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
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#include "paddle/fluid/operators/math/detail/gru_gpu_kernel.h"
#include "paddle/fluid/operators/math/detail/gru_kernel.h"
#include "paddle/fluid/operators/math/gru_compute.h"
#include "paddle/fluid/operators/math/math_function.h"
namespace paddle {
namespace operators {
namespace math {
template <typename T>
struct GRUUnitFunctor<platform::CUDADeviceContext, T> {
static void compute(const platform::CUDADeviceContext &context,
GRUMetaValue<T> value, int frame_size, int batch_size,
const detail::ActivationType active_node,
const detail::ActivationType active_gate) {
auto stream = context.stream();
dim3 threads;
dim3 grid;
if (batch_size == 1) {
int frame_per_block = frame_size <= 1024 ? frame_size : 1024;
int frame_blocks = (frame_size + 1024 - 1) / 1024;
threads = dim3(frame_per_block, 1);
grid = dim3(frame_blocks, 1);
} else {
threads = dim3(32, 32);
grid = dim3((frame_size + 32 - 1) / 32, (batch_size + 32 - 1) / 32);
}
if (value.prev_out_value) {
math::gemm<platform::CUDADeviceContext, T>(
context, false, false, batch_size, frame_size * 2, frame_size, 1,
value.prev_out_value, frame_size, value.gate_weight, frame_size * 2,
1, value.gate_value, frame_size * 3);
}
if (batch_size == 1) {
detail::KeGruForwardResetOutput<detail::forward::gru_resetOutput<T>,
/* is_batch= */ false,
T><<<grid, threads, 0, stream>>>(
detail::forward::gru_resetOutput<T>(), value.gate_value,
value.reset_output_value, value.prev_out_value, frame_size,
batch_size, active_gate);
} else {
detail::KeGruForwardResetOutput<detail::forward::gru_resetOutput<T>,
/* is_batch= */ true,
T><<<grid, threads, 0, stream>>>(
detail::forward::gru_resetOutput<T>(), value.gate_value,
value.reset_output_value, value.prev_out_value, frame_size,
batch_size, active_gate);
}
if (value.prev_out_value) {
math::gemm<platform::CUDADeviceContext, T>(
context, false, false, batch_size, frame_size, frame_size, 1,
value.reset_output_value, frame_size, value.state_weight, frame_size,
1, value.gate_value + frame_size * 2, frame_size * 3);
}
if (batch_size == 1) {
detail::KeGruForwardFinalOutput<detail::forward::gru_finalOutput<T>,
/* is_batch= */ false,
T><<<grid, threads, 0, stream>>>(
detail::forward::gru_finalOutput<T>(), value.gate_value,
value.prev_out_value, value.output_value, frame_size, batch_size,
active_node);
} else {
detail::KeGruForwardFinalOutput<detail::forward::gru_finalOutput<T>,
/* is_batch= */ true,
T><<<grid, threads, 0, stream>>>(
detail::forward::gru_finalOutput<T>(), value.gate_value,
value.prev_out_value, value.output_value, frame_size, batch_size,
active_node);
}
}
};
template <typename T>
struct GRUUnitGradFunctor<platform::CUDADeviceContext, T> {
static void compute(const platform::CUDADeviceContext &context,
GRUMetaValue<T> value, GRUMetaGrad<T> grad,
int frame_size, int batch_size,
const detail::ActivationType active_node,
const detail::ActivationType active_gate) {
auto stream = context.stream();
dim3 threads;
dim3 grid;
if (batch_size == 1) {
int frame_per_block = frame_size <= 1024 ? frame_size : 1024;
int frame_blocks = (frame_size + 1024 - 1) / 1024;
threads = dim3(frame_per_block, 1);
grid = dim3(frame_blocks, 1);
} else {
threads = dim3(32, 32);
grid = dim3((frame_size + 32 - 1) / 32, (batch_size + 32 - 1) / 32);
}
if (batch_size == 1) {
detail::KeGruBackwardStateGrad<
detail::backward::gru_stateGrad<T>,
/* is_batch= */ false><<<grid, threads, 0, stream>>>(
detail::backward::gru_stateGrad<T>(), value.gate_value,
grad.gate_grad, value.prev_out_value, grad.prev_out_grad,
grad.output_grad, frame_size, batch_size, active_node);
} else {
detail::KeGruBackwardStateGrad<
detail::backward::gru_stateGrad<T>,
/* is_batch= */ true><<<grid, threads, 0, stream>>>(
detail::backward::gru_stateGrad<T>(), value.gate_value,
grad.gate_grad, value.prev_out_value, grad.prev_out_grad,
grad.output_grad, frame_size, batch_size, active_node);
}
if (value.prev_out_value && grad.prev_out_grad) {
math::gemm<platform::CUDADeviceContext, T>(
context, false, true, batch_size, frame_size, frame_size, 1,
grad.gate_grad + frame_size * 2, frame_size * 3, value.state_weight,
frame_size, 0, grad.reset_output_grad, frame_size);
if (grad.state_weight_grad) {
math::gemm<platform::CUDADeviceContext, T>(
context, true, false, frame_size, frame_size, batch_size, 1,
value.reset_output_value, frame_size,
grad.gate_grad + frame_size * 2, frame_size * 3, 1,
grad.state_weight_grad, frame_size);
}
}
if (batch_size == 1) {
detail::KeGruBackwardResetGrad<
detail::backward::gru_resetGrad<T>,
/* is_batch= */ false><<<grid, threads, 0, stream>>>(
detail::backward::gru_resetGrad<T>(), value.gate_value,
grad.gate_grad, value.prev_out_value, grad.prev_out_grad,
grad.reset_output_grad, frame_size, batch_size, active_gate);
} else {
detail::KeGruBackwardResetGrad<
detail::backward::gru_resetGrad<T>,
/* is_batch= */ true><<<grid, threads, 0, stream>>>(
detail::backward::gru_resetGrad<T>(), value.gate_value,
grad.gate_grad, value.prev_out_value, grad.prev_out_grad,
grad.reset_output_grad, frame_size, batch_size, active_gate);
}
if (grad.prev_out_grad && value.prev_out_value) {
math::gemm<platform::CUDADeviceContext, T>(
context, false, true, batch_size, frame_size, frame_size * 2, 1,
grad.gate_grad, frame_size * 3, value.gate_weight, frame_size * 2, 1,
grad.prev_out_grad, frame_size);
if (grad.gate_weight_grad) {
math::gemm<platform::CUDADeviceContext, T>(
context, true, false, frame_size, frame_size * 2, batch_size, 1,
value.prev_out_value, frame_size, grad.gate_grad, frame_size * 3, 1,
grad.gate_weight_grad, frame_size * 2);
}
}
}
};
template struct GRUUnitFunctor<platform::CUDADeviceContext, float>;
template struct GRUUnitFunctor<platform::CUDADeviceContext, double>;
template struct GRUUnitGradFunctor<platform::CUDADeviceContext, float>;
template struct GRUUnitGradFunctor<platform::CUDADeviceContext, double>;
} // namespace math
} // namespace operators
} // namespace paddle