forked from analogdevicesinc/ai8x-training
-
Notifications
You must be signed in to change notification settings - Fork 0
/
yamlwriter.py
1238 lines (1097 loc) · 48.9 KB
/
yamlwriter.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
###################################################################################################
#
# Copyright (C) 2022-2024 Maxim Integrated Products, Inc. All Rights Reserved.
#
# Maxim Integrated Products, Inc. Default Copyright Notice:
# https://www.maximintegrated.com/en/aboutus/legal/copyrights.html
#
###################################################################################################
"""
Generate YAML template from PyTorch model
ALPHA version
TODO: Implement smart processor and data memory allocator (out_offset, in_offset)
TODO: Implement dilation
TODO: Testing
NOTE: This code partially depends on ai8x.py:
- Quantization information is expected in '.weight_bits'.
"""
import os
import warnings
from typing import Any, Callable, Dict, List, Optional, OrderedDict, Tuple, Union
import distiller
import ai8x
import devices
def allocate_offset(
_layer_name: str,
_processor_map: int,
prev_offset: int = 0,
) -> int:
"""
Find a data memory offset for the given layer. Currently just uses "ping-pong" from 0 to
half of the data memory.
TODO: Implement an algorithm that interacts with weight memories and processor selection,
and understands element-wise data.
"""
assert ai8x.dev is not None
HALF_DATA = 0x4000 if ai8x.dev.device == 85 else 0xa000
return HALF_DATA if prev_offset == 0 else 0
# pylint: disable=too-many-branches, too-many-statements, too-many-locals
def create(
model: Any,
dataset: str,
arch: str,
hwc: bool = False,
use_fifos: bool = False,
move_l0: bool = False,
filename: str = 'template.yaml',
qat_policy: Optional[str] = None,
verbose=True,
) -> None:
"""
Create YAML template
"""
assert ai8x.dev is not None
if ai8x.dev.device == 85:
MAX_PIXELS = 8192
elif ai8x.dev.device == 87:
MAX_PIXELS = 20480
else:
print(f'Unknown device {ai8x.dev.device}')
return
# Filter warnings
s0 = 'The shape inference of prim::Constant type is missing, so it may result in wrong '\
'shape inference for the exported graph. Please consider adding it in symbolic function.'
warnings.filterwarnings(action='ignore', message=s0)
s1 = 'Constant folding - Only steps=1 can be constant folded for opset >= 10 ' \
'onnx::Slice op. Constant folding not applied.'
warnings.filterwarnings(action='ignore', message=s1)
MAX_PROC = 64
model = distiller.make_non_parallel_copy(model) # type: ignore[attr-defined]
# Replace unsupported BitwiseOr and BitwiseXor with an Add() and record the locations.
# After ONNX conversion, put them back. This is necessary until PyTorch ONNX export supports
# BitwiseOr/BitwiseXor.
bitwise_replacements: List[Tuple[str, str]] = []
def replace_bitwise(m):
for attr_str in dir(m):
target_attr = getattr(m, attr_str)
if isinstance(target_attr, (ai8x.BitwiseXor, ai8x.BitwiseOr)):
print(attr_str)
source = 'bitwise_or' if isinstance(target_attr, ai8x.BitwiseOr) else 'bitwise_xor'
bitwise_replacements.append((attr_str, source))
setattr(m, attr_str, ai8x.Add())
model.apply(replace_bitwise)
# Apply the QAT policy early to set weight_bits
ai8x.fuse_bn_layers(model)
if qat_policy is not None:
ai8x.initiate_qat(model, qat_policy, export=True)
ai8x.onnx_export_prep(model, simplify=False, remove_clamp=False)
dummy_input = distiller.get_dummy_input( # type: ignore[attr-defined]
dataset=None,
device=distiller.model_device(model), # type: ignore[attr-defined]
input_shape=None,
)
g = distiller.SummaryGraph(model, dummy_input, True, '#') # type: ignore[attr-defined]
# Get the input/output dimensions
shapes: Dict[str, Tuple] = {}
for i, param in g.params.items():
shape = param['shape']
if len(shape) > 1:
shapes[i] = shape[1:] # Remove batch dimension
# all_ops is the main dictionary for the layers
all_ops: OrderedDict[str, Dict[str, Any]] = g.ops
# Put the bitwise operations back
for (attr_str, bitwise_op) in bitwise_replacements:
assert all_ops[attr_str]['type'] == 'Add'
all_ops[attr_str]['type'] = 'BitwiseOr' if bitwise_op == 'bitwise_or' else 'BitwiseXor'
# ONNX operators for convolution:
convolution_ops: Tuple[str, ...] = ('Conv', 'ConvTranspose', 'Gemm', 'MatMul')
def canonical_name(
s: str,
) -> str:
"""
Return the canonical name for this layer
"""
separator = s.rfind('_MatMul_1')
if separator > 0:
s = s[:separator]
separator = s.rfind('.')
if separator > 0:
if s[separator + 1:] == 'op':
return s[:separator]
return s
def ignore_layer(
_name: str,
layer: Dict[str, Any],
) -> bool:
"""
Remove unnecessary layers from the graph
"""
if not layer['inputs'] or not layer['outputs']:
return True # Not consuming or producing anything, useless layer
return layer['type'] not in convolution_ops and layer['type'] not in (
'Add', 'Sub', 'BitwiseOr', 'BitwiseXor', # element-wise
'MaxPool', 'AveragePool',
'Abs', 'Relu',
)
def follow_input(
trace_in: List[str],
condition: Callable[[str], bool],
replace: Tuple[str, str],
) -> List[str]:
"""
Trace a list of inputs back upstream and stop when condition is met
"""
results: List[str] = []
while len(trace_in) > 0:
next_trace: List[str] = []
for t in trace_in:
if condition(t):
results.append(t.replace(replace[0], replace[1]))
else:
for n in all_ops:
if t in all_ops[n]['outputs']:
next_trace += all_ops[n]['inputs']
trace_in = next_trace
return results
# 1 - Set output_width to 32 where needed
# Trace all "-32768" constants to the next 'Min' operation (there may be more than one)
results: List[str] = []
for n in all_ops:
if all_ops[n]['type'] in convolution_ops:
# Ignore weights/biases in the input list
# These operations have 2 or three arguments: data/weights/biases(optional)
all_ops[n]['weights'] = follow_input([all_ops[n]['inputs'][1]],
lambda s: s in shapes, ('#', '.'))
if len(all_ops[n]['inputs']) == 3:
all_ops[n]['biases'] = follow_input([all_ops[n]['inputs'][2]],
lambda s: s in shapes, ('#', '.'))
# Remove weights and biases from the inputs for this operation
all_ops[n]['inputs'] = [all_ops[n]['inputs'][0]]
if all_ops[n]['type'] != 'Constant' or 'value' not in all_ops[n]['attrs'] \
or all_ops[n]['attrs']['value'].dim() != 0 or all_ops[n]['attrs']['value'] != -32768:
continue
trace_out: List[str] = all_ops[n]['outputs']
while len(trace_out) > 0:
next_trace: List[str] = []
for t in trace_out:
for p in all_ops:
if t in all_ops[p]['inputs']:
if all_ops[p]['type'] in ('Min', 'Clip'):
results.append(p)
else:
next_trace += all_ops[p]['outputs']
trace_out = next_trace
# Trace the inputs of these layers back to the previous convolution
for n in results:
trace_in: List[str] = all_ops[n]['inputs']
while len(trace_in) > 0:
next_trace = []
for t in trace_in:
for p in all_ops:
if t in all_ops[p]['outputs']:
if all_ops[p]['type'] in convolution_ops:
all_ops[p]['wide'] = True
else:
next_trace += all_ops[p]['inputs']
trace_in = next_trace
# 2 - Remove inputs with zero dimensions
remove_list: List[str] = []
for n in all_ops:
s: List[str] = all_ops[n]['inputs']
all_ops[n]['original_inputs'] = s
remove: bool = False
new: List[str] = []
for i in s:
if i in shapes:
new.append(i)
else:
remove = True
remove_list.append(i)
if remove:
all_ops[n]['inputs'] = new
s = all_ops[n]['outputs']
remove = False
new = []
for i in s:
if i in shapes:
new.append(i)
else:
remove = True
remove_list.append(i)
if remove:
all_ops[n]['outputs'] = new
# 3 - Remove layers
ignore_layers: List[str] = []
for n in all_ops:
if ignore_layer(n, all_ops[n]):
ignore_layers.append(n)
def follow_chain(
layer: str,
remove: str,
replacements: List[str],
which: str,
) -> None:
for n in all_ops:
if layer == n:
continue
for an_output in remove:
new_inputs: List[str] = []
for an_input in all_ops[n][which]:
if an_output == an_input:
new_inputs += replacements
else:
new_inputs.append(an_input)
filtered: List[str] = []
for ie in new_inputs:
if ie not in filtered:
filtered.append(ie)
if filtered != all_ops[n][which]:
all_ops[n][which] = filtered
# Fix up inputs after removing a layer
for layer in ignore_layers:
# Look for layer's outputs in the inputs of all other layers and replace
follow_chain(
layer,
remove=all_ops[layer]['outputs'],
replacements=all_ops[layer]['inputs'],
which='inputs',
)
# 4 - Associate inputs and outputs
input_layer: Dict[str, str] = {}
output_layer: Dict[str, str] = {}
final_layer: str = ''
for name in all_ops:
for ie in all_ops[name]['inputs']:
if ie != '':
input_layer[ie] = name
for ie in all_ops[name]['outputs']:
if ie != '':
output_layer[ie] = name
final_layer = name
# 5 - Collect inputs and outputs
inputs: Dict[str, List[str]] = {}
outputs: Dict[str, List[str]] = {}
for name in all_ops:
if ignore_layer(name, all_ops[name]):
continue
# Get input names from first of the layer group
if name not in inputs:
inputs[name] = []
for ie in all_ops[name]['inputs']:
if ie != '' and (ie not in output_layer or output_layer[ie] != name):
inputs[name].append(ie)
if name not in outputs:
outputs[name] = []
for ie in all_ops[name]['outputs']:
if ie != '' and (ie not in input_layer or input_layer[ie] != name):
outputs[name].append(ie)
# 6 - Collect ops
prev_op_name: str = ''
layers: Dict[str, Dict[str, Any]] = {}
input_hwc: bool = False
input_processors: int = 0
for name in all_ops:
if ignore_layer(name, all_ops[name]):
continue
this_layer: Dict[str, Any] = {}
this_layer['name'] = name
ins = inputs[name]
# Mark output layers (the final layer is always an output layer)
if name != final_layer and any(x not in input_layer for x in outputs[name]):
this_layer['output'] = 'true'
this_layer['op'] = main_op = 'Passthrough'
operands: int = 1
op = all_ops[name]['type']
if op in ('Add', 'Sub', 'BitwiseOr', 'BitwiseXor'):
operands = len(ins)
if op.startswith('Bitwise'):
op = op[7:]
this_layer['eltwise'] = op
this_layer['operands'] = operands
elif op in ('Gemm', 'MatMul'):
this_layer['op'] = main_op = 'Linear'
this_layer['activate'] = 'None'
elif op in ('Conv', 'ConvTranspose'):
kernel_size = all_ops[name]['attrs']['kernel_shape']
this_layer['op'] = f'{op}{len(kernel_size)}d'
this_layer['activate'] = 'None'
main_op = op
elif op in ('MaxPool', 'AveragePool'):
shape = all_ops[name]['attrs']['kernel_shape']
if len(shape) == 1 or shape[0] == shape[1]:
shape = shape[0]
if all_ops[name]['type'] == 'MaxPool':
this_layer['max_pool'] = shape
else:
this_layer['avg_pool'] = shape
this_layer['pool_stride'] = all_ops[name]['attrs']['strides'][0]
if 'dilations' in all_ops[name]['attrs']:
this_layer['pool_dilation'] = all_ops[name]['attrs']['dilations']
elif op in ('Abs', 'Relu'):
this_layer['activate'] = op
else:
this_layer['op'] = f'Unknown ({op})'
this_layer['main_op'] = main_op
if op in ('Conv', 'ConvTranspose'):
if len(kernel_size) == 1:
this_layer['kernel_size'] = str(kernel_size[0])
else:
this_layer['kernel_size'] = f'{kernel_size[0]}x{kernel_size[1]}'
pad = all_ops[name]['attrs']['pads']
this_layer['pad'] = pad[0]
groups = all_ops[name]['attrs']['group']
if groups != 1:
this_layer['groups'] = groups
# Quantization uses hard-coded name from ai8x.py
quantization = 8
try:
quantization = int(model.get_parameter(name + '.weight_bits'))
except AttributeError:
try:
quantization = int(model.get_parameter(canonical_name(name) + '.weight_bits'))
except AttributeError:
pass
if quantization not in (0, 8):
assert quantization in (1, 2, 4), f'ERROR: {name}: quantization={quantization}'
this_layer['quantization'] = quantization
# Check for biases and 32-bit output using the data recorded earlier
if op in convolution_ops:
if 'wide' in all_ops[name] and all_ops[name]['wide']:
this_layer['output_width'] = 32
this_layer['have_bias'] = 'biases' in all_ops[name]
this_layer['weight_count'] = int(all_ops[name]['attrs']['weights_vol'])
# Check whether inputs need to be flattened (when they are not in 1x1 dimensions)
flatten: bool = False
if op in ('Gemm', 'MatMul'):
for ie in inputs[name]:
if ie in shapes:
mult: int = 1
for x in shapes[ie]:
mult *= x
if shapes[ie][0] != mult:
flatten = True
if flatten:
this_layer['flatten'] = 'true'
in_dim: Optional[Union[Tuple[int, ...], List[int]]] = None
prev_dim: Optional[Union[Tuple[int, ...], List[int]]] = None
# Check whether dimensions need to change
if not flatten:
mult = 1
for n in inputs[name]:
if n in shapes:
for x in shapes[n]:
mult *= x
if len(shapes[n]) > 1 and mult != shapes[n][0]:
in_dim = shapes[n][1:]
else:
in_dim = shapes[n]
break
mult = 1
for n in all_ops[name]['original_inputs']:
if n in shapes:
for x in shapes[n]:
mult *= x
if len(shapes[n]) > 1 and mult != shapes[n][0]:
prev_dim = shapes[n][1:]
else:
prev_dim = shapes[n]
break
if in_dim is not None and prev_dim is not None:
if in_dim != prev_dim \
and (in_dim[0] != prev_dim[0] or len(prev_dim) > 1 or mult != prev_dim[0]):
if len(prev_dim) > 0:
prev_dim = list(prev_dim)
this_layer['in_dim'] = prev_dim
# Find number of processors needed
processors: int = 0
for ie in inputs[name]:
if ie in shapes:
processors += shapes[ie][0]
if operands > 1: # Element-wise: data is always interleaved
processors //= operands
this_layer['proc_count'] = max(1, processors)
# Inner layers and more than 16 channels are always HWC
hwc = hwc or (processors > 16) or (prev_op_name != '')
if prev_op_name == '':
# Show input dimensions and data format for input layers
this_layer['data_format'] = hwc
input_hwc = hwc
input_processors = this_layer['proc_count']
else:
# Don't set in_sequences when using strictly sequential single inputs
ins = [output_layer[x] for x in ins]
if len(ins) > 1 or (len(ins) > 0 and ins[0] != prev_op_name):
if operands > 1:
ins.reverse() # Reverse the list since PyTorch does it backwards
this_layer['in_sequences'] = ins
prev_op_name = name
layers[name] = this_layer
del input_layer
del output_layer
prev_name: str = ''
pop_list: List[Tuple[str, str]] = []
veto: bool = False
prev: Dict[str, Any] = {}
# 7a - Merge (fuse) conv and activation layers
for count, (name, ll) in enumerate(layers.items()):
if ll['main_op'] == 'Passthrough' and 'activate' in ll and prev_name != '':
prev = layers[prev_name]
if prev['main_op'] not in ('Conv', 'ConvTranspose', 'Linear'):
print(f'ERROR: Activation layer {name} does not follow '
f'Conv/Linear layer {prev_name}!')
prev_name = name
continue
# Check that no layer other than the activation layer uses the intermediate output of
# the conv layer as an input
veto = False
for (other_name, ol) in layers.items():
if other_name != name and other_name != prev_name and 'in_sequences' in ol:
if prev_name in ol['in_sequences']:
veto = True
break
if veto:
print(f'ERROR: Activation layer {name} has inputs that are not from a directly '
f'preceding Conv/Linear layer {prev_name}!')
prev_name = name
continue
# Combine both layers
if 'comment' not in prev:
prev['comment'] = f'{prev_name} fused with {name}'
else:
prev['comment'] += f' and {name}'
pop_list.append((prev_name, name)) # Mark second layer for deletion
# Copy over convolution operation and keep the element-wise operation in place
outputs[prev_name] = outputs[name]
prev['activate'] = ll['activate']
if 'output' in ll:
prev['output'] = ll['output']
if 'quantization' in ll:
prev['quantization'] = ll['quantization']
if 'output_width' in ll:
prev['output_width'] = ll['output_width']
prev_name = name
# Delete the conv layers that were fused into the eltwise layer
for (prev_name, name) in pop_list:
# Change any dangling input sequences to the fused layer
for (other_name, ol) in layers.items():
if other_name != name and other_name != prev_name and 'in_sequences' in ol:
for i, e in enumerate(ol['in_sequences']):
if e == name:
ol['in_sequences'][i] = prev_name
# Delete the conv portion
layers.pop(name)
# 7b - Merge (fuse) pooling and conv layers
prev_name = ''
pop_list = []
veto = False
prev = {}
for count, (name, ll) in enumerate(layers.items()):
if ll['main_op'] in ('Conv', 'ConvTranspose', 'Linear') and prev_name != '':
prev = layers[prev_name]
if prev['main_op'] != 'Passthrough' or (
'max_pool' not in prev and 'avg_pool' not in prev
):
prev_name = name
continue
# Check that no layer other than the activation layer uses the intermediate output of
# the conv layer as an input
veto = False
for (other_name, ol) in layers.items():
if other_name != name and other_name != prev_name and 'in_sequences' in ol:
if prev_name in ol['in_sequences']:
veto = True
break
if veto:
prev_name = name
continue
# Combine both layers
if 'comment' not in ll:
ll['comment'] = f'{prev_name} fused with {name}'
else:
ll['comment'] = f'{prev_name} and ' + ll['comment']
pop_list.append((prev_name, name)) # Mark first layer for deletion
# Copy the pooling information into the convolution layer
inputs[name] = inputs[prev_name]
if 'in_sequences' in prev:
ll['in_sequences'] = prev['in_sequences']
if 'avg_pool' in prev:
ll['avg_pool'] = prev['avg_pool']
if 'max_pool' in prev:
ll['max_pool'] = prev['max_pool']
if 'pool_stride' in prev:
ll['pool_stride'] = prev['pool_stride']
if 'pool_dilation' in prev:
ll['pool_dilation'] = prev['pool_dilation']
if 'data_format' in prev:
ll['data_format'] = prev['data_format']
prev_name = name
# Delete the pooling layers that were fused into the conv layer
for (prev_name, _) in pop_list:
# Delete the pooling portion
layers.pop(prev_name)
# 7c - Merge (fuse) element-wise and convolution layers
prev_name = ''
pop_list = []
for count, (name, ll) in enumerate(layers.items()):
if ll['main_op'] in ('Conv', 'ConvTranspose') and prev_name != '':
prev = layers[prev_name]
# Only one pooling operation possible
pool_count: int = 0
if 'max_pool' in prev:
pool_count += 1
if 'avg_pool' in prev:
pool_count += 1
if 'max_pool' in ll:
pool_count += 1
if 'avg_pool' in ll:
pool_count += 1
# Check that no layer other than the conv layer uses the intermediate output of the
# element-wise layer as an input
veto = False
for (other_name, ol) in layers.items():
if other_name != name and other_name != prev_name and 'in_sequences' in ol:
if prev_name in ol['in_sequences']:
veto = True
break
if veto or 'in_sequences' in ll or 'in_dim' in ll or 'flatten' in ll:
prev_name = name
continue
if prev['main_op'] != 'Passthrough' or 'operands' not in prev \
or prev['operands'] == 1 or pool_count > 1:
prev_name = name
continue
# MAX78002 - avoid element-wise with bias and convolution when using multi-pass
if ai8x.dev.device == 87 and ll['have_bias'] and prev['proc_count'] > MAX_PROC:
prev_name = name
continue
# Combine both layers
if 'comment' not in ll:
ll['comment'] = f'{prev_name} fused with {name}'
else:
ll['comment'] = f'{prev_name} and ' + ll['comment']
pop_list.append((prev_name, name)) # Mark first layer for deletion
# Copy over convolution operation and keep the element-wise operation in place
inputs[name] = inputs[prev_name]
if 'in_sequences' in prev:
ll['in_sequences'] = prev['in_sequences']
if 'max_pool' in prev:
ll['max_pool'] = prev['max_pool']
ll['pool_first'] = 'true'
if 'avg_pool' in prev:
ll['avg_pool'] = prev['avg_pool']
ll['pool_first'] = 'true'
if 'max_pool' in ll or 'avg_pool' in ll:
ll['pool_first'] = 'false'
if 'pool_stride' in prev:
ll['pool_stride'] = prev['pool_stride']
if 'pool_dilation' in prev:
ll['pool_dilation'] = prev['pool_dilation']
if 'data_format' in prev:
ll['data_format'] = prev['data_format']
ll['eltwise'] = prev['eltwise']
ll['operands'] = prev['operands']
prev_name = name
# Delete the element-wise layers that were fused into the convolution layer
for (prev_name, _) in pop_list:
# Delete the element-wise portion
layers.pop(prev_name)
# 8 - Insert passthrough layers for write_gap
write_gap_list: List[Tuple[str, int]] = []
insert_list: List[Tuple[str, int]] = []
source_list: List[Tuple[str, str]] = []
prev_name = ''
for (name, ll) in layers.items():
if 'in_sequences' not in ll:
if prev_name == '':
prev_name = name
continue
sources: List[str] = [prev_name]
else:
sources = ll['in_sequences']
operands = ll['operands'] if 'operands' in ll else len(sources)
if operands < 2:
prev_name = name
continue
# For each input, check whether anybody else is using the input. If yes, insert a dummy
# layer that creates a write_gap version of the data. If no, add the write_gap to the
# producer.
for source in sources:
must_insert: bool = False
prev_name_inner = ''
for (other_name, ol) in layers.items():
if other_name not in (name, source):
if 'in_sequences' in ol:
for e in ol['in_sequences']:
if e == source:
must_insert = True
elif prev_name_inner == source:
must_insert = True
# Break the sequence
ol['in_sequences'] = [source]
prev_name_inner = other_name
if not must_insert:
# The source is used only by the element-wise layer, so we can insert the write gap
# directly
write_gap_list.append((source, operands))
else:
insert_list.append((source, operands))
# Replace source with source_gap in layers[name]['in_sequences']
source_list.append((name, source))
new_name = 'gap_' + source
# ...and insert shaope information (input and output are both the same as the
# original layer's output)
inputs[new_name] = [source + '_data']
outputs[new_name] = [name + '_data']
for ie in outputs[source]:
if ie in shapes:
shapes[name + '_data'] = shapes[ie]
shapes[source + '_data'] = shapes[ie]
prev_name = name
# Insert simple write gaps
for (name, operands) in write_gap_list:
layers[name]['write_gap'] = operands - 1
# Break sequence
for (name, source) in source_list:
seq = layers[name]['in_sequences']
for i, s in enumerate(seq):
if s == source:
seq[i] = 'gap_' + source
# Insert additional layers
for (name, operands) in insert_list:
new_layer: Dict[str, Any] = {}
new_name = 'gap_' + name
new_layer['name'] = new_name
processors = 0
for ie in inputs[new_name]:
if ie in shapes:
processors += shapes[ie][0]
new_layer['proc_count'] = max(1, processors)
new_layer['op'] = new_layer['main_op'] = 'Passthrough'
new_layer['write_gap'] = operands - 1
# Insert into dict via list
insert_pos: int = list(layers.keys()).index(name) + 1
layers_list: List[Any] = list(layers.items())
layers_list.insert(insert_pos, (new_name, new_layer))
layers = dict(layers_list)
# 9 - Record actual used processors for all layers and weight cost for all conv layers
# Add all_inputs (either in_sequence or the previous layer's name if not defined)
all_inputs: Dict[str, List[str]] = {} # Input LAYERS to the key layer (vs. shapes in inputs)
all_outputs: Dict[str, List[str]] = {} # Output LAUYERS to th key layer (vs. shapes)
prev_name = ''
for (name, ll) in layers.items():
val: Optional[List[str]] = None
if 'in_sequences' not in ll:
val = [prev_name] if prev_name != '' else None
else:
val = ll['in_sequences']
if val is not None:
all_inputs[name] = val
for ie in val:
if ie in all_outputs:
all_outputs[ie].append(name)
else:
all_outputs[ie] = [name]
prev_name = name
# There are two cases: element-wise operations ('operands' defined and > 1), and conv
# operations with multi-pass (> 64 input channels). In either case, in_sequences has
# more than one member.
if 'in_sequences' not in ll or len(ll['in_sequences']) < 2:
continue
operands = ll['operands'] if 'operands' in ll else 1
if operands == 1:
operands = len(ll['in_sequences'])
# Concat instead of interleave for small channel counts
if ll['proc_count'] <= MAX_PROC:
ll['concat'] = True
for i, ie in enumerate(ll['in_sequences']):
lin = layers[ie]
lin['concat_source'] = max(i, lin['concat_source']) \
if 'concat_source' in lin else i
else:
# Interleave the inputs in the channels
ll['proc_count'] //= operands
# Mark the size of all concat outputs
for (name, ll) in layers.items():
if 'concat_source' not in ll:
continue
if name not in outputs or len(outputs[name]) != 1 or outputs[name][0] not in shapes:
print('ERROR: Cannot derive output shape(s) for layer', name)
oshape = shapes[outputs[name][0]]
ll['concat_shape'] = oshape[0]
def calculate_processors(processors: int) -> Tuple[int, int]:
"""
Given a channel count, return the multi-pass adjusted processor count
"""
multipass: int = 1
if processors > MAX_PROC: # Multi-pass processor count
multipass = (processors + MAX_PROC - 1) // MAX_PROC
processors = (processors + multipass - 1) // multipass # Rounded up
remainder: int = processors % 4
if remainder != 0:
remainder = 4 - remainder
processors += remainder # To next multiple of 4
return processors, multipass
# cost_list: processors/weights per layer for Conv layers with 60 or fewer processors
cost_list: List[Tuple[int, int, str]] = []
for (name, ll) in layers.items():
hwc = ll['data_format'] if 'data_format' in ll else True
processors = ll['proc_count']
multipass: int = 1
# Calculate the final number of used processors
if hwc:
processors, multipass = calculate_processors(processors)
assert processors <= MAX_PROC
else:
assert processors <= MAX_PROC // 4
ll['proc_used'] = processors
ll['multipass'] = multipass
weights_per_processor: int = 0
if ll['main_op'] in ('Conv', 'ConvTranspose', 'Linear'):
# Account for kernel size for convolution operations
weights_per_processor = ll['weight_count'] // processors
if 'quantization' in ll:
weights_per_processor //= 8 // ll['quantization']
ll['weight_cost'] = weights_per_processor
if processors <= 60:
# Only add to this list if we're using less than the full processor count
# (for 64 processors, there are no options and hence no optimizations)
cost_list.append((weights_per_processor, processors, name))
else:
# Append even those layers that use 64 processors, because they may be 'concat' layer
# or it may be a layer 0 and therefore, the source may need to be arranged properly.
# But use zero cost to get them to move to the end (or ignored in the cost
# calculation).
cost_list.append((0, processors, name))
# print('\nFINAL cost_list', cost_list, '\n\n')
# Layer 0 is not the output of anything, so add it in from the beginning
bucket_groups: List[List[Tuple[int, int, str]]] = [[cost_list[0]]]
bucket: Dict[str, int] = {}
bucket[cost_list[0][2]] = 0
# Find layers where the output is used as input of more than one layer. Those layers then
# must use the same processors. This is achieved by "combining the rectangles".
for (name, ll) in layers.items():
if name not in all_outputs:
continue # This shouldn't ever happen except for the final layer
# print(name, 'grouping everything in', all_outputs[name])
# FIXME What is the output shape of this layer? May need the processor_count
merge_buckets: List[int] = []
for i, (_, _, n) in enumerate(cost_list):
if n in all_outputs[name]:
if n in bucket: # We already have a bucket that contains n.
# Merge the bucket contents
merge_buckets.append(bucket[n])
else:
# Create a new empty bucket and add the layer to it. Also add the new bucket
# to the merge list.
bucket_number = len(bucket_groups)
bucket[n] = bucket_number
merge_buckets.append(bucket_number)
bucket_groups.append([cost_list[i]])
if len(merge_buckets) > 1:
target: int = merge_buckets[0]
for i in merge_buckets[1:]:
bucket_groups[target] += bucket_groups[i]
bucket_groups[i] = [] # Rather than deleting, we just set invalid values
# For the remaining buckets, combine the weights, check the processors, and make a layer list
# print('FINAL bucket_groups', bucket_groups)
del cost_list # No longer needed
# Same as cost_list, but now with multiple entries
group_list: List[Tuple[int, int, int, int, List[Tuple[int, int, str]]]] = []
for b in bucket_groups:
if not b:
continue
# print('============ bucket group', b)
group_weights: int = 0
group_procs: int = -1
group_item: List[Tuple[int, int, str]] = []
min_shift = 0
last_processor = MAX_PROC - 1
for (weights_per_processor, processors, name) in b:
# print('examining - weights', weights_per_processor, 'procs', processors, 'name',
# name, 'concat', 'YES' if 'concat' in layers[name] else 'NO')
group_weights += weights_per_processor # Add weights
# print('processors', processors, 'vs', group_procs)
if 'concat' in layers[name]: # For concatenation, we need partials
assert group_procs < 0 or group_procs == processors \
or group_procs == processors // len(all_inputs[name])
# FIXME: a straight divide may not be correct
# FIXME: If there's a concat target, make sure to leave the extra space
# print('Found a concat, should set min_shift')
# min_shift = max(min_shift, processors // len(all_inputs[name]))
else:
# print(layers[name])
assert group_procs < 0 or group_procs == processors
group_procs = max(processors, group_procs)
group_item.append((weights_per_processor, processors, name))
if group_procs >= 0:
group_list.append((group_weights, group_procs,
min_shift, last_processor, group_item))
# print('group_list', group_list, '\n')
# 10 - Assign processors
# TODO: Real algorithm for output_offset
weights_used: List[int] = [0] * MAX_PROC
def allocate_processors(
weight_cost: int,
count: int,
min_shift: int,
data: List[Tuple[int, int, str]],
hwc: bool = True,
) -> Tuple[List[int], int]:
"""
Allocate a given count of processors and return the bit map.
TODO: This function should be expanded to evenly distribute resources, not only based on
weights but also data memory utilization.
Additionally, the weight allocator could instead use a box packing algorithm, but the
expected improvements are modest.
"""
min_cost: Tuple[int, int] = (2**63-1, -1)
shifted_map: List[int] = [0] * len(data)
processor_map: List[int] = shifted_map
# Move processor map left if needed to achieve the lowest combined 'weights_used'
for shift in range(min_shift, MAX_PROC - count + 1, 4):
for d, (item_cost, item_procs, _) in enumerate(data):
if hwc:
# Pick "count" processors starting from 0
shifted_map[d] = ((1 << item_procs) - 1) << shift
else:
# Pick the first for every quadrant (FIFO compatible) or one every 4
mult = 16 if count <= 4 else 4
shifted_map[d] = 0
for i in range(item_procs):
shifted_map[d] |= 1 << mult * i
shifted_map[d] <<= shift
if weight_cost == 0:
min_cost = (0, min_shift)
break
cost: List[int] = weights_used.copy()
for p in range(MAX_PROC):
for d, (item_cost, item_procs, _) in enumerate(data):
if shifted_map[d] & (1 << p):
cost[p] += item_cost
max_cost = max(cost) # High water mark
if max_cost < min_cost[0]: # Better option?
min_cost = (max_cost, shift)
processor_map = shifted_map.copy()
if weight_cost > 0:
# Accounting based on the result
for p in range(MAX_PROC):
proc_used: bool = False
for d, _ in enumerate(data):
if processor_map[d] & (1 << p):
proc_used = True
if proc_used:
weights_used[p] = min_cost[0]