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
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
|
#include <algorithm>
#include <cmath>
#include <vector>
#include "caffe2/core/context_gpu.h"
#include "caffe2/operators/lstm_unit_op.h"
namespace caffe2 {
namespace detail {
template <typename Dtype>
__device__ Dtype cuda_sigmoid(const Dtype x) {
return Dtype(1) / (Dtype(1) + exp(-x));
}
template <typename T, typename MATH>
__global__ void LSTMUnitKernel(
const int nthreads,
const int dim,
const int t,
const T* H_prev,
const T* C_prev,
const T* X,
const int32_t* seqLengths,
bool drop_states,
T* C,
T* H,
const MATH forget_bias) {
CUDA_1D_KERNEL_LOOP(index, nthreads) {
const int n = index / dim;
const int d = index % dim;
const bool valid = seqLengths == nullptr || t < seqLengths[n];
if (!valid) {
H[index] = convert::To<MATH, T>(convert::To<T, MATH>(H_prev[index]) * !drop_states);
C[index] = convert::To<MATH, T>(convert::To<T, MATH>(C_prev[index]) * !drop_states);
} else {
const T* X_offset = X + 4 * dim * n;
const MATH i = cuda_sigmoid(convert::To<T, MATH>(X_offset[d]));
const MATH f = cuda_sigmoid(convert::To<T, MATH>(X_offset[1 * dim + d]) + forget_bias);
const MATH o = cuda_sigmoid(convert::To<T, MATH>(X_offset[2 * dim + d]));
const MATH g = tanh(convert::To<T, MATH>(X_offset[3 * dim + d]));
const MATH c_prev = convert::To<T, MATH>(C_prev[index]);
const MATH c = f * c_prev + i * g;
C[index] = convert::To<MATH, T>(c);
const MATH tanh_c = tanh(c);
H[index] = convert::To<MATH, T>(o * tanh_c);
}
}
}
template <typename T, typename MATH>
__global__ void LSTMUnitGradientKernel(
const int nthreads,
const int dim,
const int t,
const T* C_prev,
const T* X,
const T* C,
const T* H,
const int32_t* seqLengths,
const T* C_diff,
const T* H_diff,
bool drop_states,
T* H_prev_diff,
T* C_prev_diff,
T* X_diff,
const MATH forget_bias) {
CUDA_1D_KERNEL_LOOP(index, nthreads) {
const int n = index / dim;
const bool valid = seqLengths == nullptr || t < seqLengths[n];
const int d = index % dim;
const T* X_offset = X + 4 * dim * n;
T* c_prev_diff = C_prev_diff + index;
T* h_prev_diff = H_prev_diff + index;
T* X_diff_offset = X_diff + 4 * dim * n;
T* i_diff = X_diff_offset + d;
T* f_diff = X_diff_offset + 1 * dim + d;
T* o_diff = X_diff_offset + 2 * dim + d;
T* g_diff = X_diff_offset + 3 * dim + d;
if (!valid) {
*h_prev_diff = convert::To<MATH, T>(convert::To<T, MATH>(H_diff[index]) *
!drop_states);
*c_prev_diff = convert::To<MATH, T>(convert::To<T, MATH>(C_diff[index]) *
!drop_states);
*i_diff = convert::To<MATH, T>(0);
*f_diff = convert::To<MATH, T>(0);
*o_diff = convert::To<MATH, T>(0);
*g_diff = convert::To<MATH, T>(0);
} else {
const MATH i = cuda_sigmoid(convert::To<T, MATH>(X_offset[d]));
const MATH f = cuda_sigmoid(convert::To<T, MATH>(X_offset[1 * dim + d]) + forget_bias);
const MATH o = cuda_sigmoid(convert::To<T, MATH>(X_offset[2 * dim + d]));
const MATH g = tanh(convert::To<T, MATH>(X_offset[3 * dim + d]));
const MATH c_prev = convert::To<T, MATH>(C_prev[index]);
const MATH c = convert::To<T, MATH>(C[index]);
const MATH tanh_c = tanh(c);
const MATH c_term_diff =
convert::To<T, MATH>(C_diff[index]) +
convert::To<T, MATH>(H_diff[index]) * o * (1 - tanh_c * tanh_c);
*c_prev_diff = convert::To<MATH, T>(c_term_diff * f);
*h_prev_diff = convert::To<MATH, T>(0);
*i_diff = convert::To<MATH, T>(c_term_diff * g * i * (1 - i));
*f_diff = convert::To<MATH, T>(c_term_diff * c_prev * f * (1 - f));
*o_diff = convert::To<MATH, T>(
convert::To<T, MATH>(H_diff[index]) * tanh_c * o * (1 - o));
*g_diff = convert::To<MATH, T>(c_term_diff * i * (1 - g * g));
}
}
}
template <>
void LSTMUnit<float, CUDAContext>(
int N,
int D,
int t,
const float* H_prev,
const float* C_prev,
const float* X,
const int32_t* seqLengths,
bool drop_states,
float* C,
float* H,
const float forget_bias,
CUDAContext* context) {
LSTMUnitKernel<float, float><<<
CAFFE_GET_BLOCKS(N * D),
CAFFE_CUDA_NUM_THREADS,
0,
context->cuda_stream()>>>(
N * D,
D,
t,
H_prev,
C_prev,
X,
seqLengths,
drop_states,
C,
H,
forget_bias);
C10_CUDA_KERNEL_LAUNCH_CHECK();
}
template <>
void LSTMUnit<at::Half, CUDAContext>(
int N,
int D,
int t,
const at::Half* H_prev,
const at::Half* C_prev,
const at::Half* X,
const int32_t* seqLengths,
bool drop_states,
at::Half* C,
at::Half* H,
const float forget_bias,
CUDAContext* context) {
LSTMUnitKernel<at::Half, float><<<
CAFFE_GET_BLOCKS(N * D),
CAFFE_CUDA_NUM_THREADS,
0,
context->cuda_stream()>>>(
N * D,
D,
t,
H_prev,
C_prev,
X,
seqLengths,
drop_states,
C,
H,
forget_bias);
C10_CUDA_KERNEL_LAUNCH_CHECK();
}
template <>
void LSTMUnitGradient<float, CUDAContext>(
int N,
int D,
int t,
const float* C_prev,
const float* X,
const int32_t* seqLengths,
const float* C,
const float* H,
const float* C_diff,
const float* H_diff,
bool drop_states,
float* H_prev_diff,
float* C_prev_diff,
float* X_diff,
const float forget_bias,
CUDAContext* context) {
LSTMUnitGradientKernel<float, float><<<
CAFFE_GET_BLOCKS(N * D),
CAFFE_CUDA_NUM_THREADS,
0,
context->cuda_stream()>>>(
N * D,
D,
t,
C_prev,
X,
C,
H,
seqLengths,
C_diff,
H_diff,
drop_states,
H_prev_diff,
C_prev_diff,
X_diff,
forget_bias);
C10_CUDA_KERNEL_LAUNCH_CHECK();
}
template <>
void LSTMUnitGradient<at::Half, CUDAContext>(
int N,
int D,
int t,
const at::Half* C_prev,
const at::Half* X,
const int32_t* seqLengths,
const at::Half* C,
const at::Half* H,
const at::Half* C_diff,
const at::Half* H_diff,
bool drop_states,
at::Half* H_prev_diff,
at::Half* C_prev_diff,
at::Half* X_diff,
const float forget_bias,
CUDAContext* context) {
LSTMUnitGradientKernel<at::Half, float><<<
CAFFE_GET_BLOCKS(N * D),
CAFFE_CUDA_NUM_THREADS,
0,
context->cuda_stream()>>>(
N * D,
D,
t,
C_prev,
X,
C,
H,
seqLengths,
C_diff,
H_diff,
drop_states,
H_prev_diff,
C_prev_diff,
X_diff,
forget_bias);
C10_CUDA_KERNEL_LAUNCH_CHECK();
}
}
template <>
bool LSTMUnitOp<CUDAContext>::RunOnDevice() {
return DispatchHelper<TensorTypes<float, at::Half>>::call(this, Input(0));
}
template <>
bool LSTMUnitGradientOp<CUDAContext>::RunOnDevice() {
return DispatchHelper<TensorTypes<float, at::Half>>::call(this, Input(0));
}
REGISTER_CUDA_OPERATOR(LSTMUnit, LSTMUnitOp<CUDAContext>);
REGISTER_CUDA_OPERATOR(
LSTMUnitGradient,
LSTMUnitGradientOp<CUDAContext>);
}
|
