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
|
#pragma once
#include <caffe2/core/types.h>
#ifdef __CUDA_ARCH__
#include <cuda_fp16.h>
#endif
#ifdef __CUDA_ARCH__
#define CONVERSIONS_DECL __host__ __device__ inline
#else
#define CONVERSIONS_DECL inline
#endif
namespace caffe2 {
namespace convert {
namespace {
inline float16 cpu_float2half_rn(float f) {
float16 ret;
static_assert(
sizeof(unsigned int) == sizeof(float),
"Programming error sizeof(unsigned int) != sizeof(float)");
unsigned* xp = reinterpret_cast<unsigned int*>(&f);
unsigned x = *xp;
unsigned u = (x & 0x7fffffff), remainder, shift, lsb, lsb_s1, lsb_m1;
unsigned sign, exponent, mantissa;
// Get rid of +NaN/-NaN case first.
if (u > 0x7f800000) {
ret.x = 0x7fffU;
return ret;
}
sign = ((x >> 16) & 0x8000);
// Get rid of +Inf/-Inf, +0/-0.
if (u > 0x477fefff) {
ret.x = sign | 0x7c00U;
return ret;
}
if (u < 0x33000001) {
ret.x = (sign | 0x0000);
return ret;
}
exponent = ((u >> 23) & 0xff);
mantissa = (u & 0x7fffff);
if (exponent > 0x70) {
shift = 13;
exponent -= 0x70;
} else {
shift = 0x7e - exponent;
exponent = 0;
mantissa |= 0x800000;
}
lsb = (1 << shift);
lsb_s1 = (lsb >> 1);
lsb_m1 = (lsb - 1);
// Round to nearest even.
remainder = (mantissa & lsb_m1);
mantissa >>= shift;
if (remainder > lsb_s1 || (remainder == lsb_s1 && (mantissa & 0x1))) {
++mantissa;
if (!(mantissa & 0x3ff)) {
++exponent;
mantissa = 0;
}
}
ret.x = (sign | (exponent << 10) | mantissa);
return ret;
}
inline float cpu_half2float(float16 h) {
unsigned sign = ((h.x >> 15) & 1);
unsigned exponent = ((h.x >> 10) & 0x1f);
unsigned mantissa = ((h.x & 0x3ff) << 13);
if (exponent == 0x1f) { /* NaN or Inf */
mantissa = (mantissa ? (sign = 0, 0x7fffff) : 0);
exponent = 0xff;
} else if (!exponent) { /* Denorm or Zero */
if (mantissa) {
unsigned int msb;
exponent = 0x71;
do {
msb = (mantissa & 0x400000);
mantissa <<= 1; /* normalize */
--exponent;
} while (!msb);
mantissa &= 0x7fffff; /* 1.mantissa is implicit */
}
} else {
exponent += 0x70;
}
int temp = ((sign << 31) | (exponent << 23) | mantissa);
unsigned* rp = reinterpret_cast<unsigned*>(&temp);
return *rp;
}
}; // anonymous
// general version: defer to static_cast
template <typename IN, typename OUT>
CONVERSIONS_DECL OUT To(const IN in) {
return static_cast<OUT>(in);
}
#if __CUDA_ARCH__
__device__ __inline__ __half inf_clip(__half h) {
int isi = __hisinf(h);
if (isi > 0) {
// Exponent all ones except LSB (0x1e), mantissa is all ones (0x3ff)
h.x = 0x7bffU;
} else if (isi < 0) {
// As above, negated
h.x = 0x7bffU ^ 0x8000;
}
return h;
}
#endif
// explicit for fp16
template <>
CONVERSIONS_DECL float16 To(const float in) {
#if __CUDA_ARCH__
// hacky interface between C2 fp16 and CUDA
float16 ret;
__half r;
// r.x = __float2half_rn(in);
// ret.x = inf_clip(r).x;
ret.x = __float2half(in).x;
return ret;
#else
return cpu_float2half_rn(in);
#endif
}
template <>
CONVERSIONS_DECL float To(const float16 in) {
#if __CUDA_ARCH__
__half tmp;
tmp.x = in.x;
return __half2float(tmp);
#else
return cpu_half2float(in);
#endif
};
template <>
CONVERSIONS_DECL float To(const float in) {
return in;
}
template <typename OUT, typename IN>
CONVERSIONS_DECL OUT Get(IN x) {
return static_cast<OUT>(x);
}
template <>
CONVERSIONS_DECL float Get(float16 x) {
return To<float16, float>(x);
}
template <>
CONVERSIONS_DECL float16 Get(float x) {
return To<float, float16>(x);
}
}; // namespace convert
}; // namespace caffe2
#undef CONVERSIONS_DECL
|
