forked from UCSBarchlab/OpenTPU
-
Notifications
You must be signed in to change notification settings - Fork 0
/
matrix.py
680 lines (567 loc) · 27.7 KB
/
matrix.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
from functools import reduce
from pyrtl import *
from pyrtl import rtllib
from pyrtl.rtllib import multipliers
#set_debug_mode()
globali = 0 # To give unique numbers to each MAC
def MAC(data_width, matrix_size, data_in, acc_in, switchw, weight_in, weight_we, weight_tag):
'''Multiply-Accumulate unit with programmable weight.
Inputs
data_in: The 8-bit activation value to multiply by weight.
acc_in: 32-bit value to accumulate with product.
switchw: Control signal; when 1, switch to using the other weight buffer.
weight_in: 8-bit value to write to the secondary weight buffer.
weight_we: When high, weights are being written; if tag matches, store weights.
Otherwise, pass them through with incremented tag.
weight_tag: If equal to 255, weight is for this row; store it.
Outputs
out: Result of the multiply accumulate; moves one cell down to become acc_in.
data_reg: data_in, stored in a pipeline register for cell to the right.
switch_reg: switchw, stored in a pipeline register for cell to the right.
weight_reg: weight_in, stored in a pipeline register for cell below.
weight_we_reg: weight_we, stored in a pipeline register for cell below.
weight_tag_reg: weight_tag, incremented and stored in a pipeline register for cell below
'''
global globali
# Check lengths of inupts
if len(weight_in) != len(data_in) != data_width:
raise Exception("Expected 8-bit value in MAC.")
if len(switchw) != len(weight_we) != 1:
raise Exception("Expected 1-bit control signal in MAC.")
# Should never switch weight buffers while they're changing
#rtl_assert(~(weight_we & switchw), Exception("Cannot switch weight values when they're being loaded!"))
# Use two buffers to store weight and next weight to use.
wbuf1, wbuf2 = Register(len(weight_in)), Register(len(weight_in))
# Track which buffer is current and which is secondary.
current_buffer_reg = Register(1)
with conditional_assignment:
with switchw:
current_buffer_reg.next |= ~current_buffer_reg
current_buffer = current_buffer_reg ^ switchw # reflects change in same cycle switchw goes high
# When told, store a new weight value in the secondary buffer
with conditional_assignment:
with weight_we & (weight_tag == Const(matrix_size-1)):
with current_buffer == 0: # If 0, wbuf1 is current; if 1, wbuf2 is current
wbuf2.next |= weight_in
with otherwise:
wbuf1.next |= weight_in
# Do the actual MAC operation
weight = select(current_buffer, wbuf2, wbuf1)
#probe(weight, "weight" + str(globali))
globali += 1
#inlen = max(len(weight), len(data_in))
#product = weight.sign_extended(inlen*2) * data_in.sign_extended(inlen*2)
#product = product[:inlen*2]
product = helperfuncs.mult_signed(weight, data_in)[:32]
#plen = len(weight) + len(data_in)
#product = weight.sign_extended(plen) * data_in.sign_extended(plen)
#product = product[:plen]
l = max(len(product), len(acc_in)) + 1
out = (product.sign_extended(l) + acc_in.sign_extended(l))[:-1]
#product = rtllib.multipliers.signed_tree_multiplier(weight, data_in)
#l = max(len(product), len(acc_in))
#out = product.sign_extended(l) + acc_in.sign_extended(l)
if len(out) > 32:
out = out[:32]
# For values that need to be forward to the right/bottom, store in pipeline registers
data_reg = Register(len(data_in)) # pipeline register, holds data value for cell to the right
data_reg.next <<= data_in
switch_reg = Register(1) # pipeline register, holds switch control signal for cell to the right
switch_reg.next <<= switchw
acc_reg = Register(len(out)) # output value for MAC below
acc_reg.next <<= out
weight_reg = Register(len(weight_in)) # pipeline register, holds weight input for cell below
weight_reg.next <<= weight_in
weight_we_reg = Register(1) # pipeline register, holds weight write enable signal for cell below
weight_we_reg.next <<= weight_we
weight_tag_reg = Register(len(weight_tag)) # pipeline register, holds weight tag for cell below
weight_tag_reg.next <<= (weight_tag + 1)[:len(weight_tag)] # increment tag as it passes down rows
return acc_reg, data_reg, switch_reg, weight_reg, weight_we_reg, weight_tag_reg
def MMArray(data_width, matrix_size, data_in, new_weights, weights_in, weights_we):
'''
data_in: 256-array of 8-bit activation values from systolic_setup buffer
new_weights: 256-array of 1-bit control values indicating that new weight should be used
weights_in: output of weight FIFO (8 x matsize x matsize bit wire)
weights_we: 1-bit signal to begin writing new weights into the matrix
'''
# For signals going to the right, store in a var; for signals going down, keep a list
# For signals going down, keep a copy of inputs to top row to connect to later
weights_in_top = [ WireVector(data_width) for i in range(matrix_size) ] # input weights to top row
weights_in_last = [x for x in weights_in_top]
weights_enable_top = [ WireVector(1) for i in range(matrix_size) ] # weight we to top row
weights_enable = [x for x in weights_enable_top]
weights_tag_top = [ WireVector(data_width) for i in range(matrix_size) ] # weight row tag to top row
weights_tag = [x for x in weights_tag_top]
data_out = [Const(0) for i in range(matrix_size)] # will hold output from final row
# Build array of MACs
for i in range(matrix_size): # for each row
din = data_in[i]
switchin = new_weights[i]
#probe(switchin, "switch" + str(i))
for j in range(matrix_size): # for each column
acc_out, din, switchin, newweight, newwe, newtag = MAC(data_width, matrix_size, din, data_out[j], switchin, weights_in_last[j], weights_enable[j], weights_tag[j])
#probe(data_out[j], "MACacc{}_{}".format(i, j))
#probe(acc_out, "MACout{}_{}".format(i, j))
#probe(din, "MACdata{}_{}".format(i, j))
weights_in_last[j] = newweight
weights_enable[j] = newwe
weights_tag[j] = newtag
data_out[j] = acc_out
# Handle weight reprogramming
programming = Register(1) # when 1, we're in the process of loading new weights
size = 1
while pow(2, size) < matrix_size:
size = size + 1
progstep = Register(size) # 256 steps to program new weights (also serves as tag input)
with conditional_assignment:
with weights_we & (~programming):
programming.next |= 1
with programming & (progstep == matrix_size-1):
programming.next |= 0
with otherwise:
pass
with programming: # while programming, increment state each cycle
progstep.next |= progstep + 1
with otherwise:
progstep.next |= Const(0)
# Divide FIFO output into rows (each row datawidth x matrixsize bits)
rowsize = data_width * matrix_size
weight_arr = [ weights_in[i*rowsize : i*rowsize + rowsize] for i in range(matrix_size) ]
# Mux the wire for this row
current_weights_wire = mux(progstep, *weight_arr)
# Split the wire into an array of 8-bit values
current_weights = [ current_weights_wire[i*data_width:i*data_width+data_width] for i in reversed(range(matrix_size)) ]
# Connect top row to input and control signals
for i, win in enumerate(weights_in_top):
# From the current 256-array, select the byte for this column
win <<= current_weights[i]
for we in weights_enable_top:
# Whole row gets same signal: high when programming new weights
we <<= programming
for wt in weights_tag_top:
# Tag is same for whole row; use state index (runs from 0 to 255)
wt <<= progstep
return [ x.sign_extended(32) for x in data_out ]
def accum(size, data_in, waddr, wen, wclear, raddr, lastvec):
'''A single 32-bit accumulator with 2^size 32-bit buffers.
On wen, writes data_in to the specified address (waddr) if wclear is high;
otherwise, it performs an accumulate at the specified address (buffer[waddr] += data_in).
lastvec is a control signal indicating that the operation being stored now is the
last vector of a matrix multiply instruction (at the final accumulator, this becomes
a "done" signal).
'''
mem = MemBlock(bitwidth=32, addrwidth=size)
# Writes
with conditional_assignment:
with wen:
with wclear:
mem[waddr] |= data_in
with otherwise:
mem[waddr] |= (data_in + mem[waddr])[:mem.bitwidth]
# Read
data_out = mem[raddr]
# Pipeline registers
waddrsave = Register(len(waddr))
waddrsave.next <<= waddr
wensave = Register(1)
wensave.next <<= wen
wclearsave = Register(1)
wclearsave.next <<= wclear
lastsave = Register(1)
lastsave.next <<= lastvec
return data_out, waddrsave, wensave, wclearsave, lastsave
def accumulators(accsize, datas_in, waddr, we, wclear, raddr, lastvec):
'''
Produces array of accumulators of same dimension as datas_in.
'''
#probe(we, "accum_wen")
#probe(wclear, "accum_wclear")
#probe(waddr, "accum_waddr")
accout = [ None for i in range(len(datas_in)) ]
waddrin = waddr
wein = we
wclearin = wclear
lastvecin = lastvec
for i,x in enumerate(datas_in):
#probe(x, "acc_{}_in".format(i))
#probe(wein, "acc_{}_we".format(i))
#probe(waddrin, "acc_{}_waddr".format(i))
dout, waddrin, wein, wclearin, lastvecin = accum(accsize, x, waddrin, wein, wclearin, raddr, lastvecin)
accout[i] = dout
done = lastvecin
return accout, done
def FIFO(matsize, mem_data, mem_valid, advance_fifo):
'''
matsize is the length of one row of the Matrix.
mem_data is the connection from the DRAM controller, which is assumed to be 64 bytes wide.
mem_valid is a one bit control signal from the controller indicating that the read completed and the current value is valid.
advance_fifo signals to drop the tile at the end of the FIFO and advance everything forward.
Output
tile, ready, full
tile: entire tile at the front of the queue (8 x matsize x matsize bits)
ready: the tile output is valid
full: there is no room in the FIFO
'''
#probe(mem_data, "fifo_dram_in")
#probe(mem_valid, "fifo_dram_valid")
#probe(advance_fifo, "weights_advance_fifo")
# Make some size parameters, declare state register
totalsize = matsize * matsize # total size of a tile in bytes
tilesize = totalsize * 8 # total size of a tile in bits
ddrwidth = int(len(mem_data)/8) # width from DDR in bytes (typically 64)
size = 1
while pow(2, size) < (totalsize/ddrwidth): # compute log of number of transfers required
size = size + 1
state = Register(size) # Number of reads to receive (each read is ddrwidth bytes)
startup = Register(1)
startup.next <<= 1
#probe(state, "fifo_state")
# Declare top row of buffer: need to write to it in ddrwidth-byte chunks
topbuf = [ Register(ddrwidth*8) for i in range(max(1, int(totalsize/ddrwidth))) ]
# Latch command to advance FIFO, since it may not complete immediately
droptile = Register(1)
clear_droptile = WireVector(1)
with conditional_assignment:
with advance_fifo:
droptile.next |= 1
with clear_droptile:
droptile.next |= 0
#probe(droptile, "fifo_droptile")
#probe(clear_droptile, "fifo_clear_droptile")
# When we get data from DRAM controller, write to next buffer space
with conditional_assignment:
with mem_valid:
state.next |= state + 1 # state tracks which ddrwidth-byte chunk we're writing to
for i, reg in enumerate(reversed(topbuf)): # enumerate a decoder for write-enable signals
#probe(reg, "fifo_reg{}".format(i))
with state == Const(i, bitwidth=size):
reg.next |= mem_data
# Track when first buffer is filled and when data moves out of it
full = Register(1) # goes high when last chunk of top buffer is filled
cleartop = WireVector(1)
with conditional_assignment:
with mem_valid & (state == Const(len(topbuf)-1)): # writing the last buffer spot now
full.next |= 1
with cleartop: # advancing FIFO, so buffer becomes empty
full.next |= 0
# Build buffers for remainder of FIFO
buf2, buf3, buf4 = Register(tilesize), Register(tilesize), Register(tilesize)
#probe(concat_list(topbuf), "buf1")
#probe(buf2, "buf2")
#probe(buf3, "buf3")
#probe(buf4, "buf4")
#probe(full, "buf1_full")
# If a given row is empty, track that so we can fill immediately
empty2, empty3, empty4 = Register(1), Register(1), Register(1)
#probe(empty2, "buf2_empty")
#probe(empty3, "buf3_empty")
#probe(empty4, "buf4_empty")
# Handle moving data between the buffers
with conditional_assignment:
with ~startup:
empty2.next |= 1
empty3.next |= 1
empty4.next |= 1
with full & empty2: # First buffer is full, second is empty
buf2.next |= concat_list(topbuf) # move data to second buffer
cleartop |= 1 # empty the first buffer
empty2.next |= 0 # mark the second buffer as non-empty
with empty3 & ~empty2: # Third buffer is empty and second is full
buf3.next |= buf2
empty3.next |= 0
empty2.next |= 1
with empty4 & ~empty3: # Fourth buffer is empty and third is full
buf4.next |= buf3
empty4.next |= 0
empty3.next |= 1
with droptile:
empty4.next |= 1 # mark fourth buffer as free; tiles will advance automatically
clear_droptile |= 1
ready = startup & (~empty4) & (~droptile) # there is data in final buffer and we're not about to change it
return buf4, ready, full
def systolic_setup(data_width, matsize, vec_in, waddr, valid, clearbit, lastvec, switch):
'''Buffers vectors from the unified SRAM buffer so that they can be fed along diagonals to the
Matrix Multiply array.
matsize: row size of Matrix
vec_in: row read from unified buffer
waddr: the accumulator address this vector is bound for
valid: this is a valid vector; write it when done
clearbit: if 1, store result (default accumulate)
lastvec: this is the last vector of a matrix
switch: use the next weights tile beginning with this vector
Output
next_row: diagonal cross-cut of vectors to feed to MM array
switchout: switch signals for MM array
addrout: write address for first accumulator
weout: write enable for first accumulator
clearout: clear signal for first accumulator
doneout: done signal for first accumulator
'''
# Use a diagonal set of buffer so that when a vector is read from SRAM, it "falls" into
# the correct diagonal pattern.
# The last column of buffers need extra bits for control signals, which propagate down
# and into the accumulators.
addrreg = Register(len(waddr))
addrreg.next <<= waddr
wereg = Register(1)
wereg.next <<= valid
clearreg = Register(1)
clearreg.next <<= clearbit
donereg = Register(1)
donereg.next <<= lastvec
topreg = Register(data_width)
firstcolumn = [topreg,] + [ Register(data_width) for i in range(matsize-1) ]
lastcolumn = [ None for i in range(matsize) ]
lastcolumn[0] = topreg
# Generate switch signals to matrix; propagate down
switchout = [ None for i in range(matsize) ]
switchout[0] = Register(1)
switchout[0].next <<= switch
for i in range(1, len(switchout)):
switchout[i] = Register(1)
switchout[i].next <<= switchout[i-1]
# Generate control pipeline for address, clear, and done signals
addrout = addrreg
weout = wereg
clearout = clearreg
doneout = lastvec
#probe(clearout, "sys_clear_in")
# Need one extra cycle of delay for control signals before giving them to first accumulator
# But we already did registers for first row, so cancels out
for i in range(0, matsize):
a = Register(len(addrout))
a.next <<= addrout
addrout = a
w = Register(1)
w.next <<= weout
weout = w
c = Register(1)
c.next <<= clearout
clearout = c
d = Register(1)
d.next <<= doneout
doneout = d
#probe(clearout, "sys_clear_out")
# Generate buffers in a diagonal pattern
for row in range(1, matsize): # first row is done
left = firstcolumn[row]
lastcolumn[row] = left
for column in range(0, row): # first column is done
buf = Register(data_width)
buf.next <<= left
left = buf
lastcolumn[row] = left # holds final column for output
# Connect first column to input data
datain = [ vec_in[i*data_width : i*data_width+data_width] for i in range(matsize) ]
for din, reg in zip(datain, firstcolumn):
reg.next <<= din
return lastcolumn, switchout, addrout, weout, clearout, doneout
def MMU(data_width, matrix_size, accum_size, vector_in, accum_raddr, accum_waddr, vec_valid, accum_overwrite, lastvec, switch_weights, ddr_data, ddr_valid): #, weights_in, weights_we):
logn1 = 1
while pow(2, logn1) < (matrix_size + 1):
logn1 = logn1 + 1
logn = 1
while pow(2, logn) < (matrix_size):
logn = logn + 1
programming = Register(1) # if high, we're programming new weights now
waiting = WireVector(1) # if high, a switch is underway and we're waiting
weights_wait = Register(logn1, "weights_wait") # counts cycles since last weight push
weights_count = Register(logn, "weights_count") # counts cycles of current weight push
startup = Register(1)
startup.next <<= 1 # 0 only in first cycle
weights_we = WireVector(1)
done_programming = WireVector(1)
first_tile = Register(1) # Tracks if we've programmed the first tile yet
#rtl_assert(~(switch_weights & (weights_wait != 0)), Exception("Weights are not ready to switch. Need a minimum of {} + 1 cycles since last switch.".format(matrix_size)))
# FIFO
weights_tile, tile_ready, full = FIFO(matsize=matrix_size, mem_data=ddr_data, mem_valid=ddr_valid, advance_fifo=done_programming)
#probe(tile_ready, "tile_ready")
#probe(weights_tile, "FIFO_weights_out")
matin, switchout, addrout, weout, clearout, doneout = systolic_setup(data_width=data_width, matsize=matrix_size, vec_in=vector_in, waddr=accum_waddr, valid=vec_valid, clearbit=accum_overwrite, lastvec=lastvec, switch=switch_weights)
mouts = MMArray(data_width=data_width, matrix_size=matrix_size, data_in=matin, new_weights=switchout, weights_in=weights_tile, weights_we=weights_we)
accout, done = accumulators(accsize=accum_size, datas_in=mouts, waddr=addrout, we=weout, wclear=clearout, raddr=accum_raddr, lastvec=doneout)
switchstart = switchout[0]
totalwait = Const(matrix_size + 1)
waiting <<= weights_wait != totalwait # if high, we have to wait
#probe(waiting, "waiting")
with conditional_assignment:
with ~startup: # when we start, configure values to be ready to accept a new tile
weights_wait.next |= totalwait
with waiting: # need to wait for switch to finish propagating
weights_wait.next |= weights_wait + 1
with ~first_tile & tile_ready: # start programming the first tile
weights_wait.next |= totalwait # we don't have to swait for a switch to clear
programming.next |= 1 # begin programming weights
weights_count.next |= 0
first_tile.next |= 1
with switchstart: # Weight switch initiated; begin waiting
weights_wait.next |= 0
programming.next |= 1
weights_count.next |= 0
with programming: # We're pushing new weights now
with weights_count == Const(matrix_size-1): # We've reached the end
programming.next |= 0
done_programming |= 1
with otherwise: # Still programming; increment count and keep write signal high
weights_count.next |= weights_count + 1
weights_we |= 1
'''
with conditional_assignment:
with startup == 0: # When we start, we're ready to push weights as soon as FIFO is ready
weights_wait.next |= totalwait
with switchout: # Got a switch signal; start wait count
weights_wait.next |= 1
with weights_wait != totalwait: # Stall on the final number
weights_wait.next |= weights_wait + 1
with weights_count != 0: # If we've started programming new weights, reset
weights_wait.next |= 0
with otherwise:
pass
with ~startup:
pass
with (weights_wait == totalwait) & tile_ready: # Ready to push new weights in
weights_count.next |= 1
with weights_count == Const(matrix_size): # Finished pushing new weights
done_programming |= 1
weights_count.next |= 0
with otherwise: # We're pushing weights now; increment count
weights_count.next |= weights_count + 1
weights_we |= 1
'''
return accout, done
def MMU_top(data_width, matrix_size, accum_size, ub_size, start, start_addr, nvecs, dest_acc_addr, overwrite, swap_weights, ub_rdata, accum_raddr, weights_dram_in, weights_dram_valid):
'''
Outputs
ub_raddr: read address for unified buffer
'''
#probe(ub_rdata, "ub_mm_rdata")
accum_waddr = Register(accum_size)
vec_valid = WireVector(1)
overwrite_reg = Register(1)
last = WireVector(1)
swap_reg = Register(1)
busy = Register(1)
N = Register(len(nvecs))
ub_raddr = Register(ub_size)
rtl_assert(~(start & busy), Exception("Cannot dispatch new MM instruction while previous instruction is still being issued."))
#probe(vec_valid, "MM_vec_valid_issue")
#probe(busy, "MM_busy")
# Vector issue control logic
with conditional_assignment:
with start: # new instruction being issued
accum_waddr.next |= dest_acc_addr
overwrite_reg.next |= overwrite
swap_reg.next |= swap_weights
busy.next |= 1
N.next |= nvecs
ub_raddr.next |= start_addr # begin issuing next cycle
with busy: # We're issuing a vector this cycle
vec_valid |= 1
swap_reg.next |= 0
N.next |= N - 1
with N == 1: # this was the last vector
last |= 1
overwrite_reg.next |= 0
busy.next |= 0
with otherwise: # we're going to issue a vector next cycle as well
ub_raddr.next |= ub_raddr + 1
accum_waddr.next |= accum_waddr + 1
last |= 0
acc_out, done = MMU(data_width=data_width, matrix_size=matrix_size, accum_size=accum_size, vector_in=ub_rdata, accum_raddr=accum_raddr, accum_waddr=accum_waddr, vec_valid=vec_valid, accum_overwrite=overwrite_reg, lastvec=last, switch_weights=swap_reg, ddr_data=weights_dram_in, ddr_valid=weights_dram_valid)
#probe(ub_raddr, "ub_mm_raddr")
return ub_raddr, acc_out, busy, done
'''
Do we need full/stall signal from Matrix? Would need to stop SRAM out from writing to systolic setup
Yes: MMU needs to track when both buffers used and emit such a signal
The timing systems for weights programming are wonky right now. Both rtl_asserts are failing, but the
right answer comes out if you ignore that. It looks like the state machine that counts time since the
last weights programming keeps restarting, so the MMU thinks it's always programming weights?
Control signals propagating down systolic_setup to accumulators:
-Overwrite signal (default: accumulate)
-New accumulator address value (default: add 1 to previous address)
-Done signal?
'''
def testall(input_vectors, weights_vectors):
DATWIDTH = 8
MATSIZE = 4
ACCSIZE = 8
L = len(input_vectors)
ins = [probe(Input(DATWIDTH)) for i in range(MATSIZE)]
invec = concat_list(ins)
swap = Input(1, 'swap')
waddr = Input(8)
lastvec = Input(1)
valid = Input(1)
raddr = Input(8, "raddr") # accumulator read address to read out answers
donesig = Output(1, "done")
outs = [Output(32, name="out{}".format(str(i))) for i in range(MATSIZE)]
#ws = [ Const(item, bitwidth=DATWIDTH) for sublist in weights_vectors for item in sublist ] # flatten weight matrix
#ws = concat_list(ws) # combine weights into single wire
ws = [ item for sublist in weights_vectors for item in sublist ] # flatten weight matrix
print(ws)
#ws = reduce(lambda x, y : (x<<8)+y, ws) # "concat" weights into one integer
weightsdata = Input(64*8)
weightsvalid = Input(1)
accout, done = MMU(data_width=DATWIDTH, matrix_size=MATSIZE, accum_size=ACCSIZE, vector_in=invec, accum_raddr=raddr, accum_waddr=waddr, vec_valid=valid, accum_overwrite=Const(0), lastvec=lastvec, switch_weights=swap, ddr_data=weightsdata, ddr_valid=weightsvalid)
donesig <<= done
for out, accout in zip(outs, accout):
out <<= accout
sim_trace = SimulationTrace()
sim = FastSimulation(tracer=sim_trace)
# make a default input dictionary
din = { swap:0, waddr:0, lastvec:0, valid:0, raddr:0, weightsdata:0, weightsvalid:0 }
din.update({ins[j] : 0 for j in range(MATSIZE)})
# Give a few cycles for startup
sim.step(din)
# First, simulate memory read to feed weights to FIFO
chunk = 64*8 # size of one dram read
#ws = [ ws[i*chunk:i*chunk+chunk] for i in range(max(1,len(ws)/chunk)) ] # divide weights into dram chunks
# divide weights into dram-transfer sized chunks
ws = reduce(lambda x, y : (x<<8)+y, ws) # "concat" weights into one integer
ws = [ (ws >> (64*8*i)) & pow(2, 64*8)-1 for i in range(max(1,len(weights_vectors)/64)) ]
print(ws)
for block in ws:
d = din.copy()
d.update({ins[j] : 0 for j in range(MATSIZE)})
d.update({ weightsdata:block, weightsvalid:1})
sim.step(d)
# Wait until the FIFO is ready
for i in range(10):
sim.step(din)
#din.update({ins[j]:0 for j in range(MATSIZE)})
# Send signal to write weights
#d = din.copy()
#d[weights_we] = 1
#sim.step(d)
# Wait MATSIZE cycles for weights to propagate
for i in range(MATSIZE*2):
sim.step(din)
# Send the swap signal with first row of input
d = din.copy()
d.update({ins[j] : input_vectors[0][j] for j in range(MATSIZE) })
d.update({ swap : 1, valid : 1 })
sim.step(d)
# Send rest of input
for i in range(L-1):
d = din.copy()
d.update({ins[j] : input_vectors[i+1][j] for j in range(MATSIZE) })
d.update({ waddr : i+1, lastvec : 1 if i == L-2 else 0, valid : 1 })
sim.step(d)
# Wait some cycles while it propagates
for i in range(L*2):
d = din.copy()
sim.step(d)
# Read out values
for i in range(L):
d = din.copy()
d[raddr] = i
sim.step(d)
with open('trace.vcd', 'w') as f:
sim_trace.print_vcd(f)
if __name__ == "__main__":
#weights = [[1, 10, 10, 2], [3, 9, 6, 2], [6, 8, 2, 8], [4, 1, 8, 6]] # transposed
#weights = [[4, 1, 8, 6], [6, 8, 2, 8], [3, 9, 6, 2], [1, 10, 10, 2]] # tranposed, reversed
#weights = [[1, 3, 6, 4], [10, 9, 8, 1], [10, 6, 2, 8], [2, 2, 8, 6]]
weights = [[2, 2, 8, 6], [10, 6, 2, 8], [10, 9, 8, 1], [1, 3, 6, 4]] # reversed
vectors = [[12, 7, 2, 6], [21, 21, 18, 8], [1, 4, 18, 11], [6, 3, 25, 15], [21, 12, 1, 15], [1, 6, 13, 8], [24, 25, 18, 1], [2, 5, 13, 6], [19, 3, 1, 17], [25, 10, 20, 10]]
testall(vectors, weights)