File size: 33,931 Bytes
3d4d40d 07554d1 3d4d40d cf49f13 3d4d40d cf49f13 3d4d40d cf49f13 3d4d40d cf49f13 3d4d40d cf49f13 3d4d40d 07554d1 3d4d40d 07554d1 3d4d40d 07554d1 3d4d40d cf49f13 3d4d40d cf49f13 3d4d40d 07554d1 3d4d40d 07554d1 3d4d40d eff388a 3d4d40d 07554d1 3d4d40d 07554d1 3d4d40d 07554d1 3d4d40d eff388a 3d4d40d 07554d1 3d4d40d 07554d1 3d4d40d 07554d1 3d4d40d 07554d1 3d4d40d 07554d1 3d4d40d 07554d1 3d4d40d 07554d1 eff388a 3d4d40d 07554d1 3d4d40d 07554d1 3d4d40d eff388a 3d4d40d eff388a 07554d1 eff388a 07554d1 eff388a 07554d1 eff388a 07554d1 3d4d40d 07554d1 3d4d40d 07554d1 3d4d40d 07554d1 3d4d40d 07554d1 eff388a 07554d1 3d4d40d 07554d1 3d4d40d 07554d1 3d4d40d 07554d1 3d4d40d 07554d1 3d4d40d 07554d1 3d4d40d 07554d1 3d4d40d 07554d1 3d4d40d 07554d1 3d4d40d 07554d1 3d4d40d 07554d1 3d4d40d 07554d1 3d4d40d 07554d1 3d4d40d 07554d1 3d4d40d eff388a 3d4d40d 07554d1 3d4d40d 07554d1 3d4d40d 07554d1 3d4d40d cf49f13 3d4d40d 07554d1 3d4d40d 07554d1 |
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 |
pattern_size = 6
from collections import Counter, deque
from dataclasses import dataclass
@dataclass(eq=True, frozen=True)
class ScheduledNode:
type: str
chunk: int
stage: int
minibatch: int
start_time: int
completion_time: int
def transform_schedule(schedule, f, b, w, c):
result = []
stage_order = []
local_prev = {}
stages = len(schedule)
for sid, stage in enumerate(schedule):
counter = Counter()
order = []
for p in stage:
if not p.strip():
continue
mb = counter.get(p, 0)
if order:
local_prev[(sid, p, mb)] = order[-1]
order.append((p, mb))
counter.update(p)
stage_order.append(order)
nmb = max(counter.values())
time_map = {}
cost = {
'F': f,
'B': b,
'W': w,
'f': f,
'b': b,
'w': w,
}
def get_time(stage, type, mb):
if (stage, type, mb) in time_map:
return time_map.get((stage, type, mb))
time = 0
if (stage, type, mb) in local_prev:
time = get_time(stage, *local_prev[(stage, type, mb)])
if type in "FB" and stage > 0:
time = max(time, get_time(stage - 1, type, mb) + c)
if type in "fb" and stage + 1< len(schedule):
time = max(time, get_time(stage + 1, type, mb) + c)
# print(f'{stage} {type}:{mb}', time + cost[type])
time_map[(stage, type, mb)] = time + cost[type]
return time_map[(stage, type, mb)]
r = 0
for sid, stage in enumerate(schedule):
r = max(get_time(sid, 'W', nmb - 1) - get_time(sid, 'F', 0) + f, r)
r = max(get_time(sid, 'w', nmb - 1) - get_time(sid, 'F', 0) + f, r)
for sid, stage in enumerate(stage_order):
result_stage = []
for p, mb in stage:
result_stage.append(ScheduledNode(
p.upper(),
p in "fBW",
sid,
mb,
get_time(sid, p, mb) - cost[p],
get_time(sid, p, mb)
)
)
result.append(result_stage)
return result
def evaluate_schedule(schedule, f, b, w, c):
stage_order = []
local_prev = {}
stages = len(schedule)
for sid, stage in enumerate(schedule):
counter = Counter()
order = []
for p in stage:
if not p.strip():
continue
mb = counter.get(p, 0)
if order:
local_prev[(sid, p, mb)] = order[-1]
order.append((p, mb))
counter.update(p)
stage_order.append(order)
nmb = max(counter.values())
time_map = {}
cost = {
'F': f,
'B': b,
'W': w,
'f': f,
'b': b,
'w': w,
}
def get_time(stage, type, mb):
if (stage, type, mb) in time_map:
return time_map.get((stage, type, mb))
time = 0
if (stage, type, mb) in local_prev:
time = get_time(stage, *local_prev[(stage, type, mb)])
if type in "FB" and stage > 0:
time = max(time, get_time(stage - 1, type, mb) + c)
if type in "fb" and stage + 1< len(schedule):
time = max(time, get_time(stage + 1, type, mb) + c)
# print(f'{stage} {type}:{mb}', time + cost[type])
time_map[(stage, type, mb)] = time + cost[type]
return time_map[(stage, type, mb)]
r = 0
for sid, stage in enumerate(schedule):
r = max(get_time(sid, 'W', nmb - 1) - get_time(sid, 'F', 0) + f, r)
r = max(get_time(sid, 'w', nmb - 1) - get_time(sid, 'F', 0) + f, r)
return r
debug = False
def print_schedules(schedules, msg = None, force=False):
if not debug and not force:
return
if msg is not None:
print(msg)
for seq in schedules:
_str = ""
for v in seq:
_str += v
print(_str)
def get_building_block_str(pos):
pattern = [" "] * pattern_size
notations = "FfBbWw"
for i, v in enumerate(pos):
if v < 0:
continue
pattern[v] = notations[i]
_str = ""
for v in pattern:
_str += v
return _str
def get_peak_mem(schedules, return_all=False):
max_peak = 0
all_peak = []
for schedule_ in schedules:
peak, mem = 0, 0
for v in schedule_:
if v in "Ff":
mem += 1
elif v in "Ww":
mem -= 1
peak = max(peak, mem)
all_peak.append(peak)
max_peak = max(max_peak, peak)
if return_all:
return all_peak
return max_peak
def calc_bubble(schedules):
stage_bubbles = []
for i in range(len(schedules)):
max_len = 0
count = 0
for j in range(len(schedules[i])):
if schedules[i][j] != ' ':
max_len = j + 1
count += 1
stage_bubbles.append(max_len - count - i)
return stage_bubbles
def init_repeated_schedule(p, m, building_block):
repeated = []
_len = 4 * p + m + 1
for i in range(p):
str_i = get_building_block_str(building_block[i]) * _len
repeated_i = []
for v in str_i:
repeated_i.append(v)
repeated.append(repeated_i)
return repeated
def clear_invalid(repeated, stage, pos, offset=-1):
while 0 <= pos < len(repeated[stage]):
repeated[stage][pos] = ' '
pos += offset * pattern_size
return repeated
def clear_invalid_index(repeated, m):
p = len(repeated)
index = pattern_size
for identifier in "FfBb":
if identifier in "FB":
_iter = range(p)
else:
_iter = range(p - 1, -1, -1)
for i in _iter:
for j in range(pattern_size):
if repeated[i][index] == identifier:
clear_invalid(repeated, i, index - pattern_size, offset=-1)
clear_invalid(repeated, i, index + pattern_size * m, offset=1)
index += 1
if identifier in "Bb":
w_identifier = {'B': 'W', 'b': 'w'}[identifier]
for k in range(pattern_size):
if repeated[i][index + k] == w_identifier:
clear_invalid(repeated, i, index + k - pattern_size, offset=-1)
clear_invalid(repeated, i, index + k + pattern_size * m, offset=1)
break
break
index += 1
return repeated
def process_warmup_without_increasing_peak_mem(schedules, m):
"""
FFFFFFFFFF fBWfBWfBWfBWfBW b
FFFFFFFFF f fBWfBWfBWfBWFBWb
FFFFFFFF f f fBWfBWfBWFBW b
FFFFFFF f f f fBWfBWFBW Bb
FFFFFF f f f f fBWFBWFBWb
FFFFFfFf f f f BWFBW b
FFFfFfFfFf f BW Bb
FfFfFfFfFfF BWb
We reorganize the warmup phase in the following way (i -> pipeline stage from 0):
1. Before the first B, we set #f = min(i+1, peak_mem//2), #F = peak_mem - #f
2. Before the first b, #f = peak_mem//2
3. The offset between the first B is 1
4. Before the first b, we use the pattern of (BWf)*j + (BWF)*k,
where j = max(0, peak_mem//2 - (i+1)), k = max(0, #W - j - 1)
"""
# process warmup phase (before the first b)
p = len(schedules)
peak_mem = get_peak_mem(schedules)
peak_mem = min(peak_mem, 2 * p)
cnt_f, cnt_ff = [], []
for i in range(p):
cc_ff = min(i + 1, peak_mem // 2)
cc_ff = min(cc_ff, m)
cc_f = min(peak_mem - cc_ff, m)
cnt_f.append(cc_f)
cnt_ff.append(cc_ff)
distance_b2bb = 0
for j in range(len(schedules[p - 1])):
if schedules[p - 1][j] == 'B':
for k in range(j, len(schedules[p - 1])):
if schedules[p - 1][k] == 'b':
distance_b2bb = k - j
break
break
for i in range(p):
c_f, c_ff, c_b, c_w = 0, 0, 0, 0
for j in range(len(schedules[i])):
char = schedules[i][j]
if char == 'F':
c_f += 1
elif char == 'f':
c_ff += 1
elif char == 'B':
c_b += 1
elif char == 'W':
c_w += 1
elif char == 'b':
break
# This logic can be removed because it is too complicated and should not impact the optimal solution
bj = j
while j < len(schedules[i]):
char = schedules[i][j]
if char == 'f' and c_ff < cnt_ff[p - 1]:
schedules[i][j] = ' '
c_ff += 1
if char == 'B' and c_b < c_ff:
if c_b < (2 * (p - i) + distance_b2bb) // 3 or c_b < cnt_ff[p - 1] - cnt_ff[i]:
# there is empty space, or the number of B is not enough to cover extra f
schedules[i][j] = ' '
c_b += 1
if char == 'W' and c_w < c_b:
if c_w < (2 * (p - i) + distance_b2bb - 1) // 3 or c_w < cnt_ff[p - 1] - cnt_ff[i]:
# there is empty space, or the number of W is not enough to cover extra f
schedules[i][j] = ' '
c_w += 1
j += 1
j = bj
while j < len(schedules[i]):
if schedules[i][j] == 'F':
if c_f < c_ff or c_f < cnt_f[i] or c_f - cnt_f[i] + c_ff - cnt_ff[i] < c_w - 1:
# put enough F, or there are some unused BW
schedules[i][j] = ' '
c_f += 1
j += 1
break
else:
assert char == ' '
schedules[i][j] = ' '
# assert c_f >= cnt_f[i] and c_ff >= cnt_ff[i]
# assert c_w >= cnt_ff[p - 1] - cnt_ff[i] and c_b >= cnt_ff[p - 1] - cnt_ff[i]
j = i
u_f, u_ff, u_b, u_w = 0, 0, 0, 0
for _ in range(2 * (p - 1 - i)):
if u_f < cnt_f[i] and u_f < c_f:
schedules[i][j] = 'F'
u_f += 1
j += 1
for _ in range(i + 1):
if u_f < cnt_f[i] and u_f < c_f:
schedules[i][j] = 'F'
u_f += 1
j += 1
if u_ff < cnt_ff[i] and u_ff < c_ff:
schedules[i][j] = 'f'
u_ff += 1
j += 1
while u_f < c_f or u_ff < c_ff or u_b < c_b or u_w < c_w:
if u_b < c_b:
schedules[i][j] = 'B'
u_b += 1
j += 1
if u_w < c_w:
schedules[i][j] = 'W'
u_w += 1
j += 1
if u_ff < c_ff:
assert u_ff < u_f
schedules[i][j] = 'f'
u_ff += 1
elif u_f < c_f:
schedules[i][j] = 'F'
u_f += 1
j += 1
return schedules
def squeeze_without_change_order(schedules, m):
p = len(schedules)
squeezed = [[' '] * len(schedules[_]) for _ in range(p)]
max_len = check_and_get_schedule_len(schedules)
identifier_cnt = [{_id: 0 for _id in "FfBbWw"} for _ in range(p)]
identifier_index = [{_id: -1 for _id in "FfBbWw"} for _ in range(p * m)]
stage_index = [0 for _ in range(p)]
for j in range(max_len):
for _dir in range(2):
if _dir == 0:
_iter = range(p)
else:
_iter = range(p - 1, -1, -1)
for i in _iter:
identifier = schedules[i][j]
if identifier == ' ':
continue
if _dir == 0 and identifier in "fbw":
continue
if _dir == 1 and identifier in "FBW":
continue
_cnt = identifier_cnt[i][identifier]
assert _cnt < m, "{} - {}, {}".format(i, identifier, _cnt)
if identifier in "Ww" or (i == 0 and identifier in "FB") or (i == p - 1 and identifier in "fb"):
if i == 0 and identifier == 'B':
assert identifier_index[_cnt * p + i]['f'] >= 0
if i == p - 1 and identifier == 'f':
assert identifier_index[_cnt * p + i]['F'] >= 0
if i == p - 1 and identifier == 'b':
assert identifier_index[_cnt * p + i]['B'] >= 0
index = stage_index[i]
elif identifier in "FB":
assert identifier_index[_cnt * p + i - 1][identifier] >= 0, "{} {} {}".format(i, identifier,_cnt)
index = max(identifier_index[_cnt * p + i - 1][identifier] + 1, stage_index[i])
elif identifier in "fb":
assert identifier_index[_cnt * p + i + 1][identifier] >= 0, "{} {} {}".format(i, identifier,_cnt)
index = max(identifier_index[_cnt * p + i + 1][identifier] + 1, stage_index[i])
else:
raise
squeezed[i][index] = identifier
identifier_cnt[i][identifier] = _cnt + 1
identifier_index[_cnt * p + i][identifier] = index
stage_index[i] = index + 1
new_len = max(stage_index)
for i in range(p):
squeezed[i] = squeezed[i][:new_len]
return squeezed
def process_cooldown(schedules, m):
"""
fBW bwbwbwbw
fBWBW bwbwbwbw
fBWBWBW bwbwbwbw
fBWBWBWBW bwbwbwbw
f BWBWBWBbWbwbwbww
f BWBWBbBbWbWbwwww
f BWBbBbBbWbWWwwww
f BbBbBbBbWWWWwwww
We reorganize the cooldown phase in the following way (i -> pipeline stage from 0):
1. After the last f, we set #b = (peak_mem+1)//2, and #B = min(i+1, peak_mem - #b)
2. After the last f, we make all the dependencies as tight as possible
"""
p = len(schedules)
peak_mem = get_peak_mem(schedules)
assert peak_mem <= 2 * p, peak_mem
max_bb = (peak_mem + 1) // 2
max_bb = min(max_bb, m)
max_b = min(peak_mem - max_bb, m)
# 1: reorganize B/b and remove W/w in cooldown phase
starting_index = -1
for i in range(p):
c_b, c_bb, c_w, c_ww = 0, 0, 0, 0
last_ff_index = -1
# collect B/b which can be reordered
for j in range(len(schedules[i]) - 1, -1, -1):
char = schedules[i][j]
if char == 'f' and last_ff_index == -1:
last_ff_index = j
if char == 'B' and c_b < i + 1 and c_b < max_b:
schedules[i][j] = ' '
c_b += 1
if char == 'b' and c_bb < max_bb:
schedules[i][j] = ' '
c_bb += 1
# clear W in the tail (#W + #w >= peak_mem & #W >= #B & #w >= #b)
for j in range(len(schedules[i]) - 1, -1, -1):
char = schedules[i][j]
if c_w >= c_b and c_ww >= c_bb and c_w + c_ww >= peak_mem:
break
if char == 'W':
schedules[i][j] = ' '
c_w += 1
if char == 'w':
schedules[i][j] = ' '
c_ww += 1
if i == 0:
starting_index = last_ff_index
# reorganize B/b in the tail
for k in range(c_bb):
index = starting_index - i + 2 * p - 2 * k
assert schedules[i][index] == ' ', "{} {} {}".format(schedules[i][index], k, i)
schedules[i][index] = 'b'
for k in range(c_b):
index = starting_index + 1 + i - 2 * k
# assert schedules[i][index] == ' ', schedules[i][index]
schedules[i][index] = 'B'
# 2: add W back in cooldown phase
max_len = 0
for i in range(p):
c_w, c_ww = 0, 0
last_w_index = -1
for j in range(len(schedules[i]) - 1, -1, -1):
if schedules[i][j] in "Ww":
last_w_index = j
break
for j in range(len(schedules[i])):
char = schedules[i][j]
if char == 'B':
c_w += 1
elif char == 'b':
c_ww += 1
elif char == 'W':
c_w -= 1
elif char == 'w':
c_ww -= 1
if char == ' ' and j > last_w_index:
if c_w > 0:
schedules[i][j] = 'W'
c_w -= 1
elif c_ww > 0:
schedules[i][j] = 'w'
c_ww -= 1
for _ in range(c_w):
schedules[i].append('W')
for _ in range(c_ww):
schedules[i].append('w')
max_len = max(max_len, len(schedules[i]))
for i in range(p):
for _ in range(len(schedules[i]), max_len):
schedules[i].append(' ')
schedules = squeeze_without_change_order(schedules, m)
return schedules
def check_and_get_schedule_len(schedules):
max_len = 0
for seq in schedules:
assert max_len == 0 or max_len == len(seq)
max_len = max(max_len, len(seq))
return max_len
def release_w_in_warmup_if_under_memory(schedules, peak_mem = None):
"""
FF fBWfBW bwbw -> FF fBfBWW bwbw
FF f fBW BW bwbw -> FF f fBWBW bwbw
FF f f BW BbWbww -> FF f f BWBbWbww
FfFf BbWBbwWw -> FfFf BbBbWwWw
When the number of micro-batches is too small (than mem), the warmup phase is not optimal. We simply remove some
preceding W to fully utilize the memory to reduce unnecessary bubbles.
"""
p = len(schedules)
max_len = check_and_get_schedule_len(schedules)
all_peak_mem = get_peak_mem(schedules, return_all=True)
peak_mem = peak_mem or max(all_peak_mem)
min_peak = min(all_peak_mem)
for i in range(p):
cnt = 0
padding = [" "] * (peak_mem - min_peak)
for j in range(max_len):
if all_peak_mem[i] + cnt >= peak_mem:
break
if schedules[i][j] in "Ww":
padding[cnt] = schedules[i][j]
schedules[i][j] = ' '
cnt += 1
schedules[i].extend(padding)
# max_len += peak_mem - min_peak
return schedules
def reorder_greedily_without_increasing_peak_mem(schedules, m, starting_index = None, ending_index = None, peak_mem = None):
"""
We iterate all the cells from left to right. If a vacant cell (which means a bubble) is encountered, we try to
find a computation pass to fill this bubble. We iterate all the following computation passes in the same device,
and check whether it is possible to move if we keep all other passes unchanged. If the check succeeds, we move it
to the vacant cell, and the bubble is filled.
"""
p = len(schedules)
if starting_index is not None:
assert isinstance(starting_index, list) and len(starting_index) == p
if ending_index is not None:
assert isinstance(ending_index, list) and len(ending_index) == p
peak_mem = peak_mem or get_peak_mem(schedules)
max_len = check_and_get_schedule_len(schedules)
starting_index = starting_index or [0] * p
ending_index = ending_index or [max_len] * p
last_index = [{_id: -1 for _id in "FfBbWw"} for _ in range(p)]
for i in range(p):
for j in range(max_len):
identifier = schedules[i][j]
if identifier == ' ':
continue
last_index[i][identifier] = j
stage_mem = [0] * p
def update_mem(stage_i, pass_c):
if pass_c in "Ff":
stage_mem[stage_i] += 1
elif pass_c in "Ww":
stage_mem[stage_i] -= 1
identifier_cnt = [{_id: 0 for _id in "FfBbWw"} for _ in range(p)]
identifier_index = [{_id: -1 for _id in "FfBbWw"} for _ in range(p * m)]
for j in range(0, max_len):
for i in range(p):
identifier = schedules[i][j]
if identifier in "FfBbWw":
_cnt = identifier_cnt[i][identifier]
identifier_cnt[i][identifier] = _cnt + 1
identifier_index[_cnt * p + i][identifier] = j
update_mem(i, identifier)
continue
assert identifier == ' '
if j < starting_index[i] or j >= ending_index[i]:
continue
available = set()
for c in "FfBbWw":
if last_index[i][c] > j:
available.add(c)
mem_delta, peak_delta = 0, 0
for k in range(j + 1, ending_index[i]):
if len(available) == 0:
break
identifier = schedules[i][k]
if identifier in "Ff":
mem_delta += 1
elif identifier in "Ww":
mem_delta -= 1
prev_peak = peak_delta
peak_delta = max(peak_delta, mem_delta)
if identifier == ' ' or identifier not in available:
continue
available.remove(identifier)
if identifier in "Ff" and stage_mem[i] + prev_peak >= peak_mem:
# will increase peak memory
continue
can_move = True
_cnt = identifier_cnt[i][identifier]
if identifier in "FB":
if i > 0:
_index = identifier_index[_cnt * p + i - 1][identifier]
if _index <= -1 or _index >= j:
can_move = False
elif identifier == 'B':
if identifier_cnt[i]['f'] <= _cnt:
can_move = False
elif identifier in "fb":
if i + 1 < p:
_index = identifier_index[_cnt * p + i + 1][identifier]
if _index <= -1 or _index >= j:
can_move = False
else:
_pi = 'F' if identifier == 'f' else 'B'
if identifier_cnt[i][_pi] <= _cnt:
can_move = False
elif identifier in "Ww":
_bi = 'B' if identifier == 'W' else 'b'
if identifier_cnt[i][_bi] <= _cnt:
can_move = False
else:
assert False
if not can_move:
continue
# if i == 0:
# print(peak_mem, stage_mem[i], identifier, mem_delta)
schedules[i][j] = identifier
schedules[i][k] = ' '
identifier_cnt[i][identifier] = _cnt + 1
identifier_index[_cnt * p + i][identifier] = j
update_mem(i, identifier)
break
return schedules
def check_correctness(schedules, m, raise_exception=False):
p = len(schedules)
c_index = [{_id: -1 for _id in "FfBbWw"} for _ in range(p * m)]
for i in range(p):
c_cnt = {_id: 0 for _id in "FfBbWw"}
for j in range(len(schedules[i])):
c = schedules[i][j]
if c in "FfBbWw":
_cnt = c_cnt[c]
assert _cnt < m
c_index[_cnt * p + i][c] = j
c_cnt[c] = _cnt + 1
for c in "FfBbWw":
if c_cnt[c] != m:
assert not raise_exception
return False
for i in range(p):
for j in range(m):
for c in "FfBbWw":
if c_index[j * p + i][c] == -1:
assert not raise_exception
return False
if c_index[j * p + i]['B'] >= c_index[j * p + i]['W']:
assert not raise_exception, f"{i} {j} {c}"
return False
if c_index[j * p + i]['b'] >= c_index[j * p + i]['w']:
assert not raise_exception
return False
if i == 0:
if c_index[j * p + i]['f'] >= c_index[j * p + i]['B']:
assert not raise_exception
return False
elif i == p - 1:
if c_index[j * p + i]['F'] >= c_index[j * p + i]['f']:
assert not raise_exception
return False
if c_index[j * p + i]['B'] >= c_index[j * p + i]['b']:
assert not raise_exception
return False
else:
if c_index[j * p + i - 1]['F'] >= c_index[j * p + i]['F']:
assert not raise_exception
return False
if c_index[j * p + i - 1]['B'] >= c_index[j * p + i]['B']:
assert not raise_exception
return False
if c_index[j * p + i + 1]['f'] >= c_index[j * p + i]['f']:
assert not raise_exception
return False
if c_index[j * p + i + 1]['b'] >= c_index[j * p + i]['b']:
assert not raise_exception
return False
return True
def relabel_w(schedules, m):
p = len(schedules)
c_cnt = [{_id: 0 for _id in "FfBbWw"} for _ in range(p)]
for i in range(p):
for j in range(len(schedules[i])):
if schedules[i][j] == ' ':
continue
c_cnt[i][schedules[i][j]] += 1
for c in "FfBbWw":
assert c_cnt[i][c] == m, f"{i}, {c}, {c_cnt[i][c]}"
for i in range(p):
w_queue = deque(maxlen=2 * m)
for j in range(len(schedules[i])):
identifier = schedules[i][j]
if identifier == 'B':
w_queue.append('W')
elif identifier == 'b':
w_queue.append('w')
elif identifier in "Ww":
assert len(w_queue) > 0, f"{i} {j}"
schedules[i][j] = w_queue.popleft()
assert len(w_queue) == 0
return schedules
def remove_redundancy(schedules, m):
for sid in range(len(schedules)):
cnt = {_id: 0 for _id in "FfBbWw"}
for i in range(len(schedules[sid])):
if schedules[sid][i] == ' ':
continue
if cnt[schedules[sid][i]] >= m:
schedules[sid][i] = ' '
else:
cnt[schedules[sid][i]] += 1
return schedules
def schedule_by_building_block(p, m, building_block, max_mem, keep_stable_phase=False):
# Apply the framework of repeating-squeezing-reordering
# 1. repeating
redundant_m = max(m, 2 * p) # we add some redundant micro-batches to avoid unexpected bugs
schedules = init_repeated_schedule(p, redundant_m, building_block)
schedules = clear_invalid_index(schedules, redundant_m)
init_peak_mem = get_peak_mem(schedules)
if (m == redundant_m and init_peak_mem > max_mem) or init_peak_mem > 2 * p:
return None, init_peak_mem, [6 * m] * p
print_schedules(schedules, "after repeating")
# 2. squeezing
schedules = squeeze_without_change_order(schedules, redundant_m)
print_schedules(schedules, "after squeezing")
# 3. reordering
# 3.a. reorder warm-up
schedules = process_warmup_without_increasing_peak_mem(schedules, redundant_m) # must work with m >= 2p
schedules = squeeze_without_change_order(schedules, redundant_m)
if keep_stable_phase:
ending_index = [0] * p # before second b
for i in range(p):
bb_cnt = 0
for j in range(len(schedules[i])):
if schedules[i][j] == 'b':
bb_cnt += 1
if bb_cnt >= 2:
ending_index[i] = j
break
schedules = reorder_greedily_without_increasing_peak_mem(schedules, redundant_m, ending_index=ending_index)
peak_mem = get_peak_mem(schedules)
if debug:
assert peak_mem <= init_peak_mem, f"{init_peak_mem}, {peak_mem}"
if peak_mem > init_peak_mem:
return None, init_peak_mem, [6 * m] * p
if m < redundant_m:
# 4. remove redundancy
schedules = remove_redundancy(schedules, m)
if m <= p and 2 * m <= max_mem:
schedules = release_w_in_warmup_if_under_memory(schedules, peak_mem=min(2 * p, peak_mem))
schedules = squeeze_without_change_order(schedules, m)
print_schedules(schedules, "after removing redundancy")
init_peak_mem = peak_mem = get_peak_mem(schedules)
if peak_mem > max_mem:
return None, peak_mem, [6 * m] * p
# 3.b. reorder cool-down
schedules = process_cooldown(schedules, m)
if keep_stable_phase:
starting_index = [0] * p
for i in range(p):
for j in range(len(schedules[i])):
if schedules[i][j] == 'F':
starting_index[i] = j
schedules = reorder_greedily_without_increasing_peak_mem(schedules, m, starting_index=starting_index)
if not keep_stable_phase:
reorder_greedily_without_increasing_peak_mem(schedules, m)
schedules = relabel_w(schedules, m)
print_schedules(schedules, "after reordering")
peak_mem = get_peak_mem(schedules)
if debug:
assert peak_mem <= init_peak_mem, f"{init_peak_mem}, {peak_mem}"
if peak_mem > init_peak_mem:
return None, init_peak_mem, [6 * m] * p
# return
if not check_correctness(schedules, m, raise_exception=debug):
return None, peak_mem, [6 * m] * p
stage_bubbles = calc_bubble(schedules)
if debug:
print(peak_mem, stage_bubbles)
print("-" * 100)
return schedules, peak_mem, stage_bubbles
def fill_w_in_building_block(pattern):
f, ff, b, bb, w, ww = 0, 1, 2, 3, 4, 5
vis = [False] * pattern_size
for v in pattern:
if v >= 0:
vis[v] = True
assert pattern[b] >= 0 and pattern[bb] >= 0
for v, vw in [(b, w), (bb, ww)]:
for j in range(pattern_size):
pos = (pattern[v] + j) % pattern_size
if not vis[pos]:
pattern[vw] = pos
vis[pos] = True
break
return pattern
def get_building_block(pattern_0, offset_0, offset_1, len_0, p):
# see Appendix A in the paper
build_block = [pattern_0]
for i in range(p - 1):
last_pattern = build_block[i]
new_pattern = [-1] * pattern_size
vis = [False] * pattern_size
if i < len_0:
offset = offset_0
else:
offset = offset_1
for v, v_o in enumerate(offset):
pos = (last_pattern[v] + v_o + pattern_size) % pattern_size
assert 0 <= pos < pattern_size
if vis[pos]:
return None
vis[pos] = True
new_pattern[v] = pos
new_pattern = fill_w_in_building_block(new_pattern)
build_block.append(new_pattern)
return build_block
def schedule(p, m, cost, max_mem):
f, ff, b, bb, w, ww = 0, 1, 2, 3, 4, 5
available_starting_patterns = []
# iterate available patterns for the first row/device of a building block
for ff_i in range(1, pattern_size):
for b_i in range(1, pattern_size):
for bb_i in range(1, pattern_size):
if ff_i == b_i or ff_i == bb_i or b_i == bb_i:
continue
pattern = [0, ff_i, b_i, bb_i, -1, -1]
pattern = fill_w_in_building_block(pattern)
available_starting_patterns.append(pattern)
# available uniform offsets, see Section 3.1 in the paper.
available_offsets = [
# [\delta_F^0, \delta_F^1, \delta_B^1, \delta_B^0]
[1, -1, 1, -1],
[2, -1, 2, -1],
[3, -1, 3, -1],
[4, -1, 4, -1],
[5, -1, 5, -1]
]
# available_starting_patterns = available_starting_patterns[:1]
best_schedule = None
best_bubble = None
peak_mem2min_bubble = {}
for pattern_0 in available_starting_patterns:
for i_0 in range(len(available_offsets)):
for i_1 in range(i_0 + 1):
for len_0 in range(1, p):
offset_0 = available_offsets[i_0]
offset_1 = available_offsets[i_1]
build_block = get_building_block(pattern_0, offset_0, offset_1, len_0, p)
if build_block is None:
continue
s, peak_mem, bubbles = schedule_by_building_block(p, m, build_block, min(2 * p, max_mem))
if peak_mem > 2 * p or peak_mem > max_mem:
break
if s is None:
continue
max_bubble = evaluate_schedule(s, *cost)
if best_schedule is None or max_bubble < best_bubble:
best_schedule, best_bubble = s, max_bubble
max_bubble = max(bubbles)
min_bubble = min(peak_mem2min_bubble.get(peak_mem, max_bubble), max_bubble)
peak_mem2min_bubble[peak_mem] = min_bubble
mem2bubble = {}
for peak_mem in sorted(peak_mem2min_bubble.keys()):
bubble = peak_mem2min_bubble[peak_mem]
mem2bubble[peak_mem] = bubble
# expected_bubble = max(0, 6 * p - 1 - 3 * peak_mem)
expected_bubble = 3 * p - 1 - 3 * peak_mem + max(3 * p, p - 1 + (1+(peak_mem+1)//2)*2)
# expected_bubble = 6 * p - 1 - 3 * peak_mem
print(peak_mem, bubble, expected_bubble, "|", bubble - expected_bubble)
print(mem2bubble)
res = transform_schedule(best_schedule, *cost)
return res
|