forked from SingularityKChen/dl_accelerator
-
Notifications
You must be signed in to change notification settings - Fork 0
/
ScalaModelTest.scala
631 lines (620 loc) · 34 KB
/
ScalaModelTest.scala
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
package dla.tests
import dla.cluster.{ClusterSRAMConfig, GNMFCS1Config, GNMFCS2Config}
import dla.pe.{MCRENFConfig, PESizeConfig, SPadSizeConfig}
import org.scalatest._
import Console.{MAGENTA, RESET}
import scala.math.max
import chisel3.util.log2Ceil
case class TaskMappingParam(inActWidth: Int, inActHeight: Int, inActNum: Int, inActCh: Int,
weightWidth: Int, weightHeight: Int, weightNum: Int, weightCh: Int,
cgRow: Int = 8, cgCol: Int = 2, peRow: Int = 3, peCol: Int = 4,
inActAdrSRAMSize: Int, inActDataSRAMSize: Int, inActSRAMNum: Int = 3,
pSumSRAMSize: Int,
inActAdrSPadSize: Int, inActDataSPadSize: Int,
weightAdrSPadSize: Int, weightDataSPadSize: Int,
pSumSPadSize: Int
) {
require(inActHeight > weightHeight, "the height of inAct should bigger than that of weight")
/** assume the SRAM size and SPad size is the worst case*/
val g2: Int = 1
val g1: Int = 1
/*private def getGLBLevel(): Int = {
var n2 = 1
var m2 = 1
var f2 = 1
var c2 = 1
var s2 = 1
}
private def getNoCLevel(): Int = {
var n1 = 1
var m1 = 1
var f1 = 1
var c1 = 1
var s1 = 1
}
private def getSPadLevel(): Int = {
var f0 = 1
var n0 = 1
val e = inActHeight - weightHeight //2
val r = weightHeight //4
var c0 = 1
var m0 = 1
val inActMatrixHeight: Int = r*c0 // row
val inActMatrixWidth: Int = f0*n0*e // column
val weightMatrixWidth: Int = inActMatrixHeight // column
val weightMatrixHeight: Int = m0 // row
val pSumOneSPadNum: Int = m0*e*n0*f0
val inActDataFit = inActMatrixHeight*inActMatrixWidth <= inActDataSPadSize
val weightDataFit = weightMatrixHeight*weightMatrixWidth <= weightDataSPadSize
val pSumFit = pSumOneSPadNum <= pSumSPadSize
}
*/
}
case class EyerissModelParam(
G2: Int = 1, N2: Int = 2, M2: Int = 4, F2: Int = 3, C2: Int = 3, S2: Int = 4,
G1: Int = 1, N1: Int = 4, M1: Int = 2, F1: Int = 4, C1: Int = 2, S1: Int = 3,
M0: Int = 4, C0: Int = 2, R: Int = 4, E: Int = 2, N0: Int = 3, F0: Int = 1,
cgRow: Int = 8, cgCol: Int = 2, peRow: Int = 3, peCol: Int = 4, inActSRAMNum: Int = 3
) {
/** true if Eyeriss can read from memory parallel,
* false while it only sent one memory read requirement a time*/
val parallelMemRead: Boolean = false
object physicalInfo {
val clusterNum: Int = cgRow * cgCol
private val peArraySize = peRow * peCol
val peNum: Int = peArraySize * clusterNum
// mapping <> physical
require(G1 == 1, "G1 has to be one, or you need to change the following requirements")
require(N1*C1*S1/peRow == cgRow, s"dose ${N1*C1*S1/peRow} equals to $cgRow")
require(M1*F1/peCol == cgCol, s"dose ${M1*F1/peCol} equals to $cgCol")
require(S1 % peRow == 0, s"S1 = $S1 should be multiple of peRow = $peRow") // TODO: change to S1 >= peRow // by pSum
require(F1 % peCol == 0, s"F1 = $F1 should be multiple of peCol = $peCol") // TODO: change to F1 >= peCol // by weight
def printlnPhysicalInfo(): Unit = {
println(s"[${MAGENTA}Info$RESET] the Eyeriss physical info")
println(s" |CGRow\t|CGCol\t|peRow\t|peCol\t|")
println(s" |$cgRow\t\t|$cgCol\t\t|$peRow\t\t|$peCol\t\t|")
}
}
object mappingInfo {
/**the number of different inAct in one PE array*/
private val inActNumInOnePEArray: Int = peRow + peCol - 1 //TODO: make it more fine grain
/** the read time for one SRAM to send all the PE their inAct*/
val inActSRAMReadTimes: Int = inActNumInOnePEArray/inActSRAMNum
require(inActNumInOnePEArray % inActSRAMNum == 0, s"inActNumInOnePEArray = $inActNumInOnePEArray should be" +
s"multiple of inActSRAMNum = $inActSRAMNum")
val pSumOneSPadNum: Int = M0*E*N0*F0
val inActMatrixWidth: Int = F0*N0*E // column
val inActMatrixHeight: Int = R*C0 // row
val weightMatrixWidth: Int = inActMatrixHeight // column
val weightMatrixHeight: Int = M0 // row
val inActNoCNum: Int = G1*N1*C1*(F1 + S1)
val weightNoCNum: Int = G1*M1*C1*S1
val pSumNoCNum: Int = G1*N1*M1*F1
val weightGLBNum: Int = G2*M2*C2*S2
val inActGLBNum: Int = G2*N2*C2*(F2 + S2)
val pSumGLBNum: Int = G2*N2*M2*F2
}
object nnShape {
object inAct {
/** G means the group number of the input activation. Different groups of input activation shard the same
* weight, hence increasing the data reuse ratio of weight*/
val number: Int = G2*G1*N2*N1*N0
val channel: Int = C2*C1*C0
/** although the width of inAct in RS+ data flow is (S2 + F2)*(S1 + F1)*F0,
* which is much greater than this width. It's caused by the overlap.*/
val height: Int = R + E
val width: Int = S2*S1 + F2*F1*F0
//require(height == width, s"inAct's height doesn't equal to width, $height == $width ?")
}
object weight {
val number: Int = M2*M1*M0
val channel: Int = C2*C1*C0
val height: Int = R
val width: Int = S2*S1
//require(height == width, s"weight's height doesn't equal to width, $height == $width ?")
}
object pSum {
val number: Int = G2*G1*N2*N1*N0
val channel: Int = M2*M1*M0
val height: Int = E
val width: Int = F2*F1*F0
//require(height == width, s"pSum's height doesn't equal to width, $height == $width ?")
}
require(inAct.number == pSum.number)
require(inAct.channel == weight.channel)
require(weight.number == pSum.channel)
def printNNShapeInfo(): Unit = {
println(s"[${MAGENTA}Info$RESET] the NN shape")
println(s" |\ttype\t|\tnum\t|\tchn\t|\th\t|\tw\t|")
println(s" |\tweight\t|\t${weight.number}\t|\t${weight.channel}\t|\t${weight.height}\t|\t${weight.width}\t|")
println(s" |\tinAct\t|\t${inAct.number}\t|\t${inAct.channel}\t|\t${inAct.height}\t|\t${inAct.width}\t|")
println(s" |\tpSum\t|\t${pSum.number}\t|\t${pSum.channel}\t|\t${pSum.height}\t|\t${pSum.width}\t|")
}
}
}
class EyerissModel(sequencer: GenFunc, monitor: CompareMonitor, p: EyerissModelParam, printDetails: Boolean = true) extends PESizeConfig with SPadSizeConfig
with MCRENFConfig with GNMFCS1Config with GNMFCS2Config with ClusterSRAMConfig {
/** use monitor to compare the info between different test*/
private val scoreBoard = new ScoreBoard
// TODO: change the
val pSumResult: Array[Array[Array[Array[Int]]]] = Array.fill(
p.nnShape.pSum.number,
p.nnShape.pSum.channel,
p.nnShape.pSum.height,
p.nnShape.pSum.width
) {0}
private var parallelCycle = 0
/** the first dimension is NoC level index, value true means have read this from mem, false means haven't read*/
private val weightMemReadRecord: Array[Boolean] = Array.fill(p.mappingInfo.weightNoCNum) {false}
/** the first dimension is NoC level index, value true means have read this from mem, false means haven't read*/
private val inActMemReadRecord: Array[Boolean] = Array.fill(p.mappingInfo.inActNoCNum) {false}
/** assume the data stored in Mem is pre-processed */
/** the first dimension is cgRow idx, the second is cgCol idx, the third is inActSRAMIdx, inside is a list */
private val inActAdrSRAM: Array[Array[Array[List[Int]]]] =
Array.fill(p.cgRow, p.cgCol, p.inActSRAMNum) {Nil}
/** the first dimension is cgRow idx, the second is cgCol idx, the third is inActSRAMIdx, inside is a list */
private val inActDataSRAM: Array[Array[Array[List[Int]]]] =
Array.fill(p.cgRow, p.cgCol, p.inActSRAMNum) {Nil}
/** the first dimension is cgRow idx, the second is cgCol idx, the third is inActSRAMIdx,
* the fourth is inActSRAMReadIdx. true when has written this data into inActSRAM*/
private val inActSRAMWriteRecord: Array[Array[Array[Array[Boolean]]]] =
Array.fill(p.cgRow, p.cgCol, p.inActSRAMNum, p.mappingInfo.inActSRAMReadTimes) {false}
/** the first dimension is cgRow idx, the second is cgCol idx, the third is inActSRAMIdx,
* the fourth is inActSRAMReadIdx. true when has read this data from inActSRAM*/
private val inActSRAMReadRecord: Array[Array[Array[Array[Boolean]]]] =
Array.fill(p.cgRow, p.cgCol, p.inActSRAMNum, p.mappingInfo.inActSRAMReadTimes) {false}
//private var inActSRAMBankWriteRecord: List[Seq[Int]] = Nil
/** NoC level:
* the `for loops` of NoC level is the task mapping for each PE*/
for (g1 <- 0 until p.G1) {
for (n1 <- 0 until p.N1) {
for (m1 <- 0 until p.M1) {
for (f1 <- 0 until p.F1) {
for (c1 <- 0 until p.C1) {
for (s1 <- 0 until p.S1) {
/** current physical info*/
object cPhyInfo {
/**current cluster group row idx*/
val cr: Int = n1*p.C1*p.S1/p.peRow + c1*p.S1/p.peRow + s1/p.peRow
/**current cluster group column idx*/
val cc: Int = m1*p.F1/p.peCol + f1/p.peCol
/**current pe row idx*/
val pr: Int = s1%p.peRow
/**current pe column idx*/
val pc: Int = f1%p.peCol
/**current pe's corresponding inAct SRAM idx in the GLB Cluster*/
val inActSRAMIdx: Int = (pr + pc) % p.inActSRAMNum
/** current SRAM's read/write times*/
val inActReadTimeIdx: Int = (pr + pc) / p.inActSRAMNum
}
val weightNoCLevelIdx = g1*p.M1*p.C1*p.S1 + m1*p.C1*p.S1 + c1*p.S1 + s1
val inActNoCLevelIdx = g1*p.N1*p.C1*(p.F1+p.S1) + n1*p.C1*(p.F1+p.S1) + c1*(p.F1+p.S1) + (f1+s1)
/** read inAct from main memory*/
if (!inActMemReadRecord(inActNoCLevelIdx)) {
val inActMemReadNum = sequencer.dataSequencer.glb.inAct(inActNoCLevelIdx).flatten.flatten.length
monitor.inActRead.mem += inActMemReadNum
if (p.parallelMemRead) {
monitor.cycle += inActMemReadNum*scoreBoard.accessCost.mem/(p.G1*p.N1*p.C1*(p.F1+p.S1))
} else {
/** if Eyeriss can only send one memory requirement at a time */
monitor.cycle += inActMemReadNum*scoreBoard.accessCost.mem
}
inActMemReadRecord(inActNoCLevelIdx) = true
}
/** write inAct SRAM */
//println(s"current Read Times: ${cPhyInfo.inActReadTimeIdx}")
if (!inActSRAMWriteRecord(cPhyInfo.cr)(cPhyInfo.cc)(cPhyInfo.inActSRAMIdx)(cPhyInfo.inActReadTimeIdx)){
inActAdrSRAM(cPhyInfo.cr)(cPhyInfo.cc)(cPhyInfo.inActSRAMIdx) ++=
sequencer.dataSequencer.glb.cscData.inActAdr(inActNoCLevelIdx)
inActDataSRAM(cPhyInfo.cr)(cPhyInfo.cc)(cPhyInfo.inActSRAMIdx) ++=
sequencer.dataSequencer.glb.cscData.inActData(inActNoCLevelIdx)
val inActAdrGLBWriteNum = sequencer.dataSequencer.glb.cscData.inActAdr(inActNoCLevelIdx).length
val inActDataGLBWriteNum = sequencer.dataSequencer.glb.cscData.inActData(inActNoCLevelIdx).length
monitor.inActWrite.adr.glb += inActAdrGLBWriteNum
monitor.inActWrite.data.glb += inActDataGLBWriteNum
inActSRAMWriteRecord(cPhyInfo.cr)(cPhyInfo.cc)(cPhyInfo.inActSRAMIdx)(cPhyInfo.inActReadTimeIdx) = true
}
/** GLB level */
for (g2 <- 0 until p.G2) {
for (n2 <- 0 until p.N2) {
for (m2 <- 0 until p.M2) {
for (f2 <- 0 until p.F2) {
for (c2 <- 0 until p.C2) {
for (s2 <- 0 until p.S2) {
val weightGLBLevelIdx = g2*p.M2*p.C2*p.S2 + m2*p.C2*p.S2 + c2*p.S2 + s2
val inActGLBLevelIdx = g2*p.N2*p.C2*(p.F2 + p.S2) + n2*p.C2*(p.F2+p.S2) + c2*(p.F2+p.S2) + f2+s2
/** SPad level */
val inActAdrSPad = sequencer.dataSequencer.glb.separatedSPadCSCData.
inActAdr(inActNoCLevelIdx)(inActGLBLevelIdx)
val inActDataSPad = sequencer.dataSequencer.glb.separatedSPadCSCData.
inActData(inActNoCLevelIdx)(inActGLBLevelIdx)
val weightAdrSPad = sequencer.dataSequencer.glb.separatedSPadCSCData.
weightAdr(weightNoCLevelIdx)(weightGLBLevelIdx)
val weightDataSPad = sequencer.dataSequencer.glb.separatedSPadCSCData.
weightData(weightNoCLevelIdx)(weightGLBLevelIdx)
val pSumSPad: Array[Array[Int]] = Array.fill(p.F0*p.N0*p.E, p.M0) {0}
val inActGLBReadNum = max(inActAdrSPad.length, inActDataSPad.length)
val weightMemReadNum = sequencer.dataSequencer.glb.
weight(weightNoCLevelIdx)(weightGLBLevelIdx).flatten.length
/** the same weight value only be read once from Mem*/
if (!weightMemReadRecord(weightNoCLevelIdx)) {
monitor.weightRead.mem += weightMemReadNum
if (!p.parallelMemRead) {
/**if Eyeriss can not read parallel, then assume that read weight from memory
* for GLB levels times will cost more than read inAct from SRAM*/
monitor.cycle += weightMemReadNum*scoreBoard.accessCost.mem
}
}
if (p.parallelMemRead) {
parallelCycle += max(inActGLBReadNum*scoreBoard.accessCost.glb,
weightMemReadNum*scoreBoard.accessCost.mem)
}
/** read from GLB once and send into diagonal PEs
* and if this has been read, then other clusters can receive inAct via Router */
if (!inActSRAMReadRecord(cPhyInfo.cr)(cPhyInfo.cc)(cPhyInfo.inActSRAMIdx)(cPhyInfo.inActReadTimeIdx)) {
monitor.inActRead.adr.glb += inActAdrSPad.length
monitor.inActRead.data.glb += inActDataSPad.length
}
monitor.inActWrite.adr.sPad += inActAdrSPad.length
monitor.inActWrite.data.sPad += inActDataSPad.length
monitor.weightWrite.adr.sPad += weightAdrSPad.length
monitor.weightWrite.data.sPad += weightDataSPad.length
var inActDataSPadIdx = 0
for (inActAdrSPadIdx <- inActAdrSPad.indices) {
/** padInActAdr: read each column of current inAct Matrix */
val inActAdr = inActAdrSPad(inActAdrSPadIdx)
monitor.inActRead.adr.sPad += 1
parallelCycle += 1
if (inActAdr != inActZeroColumnCode || inActAdr == 0) {
/** padInActData: read each row of current column */
while (inActDataSPadIdx < inActAdr) {
val inActDataRead = inActDataSPad(inActDataSPadIdx)
val inActData = BigInt(inActDataRead.toBinaryString.take(cscDataWidth), 2).toInt
val inActRow = BigInt(inActDataRead.toBinaryString.takeRight(cscCountWidth), 2).toInt
monitor.inActRead.data.sPad += 1
parallelCycle += 1
if (inActDataRead != 0) {
/** padWeightAdr */
val weightAdr = weightAdrSPad(inActRow)
val weightDataSPadStartIdx = if (inActRow == 0) 0 else weightAdrSPad(inActRow - 1)
monitor.weightRead.adr.sPad += 1
parallelCycle += 1
if (weightAdr != weightZeroColumnCode || weightAdr == 0) {
/** padWeightData */
for (weightDataSPadIdx <- weightDataSPadStartIdx until weightAdr) {
val weightDataRead = weightDataSPad(weightDataSPadIdx)
val weightData = BigInt(weightDataRead.toBinaryString.take(cscDataWidth), 2).toInt
val weightRow = BigInt(weightDataRead.toBinaryString.takeRight(cscCountWidth), 2).toInt
monitor.weightRead.data.sPad += 1
parallelCycle += 2 // need 2 cycles to read from SRAM
pSumSPad(inActAdrSPadIdx)(weightRow) += weightData * inActData
monitor.macNum += 1
parallelCycle += 2 // one for mpy, one for write back
}
}
}
inActDataSPadIdx += 1
}
}
}
//print(".") // finish SPad Level
/** accumulate PSum vertically */
// TODO
}
}
}
}
}
}
//print("*\n") // finish GLB Level
inActSRAMReadRecord(cPhyInfo.cr)(cPhyInfo.cc)(cPhyInfo.inActSRAMIdx)(cPhyInfo.inActReadTimeIdx) = true
weightMemReadRecord(weightNoCLevelIdx) = true
}
}
}
}
}
}
monitor.cycle += parallelCycle/p.physicalInfo.peNum
if (printDetails) {
monitor.printMonitorInfo(p.physicalInfo.peNum)
p.nnShape.printNNShapeInfo()
}
}
class CommonModel(sequencer: GenFunc, monitor: CompareMonitor, p: EyerissModelParam,
printDetails: Boolean = true, needPSum: Boolean)
extends PESizeConfig with SPadSizeConfig with MCRENFConfig with GNMFCS1Config with GNMFCS2Config with ClusterSRAMConfig {
private val scoreBoard = new ScoreBoard
val pSumResult: Array[Array[Array[Array[Int]]]] = Array.fill(
p.nnShape.pSum.number,
p.nnShape.pSum.channel,
p.nnShape.pSum.height,
p.nnShape.pSum.width
) {0}
if (needPSum) {
/** each PSum number */
for (n <- 0 until p.nnShape.pSum.number) {
/** each PSum channel*/
for (m <- 0 until p.nnShape.pSum.channel) {
/** PSum height */
for (f <- 0 until p.nnShape.pSum.width) {
/** PSum width*/
for (e <- 0 until p.nnShape.pSum.height) {
/** inside this for loop, do mac, for the size of weight matrix */
/** weight channel */
for (c <- 0 until p.nnShape.weight.channel) {
/** weight height */
for (s <- 0 until p.nnShape.weight.width) {
/** weight width */
for (r <- 0 until p.nnShape.weight.height) {
pSumResult(n)(m)(e)(f) += sequencer.dataSequencer.dram.weight(m)(c)(r)(s) *
sequencer.dataSequencer.dram.inAct(n)(c)(r+e)(s+f)
monitor.inActRead.mem += 1
monitor.weightRead.mem += 1
monitor.macNum += 1
}
}
}
//print(".") // finish one PSum
}
}
//print("*") // finish one PSum matrix
}
/*println("\n[INFO] finish one batch of PSum " +
f"${((n + 1).toFloat/p.nnShape.pSum.number.toFloat)*100}%.2f%%")*/
}
} else {
val totalNum = p.nnShape.pSum.number * p.nnShape.pSum.channel * p.nnShape.pSum.width * p.nnShape.pSum.height *
p.nnShape.weight.channel * p.nnShape.weight.width * p.nnShape.weight.height
monitor.inActRead.mem = totalNum
monitor.weightRead.mem = totalNum
monitor.macNum = totalNum
}
monitor.cycle = scoreBoard.totalCycles(monitor.macNum, p.physicalInfo.peNum, monitor.inActRead.mem, 0, 0)
if (printDetails) {
monitor.printMonitorInfo(p.physicalInfo.peNum)
p.nnShape.printNNShapeInfo()
}
}
class IndVarTmp(val start: Int, val end: Int) {
val during: Int = end - start
require(end > start, s"we need end = $end > start = $start")
}
class ScalaModelDriver(indVar1: IndVarTmp, indVar2: IndVarTmp, indVarString: String,
paramFunc: (Int, Int) => EyerissModelParam,
sequencerFunc: (Int, Int, EyerissModelParam) => GenFunc,
printEndInfo: (Int, Int) => Unit,
getPrintVarInfo: (Int, Int) => String) {
val monitorSeq: Seq[Seq[Seq[CompareMonitor]]] = Seq.fill(indVar1.during, indVar2.during, 2) {new CompareMonitor}
/** the min size of inAct adr SPad and data SPad to meet the requirement.
* [[SPadSizeConfig]].[[inActAdrSPadSize]] and [[SPadSizeConfig]].[[inActDataSPadSize]]*/
val inActSPadSizeNeed: Array[Array[Array[Int]]] = Array.fill(indVar1.during, indVar2.during, 2) {0}
/** the min size of inAct adr SRAM and data SRAM to meet the requirement.
* [[ClusterSRAMConfig]].[[inActAdrSRAMSize]] and [[ClusterSRAMConfig]].[[inActDataSRAMSize]]*/
val inActSRAMSizeNeed: Array[Array[Array[Int]]] = Array.fill(indVar1.during, indVar2.during, 2) {0}
/** the min bits of inAct adr to meet the requirement. [[PESizeConfig]].[[inActAdrWidth]]*/
val inActAdrWidthNeed: Array[Array[Int]] = Array.fill(indVar1.during, indVar2.during) {0}
/** the min bits of inAct data to meet the requirement. [[PESizeConfig]].[[inActDataWidth]]*/
val inActDataWidthNeed: Array[Array[Int]] = Array.fill(indVar1.during, indVar2.during) {0}
/** the min size of weight adr SPad and data SPad to meet the requirement.
* [[SPadSizeConfig]].[[weightAdrSPadSize]] and [[SPadSizeConfig]].[[weightDataSPadSize]]*/
val weightSPadSizeNeed: Array[Array[Array[Int]]] = Array.fill(indVar1.during, indVar2.during, 2) {0}
/** the min bits of weight adr to meet the requirement. [[PESizeConfig]].[[weightAdrWidth]]*/
val weightAdrWidthNeed: Array[Array[Int]] = Array.fill(indVar1.during, indVar2.during) {0}
/** the min bits of weight data to meet the requirement. [[PESizeConfig]].[[weightDataWidth]]*/
val weightDataWidthNeed: Array[Array[Int]] = Array.fill(indVar1.during, indVar2.during) {0}
for (indvar1 <- indVar1.start until indVar1.end) {
for (indvar2 <- indVar2.start until indVar2.end) {
val indVar1Idx = indvar1 - indVar1.start
val indVar2Idx = indvar2 - indVar2.start
val param = paramFunc(indvar1, indvar2)
val sequencer = sequencerFunc(indvar1, indvar2, param)
new EyerissModel(sequencer = sequencer,
monitor = monitorSeq(indVar1Idx)(indVar2Idx).head,
p = param, printDetails = false)
new CommonModel(sequencer = sequencer,
monitor = monitorSeq(indVar1Idx)(indVar2Idx)(1),
p = param, printDetails = false, needPSum = false)
inActSPadSizeNeed(indVar1Idx)(indVar2Idx)(0) =
sequencer.dataSequencer.glb.separatedSPadCSCData.inActAdr.map(x => x.map(y => y.length).max).max
inActSPadSizeNeed(indVar1Idx)(indVar2Idx)(1) =
sequencer.dataSequencer.glb.separatedSPadCSCData.inActData.map(x => x.map(y => y.length).max).max
inActSRAMSizeNeed(indVar1Idx)(indVar2Idx)(0) =
sequencer.dataSequencer.glb.cscData.inActAdr.map(x => x.length).max
inActSRAMSizeNeed(indVar1Idx)(indVar2Idx)(1) =
sequencer.dataSequencer.glb.cscData.inActData.map(x => x.length).max
inActAdrWidthNeed(indVar1Idx)(indVar2Idx) =
log2Ceil(sequencer.dataSequencer.glb.cscData.inActAdr.flatten.filter(x => x != scala.math.pow(2,7)-1).max)
inActDataWidthNeed(indVar1Idx)(indVar2Idx) =
log2Ceil(sequencer.dataSequencer.glb.cscData.inActData.flatten.filter(x => x != scala.math.pow(2,12)-1).max)
weightSPadSizeNeed(indVar1Idx)(indVar2Idx)(0) =
sequencer.dataSequencer.glb.separatedSPadCSCData.weightAdr.map(x => x.map(y => y.length).max).max
weightSPadSizeNeed(indVar1Idx)(indVar2Idx)(1) =
sequencer.dataSequencer.glb.separatedSPadCSCData.weightData.map(x => x.map(y => y.length).max).max
weightAdrWidthNeed(indVar1Idx)(indVar2Idx) =
log2Ceil(sequencer.dataSequencer.glb.cscData.weightAdr.flatten.filter(x => x != scala.math.pow(2,7)-1).max)
weightDataWidthNeed(indVar1Idx)(indVar2Idx) =
log2Ceil(sequencer.dataSequencer.glb.cscData.weightData.flatten.filter(x => x != scala.math.pow(2,12)-1).max)
printEndInfo(indvar1, indvar2)
if (indvar1 == indVar1.end - 1 && indvar2 == indVar2.end - 1) {
param.nnShape.printNNShapeInfo()
param.physicalInfo.printlnPhysicalInfo()
}
}
}
println(s"$indVarString|cycle%\t\t|mac%\t\t|iMem%\t\t|wMem%\t\t|iGLB%\t\t|iSPad%\t\t|")
for (indvar1 <- indVar1.start until indVar1.end) {
for (indvar2 <- indVar2.start until indVar2.end) {
val indVar1Idx = indvar1 - indVar1.start
val indVar2Idx = indvar2 - indVar2.start
val eyerissMonitor = monitorSeq(indVar1Idx)(indVar2Idx).head
val commonMonitor = monitorSeq(indVar1Idx)(indVar2Idx).last
val inActGLBWriteTotal = eyerissMonitor.inActWrite.adr.glb + eyerissMonitor.inActWrite.data.glb
val inActGLBReadTotal = eyerissMonitor.inActRead.adr.glb + eyerissMonitor.inActRead.data.glb
val inActSPadWriteTotal = eyerissMonitor.inActWrite.adr.sPad + eyerissMonitor.inActWrite.data.sPad
val inActSPadReadTotal = eyerissMonitor.inActRead.adr.sPad + eyerissMonitor.inActRead.data.sPad
/**inAct GLB R / GLB W */
val eyerissInActGLBRW = f"${(inActGLBReadTotal.toFloat/inActGLBWriteTotal.toFloat)*100}%.2f%%"
/**inAct SPad W / GLB R */
val eyerissInActSPadReuse = f"${(inActSPadWriteTotal.toFloat/inActGLBReadTotal.toFloat)*100}%.2f%%"
val cycleEfficiency = f"${(eyerissMonitor.cycle.toFloat / commonMonitor.cycle.toFloat)*100}%.4f%%"
val macEfficiency = f"${(eyerissMonitor.macNum.toFloat / commonMonitor.macNum.toFloat)*100}%.4f%%"
val inActMemReadEfficiency =
f"${(eyerissMonitor.inActRead.mem.toFloat / commonMonitor.inActRead.mem.toFloat)*100}%.4f%%"
val weightMemReadEfficiency =
f"${(eyerissMonitor.weightRead.mem.toFloat / commonMonitor.weightRead.mem.toFloat)*100}%.4f%%"
println(getPrintVarInfo(indvar1, indvar2) + s"$cycleEfficiency\t|" +
s"$macEfficiency\t|$inActMemReadEfficiency\t|$weightMemReadEfficiency\t|" +
s"$eyerissInActGLBRW\t|$eyerissInActSPadReuse\t|")
}
}
println(s"\n$indVarString|inActAdrSPad\t|inActDataSPad\t|weightAdrSPad\t|weightDataSPad\t|inActAdrSRAM\t|inActDataSRAM\t|")
for (indvar1 <- indVar1.start until indVar1.end) {
for (indvar2 <- indVar2.start until indVar2.end) {
val indVar1Idx = indvar1 - indVar1.start
val indVar2Idx = indvar2 - indVar2.start
println(getPrintVarInfo(indvar1, indvar2) +
s"${inActSPadSizeNeed(indVar1Idx)(indVar2Idx)(0)}\t${inActAdrWidthNeed(indVar1Idx)(indVar2Idx)}-bit\t|" +
s"${inActSPadSizeNeed(indVar1Idx)(indVar2Idx)(1)}\t${inActDataWidthNeed(indVar1Idx)(indVar2Idx)}-bit\t|" +
s"${weightSPadSizeNeed(indVar1Idx)(indVar2Idx)(0)}\t${weightAdrWidthNeed(indVar1Idx)(indVar2Idx)}-bit\t|" +
s"${weightSPadSizeNeed(indVar1Idx)(indVar2Idx)(1)}\t${weightDataWidthNeed(indVar1Idx)(indVar2Idx)}-bit\t|" +
s"${inActSRAMSizeNeed(indVar1Idx)(indVar2Idx)(0)}\t|" +
s"${inActSRAMSizeNeed(indVar1Idx)(indVar2Idx)(1)}\t|"
)
}
}
}
class ScalaModelTest extends FlatSpec {
behavior of "compare the efficiency of Eyeriss"
/** model the behavior of Eyeriss cluster group */
it should "changing mapping parameters" in {
def getSRAMNum(peRowFactor: Int, peColFactor: Int): Int = {
val peRowNum = 3*peRowFactor
val peColNum = 4*peColFactor
var readTime = 2
while ((peRowNum + peColNum - 1) % readTime != 0) {
readTime += 1
}
val inActSRAMNum = (peRowNum + peColNum - 1) / readTime
inActSRAMNum
}
def paramFunc(peRowFactor: Int, peColFactor: Int): EyerissModelParam = {
val peRowNum = 3*peRowFactor
val peColNum = 4*peColFactor
val inActSRAMNum = getSRAMNum(peRowFactor, peColFactor)
val param = EyerissModelParam(peRow = peRowNum, peCol = peColNum,
S1 = peRowNum, F1 = peColNum, inActSRAMNum = inActSRAMNum)
param
}
def printFunc(peRowFactor: Int, peColFactor: Int): Unit = {
val peRowNum = 3*peRowFactor
val peColNum = 4*peColFactor
val inActSRAMNum = getSRAMNum(peRowFactor, peColFactor)
println(s"[${MAGENTA}Info$RESET] current peRow = $peRowNum, peCol = $peColNum, inActSRAM = $inActSRAMNum")
}
def sequencerFunc(peRowFactor: Int, peColFactor: Int, p: EyerissModelParam): GenFunc = {
new GenFunc(inActSparseRatio = 0.6, weightSparseRatio = 0.6, p = p)
}
def getPrintVarInfo(peRowFactor: Int, peColFactor: Int): String = {
s"|${peRowFactor*3}\t\t|${peColFactor*4}\t\t|"
}
val peRow = new IndVarTmp(start = 1, end = 4)
val peCol = new IndVarTmp(start = 1, end = 4)
new ScalaModelDriver(indVar1 = peRow, indVar2 = peCol, indVarString = "|pRow\t|pCol\t",
paramFunc = paramFunc, sequencerFunc = sequencerFunc,
printEndInfo = printFunc, getPrintVarInfo = getPrintVarInfo)
}
it should "compare the info across sparse ratio between eyeriss and common device" in {
def printFunc(inActRatio: Int, weightRatio: Int): Unit = {
println(s"[${MAGENTA}Info$RESET] current sparse ratio: 0.$inActRatio, 0.$weightRatio")
}
def paramFunc(inActRatio: Int, weightRatio: Int): EyerissModelParam = {
EyerissModelParam()
}
def sequencerFunc(inActRatio: Int, weightRatio: Int, param: EyerissModelParam): GenFunc = {
new GenFunc(inActSparseRatio = inActRatio.toDouble/10,
weightSparseRatio = weightRatio.toDouble/10, p = param)
}
def getPrintVarInfo(inActRatio: Int, weightRatio: Int): String = {
s"|${inActRatio.toDouble/10}\t|${weightRatio.toDouble/10}\t|"
}
val inActRatioSeq = new IndVarTmp(start = 4, end = 10)
val weightRatioSeq = new IndVarTmp(start = 4, end = 10)
new ScalaModelDriver(indVar1 = inActRatioSeq, indVar2 = weightRatioSeq, indVarString = "|iRa\t|wRa\t",
paramFunc = paramFunc, sequencerFunc = sequencerFunc,
printEndInfo = printFunc, getPrintVarInfo = getPrintVarInfo)
}
}
class CompareMonitor extends ClusterSRAMConfig {
var cycle: BigInt = 0 // the number of clock cycles
var macNum: BigInt = 0 // the number of mac
class CSCAccess {
var glb: BigInt = 0 // the times to access glb sram
var sPad: BigInt = 0 // the times to access sPad register
}
class MemHierarchyAccess {
var mem: BigInt = 0 // the times to access memory
val adr = new CSCAccess
val data = new CSCAccess
}
val inActRead = new MemHierarchyAccess
val inActWrite = new MemHierarchyAccess
val weightRead = new MemHierarchyAccess
val weightWrite = new MemHierarchyAccess
def printMonitorInfo(peNum: Int): Unit = {
val inActGLBWriteTotal = inActWrite.adr.glb + inActWrite.data.glb
val inActGLBReadTotal = inActRead.adr.glb + inActRead.data.glb
val inActSPadWriteTotal = inActWrite.adr.sPad + inActWrite.data.sPad
val inActSPadReadTotal = inActRead.adr.sPad + inActRead.data.sPad
println(s"[${MAGENTA}Info$RESET] computation finishes, using $peNum PEs")
println(s"------ time = $cycle cycles")
println(s"------ mac num = $macNum")
println(s"------ inActAccess ")
println(s" | memRead: ${inActRead.mem}")
println(s" | glb")
println(s" | glbWrite: $inActGLBWriteTotal")
println(s" | glbAdrWrite: ${inActWrite.adr.glb}")
println(s" | glbDataWrite: ${inActWrite.data.glb}")
println(s" | glbRead: $inActGLBReadTotal")
println(s" | glbAdrRead: ${inActRead.adr.glb}")
println(s" | glbDataRead: ${inActRead.data.glb}")
println(s" | sPad")
println(s" | sPadWrite: $inActSPadWriteTotal")
println(s" | sPadAdrWrite: ${inActWrite.adr.sPad}")
println(s" | sPadDataWrite: ${inActWrite.data.sPad}")
println(s" | sPadRead: $inActSPadReadTotal")
println(s" | sPadAdrRead: ${inActRead.adr.sPad}")
println(s" | sPadDataRead: ${inActRead.data.sPad}")
println(s"------ weightAccess")
println(s" | mem: ${weightRead.mem}")
println(s" | sPad")
println(s" | sPadWrite: ${weightWrite.adr.sPad + weightWrite.data.sPad}")
println(s" | sPadAdrWrite: ${weightWrite.adr.sPad}")
println(s" | sPadDataWrite: ${weightWrite.data.sPad}")
println(s" | sPadRead: ${weightRead.adr.sPad + weightRead.data.sPad}")
println(s" | sPadAdrRead: ${weightRead.adr.sPad}")
println(s" | sPadDataRead: ${weightRead.data.sPad}")
println(s"------ dataReuse")
println(" | inAct GLB R / GLB W: " +
f"${(inActGLBReadTotal.toFloat/inActGLBWriteTotal.toFloat)*100}%.2f%%")
println(" | inAct SPad W / GLB R: " +
f"${(inActSPadWriteTotal.toFloat/inActGLBReadTotal.toFloat)*100}%.2f%%")
}
}
class ScoreBoard {
/** [[macCost]]: every mac will cost 3 clock cycle
* 0: multiply
* 1: accumulate
* 2: write back*/
val macCost: Int = 3
object accessCost { // every access will cost ? clock cycles
val mem: Int = 60
val glb: Int = 2
val sPad: Int = 1
}
def totalCycles(macNum: BigInt, peNum: Int, memNum: BigInt, glbNum: BigInt, sPadNum: BigInt ): BigInt = {
val cycles: BigInt = (macNum/peNum) * macCost + memNum * accessCost.mem +
glbNum * accessCost.glb + sPadNum * accessCost.sPad
cycles
}
}