Newer
Older
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
'''
Cost model.
'''
#import numpy as np
from operator import mul
from operator import add
from functools import reduce
import copy
import math
from . import loop_enum as le
from . import buffer_enum as be
def get_comp_cost(layer):
'''
Compute the total # of MAC computation, it is independent of other optimizations
Also it is independent of input size and input/filter stride
Total # of computation = OX*OY*IC*OC*ON*FX*FY
'''
cost = layer.wofm * layer.hofm * layer.nifm * layer.nofm \
* layer.nimg * layer.wfil * layer.hfil
return cost
def get_ideal_performance(layer, resource):
'''
Compute the ideal runtime in cycles by assuming 100% PE array utilization
Ideal # of cycles = Total # of MAC computation / Total # of PEs
#LMEI Need to be modified if later when adding precision-scalable PE.
# of functional PE will change depending on different precision modes.
'''
total_comp = get_comp_cost(layer)
number_pe = reduce(mul, resource.para_count_list, 1)
runtime = math.ceil(total_comp *1.0 / number_pe)
return runtime
def get_layer_size(layer):
'''
Get size of ifmap, ofmap, filter of the layer
#LMEI ifmap_size should be able to calculate based on ofmap_size and input stride(IS) /filter stride(FS)
IX = IS*(OX-1) + FS*(FX-1) + 1
wifm = wistd*(wofm-1) + wfstd*(wfil-1) + 1
'''
ifmap_size = layer.wifm * layer.hifm * layer.nifm * layer.nimg
ofmap_size = layer.wofm * layer.hofm * layer.nofm * layer.nimg
flmap_size = layer.wfil * layer.hfil * layer.nifm * layer.nofm
return [ifmap_size, ofmap_size, flmap_size]
def get_hinted_para(level, hint):
'''
Get the actual total spatial unrolling size from loop schedule
'''
assert hint
hinted_para = 1
for loop in range(le.NUM):
if loop in hint:
hinted_loop_para = hint[loop][level][2]
hinted_para *= hinted_loop_para
return hinted_para
def valid_dataflow(resource, hint):
'''
Check if the actual spatial unrolling size from loop schedule meets the HW utilization requirement
by comparing it with real HW parallelism size * utilization threshold.
'''
num_levels = resource.buffer_levels()
for level in range(num_levels):
if resource.paras[level].count != 1 and \
get_hinted_para(level, hint) < (resource.paras[level].count * resource.utilization_threshold):
return False
return True
def get_if_access(level, point, layer, mac_capacity = 1):
'''
Get per element # of access of Input at current level
Not accurate because [FX, FY] is not totally irrelevant terms for ifmap..
#LMEI Need to be modified by using the concept of the dataset.
'''
if level == 0 and mac_capacity == 0:
return layer.wfil * layer.hfil * layer.nofm / (layer.wstd * layer.hstd)
ex_order_index = min(point.loop_orders[le.OX][level],
point.loop_orders[le.OY][level],
point.loop_orders[le.IC][level],
point.loop_orders[le.ON][level])
fx_exclusive = point.loop_orders[le.FX][level] < ex_order_index
fy_exclusive = point.loop_orders[le.FY][level] < ex_order_index
oc_exclusive = point.loop_orders[le.OC][level] < ex_order_index
fx_acc = reduce(mul, point.loop_blockings[le.FX][level+fx_exclusive:], 1)
fy_acc = reduce(mul, point.loop_blockings[le.FY][level+fy_exclusive:], 1)
oc_acc = reduce(mul, point.loop_blockings[le.OC][level+oc_exclusive:], 1)
# No loop orders among unrolled loops, they have the same order
fx_par = reduce(mul, point.loop_partitionings[le.FX][level:], 1)
fy_par = reduce(mul, point.loop_partitionings[le.FY][level:], 1)
oc_par = reduce(mul, point.loop_partitionings[le.OC][level:], 1)
return fx_acc * fy_acc * oc_acc * fx_par * fy_par * oc_par / (layer.wstd * layer.hstd)
def get_of_access(level, point, layer, mac_capacity = 1):
'''
Get per element # of access of Output at current level
For output:
Relevant terms [OX, OY, OC, ON]
irrelevant terms [FX, FY, IC]
Calculating rule:
At lowest mem level (directly talk to MAC), calculate per element access
by timing all irrelevant terms [FX, FY, IC] together
For the rest higher mem levels,
firstly, check if there is stationary possibility
(irrelevant loops for filter [FX, FY, IC] are at the innermost position of this level)
if there is, exclude the irrelevant loop(s) from the current level's # of per element access computing
because they have been taken into account in lower level's # of per element access computing
secondly, calculate the current level's # of per element access
by multiplying all the irrelevant terms from current level to the highest level
including both temporal unrolling part and spatial unrolling part (parallelism).
'''
if level == 0 and mac_capacity == 0 :
return layer.wfil * layer.hfil * layer.nifm
ex_order_index = min(point.loop_orders[le.OX][level],
point.loop_orders[le.OY][level],
point.loop_orders[le.OC][level],
point.loop_orders[le.ON][level])
fx_exclusive = point.loop_orders[le.FX][level] < ex_order_index
fy_exclusive = point.loop_orders[le.FY][level] < ex_order_index
ic_exclusive = point.loop_orders[le.IC][level] < ex_order_index
fx_acc = reduce(mul, point.loop_blockings[le.FX][level+fx_exclusive:], 1)
fy_acc = reduce(mul, point.loop_blockings[le.FY][level+fy_exclusive:], 1)
ic_acc = reduce(mul, point.loop_blockings[le.IC][level+ic_exclusive:], 1)
fx_par = reduce(mul, point.loop_partitionings[le.FX][level:], 1)
fy_par = reduce(mul, point.loop_partitionings[le.FY][level:], 1)
ic_par = reduce(mul, point.loop_partitionings[le.IC][level:], 1)
return fx_acc * fy_acc * ic_acc * fx_par * fy_par * ic_par
def get_fl_access(level, point, layer, mac_capacity = 1):
'''
Get per element # of access of Weight at current level
For filter:
Relevant terms [FX, FY, IC, OC]
irrelevant terms [OX, OY, ON]
Calculating rule:
At lowest mem level (directly talk to MAC), calculate per element access
by timing all irrelevant terms [OX, OY, ON] together
For the rest higher mem levels,
firstly, check if there is stationary possibility
(irrelevant loops for filter [OX, OY, ON] are at the innermost position of this level)
if there is, exclude the irrelevant loop(s) from the current level's # of per element access computing
because they have been taken into account in lower level's # of per element access computing
secondly, calculate the current level's # of per element access
by multiplying all the irrelevant terms from current level to the highest level
including both temporal unrolling part and spatial unrolling part (parallelism).
'''
if level == 0 and mac_capacity == 0:
return layer.wofm * layer.hofm * layer.nimg
ex_order_index = min(point.loop_orders[le.FX][level],
point.loop_orders[le.FY][level],
point.loop_orders[le.IC][level],
point.loop_orders[le.OC][level])
ox_exclusive = point.loop_orders[le.OX][level] < ex_order_index
oy_exclusive = point.loop_orders[le.OY][level] < ex_order_index
on_exclusive = point.loop_orders[le.ON][level] < ex_order_index
ox_acc = reduce(mul, point.loop_blockings[le.OX][level+ox_exclusive:], 1)
oy_acc = reduce(mul, point.loop_blockings[le.OY][level+oy_exclusive:], 1)
on_acc = reduce(mul, point.loop_blockings[le.ON][level+on_exclusive:], 1)
ox_par = reduce(mul, point.loop_partitionings[le.OX][level:], 1)
oy_par = reduce(mul, point.loop_partitionings[le.OY][level:], 1)
on_par = reduce(mul, point.loop_partitionings[le.ON][level:], 1)
return ox_acc * oy_acc * on_acc * ox_par * oy_par * on_par
def opt_get_if_access(level, point, ba_arr, pa_arr):
'''
Get # access of if block at current level
The repeated access to ifmap is determined by the blocking factors and
parallelism counts of those loops other than ifmap-related loops outside of
this level.
At the same buffer level, if the other loops are outside of the innermost
loop of ifmap-related loops, their blocking factors and parallelism counts
at this level should also contribute to the number of accesses.
'''
ex_order_index = min(point.loop_orders[le.OX][level],
point.loop_orders[le.OY][level],
point.loop_orders[le.IC][level],
point.loop_orders[le.ON][level])
fx_exclusive = point.loop_orders[le.FX][level] < ex_order_index
fy_exclusive = point.loop_orders[le.FY][level] < ex_order_index
oc_exclusive = point.loop_orders[le.OC][level] < ex_order_index
fx_acc = ba_arr[le.FX][level+fx_exclusive] #reduce(mul, point.loop_blockings[le.FX][level+fx_exclusive:], 1)
fy_acc = ba_arr[le.FY][level+fy_exclusive] #reduce(mul, point.loop_blockings[le.FY][level+fy_exclusive:], 1)
oc_acc = ba_arr[le.OC][level+oc_exclusive] #reduce(mul, point.loop_blockings[le.OC][level+oc_exclusive:], 1)
fx_par = pa_arr[le.FX][level] #reduce(mul, point.loop_partitionings[le.FX][level+fx_exclusive:], 1)
fy_par = pa_arr[le.FY][level] #reduce(mul, point.loop_partitionings[le.FY][level+fy_exclusive:], 1)
oc_par = pa_arr[le.OC][level] #reduce(mul, point.loop_partitionings[le.OC][level+oc_exclusive:], 1)
return fx_acc * fy_acc * oc_acc * fx_par * fy_par * oc_par
def opt_get_of_access(level, point, ba_arr, pa_arr):
'''
Get # access of of block at current level
See comments in routine for ifmap.
'''
ex_order_index = min(point.loop_orders[le.OX][level],
point.loop_orders[le.OY][level],
point.loop_orders[le.OC][level],
point.loop_orders[le.ON][level])
fx_exclusive = point.loop_orders[le.FX][level] < ex_order_index
fy_exclusive = point.loop_orders[le.FY][level] < ex_order_index
ic_exclusive = point.loop_orders[le.IC][level] < ex_order_index
#TODO
fx_acc = ba_arr[le.FX][level+fx_exclusive] #reduce(mul, point.loop_blockings[le.FX][level+fx_exclusive:], 1)
fy_acc = ba_arr[le.FY][level+fy_exclusive] #reduce(mul, point.loop_blockings[le.FY][level+fy_exclusive:], 1)
ic_acc = ba_arr[le.IC][level+ic_exclusive] #reduce(mul, point.loop_blockings[le.OC][level+oc_exclusive:], 1)
fx_par = pa_arr[le.FX][level] #reduce(mul, point.loop_partitionings[le.FX][level+fx_exclusive:], 1)
fy_par = pa_arr[le.FY][level] #reduce(mul, point.loop_partitionings[le.FY][level+fy_exclusive:], 1)
ic_par = pa_arr[le.IC][level] #reduce(mul, point.loop_partitionings[le.OC][level+oc_exclusive:], 1)
return fx_acc * fy_acc * ic_acc * fx_par * fy_par * ic_par
def opt_get_fl_access(level, point, ba_arr, pa_arr):
'''
Get # access of fl block at current level
See comments in routine for ifmap.
'''
ex_order_index = min(point.loop_orders[le.FX][level],
point.loop_orders[le.FY][level],
point.loop_orders[le.IC][level],
point.loop_orders[le.OC][level])
ox_exclusive = point.loop_orders[le.OX][level] < ex_order_index
oy_exclusive = point.loop_orders[le.OY][level] < ex_order_index
on_exclusive = point.loop_orders[le.ON][level] < ex_order_index
ox_acc = ba_arr[le.OX][level+ox_exclusive] #reduce(mul, point.loop_blockings[le.OX][level+ox_exclusive:], 1)
oy_acc = ba_arr[le.OY][level+oy_exclusive] #reduce(mul, point.loop_blockings[le.OY][level+oy_exclusive:], 1)
on_acc = ba_arr[le.ON][level+on_exclusive] #reduce(mul, point.loop_blockings[le.ON][level+on_exclusive:], 1)
ox_par = pa_arr[le.OX][level] #reduce(mul, point.loop_partitionings[le.OX][level+ox_exclusive:], 1)
oy_par = pa_arr[le.OY][level] #reduce(mul, point.loop_partitionings[le.OY][level+oy_exclusive:], 1)
on_par = pa_arr[le.ON][level] #reduce(mul, point.loop_partitionings[le.ON][level+on_exclusive:], 1)
return ox_acc * oy_acc * on_acc * ox_par * oy_par * on_par
def get_if_size(blocking_accum_list, partitioning_accum_list, partitioning_list, layer):
'''
Get size of if block at current level including both temporal and spatial loop part
blocking -> temporal loop part
partitioning -> spatial loop part
#LMEI to support filter stride(FS) later
right now, FS/wfstd = 1 in
IX = IS*(OX-1) + FS*(FX-1) + 1 or
wifm = wistd*(wofm-1) + wfstd*(wfil-1) + 1
#LMEI (new HW template) no need for Input Duplication when OC partitions
by letting one reg broadcast Input to a row of OC partitioned PE
and remove inner PE ifamp register
'''
fx_acc = blocking_accum_list[le.FX] * partitioning_accum_list[le.FX]
fy_acc = blocking_accum_list[le.FY] * partitioning_accum_list[le.FY]
ox_acc = blocking_accum_list[le.OX] * partitioning_accum_list[le.OX]
oy_acc = blocking_accum_list[le.OY] * partitioning_accum_list[le.OY]
width = fx_acc + (ox_acc - 1) * layer.wstd
height = fy_acc + (oy_acc - 1) * layer.hstd
return width * height * \
blocking_accum_list[le.IC] * partitioning_accum_list[le.IC] * \
blocking_accum_list[le.ON] * partitioning_accum_list[le.ON] * \
partitioning_list[le.OC] # Duplication when OC partitions
def get_of_size(blocking_accum_list, partitioning_accum_list, partitioning_list):
'''
Get size of of block at current level including both temporal and spatial loop part
#LMEI (new HW template) no need for Output Duplication when IC, FX or FY partitions
by letting output data from a row of IC, FX or FY partitioned PE add together
and remove inner PE ofamp register
'''
return blocking_accum_list[le.OX] * partitioning_accum_list[le.OX] * \
blocking_accum_list[le.OY] * partitioning_accum_list[le.OY] * \
blocking_accum_list[le.OC] * partitioning_accum_list[le.OC] * \
blocking_accum_list[le.ON] * partitioning_accum_list[le.ON] * \
partitioning_list[le.IC] * partitioning_list[le.FX] * \
partitioning_list[le.FY] # Duplication when IC, FX or FY partitions
def get_fl_size(blocking_accum_list, partitioning_accum_list, partitioning_list):
'''
Get size of fl block at current level
#LMEI (new HW template) no need for Weight Duplication when OX, OY or ON partitions
by letting one reg broadcast Weight to a row of OX, OY or ON partitioned PE
and remove inner PE weight register
'''
return blocking_accum_list[le.FX] * partitioning_accum_list[le.FX] * \
blocking_accum_list[le.FY] * partitioning_accum_list[le.FY] * \
blocking_accum_list[le.IC] * partitioning_accum_list[le.IC] * \
blocking_accum_list[le.OC] * partitioning_accum_list[le.OC] * \
partitioning_list[le.OX] * partitioning_list[le.OY] *\
partitioning_list[le.ON] # Duplication when OX, OY or ON partitions
def get_if_bank_size(blocking_accum_list, layer):
'''
Get size of if block at current level
blocking -> temporal loop part
#LMEI to support filter stride(FS) later
right now, FS/wfstd = 1 in
IX = IS*(OX-1) + FS*(FX-1) + 1 or
wifm = wistd*(wofm-1) + wfstd*(wfil-1) + 1
'''
fx_acc = blocking_accum_list[le.FX]
fy_acc = blocking_accum_list[le.FY]
ox_acc = blocking_accum_list[le.OX]
oy_acc = blocking_accum_list[le.OY]
width = fx_acc + (ox_acc - 1) * layer.wstd
height = fy_acc + (oy_acc - 1) * layer.hstd
return width * height * \
blocking_accum_list[le.IC] * blocking_accum_list[le.ON]
def get_of_bank_size(blocking_accum_list):
'''
Get size of of block at current level
blocking -> temporal loop part
'''
return blocking_accum_list[le.OX] * blocking_accum_list[le.OY] * \
blocking_accum_list[le.OC] * blocking_accum_list[le.ON]
def get_fl_bank_size(blocking_accum_list):
'''
Get size of fl block at current level
blocking -> temporal loop part
'''
return blocking_accum_list[le.FX] * blocking_accum_list[le.FY] * \
blocking_accum_list[le.IC] * blocking_accum_list[le.OC]
def get_array_access_and_cost(level, para, access_list, point):
'''
Get the access at array level from the access at the
lower level of memory hierarchy
'''
para_mode = para.access_mode
assert para_mode == 1 or para_mode == 2 # Don't get it
array_dim = para.array_dim
para_count = para.array_width
para_cost = para.array_access_cost * 1.0
nearest_pe_cost = para_cost
[if_block_access, of_block_access, fl_block_access] = access_list
partitions = list(zip(*point.loop_partitionings))[level]
para_dim = point.para_loop_dim[level]
partitions_nearest = [1,]*le.NUM
partitions_far = []
across_block_cost = [0]*array_dim
if para_mode == 1:
for i in range(len(para_dim)):
para_index = para_dim[i]
partitions_far.append([1,]*le.NUM)
if len(para_index) == 1:
partitions_nearest[para_index[0]] = partitions[para_index[0]]
else:
inner_loop, outer_loop = para_index
partitions_nearest[inner_loop] = partitions[inner_loop]
partitions_far[i][outer_loop] = partitions[outer_loop]
across_block_cost[i] = para_cost * partitions[inner_loop]
array_if_block_access_nearest = if_block_access * partitions_nearest[le.FX] * \
partitions_nearest[le.FY] * partitions_nearest[le.OC]
array_of_block_access_nearest = of_block_access * partitions_nearest[le.FX] * \
partitions_nearest[le.FY] * partitions_nearest[le.IC]
array_fl_block_access_nearest = fl_block_access * partitions_nearest[le.OX] * \
partitions_nearest[le.OY] * partitions_nearest[le.ON]
array_access = [[array_if_block_access_nearest, array_of_block_access_nearest, array_fl_block_access_nearest]]
for i in range(array_dim): # Don't get it
if_partitions_far = partitions_far[i][le.FX] * partitions_far[i][le.FY] * partitions_far[i][le.OC]
if_partitions_far = if_partitions_far if if_partitions_far != 1 else 0
of_partitions_far = partitions_far[i][le.FX] * partitions_far[i][le.FY] * partitions_far[i][le.IC]
of_partitions_far = of_partitions_far if of_partitions_far != 1 else 0
fl_partitions_far = partitions_far[i][le.OX] * partitions_far[i][le.OY] * partitions_far[i][le.ON]
fl_partitions_far = fl_partitions_far if fl_partitions_far != 1 else 0
if_array_block_access = if_block_access * if_partitions_far
of_array_block_access = of_block_access * of_partitions_far
fl_array_block_access = fl_block_access * fl_partitions_far
array_access.append([if_array_block_access, of_array_block_access, fl_array_block_access])
return [array_access, [nearest_pe_cost] + across_block_cost]
elif para_mode == 2:
for i in range(len(para_dim)):
para_index = para_dim[i]
for j in para_index:
partitions_nearest[j] = partitions[j]
array_if_block_access_nearest = if_block_access * partitions_nearest[le.FX] * \
partitions_nearest[le.FY] * partitions_nearest[le.OC]
array_of_block_access_nearest = of_block_access * partitions_nearest[le.FX] * \
partitions_nearest[le.FY] * partitions_nearest[le.IC]
array_fl_block_access_nearest = fl_block_access * partitions_nearest[le.OX] * \
partitions_nearest[le.OY] * partitions_nearest[le.ON]
array_access = [[array_if_block_access_nearest, array_of_block_access_nearest, array_fl_block_access_nearest]]
return [array_access, [nearest_pe_cost]]
def get_access(point, layer, resource):
'''
Get the total access of each block at each level,
return the list as
[[if_block_access, of_block_access, fl_block_access], ...].
Assume all the buffers are inclusive, so buffers in lower level
appear in higher level as well.
For the parallelism case assume read from next memory level,
Support more access modes in parallelism case
'''
#TODO support more customized memory
#TODO more access at overlapped boundary
num_levels = resource.buffer_levels()
mac_capacity = resource.mac_capacity
access_list = []
for level in range(num_levels):
if_block_access = get_if_access(level, point, layer, mac_capacity)
of_block_access = 2 * get_of_access(level, point, layer, mac_capacity) - 1
fl_block_access = get_fl_access(level, point, layer, mac_capacity)
access_list.append([if_block_access, of_block_access, fl_block_access])
#para_mode = [e.access_mode for i, e in enumerate(resource.paras) if e.access_mode != 0]
para_mode_level = [i for i, e in enumerate(resource.paras) if e.access_mode != 0]
partitions = list(zip(*point.loop_partitionings))
array_costs = []
if para_mode_level:
# access at array level
#para_mode_level = [i for i, e in enumerate(resource.paras) if e.access_mode != 0]
delta = 0
for level in para_mode_level:
if level + delta + 1 >= num_levels :
next_level_access = [1, 1, 1]
else:
next_level_access = copy.copy(access_list[level + delta + 1])
next_level_access[1] = (next_level_access[1] + 1)/2
array_access, array_cost = get_array_access_and_cost(level, resource.paras[level], next_level_access, point)
array_costs.append(array_cost)
access_list.insert(level + delta + 1, array_access)
delta += 1
return [access_list, array_costs]
def opt_get_access(num_levels, point, mac_capacity):
'''
See the above function's comments. This function is just an
optimized version of the above function
'''
''' blocking_accum_arr is reversed cumprod numpy array '''
#TODO support mac_capacity
#blocking_arr = np.ones((le.NUM, num_levels+1))
#partitioning_arr = np.ones((le.NUM, num_levels+1))
#blocking_arr[:,:-1] = np.array(point.loop_blockings)
#partitioning_arr[:,:-1] = np.array(point.loop_partitionings)
#blocking_accum_arr = np.ones((le.NUM, num_levels+1))
#partitioning_accum_arr = np.ones((le.NUM, num_levels+1))
#for i in range(le.NUM):
# blocking_accum_arr[i][:-1] = np.cumprod(blocking_arr[i][::-1])[::-1]
# partitioning_accum_arr[i][:-1] = np.cumprod(partitioning_arr[i][::-1])[::-1]
#blocking_accum_arr = blocking_arr[...,::-1].cumprod(axis=-1)[...,::-1]
#partitioning_accum_arr = partitioning_arr[...,::-1].cumprod(axis=-1)[...,::-1]
#blocking_accum_arr = np.hstack((blocking_accum_arr, np.ones((le.NUM, 1))))
#partitioning_accum_arr = np.hstack((partitioning_accum_arr, np.ones((le.NUM, 1))))
blocking_accum_arr = []
partitioning_accum_arr = []
for i in range(le.NUM):
ba_current_level = [1]
pa_current_level = [1]
ba_tmp = 1
pa_tmp = 1
for level in range(num_levels-1, -1, -1):
ba_tmp = ba_tmp * point.loop_blockings[i][level]
pa_tmp = pa_tmp * point.loop_partitionings[i][level]
ba_current_level.append(ba_tmp)
pa_current_level.append(pa_tmp)
blocking_accum_arr.append(ba_current_level[::-1])
partitioning_accum_arr.append(pa_current_level[::-1])
access_arr = np.zeros((num_levels, 3))
for level in range(num_levels):
access_arr[level][0] = opt_get_if_access(level, point, blocking_accum_arr, partitioning_accum_arr)
access_arr[level][1] = 2 * opt_get_of_access(level, point, blocking_accum_arr, partitioning_accum_arr) - 1
access_arr[level][2] = opt_get_fl_access(level, point, blocking_accum_arr, partitioning_accum_arr)
return access_arr
def get_bank_size(point, layer, level):
blocking_accum_list = []
for i in range(le.NUM):
blocking_accum_list.append(reduce(mul, point.loop_blocking(i)[:level+1], 1))
if_bank_size = get_if_bank_size(blocking_accum_list, layer)
of_bank_size = get_of_bank_size(blocking_accum_list)
fl_bank_size = get_fl_bank_size(blocking_accum_list)
return (if_bank_size, of_bank_size, fl_bank_size)
def get_block_size(point, layer, level):
blocking_accum_list = []
partitioning_accum_list = []
partitioning_reshape = list(zip(*point.loop_partitionings))
partitioning_list = partitioning_reshape[level]
for i in range(le.NUM):
blocking_accum_list.append(reduce(mul, point.loop_blocking(i)[:level+1], 1))
partitioning_accum_list.append(reduce(mul, point.loop_partitioning(i)[:level+1], 1)) #FIXME inclusive mode also duplicates data
if_block_size = get_if_size(blocking_accum_list, partitioning_accum_list, partitioning_list, layer)
of_block_size = get_of_size(blocking_accum_list, partitioning_accum_list, partitioning_list)
fl_block_size = get_fl_size(blocking_accum_list, partitioning_accum_list, partitioning_list)
return (if_block_size, of_block_size, fl_block_size)
def get_block_sizes(num_levels, point, layer):
'''
Get size of ifmap, ofmap, filter
'''
bank_list = []
block_list = []
for level in range(num_levels):
block_list.append(get_block_size(point, layer, level))
bank_list.append(get_bank_size(point, layer, level))
return [bank_list, block_list]
def fit_in_level(cap, blocks, invalid_underutilized, level,memory_partitions):
'''
Check if the current level mem size >= current level loop blocking size
invalid_underutilized is used to exclude mapping points with too low memory utilization (< 50%)
#LMEI can later put the memory utilization threshold as a user defined parameter
'''
if type(cap) is list:
#I/O/W example: [0,0,1] I is stored in memory 0, O is stored in memory 0, W is stored in memory 1
#leave last empty
#memory_partitions = [[0,1, 2],[0,0,1],[0,0,None]] #if 3 level do not contain weights [0, 0, None]
#capacity = [[2,2], [30000,30000], [1000000,1000000]]
for i in range(len(cap)):
indices = [index for index,partition in enumerate(memory_partitions[level]) if partition == i]
size = sum([blocks[j] for j in indices])
if size == 0:
continue
if (size > cap[i]) == True:
return False #it does not fit
check_if_underutilized = 0
#print level, i, invalid_underutilized, memory_partitions[level+1][i], size, cap[i]
if invalid_underutilized:
last_layer = []
for mem in indices:
last_layer.append(memory_partitions[level+1][mem])
if None not in last_layer:
if ((size <= cap[i]) and (2*size <= cap[i])) == True: #if double the size fit then there will be a better to block partition that will utilized all memory,
#print "NO level: ", level,"blocks: ", blocks, "size: ", size, "cap: ", cap, "indices: ", indices, "last_layer", last_layer
check_if_underutilized += 1
else:
test = 1
else:
#print "OK level: ", level,"blocks: ", blocks, "size: ", size, "cap: ", cap, "indices: ", indices, "last_layer", last_layer
test =2
if check_if_underutilized == len(cap):
return False
return True
else:
total_size = sum(blocks)
# for size,contain in zip(blocks, contains):
# if contain:
# total_size += size
# total_capacity = 0
# for size,contain in zip(cap, contains):
# if contain:
# total_capacity += size
# total_size = sum(blocks)
if invalid_underutilized:
return (total_size <= cap) and (2*total_size >= cap)
else:
return (total_size <= cap)
def valid_partition_number(resource, partitioning, level):
max_parallelism = resource.parallelism(level).count
actual_parallelism = reduce(mul, partitioning[level], 1)
return actual_parallelism <= max_parallelism
def valid_partitioning_current_level(resource, point, layer, level, verbose=False):
valid_size = fit_in_level(resource.buffer(level).capacity, \
get_bank_size(point, layer, level), resource.invalid_underutilized, level,resource.memory_partitions)
return valid_size
def valid_mapping_point_current_level(resource, point, layer, level, verbose=False):
if resource.paras[level].count > 1:
valid_size = fit_in_level(resource.buffer(level).capacity,
get_bank_size(point, layer, level), resource.invalid_underutilized,level,resource.memory_partitions)
else :
valid_size = fit_in_level(resource.buffer(level).capacity,
get_block_size(point, layer, level), resource.invalid_underutilized,level,resource.memory_partitions)
partitioning = list(zip(*(point.loop_partitionings)))
valid_para = valid_partition_number(resource, partitioning, level)
if verbose == 3:
print("Level ", level, ": Partitioned block size fit in bank: ", valid_size)
print("Level ", level, ": Partition number is valid: ", valid_para)
return valid_size and valid_para
def valid_partitioning(resource, point, layer, verbose=False):
para_level = resource.para_index
for level in para_level:
if not valid_partitioning_current_level(resource, point, layer, level, verbose):
return False
return True
def valid_blocking_size_current_level(resource, point, layer, level, verbose=False):
if level == resource.buffer_levels()-1:
return True
if type(resource.buffer(level).capacity) is list:
capacity = copy.deepcopy(resource.buffer(level).capacity)
for i in range(len(capacity)):
capacity[i] =capacity[ i]* resource.paras[level].count
return fit_in_level(capacity,get_block_size(point, layer, level), (resource.invalid_underutilized and (level not in resource.para_index)),level,resource.memory_partitions)
else:
return fit_in_level(resource.buffer(level).capacity * resource.paras[level].count,
get_block_size(point, layer, level), (resource.invalid_underutilized and (level not in resource.para_index)),level,resource.memory_partitions)
#get_block_size(point, layer, level), (level > min(resource.para_index)))
def valid_blocking_size(resource, point, layer, verbose=False):
for level in range(resource.buffer_levels()):
if not valid_blocking_size_current_level(resource, point, layer, level, verbose):
return False
return True
def valid_mapping_point(resource, point, layer, verbose=False):
for i in range(resource.buffer_levels()):
if not valid_mapping_point_current_level(resource, point, layer, i, verbose):
return False
return True
def get_total_access_cost(resource, array_cost):
total_access_cost = copy.deepcopy(resource.access_cost)
if not resource.array_access_cost:
return total_access_cost
para_index = [i for i, e in enumerate(resource.paras) if e.access_mode != 0]
addition_levels = len(para_index)
delta = 1
for i in range(addition_levels):
index = para_index[i]
total_access_cost.insert(index+delta, array_cost[i])
delta += 1
return total_access_cost
def get_array_level_cost(resource, point, layer_size, level, next_level_access, verbose=False):
'''
Given next_level_access (above-level memory access)
calculate the current level (paralleled level) inter-PE data access
thus calculate the current level (paralleled level) inter-PE communication energy
i.e. the energy spent on interconnection
Specific to Systolic Array template.
level_access: [[close access for I/O/W],[far access on one dimension for I/O/W],[far access on another dimension]]
close access means data are passing from one PE to its neighbour PE
Far access means data need to jump from one PE to PEs far away from it.
Far jump happens because of dataflow spatial replication (e.g. 2D array -> kinds of 3D array)
'''
# TODO add support for other access_mode # don't get it
# LMEI to distinguish O (partial sum) in buffer_access from A and W
assert resource.paras[level].count and resource.paras[level].access_mode
level_access, level_cost = get_array_access_and_cost(level, resource.paras[level], next_level_access, point)
total_cost = 0
for i in range(len(level_access)):
buffer_access = list(map(mul, level_access[i], layer_size))
total_cost += sum(buffer_access) *level_cost[i]
if verbose >= 3:
print("Level ", level, " array level access: ", level_access)
return total_cost
def get_array_and_curr_level_cost(resource, point, layer, level, verbose=False):
'''
Get the energy from current level of memory access + inter-PE access
'''
# LMEI to distinguish O (partial sum) in buffer_access from A and W
layer_size = get_layer_size(layer)
mac_capacity = resource.mac_capacity
level_access = [get_if_access(level, point, layer, mac_capacity), \
get_of_access(level, point, layer, mac_capacity), \
get_fl_access(level, point, layer, mac_capacity)]
[if_access, of_access, fl_access] = level_access
buffer_level_access = [if_access, 2*of_access-1, fl_access]
total_buffer_access = list(map(mul, buffer_level_access, layer_size))
# level_cost = sum(total_buffer_access) * resource.access_cost[level]
level_cost = 0
for i in range(len(total_buffer_access)):
index = resource.memory_partitions[level][i]
if index is not None:
level_cost += total_buffer_access[i] * resource.access_cost[level][index]
# operand_costs = [access_cost * num_accesses for access_cost,num_accesses in zip(total_buffer_access,resource.access_cost[level]) ]
# level_cost = sum(operand_costs)
if verbose >= 3:
print("Level ", level, " access: ", buffer_level_access)
level_cost += get_array_level_cost(resource, point, layer_size, level-1, level_access, verbose)
return level_cost
def get_level_cost(resource, point, layer, level, verbose=False):
'''
Get the energy from current level of memory access
#LMEI to distinguish O (partial sum) in buffer_access from A and W
'''
layer_size = get_layer_size(layer)
mac_capacity = resource.mac_capacity
level_access = [get_if_access(level, point, layer, mac_capacity), \
2 * get_of_access(level, point, layer, mac_capacity) - 1, \
get_fl_access(level, point, layer, mac_capacity)]
buffer_access = list(map(mul, level_access, layer_size))
# Inputs, weights, and outputs may have different costs
# level_cost = sum(buffer_access) * resource.access_cost[level]
level_cost = 0
for i in range(len(buffer_access)):
index = resource.memory_partitions[level][i]
if index is not None:
level_cost += buffer_access[i] * resource.access_cost[level][index]
# resouce.memory_partitions
# operand_costs = [access_cost * num_accesses for access_cost,num_accesses in zip(buffer_access,resource.access_cost[level]) ]
# level_cost = sum(operand_costs)
if verbose >= 3:
print("Level", level, " access: ", level_access)
return level_cost
def get_total_access(resource, point, layer, verbose=False):
layer_size = get_layer_size(layer)
access_list, array_cost = get_access(point, layer, resource)
if verbose >= 3:
print("access breakdown: ", access_list)
total_level_access = []
for i in range(len(access_list)):
''' List of total access of each buffer at level i'''
if not isinstance(access_list[i][0], list):
buffer_access = list(map(mul, access_list[i], layer_size))
total_level_access.append(sum(buffer_access))
else :
for j in range(len(access_list[i])):
buffer_access = list(map(mul, access_list[i][j], layer_size))
total_level_access.append(sum(buffer_access))
return total_level_access
def get_level_costs(resource, point, layer, verbose=False):
num_levels = resource.buffer_levels()
level_energy = []
for level in range(num_levels):
level_energy.append(get_level_cost(resource, point, layer, level))
para_index = [i for i, e in enumerate(resource.paras) if e.access_mode != 0]
delta = 1
for index in para_index:
array_energy = get_array_and_curr_level_cost(resource, point, layer, index+1) - level_energy[index+delta]
level_energy.insert(index+delta, array_energy)
delta += 1
return level_energy
#FIXME
def get_block_cost(resource, point, layer, verbose=False):
'''
Get the cost of the given mapping point on given resource.
If the point is not feasible on the resource, return inf.
'''
#TODO include static energy
num_levels = resource.buffer_levels()
access_list, array_cost = get_access(point, layer, resource)
layer_size = get_layer_size(layer)
total_access_cost = get_total_access_cost(resource, array_cost)
assert len(total_access_cost) == len(access_list)
block_costs = [0.0, 0.0, 0.0]
for i in range(len(total_access_cost)):
buffer_access = [a*b for a,b in list(zip(access_list[i], layer_size))]
block_cost = [x * total_access_cost[i] for x in buffer_access]
block_costs = list(map(add, block_cost, block_costs))
if verbose:
print('access_list: ', access_list)
bank_size_list, block_size_list = get_block_sizes(num_levels, point, layer)
print('bank_size_list: ', bank_size_list)
print('block_size_list: ', block_size_list)
print('layer_size: ', layer_size)
print('block costs: ', block_costs)
return block_costs
def get_cost(resource, point, layer, verbose=False):
'''
Get the cost of the given mapping point on given resource.
If the point is not feasible on the resource, return inf.
'''
#TODO include static energy
#TODO support other access_mode
num_levels = resource.buffer_levels()
assert len(point.loop_blockings[0]) == num_levels, \
"number of blockings does not match with number of memory " \
"levels: %d" % num_levels
access_list, array_cost = get_access(point, layer, resource)
layer_size = get_layer_size(layer)
total_access_cost = get_total_access_cost(resource, array_cost)
assert len(total_access_cost) == len(access_list)
total_cost = 0.0
for i in range(len(total_access_cost)):
''' List of total access of each buffer at level i'''
if not isinstance(access_list[i][0], list):
buffer_access = list(map(mul, access_list[i], layer_size))
total_cost += (sum(buffer_access) * total_access_cost[i][0])
else :
for j in range(len(access_list[i])):
buffer_access = list(map(mul, access_list[i][j], layer_size))
total_cost += sum(buffer_access) * total_access_cost[i][j]
if verbose:
#print("total_access_cost", total_access_cost)
#print("access_list", access_list)
#print("layer_size",layer_size)
idx_adjust = 0
if len(total_access_cost) > 4:
idx_adjust = 1
layer_access_cost = total_access_cost[:1 + idx_adjust] + total_access_cost[2 + idx_adjust:]
print('16b_Access_Energy_[RegisterFile(s),Buffer,DRAM]_(pJ): \n\tifmap: {}\n\tofmap: {}\n\tfilter: {}'\
.format([item[0] for item in layer_access_cost], [item[1] for item in layer_access_cost], [item[2] for item in layer_access_cost]))
print('PE_Access_Cost_(pJ): \n\tifmap: {}\n\tofmap: {}\n\tfilter: {}'\
.format(total_access_cost[1 + idx_adjust][0], total_access_cost[1 + idx_adjust][1], total_access_cost[1 + idx_adjust][2]))
layer_num_access = access_list[:1 + idx_adjust] + access_list[2 + idx_adjust:]
print('Tiles_Accessed_from_[RegisterFile(s),Buffer,DRAM]_in_Layer: \n\tifmap: {}\n\tofmap: {}\n\tfilter: {}'\
.format([item[0] for item in layer_num_access], [item[1] for item in layer_num_access], [item[2] for item in layer_num_access]))
print("Tiles_Accessed_from_[RegisterFile(s),Buffer,DRAM]_PEs_in_Layer: \n\tifmap: {}\n\tofmap: {}\n\tfilter: {}"\
.format(access_list[1 + idx_adjust][0], access_list[1 + idx_adjust][1], access_list[1 + idx_adjust][2]))