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
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
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. */
#pragma once
#include <math.h> // for sqrt in CPU and CUDA
#include <Eigen/Dense>
#include <vector>
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/operators/detail/safe_ref.h"
#include "paddle/fluid/operators/math/algorithm.h"
#include "paddle/fluid/operators/math/selected_rows_functor.h"
#include "paddle/fluid/platform/for_range.h"
namespace paddle {
namespace operators {
namespace scatter = paddle::operators::math::scatter;
struct GPUAdam;
struct CPUAdam;
template <typename T, typename Flavour>
struct AdamFunctor;
template <typename T>
struct AdamFunctor<T, GPUAdam> {
T beta1_;
T beta2_;
T epsilon_;
const T* beta1_pow_;
const T* beta2_pow_;
const T* moment1_;
T* moment1_out_;
const T* moment2_;
T* moment2_out_;
const T* lr_;
const T* grad_;
const T* param_;
T* param_out_;
AdamFunctor(T beta1, T beta2, T epsilon, const T* beta1_pow,
const T* beta2_pow, const T* mom1, T* mom1_out, const T* mom2,
T* mom2_out, const T* lr, const T* grad, const T* param,
T* param_out)
: beta1_(beta1),
beta2_(beta2),
epsilon_(epsilon),
beta1_pow_(beta1_pow),
beta2_pow_(beta2_pow),
moment1_(mom1),
moment1_out_(mom1_out),
moment2_(mom2),
moment2_out_(mom2_out),
lr_(lr),
grad_(grad),
param_(param),
param_out_(param_out) {}
inline HOSTDEVICE void operator()(size_t i) const {
// Merge all memory access together.
T g = grad_[i];
T mom1 = moment1_[i];
T mom2 = moment2_[i];
T lr = *lr_;
T beta1_pow = *beta1_pow_;
T beta2_pow = *beta2_pow_;
T p = param_[i];
// Calculation
lr *= sqrt(1 - beta2_pow) / (1 - beta1_pow);
mom1 = beta1_ * mom1 + (1 - beta1_) * g;
mom2 = beta2_ * mom2 + (1 - beta2_) * g * g;
p -= lr * (mom1 / (sqrt(mom2) + epsilon_));
// Write back to global memory
moment1_out_[i] = mom1;
moment2_out_[i] = mom2;
param_out_[i] = p;
}
};
template <typename T>
struct AdamFunctor<T, CPUAdam> {
T beta1_;
T beta2_;
T epsilon_;
const T* beta1_pow_;
const T* beta2_pow_;
const T* moment1_;
T* moment1_out_;
const T* moment2_;
T* moment2_out_;
const T* lr_;
const T* grad_;
const T* param_;
T* param_out_;
AdamFunctor(T beta1, T beta2, T epsilon, const T* beta1_pow,
const T* beta2_pow, const T* mom1, T* mom1_out, const T* mom2,
T* mom2_out, const T* lr, const T* grad, const T* param,
T* param_out)
: beta1_(beta1),
beta2_(beta2),
epsilon_(epsilon),
beta1_pow_(beta1_pow),
beta2_pow_(beta2_pow),
moment1_(mom1),
moment1_out_(mom1_out),
moment2_(mom2),
moment2_out_(mom2_out),
lr_(lr),
grad_(grad),
param_(param),
param_out_(param_out) {}
void operator()(size_t numel) const {
Eigen::Map<const Eigen::Array<T, 1, Eigen::Dynamic>> g{
grad_, static_cast<Eigen::Index>(numel)};
Eigen::Map<const Eigen::Array<T, 1, Eigen::Dynamic>> mom1{
moment1_, static_cast<Eigen::Index>(numel)};
Eigen::Map<const Eigen::Array<T, 1, Eigen::Dynamic>> mom2{
moment2_, static_cast<Eigen::Index>(numel)};
Eigen::Map<const Eigen::Array<T, 1, Eigen::Dynamic>> param{
param_, static_cast<Eigen::Index>(numel)};
Eigen::Map<Eigen::Array<T, 1, Eigen::Dynamic>> param_out{
param_out_, static_cast<Eigen::Index>(numel)};
Eigen::Map<Eigen::Array<T, 1, Eigen::Dynamic>> moment1_out{
moment1_out_, static_cast<Eigen::Index>(numel)};
Eigen::Map<Eigen::Array<T, 1, Eigen::Dynamic>> moment2_out{
moment2_out_, static_cast<Eigen::Index>(numel)};
T lr = *lr_;
T beta1_pow = *beta1_pow_;
T beta2_pow = *beta2_pow_;
// Calculation
lr *= sqrt(1 - beta2_pow) / (1 - beta1_pow);
moment1_out = beta1_ * mom1 + (1 - beta1_) * g;
moment2_out = beta2_ * mom2 + (1 - beta2_) * g * g;
param_out = param - lr * (moment1_out / (moment2_out.sqrt() + epsilon_));
}
};
template <typename T>
struct SparseAdamFunctor {
T beta1_;
T beta2_;
T epsilon_;
const T* beta1_pow_;
const T* beta2_pow_;
const T* moment1_;
T* moment1_out_;
const T* moment2_;
T* moment2_out_;
const T* lr_;
const T* grad_;
const T* param_;
T* param_out_;
const int64_t* rows_;
int64_t row_numel_;
int64_t row_count_;
SparseAdamFunctor(T beta1, T beta2, T epsilon, const T* beta1_pow,
const T* beta2_pow, const T* mom1, T* mom1_out,
const T* mom2, T* mom2_out, const T* lr, const T* grad,
const T* param, T* param_out, const int64_t* rows,
int64_t row_numel, int64_t row_count)
: beta1_(beta1),
beta2_(beta2),
epsilon_(epsilon),
beta1_pow_(beta1_pow),
beta2_pow_(beta2_pow),
moment1_(mom1),
moment1_out_(mom1_out),
moment2_(mom2),
moment2_out_(mom2_out),
lr_(lr),
grad_(grad),
param_(param),
param_out_(param_out),
rows_(rows),
row_numel_(row_numel),
row_count_(row_count) {}
inline HOSTDEVICE void operator()(size_t i) const {
auto row_idx =
math::BinarySearch<int64_t>(rows_, row_count_, i / row_numel_);
T g = row_idx >= 0 ? grad_[row_idx * row_numel_ + i % row_numel_] : 0;
// The following code is the same as dense
T mom1 = moment1_[i];
T mom2 = moment2_[i];
T lr = *lr_;
T beta1_pow = *beta1_pow_;
T beta2_pow = *beta2_pow_;
T p = param_[i];
// Calculation
lr *= sqrt(1 - beta2_pow) / (1 - beta1_pow);
mom1 = beta1_ * mom1 + (1 - beta1_) * g;
mom2 = beta2_ * mom2 + (1 - beta2_) * g * g;
p -= lr * (mom1 / (sqrt(mom2) + epsilon_));
// Write back to global memory
moment1_out_[i] = mom1;
moment2_out_[i] = mom2;
param_out_[i] = p;
}
};
template <typename DeviceContext, typename T>
class AdamOpKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext& ctx) const override {
const auto* param_var = ctx.InputVar("Param");
PADDLE_ENFORCE(param_var->IsType<framework::LoDTensor>(),
"The Var(%s)'s type should be LoDTensor, "
"but the received is %s",
ctx.Inputs("Param").front(), param_var->Type().name());
using paddle::framework::LoDTensor;
using paddle::operators::detail::Ref;
T beta1 = static_cast<T>(ctx.Attr<float>("beta1"));
T beta2 = static_cast<T>(ctx.Attr<float>("beta2"));
T epsilon = static_cast<T>(ctx.Attr<float>("epsilon"));
auto& param = Ref(ctx.Input<LoDTensor>("Param"), "Must set Param");
// auto& grad = Ref(ctx.Input<LoDTensor>("Grad"), "Must set Grad");
auto* grad_var = ctx.InputVar("Grad");
auto& mom1 = Ref(ctx.Input<LoDTensor>("Moment1"), "Must set Moment1");
auto& mom2 = Ref(ctx.Input<LoDTensor>("Moment2"), "Must set Moment2");
auto& lr =
Ref(ctx.Input<LoDTensor>("LearningRate"), "Must set LearningRate");
auto& beta1_pow =
Ref(ctx.Input<LoDTensor>("Beta1Pow"), "Must set Beta1Pow");
auto& beta2_pow =
Ref(ctx.Input<LoDTensor>("Beta2Pow"), "Must set Beta2Pow");
auto& param_out =
Ref(ctx.Output<LoDTensor>("ParamOut"), "Must set ParamOut");
auto& mom1_out =
Ref(ctx.Output<LoDTensor>("Moment1Out"), "Must set Moment1Out");
auto& mom2_out =
Ref(ctx.Output<LoDTensor>("Moment2Out"), "Must set Moment1Out");
if (grad_var->IsType<framework::LoDTensor>()) {
auto& grad = Ref(ctx.Input<LoDTensor>("Grad"), "Must set Grad");
if (platform::is_cpu_place(ctx.GetPlace())) {
AdamFunctor<T, CPUAdam> functor(
beta1, beta2, epsilon, beta1_pow.template data<T>(),
beta2_pow.template data<T>(), mom1.template data<T>(),
mom1_out.template mutable_data<T>(ctx.GetPlace()),
mom2.template data<T>(),
mom2_out.template mutable_data<T>(ctx.GetPlace()),
lr.template data<T>(), grad.template data<T>(),
param.template data<T>(),
param_out.template mutable_data<T>(ctx.GetPlace()));
functor(param.numel());
} else if (platform::is_gpu_place(ctx.GetPlace())) {
AdamFunctor<T, GPUAdam> functor(
beta1, beta2, epsilon, beta1_pow.template data<T>(),
beta2_pow.template data<T>(), mom1.template data<T>(),
mom1_out.template mutable_data<T>(ctx.GetPlace()),
mom2.template data<T>(),
mom2_out.template mutable_data<T>(ctx.GetPlace()),
lr.template data<T>(), grad.template data<T>(),
param.template data<T>(),
param_out.template mutable_data<T>(ctx.GetPlace()));
platform::ForRange<DeviceContext> for_range(
static_cast<const DeviceContext&>(ctx.device_context()),
param.numel());
for_range(functor);
}
} else if (grad_var->IsType<framework::SelectedRows>()) {
auto& grad =
Ref(ctx.Input<framework::SelectedRows>("Grad"), "Must set Grad");
if (grad.rows().size() == 0) {
VLOG(30) << "grad row size is 0!!";
return;
}
std::vector<int64_t> cpu_rows(grad.rows().begin(), grad.rows().end());
bool is_strict_sorted = true;
for (size_t i = 1; i < cpu_rows.size(); ++i) {
if (cpu_rows[i - 1] >= cpu_rows[i]) {
is_strict_sorted = false;
break;
}
}
const framework::SelectedRows* grad_merge_ptr;
if (is_strict_sorted) {
grad_merge_ptr = &grad;
} else {
// merge duplicated rows if any.
// The rows of grad_merge have been sorted inside MergeAdd functor
scatter::MergeAdd<DeviceContext, T> merge_func;
auto* grad_merge_var = const_cast<framework::Scope&>(ctx.scope())
.Var()
->GetMutable<framework::SelectedRows>();
merge_func(ctx.template device_context<DeviceContext>(), grad,
grad_merge_var);
grad_merge_ptr = grad_merge_var;
}
auto& grad_merge = *grad_merge_ptr;
auto& grad_tensor = grad_merge.value();
const T* grad_data = grad_tensor.template data<T>();
const int64_t* rows = nullptr;
// When compiled without CUDA, the CUDAData() interface should not be
// provided.
#if defined(PADDLE_WITH_CUDA)
if (platform::is_gpu_place(ctx.GetPlace())) {
rows = grad_merge.rows().CUDAData(ctx.GetPlace());
} else {
#endif
rows = grad_merge.rows().data();
#if defined(PADDLE_WITH_CUDA)
}
#endif
auto row_numel = grad_tensor.numel() / grad_merge.rows().size();
SparseAdamFunctor<T> functor(
beta1, beta2, epsilon, beta1_pow.template data<T>(),
beta2_pow.template data<T>(), mom1.template data<T>(),
mom1_out.template mutable_data<T>(ctx.GetPlace()),
mom2.template data<T>(),
mom2_out.template mutable_data<T>(ctx.GetPlace()),
lr.template data<T>(), grad_data, param.template data<T>(),
param_out.template mutable_data<T>(ctx.GetPlace()), rows, row_numel,
grad_merge.rows().size());
platform::ForRange<DeviceContext> for_range(
static_cast<const DeviceContext&>(ctx.device_context()),
param.numel());
for_range(functor);
} else {
PADDLE_THROW("Variable type not supported by adam_op");
}
}
};
} // namespace operators
} // namespace paddle