forked from neurosim/MLP_NeuroSim_V1.0
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Test.cpp
252 lines (240 loc) · 12.1 KB
/
Test.cpp
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
/*******************************************************************************
* Copyright (c) 2015-2017
* School of Electrical, Computer and Energy Engineering, Arizona State University
* PI: Prof. Shimeng Yu
* All rights reserved.
*
* This source code is part of NeuroSim - a device-circuit-algorithm framework to benchmark
* neuro-inspired architectures with synaptic devices(e.g., SRAM and emerging non-volatile memory).
* Copyright of the model is maintained by the developers, and the model is distributed under
* the terms of the Creative Commons Attribution-NonCommercial 4.0 International Public License
* http://creativecommons.org/licenses/by-nc/4.0/legalcode.
* The source code is free and you can redistribute and/or modify it
* by providing that the following conditions are met:
*
* 1) Redistributions of source code must retain the above copyright notice,
* this list of conditions and the following disclaimer.
*
* 2) Redistributions in binary form must reproduce the above copyright notice,
* this list of conditions and the following disclaimer in the documentation
* and/or other materials provided with the distribution.
*
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
* ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
* WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
* DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
* FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
* DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
* SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
* CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
* OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
*
* Developer list:
* Pai-Yu Chen Email: pchen72 at asu dot edu
*
* Xiaochen Peng Email: xpeng15 at asu dot edu
********************************************************************************/
#include <cstdio>
#include <iostream>
#include <vector>
#include <random>
#include "formula.h"
#include "Param.h"
#include "Array.h"
#include "Mapping.h"
#include "NeuroSim.h"
extern Param *param;
extern std::vector< std::vector<double> > testInput;
extern std::vector< std::vector<int> > dTestInput;
extern std::vector< std::vector<double> > testOutput;
extern std::vector< std::vector<double> > weight1;
extern std::vector< std::vector<double> > weight2;
extern Technology techIH;
extern Technology techHO;
extern Array *arrayIH;
extern Array *arrayHO;
extern SubArray *subArrayIH;
extern SubArray *subArrayHO;
extern Adder adderIH;
extern Mux muxIH;
extern RowDecoder muxDecoderIH;
extern DFF dffIH;
extern Adder adderHO;
extern Mux muxHO;
extern RowDecoder muxDecoderHO;
extern DFF dffHO;
extern int correct; // # of correct prediction
/* Validation */
void Validate() {
int numBatchReadSynapse; // # of read synapses in a batch read operation (decide later)
double outN1[param->nHide]; // Net input to the hidden layer [param->nHide]
double a1[param->nHide]; // Net output of hidden layer [param->nHide] also the input of hidden layer to output layer
int da1[param->nHide]; // Digitized net output of hidden layer [param->nHide] also the input of hidden layer to output layer
double outN2[param->nOutput]; // Net input to the output layer [param->nOutput]
double a2[param->nOutput]; // Net output of output layer [param->nOutput]
double tempMax;
int countNum;
correct = 0;
double sumArrayReadEnergyIH = 0; // Use a temporary variable here since OpenMP does not support reduction on class member
double sumNeuroSimReadEnergyIH = 0; // Use a temporary variable here since OpenMP does not support reduction on class member
double sumReadLatencyIH = 0; // Use a temporary variable here since OpenMP does not support reduction on class member
double readVoltageIH = static_cast<eNVM*>(arrayIH->cell[0][0])->readVoltage;
double readPulseWidthIH = static_cast<eNVM*>(arrayIH->cell[0][0])->readPulseWidth;
double sumArrayReadEnergyHO = 0; // Use a temporary variable here since OpenMP does not support reduction on class member
double sumNeuroSimReadEnergyHO = 0; // Use a temporary variable here since OpenMP does not support reduction on class member
double sumReadLatencyHO = 0; // Use a temporary variable here since OpenMP does not support reduction on class member
double readVoltageHO = static_cast<eNVM*>(arrayHO->cell[0][0])->readVoltage;
double readPulseWidthHO = static_cast<eNVM*>(arrayHO->cell[0][0])->readPulseWidth;
#pragma omp parallel for private(outN1, a1, da1, outN2, a2, tempMax, countNum, numBatchReadSynapse) reduction(+: correct, sumArrayReadEnergyIH, sumNeuroSimReadEnergyIH, sumArrayReadEnergyHO, sumNeuroSimReadEnergyHO, sumReadLatencyIH, sumReadLatencyHO)
for (int i = 0; i < param->numMnistTestImages; i++)
{
// Forward propagation
/* First layer from input layer to the hidden layer */
std::fill_n(outN1, param->nHide, 0);
std::fill_n(a1, param->nHide, 0);
if (param->useHardwareInTestingFF) { // Hardware
for (int j=0; j<param->nHide; j++) {
if (AnalogNVM *temp = dynamic_cast<AnalogNVM*>(arrayIH->cell[0][0])) { // Analog eNVM
if (static_cast<eNVM*>(arrayIH->cell[0][0])->cmosAccess) { // 1T1R
sumArrayReadEnergyIH += arrayIH->wireGateCapRow * techIH.vdd * techIH.vdd * param->nInput; // All WLs open
}
} else if (DigitalNVM *temp = dynamic_cast<DigitalNVM*>(arrayIH->cell[0][0])) { // Digital eNVM
// XXX: To be released
}
for (int n=0; n<param->numBitInput; n++) {
double pSumMaxAlgorithm = pow(2, n) / (param->numInputLevel - 1) * arrayIH->arrayRowSize; // Max algorithm partial weighted sum for the nth vector bit (if both max input value and max weight are 1)
if (AnalogNVM *temp = dynamic_cast<AnalogNVM*>(arrayIH->cell[0][0])) { // Analog eNVM
double Isum = 0; // weighted sum current
double IsumMax = 0; // Max weighted sum current
double inputSum = 0; // Weighted sum current of input vector * weight=1 column
for (int k=0; k<param->nInput; k++) {
if ((dTestInput[i][k]>>n) & 1) { // if the nth bit of dTestInput[i][k] is 1
Isum += arrayIH->ReadCell(j,k);
inputSum += arrayIH->GetMaxCellReadCurrent(j,k);
sumArrayReadEnergyIH += arrayIH->wireCapRow * readVoltageIH * readVoltageIH; // Selected BLs (1T1R) or Selected WLs (cross-point)
}
IsumMax += arrayIH->GetMaxCellReadCurrent(j,k);
}
sumArrayReadEnergyIH += Isum * readVoltageIH * readPulseWidthIH;
int outputDigits = 2 * CurrentToDigits(Isum, IsumMax) - CurrentToDigits(inputSum, IsumMax);
outN1[j] += DigitsToAlgorithm(outputDigits, pSumMaxAlgorithm);
} else { // SRAM or digital eNVM
// XXX: To be released
}
}
a1[j] = sigmoid(outN1[j]);
da1[j] = round_th(a1[j]*(param->numInputLevel-1), param->Hthreshold);
}
numBatchReadSynapse = (int)ceil((double)param->nHide/param->numColMuxed);
#pragma omp critical // Use critical here since NeuroSim class functions may update its member variables
for (int j=0; j<param->nHide; j+=numBatchReadSynapse) {
int numActiveRows = 0; // Number of selected rows for NeuroSim
for (int n=0; n<param->numBitInput; n++) {
for (int k=0; k<param->nInput; k++) {
if ((dTestInput[i][k]>>n) & 1) { // if the nth bit of dTestInput[i][k] is 1
numActiveRows++;
}
}
}
subArrayIH->activityRowRead = (double)numActiveRows/param->nInput/param->numBitInput;
sumNeuroSimReadEnergyIH += NeuroSimSubArrayReadEnergy(subArrayIH);
sumNeuroSimReadEnergyIH += NeuroSimNeuronReadEnergy(subArrayIH, adderIH, muxIH, muxDecoderIH, dffIH);
sumReadLatencyIH += NeuroSimSubArrayReadLatency(subArrayIH);
sumReadLatencyIH += NeuroSimNeuronReadLatency(subArrayIH, adderIH, muxIH, muxDecoderIH, dffIH);
}
} else { // Algorithm
for (int j=0; j<param->nHide; j++){
for (int k=0; k<param->nInput; k++){
outN1[j] += 2 * testInput[i][k] * weight1[j][k] - testInput[i][k];
}
a1[j] = sigmoid(outN1[j]);
}
}
/* Second layer from hidden layer to the output layer */
tempMax = 0;
countNum = 0;
std::fill_n(outN2, param->nOutput, 0);
std::fill_n(a2, param->nOutput, 0);
if (param->useHardwareInTestingFF) { // Hardware
for (int j=0; j<param->nOutput; j++) {
if (AnalogNVM *temp = dynamic_cast<AnalogNVM*>(arrayHO->cell[0][0])) { // Analog eNVM
if (static_cast<eNVM*>(arrayHO->cell[0][0])->cmosAccess) { // 1T1R
sumArrayReadEnergyHO += arrayHO->wireGateCapRow * techHO.vdd * techHO.vdd * param->nHide; // All WLs open
}
} else if (DigitalNVM *temp = dynamic_cast<DigitalNVM*>(arrayHO->cell[0][0])) {
if (static_cast<eNVM*>(arrayHO->cell[0][0])->cmosAccess) { // 1T1R
sumArrayReadEnergyHO += arrayHO->wireGateCapRow * techHO.vdd * techHO.vdd; // Selected WL
} else { // Cross-point
sumArrayReadEnergyHO += arrayHO->wireCapRow * techHO.vdd * techHO.vdd * (param->nHide - 1); // Unselected WLs
}
}
for (int n=0; n<param->numBitInput; n++) {
double pSumMaxAlgorithm = pow(2, n) / (param->numInputLevel - 1) * arrayHO->arrayRowSize; // Max algorithm partial weighted sum for the nth vector bit (if both max input value and max weight are 1)
if (AnalogNVM *temp = dynamic_cast<AnalogNVM*>(arrayHO->cell[0][0])) { // Analog NVM
double Isum = 0; // weighted sum current
double IsumMax = 0; // Max weighted sum current
double a1Sum = 0; // Weighted sum current of a1 vector * weight=1 column
for (int k=0; k<param->nHide; k++) {
if ((da1[k]>>n) & 1) { // if the nth bit of da1[k] is 1
Isum += arrayHO->ReadCell(j,k);
a1Sum += arrayHO->GetMaxCellReadCurrent(j,k);
sumArrayReadEnergyHO += arrayHO->wireCapRow * readVoltageHO * readVoltageHO; // Selected BLs (1T1R) or Selected WLs (cross-point)
}
IsumMax += arrayHO->GetMaxCellReadCurrent(j,k);
}
sumArrayReadEnergyHO += Isum * readVoltageHO * readPulseWidthHO;
int outputDigits = 2 * CurrentToDigits(Isum, IsumMax) - CurrentToDigits(a1Sum, IsumMax);
outN2[j] += DigitsToAlgorithm(outputDigits, pSumMaxAlgorithm);
} else { // SRAM or digital eNVM
// XXX: To be released
}
}
a2[j] = sigmoid(outN2[j]);
if (a2[j] > tempMax) {
tempMax = a2[j];
countNum = j;
}
}
numBatchReadSynapse = (int)ceil((double)param->nOutput/param->numColMuxed);
#pragma omp critical // Use critical here since NeuroSim class functions may update its member variables
for (int j=0; j<param->nOutput; j+=numBatchReadSynapse) {
int numActiveRows = 0; // Number of selected rows for NeuroSim
for (int n=0; n<param->numBitInput; n++) {
for (int k=0; k<param->nHide; k++) {
if ((da1[k]>>n) & 1) { // if the nth bit of da1[k] is 1
numActiveRows++;
}
}
}
subArrayHO->activityRowRead = (double)numActiveRows/param->nHide/param->numBitInput;
sumNeuroSimReadEnergyHO += NeuroSimSubArrayReadEnergy(subArrayHO);
sumNeuroSimReadEnergyHO += NeuroSimNeuronReadEnergy(subArrayHO, adderHO, muxHO, muxDecoderHO, dffHO);
sumReadLatencyHO += NeuroSimSubArrayReadLatency(subArrayHO);
sumReadLatencyHO += NeuroSimNeuronReadLatency(subArrayHO, adderHO, muxHO, muxDecoderHO, dffHO);
}
} else { // Algorithm
for (int j=0; j<param->nOutput; j++) {
for (int k=0; k<param->nHide; k++) {
outN2[j] += 2 * a1[k] * weight2[j][k] - a1[k];
}
a2[j] = sigmoid(outN2[j]);
if (a2[j] > tempMax) {
tempMax = a2[j];
countNum = j;
}
}
}
if (testOutput[i][countNum] == 1) {
correct++;
}
}
if (!param->useHardwareInTraining) { // Calculate the classification latency and energy only for offline classification
arrayIH->readEnergy += sumArrayReadEnergyIH;
subArrayIH->readDynamicEnergy += sumNeuroSimReadEnergyIH;
arrayHO->readEnergy += sumArrayReadEnergyHO;
subArrayHO->readDynamicEnergy += sumNeuroSimReadEnergyHO;
subArrayIH->readLatency += sumReadLatencyIH;
subArrayHO->readLatency += sumReadLatencyHO;
}
}