-
Notifications
You must be signed in to change notification settings - Fork 240
/
group.py
2172 lines (1829 loc) · 91.7 KB
/
group.py
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
"""Define the Group class."""
from __future__ import division
import os
from collections import Iterable, Counter, OrderedDict, defaultdict
from itertools import product, chain
from numbers import Number
import inspect
from fnmatch import fnmatchcase
import copy
from six import iteritems, string_types, itervalues
from six.moves import range
import numpy as np
import networkx as nx
from openmdao.jacobians.dictionary_jacobian import DictionaryJacobian
from openmdao.approximation_schemes.complex_step import ComplexStep
from openmdao.approximation_schemes.finite_difference import FiniteDifference
from openmdao.core.system import System, INT_DTYPE, get_relevant_vars
from openmdao.core.component import Component, _DictValues, _full_slice
from openmdao.proc_allocators.default_allocator import DefaultAllocator, ProcAllocationError
from openmdao.jacobians.jacobian import SUBJAC_META_DEFAULTS
from openmdao.recorders.recording_iteration_stack import Recording
from openmdao.solvers.nonlinear.nonlinear_runonce import NonlinearRunOnce
from openmdao.solvers.linear.linear_runonce import LinearRunOnce
from openmdao.utils.array_utils import convert_neg, array_connection_compatible, \
_flatten_src_indices
from openmdao.utils.general_utils import warn_deprecation, ContainsAll, all_ancestors, \
simple_warning
from openmdao.utils.units import is_compatible, get_conversion
from openmdao.utils.mpi import MPI
from openmdao.utils.coloring import Coloring, _STD_COLORING_FNAME, _DYN_COLORING
import openmdao.utils.coloring as coloring_mod
# regex to check for valid names.
import re
namecheck_rgx = re.compile('[a-zA-Z][_a-zA-Z0-9]*')
class Group(System):
"""
Class used to group systems together; instantiate or inherit.
Attributes
----------
_mpi_proc_allocator : ProcAllocator
Object used to allocate MPI processes to subsystems.
_proc_info : dict of subsys_name: (min_procs, max_procs, weight)
Information used to determine MPI process allocation to subsystems.
_local_system_set : set or None
Set of pathnames of all fully local (not remote or distributed)
direct or indirect subsystems.
_subgroups_myproc : list
List of local subgroups.
_manual_connections : dict
Dictionary of input_name: (output_name, src_indices) connections.
_static_manual_connections : dict
Dictionary that stores all explicit connections added outside of setup.
_conn_abs_in2out : {'abs_in': 'abs_out'}
Dictionary containing all explicit & implicit connections owned
by this system only. The data is the same across all processors.
_conn_discrete_in2out : {'abs_in': 'abs_out'}
Dictionary containing all explicit & implicit discrete var connections owned
by this system only. The data is the same across all processors.
_transfers : dict of dict of Transfers
First key is the vec_name, second key is (mode, isub) where
mode is 'fwd' or 'rev' and isub is the subsystem index among allprocs subsystems
or isub can be None for the full, simultaneous transfer.
_discrete_transfers : dict of discrete transfer metadata
Key is system pathname or None for the full, simultaneous transfer.
_loc_subsys_map : dict
Mapping of local subsystem names to their corresponding System.
_approx_subjac_keys : list
List of subjacobian keys used for approximated derivatives.
"""
def __init__(self, **kwargs):
"""
Set the solvers to nonlinear and linear block Gauss--Seidel by default.
Parameters
----------
**kwargs : dict
dict of arguments available here and in all descendants of this
Group.
"""
self._mpi_proc_allocator = DefaultAllocator()
self._proc_info = {}
super(Group, self).__init__(**kwargs)
self._local_system_set = None
self._subgroups_myproc = None
self._manual_connections = {}
self._static_manual_connections = {}
self._conn_abs_in2out = {}
self._conn_discrete_in2out = {}
self._transfers = {}
self._discrete_transfers = {}
self._approx_subjac_keys = None
# TODO: we cannot set the solvers with property setters at the moment
# because our lint check thinks that we are defining new attributes
# called nonlinear_solver and linear_solver without documenting them.
if not self._nonlinear_solver:
self._nonlinear_solver = NonlinearRunOnce()
if not self._linear_solver:
self._linear_solver = LinearRunOnce()
def setup(self):
"""
Build this group.
This method should be overidden by your Group's method. The reason for using this
method to add subsystem is to save memory and setup time when using your Group
while running under MPI. This avoids the creation of systems that will not be
used in the current process.
You may call 'add_subsystem' to add systems to this group. You may also issue connections,
and set the linear and nonlinear solvers for this group level. You cannot safely change
anything on children systems; use the 'configure' method instead.
Available attributes:
name
pathname
comm
options
"""
pass
def configure(self):
"""
Configure this group to assign children settings.
This method may optionally be overidden by your Group's method.
You may only use this method to change settings on your children subsystems. This includes
setting solvers in cases where you want to override the defaults.
You can assume that the full hierarchy below your level has been instantiated and has
already called its own configure methods.
Available attributes:
name
pathname
comm
options
system hieararchy with attribute access
"""
pass
def _get_scope(self, excl_sub=None):
"""
Find the input and output variables that are needed for a particular matvec product.
Parameters
----------
excl_sub : <System>
A subsystem whose variables should be excluded from the matvec product.
Returns
-------
(set, set)
Sets of output and input variables.
"""
try:
return self._scope_cache[excl_sub]
except KeyError:
pass
if excl_sub is None:
# All myproc outputs
scope_out = frozenset(self._var_abs_names['output'])
# All myproc inputs connected to an output in this system
scope_in = frozenset(self._conn_global_abs_in2out).intersection(
self._var_abs_names['input'])
else:
# All myproc outputs not in excl_sub
scope_out = frozenset(self._var_abs_names['output']).difference(
excl_sub._var_abs_names['output'])
# All myproc inputs connected to an output in this system but not in excl_sub
scope_in = set()
for abs_in in self._var_abs_names['input']:
if abs_in in self._conn_global_abs_in2out:
abs_out = self._conn_global_abs_in2out[abs_in]
if abs_out not in excl_sub._var_allprocs_abs2idx['linear']:
scope_in.add(abs_in)
scope_in = frozenset(scope_in)
self._scope_cache[excl_sub] = (scope_out, scope_in)
return scope_out, scope_in
def _compute_root_scale_factors(self):
"""
Compute scale factors for all variables.
Returns
-------
dict
Mapping of each absolute var name to its corresponding scaling factor tuple.
"""
scale_factors = super(Group, self)._compute_root_scale_factors()
if self._has_input_scaling:
abs2meta_in = self._var_abs2meta
allprocs_meta_out = self._var_allprocs_abs2meta
for abs_in, abs_out in iteritems(self._conn_global_abs_in2out):
if abs_in not in abs2meta_in:
# we only perform scaling on local, non-discrete arrays, so skip
continue
meta_in = abs2meta_in[abs_in]
meta_out = allprocs_meta_out[abs_out]
ref = meta_out['ref']
ref0 = meta_out['ref0']
src_indices = meta_in['src_indices']
if src_indices is not None:
if not (np.isscalar(ref) and np.isscalar(ref0)):
# TODO: if either ref or ref0 are not scalar and the output is
# distributed, we need to do a scatter
# to obtain the values needed due to global src_indices
if meta_out['distributed']:
raise RuntimeError("{}: vector scalers with distrib vars "
"not supported yet.".format(self.msginfo))
if src_indices.ndim != 1:
src_indices = _flatten_src_indices(src_indices, meta_in['shape'],
meta_out['global_shape'],
meta_out['global_size'])
ref = ref[src_indices]
ref0 = ref0[src_indices]
# Compute scaling arrays for inputs using a0 and a1
# Example:
# Let x, x_src, x_tgt be the dimensionless variable,
# variable in source units, and variable in target units, resp.
# x_src = a0 + a1 x
# x_tgt = b0 + b1 x
# x_tgt = g(x_src) = d0 + d1 x_src
# b0 + b1 x = d0 + d1 a0 + d1 a1 x
# b0 = d0 + d1 a0
# b0 = g(a0)
# b1 = d0 + d1 a1 - d0
# b1 = g(a1) - g(0)
units_in = meta_in['units']
units_out = meta_out['units']
if units_in is None or units_out is None or units_in == units_out:
a0 = ref0
a1 = ref - ref0
else:
factor, offset = get_conversion(units_out, units_in)
a0 = (ref0 + offset) * factor
a1 = (ref - ref0) * factor
scale_factors[abs_in] = {
('input', 'phys'): (a0, a1),
('input', 'norm'): (-a0 / a1, 1.0 / a1)
}
return scale_factors
def _configure(self):
"""
Configure our model recursively to assign any children settings.
Highest system's settings take precedence.
"""
for subsys in self._subsystems_myproc:
subsys._configure()
if subsys._has_guess:
self._has_guess = True
if subsys._has_bounds:
self._has_bounds = True
if subsys.matrix_free:
self.matrix_free = True
self._static_mode = False
try:
self.configure()
finally:
self._static_mode = True
def _setup_procs(self, pathname, comm, mode, prob_options):
"""
Execute first phase of the setup process.
Distribute processors, assign pathnames, and call setup on the group. This method recurses
downward through the model.
Parameters
----------
pathname : str
Global name of the system, including the path.
comm : MPI.Comm or <FakeComm>
MPI communicator object.
mode : string
Derivatives calculation mode, 'fwd' for forward, and 'rev' for
reverse (adjoint). Default is 'rev'.
prob_options : OptionsDictionary
Problem level options.
"""
self.pathname = pathname
self._problem_options = prob_options
self.options._parent_name = self.msginfo
self.recording_options._parent_name = self.msginfo
if self._num_par_fd > 1:
info = self._coloring_info
if comm.size > 1:
# if approx_totals has been declared, or there is an approx coloring, setup par FD
if self._owns_approx_jac or info['coloring'] is not None:
comm = self._setup_par_fd_procs(comm)
else:
msg = "%s: num_par_fd = %d but FD is not active." % (self.msginfo,
self._num_par_fd)
raise RuntimeError(msg)
elif not MPI:
msg = ("%s: MPI is not active but num_par_fd = %d. No parallel finite difference "
"will be performed." % (self.msginfo, self._num_par_fd))
simple_warning(msg)
self.comm = comm
self._mode = mode
self._subsystems_allprocs = []
self._manual_connections = {}
self._design_vars = OrderedDict()
self._responses = OrderedDict()
self._first_call_to_linearize = True
self._approx_subjac_keys = None
self._static_mode = False
self._subsystems_allprocs.extend(self._static_subsystems_allprocs)
self._manual_connections.update(self._static_manual_connections)
self._design_vars.update(self._static_design_vars)
self._responses.update(self._static_responses)
# Call setup function for this group.
self.setup()
self._static_mode = True
if MPI:
proc_info = [self._proc_info[s.name] for s in self._subsystems_allprocs]
# Call the load balancing algorithm
try:
sub_inds, sub_comm, sub_proc_range = self._mpi_proc_allocator(
proc_info, len(self._subsystems_allprocs), comm)
except ProcAllocationError as err:
subs = self._subsystems_allprocs
if err.sub_inds is None:
raise RuntimeError("%s: %s" % (self.msginfo, err.msg))
else:
raise RuntimeError("%s: MPI process allocation failed: %s for the following "
"subsystems: %s" % (self.msginfo, err.msg,
[subs[i].name for i in err.sub_inds]))
self._subsystems_myproc = [self._subsystems_allprocs[ind] for ind in sub_inds]
# Define local subsystems
if np.sum([minp for minp, _, _ in proc_info]) <= comm.size:
self._subsystems_myproc_inds = sub_inds
else:
# reorder the subsystems_allprocs based on which procs they live on. If we don't
# do this, we can get ordering mismatches in some of our data structures.
new_allsubs = []
seen = set()
gathered = self.comm.allgather(sub_inds)
for rank, inds in enumerate(gathered):
for ind in inds:
if ind not in seen:
new_allsubs.append(self._subsystems_allprocs[ind])
seen.add(ind)
self._subsystems_allprocs = new_allsubs
sub_idxs = {s.name: i for i, s in enumerate(self._subsystems_allprocs)}
# since the subsystems_allprocs order changed, we also have to update
# subsystems_myproc_inds
self._subsystems_myproc_inds = [sub_idxs[s.name] for s in self._subsystems_myproc]
else:
sub_comm = comm
self._subsystems_myproc = self._subsystems_allprocs
self._subsystems_myproc_inds = list(range(len(self._subsystems_myproc)))
sub_proc_range = (0, 1)
# Compute _subsystems_proc_range
self._subsystems_proc_range = [sub_proc_range] * len(self._subsystems_myproc)
self._local_system_set = set()
# Perform recursion
for subsys in self._subsystems_myproc:
subsys._local_vector_class = self._local_vector_class
subsys._distributed_vector_class = self._distributed_vector_class
subsys.force_alloc_complex = self.force_alloc_complex
subsys._use_derivatives = self._use_derivatives
subsys._solver_info = self._solver_info
subsys._recording_iter = self._recording_iter
if self.pathname:
subsys._setup_procs('.'.join((self.pathname, subsys.name)), sub_comm, mode,
prob_options)
else:
subsys._setup_procs(subsys.name, sub_comm, mode, prob_options)
# build a list of local subgroups to speed up later loops
self._subgroups_myproc = [s for s in self._subsystems_myproc if isinstance(s, Group)]
self._loc_subsys_map = {s.name: s for s in self._subsystems_myproc}
def _check_child_reconf(self, subsys=None):
"""
Check if any subsystem has reconfigured and if so, perform the necessary update setup.
Parameters
----------
subsys : System or None
If not None, check only if the given subsystem has reconfigured.
"""
if subsys is None:
# See if any local subsystem has reconfigured
for subsys in self._subgroups_myproc:
if subsys._reconfigured:
reconf = 1
break
else:
reconf = 0
else:
reconf = int(subsys._reconfigured) if subsys.name in self._loc_subsys_map else 0
# See if any subsystem on this or any other processor has configured
if self.comm.size > 1:
reconf = self.comm.allreduce(reconf) > 0
if reconf:
# Perform an update setup
with self._unscaled_context_all():
self.resetup('update')
# Reset the _reconfigured attribute to False
for subsys in self._subsystems_myproc:
subsys._reconfigured = False
self._reconfigured = True
def _list_states(self):
"""
Return list of all local states at and below this system.
Returns
-------
list
List of all states.
"""
states = []
for subsys in self._subsystems_myproc:
states.extend(subsys._list_states())
return sorted(states)
def _list_states_allprocs(self):
"""
Return list of all states at and below this system across all procs.
Returns
-------
list
List of all states.
"""
if MPI:
all_states = set()
byproc = self.comm.allgather(self._list_states())
for proc_states in byproc:
all_states.update(proc_states)
return sorted(all_states)
else:
return self._list_states()
def _setup_var_index_ranges(self, recurse=True):
"""
Compute the division of variables by subsystem.
Parameters
----------
recurse : bool
Whether to call this method in subsystems.
"""
nsub_allprocs = len(self._subsystems_allprocs)
subsystems_var_range = self._subsystems_var_range = {}
vec_names = self._lin_rel_vec_name_list if self._use_derivatives else self._vec_names
# First compute these on one processor for each subsystem
for vec_name in vec_names:
# Here, we count the number of variables in each subsystem.
# We do this so that we can compute the offset when we recurse into each subsystem.
allprocs_counters = {}
for type_ in ['input', 'output']:
allprocs_counters[type_] = np.zeros(nsub_allprocs, INT_DTYPE)
for subsys, isub in zip(self._subsystems_myproc, self._subsystems_myproc_inds):
comm = subsys.comm if subsys._full_comm is None else subsys._full_comm
if comm.rank == 0 and vec_name in subsys._rel_vec_names:
allprocs_counters[type_][isub] = \
len(subsys._var_allprocs_relevant_names[vec_name][type_])
# If running in parallel, allgather
if self.comm.size > 1:
gathered = self.comm.allgather(allprocs_counters)
allprocs_counters = {
type_: np.zeros(nsub_allprocs, INT_DTYPE) for type_ in ['input', 'output']}
for myproc_counters in gathered:
for type_ in ['input', 'output']:
allprocs_counters[type_] += myproc_counters[type_]
# Compute _subsystems_var_range
subsystems_var_range[vec_name] = {}
for type_ in ['input', 'output']:
subsystems_var_range[vec_name][type_] = {}
for subsys, isub in zip(self._subsystems_myproc, self._subsystems_myproc_inds):
if vec_name not in subsys._rel_vec_names:
continue
start = np.sum(allprocs_counters[type_][:isub])
subsystems_var_range[vec_name][type_][subsys.name] = (
start, start + allprocs_counters[type_][isub]
)
if self._use_derivatives:
subsystems_var_range['nonlinear'] = subsystems_var_range['linear']
self._setup_var_index_maps(recurse=recurse)
# Recursion
if recurse:
for subsys in self._subsystems_myproc:
subsys._setup_var_index_ranges(recurse)
def _setup_var_data(self, recurse=True):
"""
Compute the list of abs var names, abs/prom name maps, and metadata dictionaries.
Parameters
----------
recurse : bool
Whether to call this method in subsystems.
"""
super(Group, self)._setup_var_data()
allprocs_abs_names = self._var_allprocs_abs_names
allprocs_discrete = self._var_allprocs_discrete
abs_names = self._var_abs_names
var_discrete = self._var_discrete
allprocs_prom2abs_list = self._var_allprocs_prom2abs_list
abs2prom = self._var_abs2prom
allprocs_abs2meta = self._var_allprocs_abs2meta
allprocs_abs2prom = self._var_allprocs_abs2prom
abs2meta = self._var_abs2meta
for subsys in self._subsystems_myproc:
if recurse:
subsys._setup_var_data(recurse)
self._has_output_scaling |= subsys._has_output_scaling
self._has_resid_scaling |= subsys._has_resid_scaling
var_maps = subsys._get_maps(subsys._var_allprocs_prom2abs_list)
# Assemble allprocs_abs2meta and abs2meta
allprocs_abs2meta.update(subsys._var_allprocs_abs2meta)
abs2meta.update(subsys._var_abs2meta)
sub_prefix = subsys.name + '.'
for type_ in ['input', 'output']:
# Assemble abs_names and allprocs_abs_names
allprocs_abs_names[type_].extend(subsys._var_allprocs_abs_names[type_])
allprocs_discrete[type_].update({k: v for k, v in
iteritems(subsys._var_allprocs_discrete[type_])})
abs_names[type_].extend(subsys._var_abs_names[type_])
var_discrete[type_].update({sub_prefix + k: v for k, v in
iteritems(subsys._var_discrete[type_])})
# Assemble abs2prom
for abs_name in subsys._var_abs_names[type_]:
sub_prom_name = subsys._var_abs2prom[type_][abs_name]
abs2prom[type_][abs_name] = var_maps[type_][sub_prom_name]
allprocs_abs2prom[type_] = abs2prom[type_]
# Assemble allprocs_prom2abs_list
for sub_prom, sub_abs in iteritems(subsys._var_allprocs_prom2abs_list[type_]):
prom_name = var_maps[type_][sub_prom]
if prom_name not in allprocs_prom2abs_list[type_]:
allprocs_prom2abs_list[type_][prom_name] = []
allprocs_prom2abs_list[type_][prom_name].extend(sub_abs)
for prom_name, abs_list in iteritems(allprocs_prom2abs_list['output']):
if len(abs_list) > 1:
raise RuntimeError("{}: Output name '{}' refers to "
"multiple outputs: {}.".format(self.msginfo, prom_name,
sorted(abs_list)))
# If running in parallel, allgather
if self.comm.size > 1:
mysub = self._subsystems_myproc[0] if self._subsystems_myproc else False
if (mysub and mysub.comm.rank == 0 and (mysub._full_comm is None or
mysub._full_comm.rank == 0)):
raw = (allprocs_abs_names, allprocs_discrete, allprocs_prom2abs_list, abs2prom,
allprocs_abs2meta, self._has_output_scaling, self._has_resid_scaling)
else:
raw = (
{'input': [], 'output': []},
{'input': {}, 'output': {}},
{'input': {}, 'output': {}},
{'input': {}, 'output': {}},
{},
False,
False
)
gathered = self.comm.allgather(raw)
for type_ in ['input', 'output']:
allprocs_abs_names[type_] = []
allprocs_abs2prom[type_] = {}
allprocs_prom2abs_list[type_] = OrderedDict()
for (myproc_abs_names, myproc_discrete, myproc_prom2abs_list, myproc_abs2prom,
myproc_abs2meta, oscale, rscale) in gathered:
self._has_output_scaling |= oscale
self._has_resid_scaling |= rscale
# Assemble in parallel allprocs_abs2meta
allprocs_abs2meta.update(myproc_abs2meta)
for type_ in ['input', 'output']:
# Assemble in parallel allprocs_abs_names
allprocs_abs_names[type_].extend(myproc_abs_names[type_])
allprocs_discrete[type_].update(myproc_discrete[type_])
allprocs_abs2prom[type_].update(myproc_abs2prom[type_])
# Assemble in parallel allprocs_prom2abs_list
for prom_name, abs_names_list in iteritems(myproc_prom2abs_list[type_]):
if prom_name not in allprocs_prom2abs_list[type_]:
allprocs_prom2abs_list[type_][prom_name] = []
allprocs_prom2abs_list[type_][prom_name].extend(abs_names_list)
if self._var_discrete['input'] or self._var_discrete['output']:
self._discrete_inputs = _DictValues(self._var_discrete['input'])
self._discrete_outputs = _DictValues(self._var_discrete['output'])
else:
self._discrete_inputs = self._discrete_outputs = ()
def _setup_var_sizes(self, recurse=True):
"""
Compute the arrays of local variable sizes for all variables/procs on this system.
Parameters
----------
recurse : bool
Whether to call this method in subsystems.
"""
super(Group, self)._setup_var_sizes()
self._var_offsets = None
iproc = self.comm.rank
nproc = self.comm.size
subsystems_proc_range = self._subsystems_proc_range
# Recursion
if recurse:
for subsys in self._subsystems_myproc:
subsys._setup_var_sizes(recurse)
sizes = self._var_sizes
relnames = self._var_allprocs_relevant_names
vec_names = self._lin_rel_vec_name_list if self._use_derivatives else self._vec_names
# Compute _var_sizes
for vec_name in vec_names:
sizes[vec_name] = {}
subsystems_var_range = self._subsystems_var_range[vec_name]
for type_ in ['input', 'output']:
sizes[vec_name][type_] = sz = np.zeros((nproc, len(relnames[vec_name][type_])),
INT_DTYPE)
for ind, subsys in enumerate(self._subsystems_myproc):
if vec_name not in subsys._rel_vec_names:
continue
proc_slice = slice(*subsystems_proc_range[ind])
var_slice = slice(*subsystems_var_range[type_][subsys.name])
if proc_slice.stop - proc_slice.start > subsys.comm.size:
# in this case, we've split the proc for parallel FD, so subsys doesn't
# have var_sizes for all the ranks we need. Since each parallel FD comm
# has the same size distribution (since all are identical), just 'tile'
# the var_sizes from the subsystem to fill in the full rank range we need
# at this level.
assert (proc_slice.stop - proc_slice.start) % subsys.comm.size == 0, \
"%s comm size (%d) is not an exact multiple of %s comm size (%d)" % (
self.pathname, self.comm.size, subsys.pathname, subsys.comm.size)
proc_i = proc_slice.start
while proc_i < proc_slice.stop:
sz[proc_i:proc_i + subsys.comm.size, var_slice] = \
subsys._var_sizes[vec_name][type_]
proc_i += subsys.comm.size
else:
sz[proc_slice, var_slice] = subsys._var_sizes[vec_name][type_]
# If parallel, all gather
if self.comm.size > 1:
for vec_name in self._lin_rel_vec_name_list:
sizes = self._var_sizes[vec_name]
for type_ in ['input', 'output']:
sizes_in = sizes[type_][iproc, :].copy()
self.comm.Allgather(sizes_in, sizes[type_])
# compute owning ranks
owns = self._owning_rank
for type_ in ('input', 'output'):
sizes = self._var_sizes[vec_names[0]][type_]
for i, name in enumerate(self._var_allprocs_abs_names[type_]):
for rank in range(self.comm.size):
if sizes[rank, i] > 0:
owns[name] = rank
break
if self._var_allprocs_discrete[type_]:
local = list(self._var_discrete[type_])
for i, names in enumerate(self.comm.allgather(local)):
for n in names:
if n not in owns:
owns[n] = i
if self._use_derivatives:
self._var_sizes['nonlinear'] = self._var_sizes['linear']
self._setup_global_shapes()
def _setup_global_connections(self, recurse=True, conns=None):
"""
Compute dict of all connections between this system's inputs and outputs.
The connections come from 4 sources:
1. Implicit connections owned by the current system
2. Explicit connections declared by the current system
3. Explicit connections declared by parent systems
4. Implicit / explicit from subsystems
Parameters
----------
recurse : bool
Whether to call this method in subsystems.
conns : dict
Dictionary of connections passed down from parent group.
"""
global_abs_in2out = self._conn_global_abs_in2out
allprocs_prom2abs_list_in = self._var_allprocs_prom2abs_list['input']
allprocs_prom2abs_list_out = self._var_allprocs_prom2abs_list['output']
abs2meta = self._var_abs2meta
pathname = self.pathname
abs_in2out = {}
if pathname == '':
path_len = 0
nparts = 0
else:
path_len = len(pathname) + 1
nparts = len(pathname.split('.'))
new_conns = defaultdict(dict)
if conns is not None:
for abs_in, abs_out in iteritems(conns):
inparts = abs_in.split('.')
outparts = abs_out.split('.')
if inparts[:nparts] == outparts[:nparts]:
global_abs_in2out[abs_in] = abs_out
# if connection is contained in a subgroup, add to conns
# to pass down to subsystems.
if inparts[:nparts + 1] == outparts[:nparts + 1]:
new_conns[inparts[nparts]][abs_in] = abs_out
# Add implicit connections (only ones owned by this group)
for prom_name in allprocs_prom2abs_list_out:
if prom_name in allprocs_prom2abs_list_in:
abs_out = allprocs_prom2abs_list_out[prom_name][0]
out_subsys = abs_out[path_len:].split('.', 1)[0]
for abs_in in allprocs_prom2abs_list_in[prom_name]:
in_subsys = abs_in[path_len:].split('.', 1)[0]
if out_subsys != in_subsys:
abs_in2out[abs_in] = abs_out
# Add explicit connections (only ones declared by this group)
for prom_in, (prom_out, src_indices, flat_src_indices) in \
iteritems(self._manual_connections):
# throw an exception if either output or input doesn't exist
# (not traceable to a connect statement, so provide context)
if (prom_out not in allprocs_prom2abs_list_out and
prom_out not in self._var_allprocs_discrete['output']):
raise NameError(
"%s: Output '%s' does not exist for connection in '%s' from '%s' to '%s'." %
(self.msginfo, prom_out, self.pathname, prom_out, prom_in))
if (prom_in not in allprocs_prom2abs_list_in and
prom_in not in self._var_allprocs_discrete['input']):
raise NameError(
"%s: Input '%s' does not exist for connection from '%s' to '%s'." %
(self.msginfo, prom_in, prom_out, prom_in))
# Throw an exception if output and input are in the same system
# (not traceable to a connect statement, so provide context)
# and check if src_indices is defined in both connect and add_input.
abs_out = allprocs_prom2abs_list_out[prom_out][0]
outparts = abs_out.split('.')
out_subsys = outparts[:-1]
for abs_in in allprocs_prom2abs_list_in[prom_in]:
inparts = abs_in.split('.')
in_subsys = inparts[:-1]
if out_subsys == in_subsys:
raise RuntimeError("{}: Output and input are in the same System "
"for connection from '{}' to '{}'.".format(self.msginfo,
prom_out,
prom_in))
if src_indices is not None and abs_in in abs2meta:
meta = abs2meta[abs_in]
if meta['src_indices'] is not None:
raise RuntimeError("{}: src_indices has been defined "
"in both connect('{}', '{}') "
"and add_input('{}', ...).".format(self.msginfo,
prom_out, prom_in,
prom_in))
meta['src_indices'] = np.atleast_1d(src_indices)
meta['flat_src_indices'] = flat_src_indices
if abs_in in abs_in2out:
raise RuntimeError("%s: Input '%s' cannot be connected to '%s' because it's "
"already connected to '%s'" % (self.msginfo, abs_in,
abs_out, abs_in2out[abs_in]))
abs_in2out[abs_in] = abs_out
# if connection is contained in a subgroup, add to conns to pass down to subsystems.
if inparts[:nparts + 1] == outparts[:nparts + 1]:
new_conns[inparts[nparts]][abs_in] = abs_out
# Recursion
if recurse:
for subsys in self._subgroups_myproc:
if subsys.name in new_conns:
subsys._setup_global_connections(recurse=recurse,
conns=new_conns[subsys.name])
else:
subsys._setup_global_connections(recurse=recurse)
# Compute global_abs_in2out by first adding this group's contributions,
# then adding contributions from systems above/below, then allgathering.
conn_list = list(iteritems(global_abs_in2out))
conn_list.extend(iteritems(abs_in2out))
global_abs_in2out.update(abs_in2out)
for subsys in self._subgroups_myproc:
global_abs_in2out.update(subsys._conn_global_abs_in2out)
conn_list.extend(iteritems(subsys._conn_global_abs_in2out))
if len(conn_list) > len(global_abs_in2out):
dupes = [n for n, val in iteritems(Counter(tgt for tgt, src in conn_list)) if val > 1]
dup_info = defaultdict(set)
for tgt, src in conn_list:
for dup in dupes:
if tgt == dup:
dup_info[tgt].add(src)
dup_info = [(n, srcs) for n, srcs in iteritems(dup_info) if len(srcs) > 1]
if dup_info:
msg = ["%s from %s" % (tgt, sorted(srcs)) for tgt, srcs in dup_info]
raise RuntimeError("%s: The following inputs have multiple connections: %s" %
(self.msginfo, ", ".join(msg)))
# If running in parallel, allgather
if self.comm.size > 1:
if self._subsystems_myproc and self._subsystems_myproc[0].comm.rank == 0:
raw = global_abs_in2out
else:
raw = {}
gathered = self.comm.allgather(raw)
for myproc_global_abs_in2out in gathered:
global_abs_in2out.update(myproc_global_abs_in2out)
def _setup_connections(self, recurse=True):
"""
Compute dict of all implicit and explicit connections owned by this Group.
Parameters
----------
recurse : bool
Whether to call this method in subsystems.
"""
abs_in2out = self._conn_abs_in2out = {}
global_abs_in2out = self._conn_global_abs_in2out
pathname = self.pathname
allprocs_discrete_in = self._var_allprocs_discrete['input']
allprocs_discrete_out = self._var_allprocs_discrete['output']
# Recursion
if recurse:
for subsys in self._subsystems_myproc:
subsys._setup_connections(recurse)
if MPI:
# collect set of local (not remote, not distributed) subsystems so we can
# identify cross-process connections, which require the use of distributed
# instead of purely local vector and transfer objects.
self._local_system_set = set()
for s in self._subsystems_myproc:
if isinstance(s, Group):
self._local_system_set.update(s._local_system_set)
elif not s.options['distributed']:
self._local_system_set.add(s.pathname)
path_dot = pathname + '.' if pathname else ''
path_len = len(path_dot)
allprocs_abs2meta = self._var_allprocs_abs2meta
self._vector_class = None
# Check input/output units here, and set _has_input_scaling
# to True for this Group if units are defined and different, or if
# ref or ref0 are defined for the output.
for abs_in, abs_out in iteritems(global_abs_in2out):
# First, check that this system owns both the input and output.
if abs_in[:path_len] == path_dot and abs_out[:path_len] == path_dot:
# Second, check that they are in different subsystems of this system.
out_subsys = abs_out[path_len:].split('.', 1)[0]
in_subsys = abs_in[path_len:].split('.', 1)[0]
if out_subsys != in_subsys:
if abs_in in allprocs_discrete_in:
self._conn_discrete_in2out[abs_in] = abs_out
elif abs_out in allprocs_discrete_out:
raise RuntimeError("%s: Can't connect discrete output '%s' to continuous "
"input '%s'." % (self.msginfo, abs_out, abs_in))
else:
abs_in2out[abs_in] = abs_out
if MPI and self._vector_class is None:
# check for any cross-process data transfer. If found, use
# self._distributed_vector_class as our vector class.
in_path = abs_in.rsplit('.', 1)[0]
if in_path not in self._local_system_set:
self._vector_class = self._distributed_vector_class
else:
out_path = abs_out.rsplit('.', 1)[0]
if out_path not in self._local_system_set:
self._vector_class = self._distributed_vector_class
# if connected output has scaling then we need input scaling
if not self._has_input_scaling and not (abs_in in allprocs_discrete_in or
abs_out in allprocs_discrete_out):
out_units = allprocs_abs2meta[abs_out]['units']
in_units = allprocs_abs2meta[abs_in]['units']
# if units are defined and different, we need input scaling.
needs_input_scaling = (in_units and out_units and in_units != out_units)
# we also need it if a connected output has any scaling.
if not needs_input_scaling:
out_meta = allprocs_abs2meta[abs_out]
ref = out_meta['ref']
if np.isscalar(ref):
needs_input_scaling = ref != 1.0