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
|
#include "caffe2/core/common_gpu.h"
#include "caffe2/core/context_gpu.h"
#include "caffe2/operators/fully_connected_op.h"
namespace caffe2 {
namespace {
template <class FullyConnectedOp>
bool RunFullyConnectedOpOnCUDADevice(
const bool float16_compute,
FullyConnectedOp* op) {
if (op->Input(0).template IsType<float>()) {
return op->template DoRunWithType<
float, // X
float, // W
float, // B
float, // Y
float>(); // Math
} else if (op->Input(0).template IsType<at::Half>()) {
if (float16_compute) {
const cudaDeviceProp& prop = GetDeviceProperty(0);
if (prop.major >= kFp16CUDADevicePropMajor) {
return op->template DoRunWithType<
at::Half, // X
at::Half, // W
at::Half, // B
at::Half, // Y
at::Half>(); // Math
} else {
LOG(INFO) << "CUDA Device does not support FP16 computation, "
"falling back to FP32.";
return op->template DoRunWithType<
at::Half, // X
at::Half, // W
at::Half, // B
at::Half, // Y
float>(); // Math
}
} else {
return op->template DoRunWithType<
at::Half, // X
at::Half, // W
at::Half, // B
at::Half, // Y
float>(); // Math
}
} else {
CAFFE_THROW("Unsupported type");
}
return false;
}
template <class FullyConnectedGradientOp>
bool RunFullyConnectedGradientOpOnCUDADevice(
const bool float16_compute,
FullyConnectedGradientOp* op) {
if (op->Input(0).template IsType<float>()) {
return op->template DoRunWithType<
float, // X
float, // W
float, // dY
float, // B
float, // dX
float, // dW
float, // dB
float>(); // Math
} else if (op->Input(0).template IsType<at::Half>()) {
if (float16_compute) {
const cudaDeviceProp& prop = GetDeviceProperty(0);
if (prop.major >= kFp16CUDADevicePropMajor) {
return op->template DoRunWithType<
at::Half, // X
at::Half, // W
at::Half, // dY
at::Half, // B
at::Half, // dX
at::Half, // dW
at::Half, // dB
at::Half>(); // Math
} else {
LOG(INFO) << "CUDA Device does not support FP16 computation, "
"falling back to FP32.";
return op->template DoRunWithType<
at::Half, // X
at::Half, // W
at::Half, // dY
at::Half, // B
at::Half, // dX
at::Half, // dW
at::Half, // dB
float>(); // Math
}
} else {
return op->template DoRunWithType<
at::Half, // X
at::Half, // W
at::Half, // dY
at::Half, // B
at::Half, // dX
at::Half, // dW
at::Half, // dB
float>(); // Math
}
} else {
CAFFE_THROW("Unsupported type");
}
return false;
}
} // namespace
// The RunFullyConnectedOpOnCUDADevice Function will use the pointer of current
// op and the DoRunWithType will make sure to run the correct things.
template <>
bool FullyConnectedOp<CUDAContext>::RunOnDevice() {
return RunFullyConnectedOpOnCUDADevice(float16_compute_, this);
}
template <>
bool FullyConnectedOp<
CUDAContext,
DefaultEngine,
false /* don't transpose weight */>::RunOnDevice() {
return RunFullyConnectedOpOnCUDADevice(float16_compute_, this);
}
template <>
bool FullyConnectedGradientOp<CUDAContext>::RunOnDevice() {
return RunFullyConnectedGradientOpOnCUDADevice(float16_compute_, this);
}
template <>
bool FullyConnectedGradientOp<
CUDAContext,
DefaultEngine,
false /* don't transpose weight */>::RunOnDevice() {
return RunFullyConnectedGradientOpOnCUDADevice(float16_compute_, this);
}
#if !defined(USE_ROCM)
// Require these to be defined otherwise TensorCore FC ops will end
// up calling the default FC implementation which doesn't have
// fp16 support...
template <>
bool FullyConnectedOp<CUDAContext, TensorCoreEngine>::RunOnDevice() {
return RunFullyConnectedOpOnCUDADevice(false /* float16_compute */, this);
}
template <>
bool FullyConnectedOp<
CUDAContext,
TensorCoreEngine,
false /* don't transpose weight */>::RunOnDevice() {
return RunFullyConnectedOpOnCUDADevice(false /* float16_compute */, this);
}
template <>
bool FullyConnectedGradientOp<CUDAContext, TensorCoreEngine>::RunOnDevice() {
return RunFullyConnectedGradientOpOnCUDADevice(
false /* float16_compute */, this);
}
template <>
bool FullyConnectedGradientOp<
CUDAContext,
TensorCoreEngine,
false /* don't transpose weight */>::RunOnDevice() {
return RunFullyConnectedGradientOpOnCUDADevice(
false /* float16_compute */, this);
}
#endif
REGISTER_CUDA_OPERATOR(FC, FullyConnectedOp<CUDAContext>);
REGISTER_CUDA_OPERATOR(FCGradient, FullyConnectedGradientOp<CUDAContext>);
REGISTER_CUDA_OPERATOR(
FCTransposed,
FullyConnectedOp<
CUDAContext,
DefaultEngine,
false /* don't transpose weight */>);
REGISTER_CUDA_OPERATOR(
FCTransposedGradient,
FullyConnectedGradientOp<
CUDAContext,
DefaultEngine,
false /* don't transpose weight */>);
#if !defined(USE_ROCM)
REGISTER_CUDA_OPERATOR_WITH_ENGINE(
FC,
TENSORCORE,
FullyConnectedOp<CUDAContext, TensorCoreEngine>);
REGISTER_CUDA_OPERATOR_WITH_ENGINE(
FCGradient,
TENSORCORE,
FullyConnectedGradientOp<CUDAContext, TensorCoreEngine>);
REGISTER_CUDA_OPERATOR_WITH_ENGINE(
FCTransposed,
TENSORCORE,
FullyConnectedOp<
CUDAContext,
TensorCoreEngine,
false /* don't transpose weight */>);
REGISTER_CUDA_OPERATOR_WITH_ENGINE(
FCTransposedGradient,
TENSORCORE,
FullyConnectedGradientOp<
CUDAContext,
TensorCoreEngine,
false /* don't transpose weight */>);
#endif
} // namespace caffe2
|
