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
|
#ifndef TH_GENERIC_FILE
#define TH_GENERIC_FILE "generic/unfold.c"
#else
/* note: due to write issues, this one cannot be parallelized as well as unfolded_copy */
void THNN_(unfolded_acc)(
THTensor *finput,
THTensor *input,
int kW,
int kH,
int dW,
int dH,
int padW,
int padH,
int nInputPlane,
int inputWidth,
int inputHeight,
int outputWidth,
int outputHeight)
{
// This function assumes that
// outputHeight*dH does not overflow a long
// outputWidth*dW does not overflow a long
int nip;
real *input_data = THTensor_(data)(input);
real *finput_data = THTensor_(data)(finput);
#pragma omp parallel for private(nip)
for(nip = 0; nip < nInputPlane; nip++)
{
int kw, kh, y, x;
long ix, iy;
for(kh = 0; kh < kH; kh++)
{
for(kw = 0; kw < kW; kw++)
{
real *src = finput_data + nip*((size_t)kH*kW*outputHeight*outputWidth) + kh*((size_t)kW*outputHeight*outputWidth) + kw*((size_t)outputHeight*outputWidth);
real *dst = input_data + nip*((size_t)inputHeight*inputWidth);
if (padW > 0 || padH > 0) {
int lpad,rpad;
for(y = 0; y < outputHeight; y++) {
iy = (long)y*dH - padH + kh;
if (iy < 0 || iy >= inputHeight) {
} else {
if (dW==1){
ix = 0 - padW + kw;
lpad = fmaxf(0,padW-kw);
rpad = fmaxf(0,padW-(kW-kw-1));
real *dst_slice = dst+(size_t)iy*inputWidth+ix+lpad;
THVector_(cadd)(dst_slice, dst_slice, src+(size_t)y*outputWidth+lpad, 1, outputWidth - lpad - rpad); /* note: THVector_add could handle 1 value better */
}
else{
for (x=0; x<outputWidth; x++){
ix = (long)x*dW - padW + kw;
if (ix < 0 || ix >= inputWidth){
}else{
real *dst_slice = dst+(size_t)iy*inputWidth+ix;
THVector_(cadd)(dst_slice, dst_slice, src+(size_t)y*outputWidth+x, 1, 1);
}
}
}
}
}
} else {
for(y = 0; y < outputHeight; y++) {
iy = (long)y*dH + kh;
ix = 0 + kw;
if (dW == 1 ) {
real *dst_slice = dst+(size_t)iy*inputWidth+ix;
THVector_(cadd)(dst_slice, dst_slice, src+(size_t)y*outputWidth, 1, outputWidth); /* note: THVector_add could handle 1 value better */
}else{
for(x = 0; x < outputWidth; x++) {
real *dst_slice = dst+(size_t)iy*inputWidth+ix+x*dW;
THVector_(cadd)(dst_slice, dst_slice, src+(size_t)y*outputWidth+x, 1, 1);
}
}
}
}
}
}
}
}
void THNN_(unfolded_copy)(
THTensor *finput,
THTensor *input,
int kW,
int kH,
int dW,
int dH,
int padW,
int padH,
int nInputPlane,
int inputWidth,
int inputHeight,
int outputWidth,
int outputHeight)
{
// This function assumes that
// kH*kW does not overflow an int
// nInputPlane*kH*kW does not overflow a long
// outputHeight*dH does not overflow a long
// outputWidth*dW does not overflow a long
long k;
real *input_data = THTensor_(data)(input);
real *finput_data = THTensor_(data)(finput);
#pragma omp parallel for private(k)
for(k = 0; k < (long)nInputPlane*kH*kW; k++) {
long nip = k / (kH*kW);
long rest = k % (kH*kW);
long kh = rest / kW;
long kw = rest % kW;
int x, y;
long ix, iy;
real *dst = finput_data + nip*((size_t)kH*kW*outputHeight*outputWidth) + kh*((size_t)kW*outputHeight*outputWidth) + kw*((size_t)outputHeight*outputWidth);
real *src = input_data + nip*((size_t)inputHeight*inputWidth);
if (padW > 0 || padH > 0) {
long lpad,rpad;
for(y = 0; y < outputHeight; y++) {
iy = (long)y*dH - padH + kh;
if (iy < 0 || iy >= inputHeight) {
memset(dst+(size_t)y*outputWidth, 0, sizeof(real)*outputWidth);
} else {
if (dW==1){
ix = 0 - padW + kw;
lpad = fmaxf(0,padW-kw);
rpad = fmaxf(0,padW-(kW-kw-1));
if (outputWidth-rpad-lpad <= 0) {
memset(dst+(size_t)y*outputWidth, 0, sizeof(real)*outputWidth);
} else {
if (lpad > 0) memset(dst+(size_t)y*outputWidth, 0, sizeof(real)*lpad);
memcpy(dst+(size_t)y*outputWidth+lpad, src+(size_t)iy*inputWidth+ix+lpad, sizeof(real)*(outputWidth-rpad-lpad));
if (rpad > 0) memset(dst+(size_t)y*outputWidth + outputWidth - rpad, 0, sizeof(real)*rpad);
}
}
else{
for (x=0; x<outputWidth; x++){
ix = (long)x*dW - padW + kw;
if (ix < 0 || ix >= inputWidth)
memset(dst+(size_t)y*outputWidth+x, 0, sizeof(real)*1);
else
memcpy(dst+(size_t)y*outputWidth+x, src+(size_t)iy*inputWidth+ix, sizeof(real)*(1));
}
}
}
}
} else {
for(y = 0; y < outputHeight; y++) {
iy = (long)y*dH + kh;
ix = 0 + kw;
if (dW == 1)
memcpy(dst+(size_t)y*outputWidth, src+(size_t)iy*inputWidth+ix, sizeof(real)*outputWidth);
else{
for (x=0; x<outputWidth; x++)
memcpy(dst+(size_t)y*outputWidth+x, src+(size_t)iy*inputWidth+ix+(long)x*dW, sizeof(real)*(1));
}
}
}
}
}
#endif
|
