-
Notifications
You must be signed in to change notification settings - Fork 11
/
convolution.cpp
179 lines (140 loc) · 4.83 KB
/
convolution.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
#include <iostream>
#include <string>
#include "dnn.hpp"
using namespace std;
//Define the parameters if not defined externally
#ifndef Sy
#define Sy 1
#define Sx 1
#endif
#ifndef Tnn
//Tiling Sizes
#define Tnn 32
#define Tn 16
#define Ti 16
#define Ty 8
#define Tx 8
#endif
#define NYPAD (Ny+Ky)
#define NXPAD (Nx+Kx)
#define NYSCL (Ny/Sy)
#define NXSCL (Nx/Sx)
#define SYNAPSE_SIZE (1L*Ky*Kx*Nn*Ni)
VTYPE (*synapse)[Ky][Kx][Nn][Ni];
VTYPE (*neuron_i)[NYPAD][NXPAD][Ni];
VTYPE (*neuron_n)[NYSCL][NXSCL][Nn];
VTYPE (*neuron_n2)[NYSCL][NXSCL][Nn];
void fill_convolution_shared_simple(VTYPE (&synapse)[Ky][Kx][Nn][Ni],
VTYPE (&neuron_i)[NYPAD][NXPAD][Ni]) {
for(int yy = 0; yy < Ky; ++yy) {
for(int xx = 0; xx < Kx; ++xx) {
for(int nn = 0; nn < Nn; ++nn) {
for(int ni = 0; ni < Ni; ++ni) {
synapse[yy][xx][nn][ni] = static_cast <float> (rand()) / static_cast <float> (RAND_MAX) - 0.5f;
} } } }
for(int yy = 0; yy < NYPAD; ++yy) {
for(int xx = 0; xx < NXPAD; ++xx) {
for(int ni = 0; ni < Ni; ++ni) {
neuron_i[yy][xx][ni] = static_cast <float> (rand()) / static_cast <float> (RAND_MAX) - 0.5f;
} } }
}
std::pair<int,int> convolution_layer_blocked(
VTYPE (&synapse)[Ky][Kx][Nn][Ni],
VTYPE (&neuron_i)[NYPAD][NXPAD][Ni],
VTYPE (&neuron_n)[NYSCL][NXSCL][Nn]) {
int c1=0,c2=0;
VTYPE sum[Nn]={0};
for (int yy = 0; yy < Ny; yy += Ty) {
for (int xx = 0; xx < Nx; xx += Tx) {
for (int nnn = 0; nnn < Nn; nnn += Tnn) {
int yout = yy/Sy;
for (int y = yy; y < yy + Ty; y += Sy) { // tiling for y;
int xout = xx/Sx;
for (int x = xx; x < xx + Tx; x += Sx) { // tiling for x;
for (int nn = nnn; nn < nnn + Tnn; nn += Tn) {
for (int n = nn; n < nn + Tn; n++) {
sum[n] = 0;
}
for (int ky = 0; ky < Ky; ky++) { // sliding window;
for (int kx = 0; kx < Kx; kx++) {
int ii = 0;
VTYPE sum_sc;
for (; ii < Ni -Ti+1; ii += Ti) {
for (int n = nn; n < nn + Tn; n++) {
sum_sc=0;
for (int i = ii; i < ii + Ti; i++) {
VTYPE sv = synapse[ky][kx][n][i];
VTYPE nv = neuron_i[ky + y][kx + x][i];
sum_sc+=sv*nv;
}
sum[n]+=sum_sc;
}
}
}
}
//transfer
for (int n = nn; n < nn + Tn; n++) {
neuron_n[yout][xout][n] = transfer(sum[n]);
}
}
xout++;
}
yout++;
}
}
}
}
}
void convolution_layer(VTYPE (&synapse)[Ky][Kx][Nn][Ni],
VTYPE (&neuron_i)[NYPAD][NXPAD][Ni],
VTYPE (&neuron_n)[NYSCL][NXSCL][Nn]) {
VTYPE sum[Nn]={0};
// — Original code — (excluding nn, ii loops)
int yout = 0;
for (int y = 0; y < Ny; y += Sy) { // tiling for y;
int xout = 0;
for (int x = 0; x < Ny; x += Sx) { // tiling for x;
for (int nn = 0; nn < Nn; nn += Tn) {
for (int n = nn; n < nn + Tn; n++) {
sum[n]=0;
}
// sliding window;
for (int ky = 0; ky < Ky; ky++)
for (int kx = 0; kx < Kx; kx++)
for (int n = nn; n < nn + Tn; n++)
for (int i = 0; i < Ni; i++) {
VTYPE sv = synapse[ky][kx][n][i];
VTYPE nv = neuron_i[ky + y][kx + x][i];
sum[n]+=sv*nv;
}
for (int n = nn; n < nn + Tn; n++) {
neuron_n[yout][xout][n] = transfer(sum[n]);
}
}
xout++;
}
yout++;
}
}
int main(const int argc, const char** argv) {
cout << "allocating memory\n";
synapse = (VTYPE (*)[Ky][Kx][Nn][Ni]) aligned_malloc(64, SYNAPSE_SIZE*sizeof(VTYPE));
neuron_i = (VTYPE (*)[NYPAD][NXPAD][Ni])aligned_malloc(64,NYPAD*NXPAD*Ni*sizeof(VTYPE));
neuron_n = (VTYPE (*)[NYSCL][NXSCL][Nn])aligned_malloc(64,NYSCL*NXSCL*Nn*sizeof(VTYPE));
neuron_n2 = (VTYPE (*)[NYSCL][NXSCL][Nn])aligned_malloc(64,NYSCL*NXSCL*Nn*sizeof(VTYPE));
cout << "initializing arrays\n";
fill_convolution_shared_simple(*synapse,*neuron_i);
cout << "starting computation\n";
//Simple Version
begin_roi();
convolution_layer(*synapse,*neuron_i,*neuron_n);
end_roi();
cout << "simple version complete!\n";
//Blocked Version
begin_roi();
convolution_layer_blocked(*synapse,*neuron_i,*neuron_n2);
end_roi();
cout << "blocked computation complete!\n";
compare((VTYPE*)*neuron_n,(VTYPE*)*neuron_n2,NYSCL*NXSCL*Nn);
cout << "done\n";
}