:
kws['header'] = header_var
rhs.kws = kws
# fix to_csv() type to have None as 1st arg
call_type = self.state.calltypes.pop(rhs)
arg_typs = list((types.none,) + call_type.args[1:])
arg_typs[5] = types.bool_
arg_typs = tuple(arg_typs)
# self.state.calltypes[rhs] = self.state.typemap[rhs.func.name].get_call_type(
# self.state.typingctx, arg_typs, {})
self.state.calltypes[rhs] = numba.core.typing.Signature(
string_type, arg_typs, new_df_typ,
call_type.pysig)
# None as 1st arg
none_var = ir.Var(assign.target.scope, mk_unique_var('none'), rhs.loc)
self.state.typemap[none_var.name] = types.none
none_assign = ir.Assign(ir.Const(None, rhs.loc), none_var, rhs.loc)
nodes.append(none_assign)
rhs.args[0] = none_var
# str_out = df.to_csv(None)
str_out = ir.Var(assign.target.scope, mk_unique_var('write_csv'), rhs.loc)