activation_op.h 65.7 KB
Newer Older
1
/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
2

L
Luo Tao 已提交
3 4 5 6 7 8 9 10 11
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. */
Q
qijun 已提交
12 13

#pragma once
D
dzhwinter 已提交
14
#include <glog/logging.h>
Y
Yihua Xu 已提交
15
#include <algorithm>
16
#include <memory>
D
dzhwinter 已提交
17 18
#include <string>
#include <unordered_set>
19 20
#include <utility>
#include <vector>
21

C
Clementine 已提交
22 23 24 25 26
#include <cmath>
#ifndef _USE_MATH_DEFINES
#define _USE_MATH_DEFINES
#endif

Y
Yi Wang 已提交
27 28
#include "paddle/fluid/framework/eigen.h"
#include "paddle/fluid/framework/op_registry.h"
Y
Yihua Xu 已提交
29
#include "paddle/fluid/operators/math/blas.h"
30
#include "paddle/fluid/platform/enforce.h"
31
#include "paddle/fluid/platform/float16.h"
Q
qijun 已提交
32

33 34 35 36
#ifdef PADDLE_WITH_MKLDNN
#include "paddle/fluid/platform/mkldnn_helper.h"
#endif

Q
qijun 已提交
37 38 39
namespace paddle {
namespace operators {

40 41
using framework::To32BitIndex;

42 43 44 45 46 47
enum ActBwdOpFwdDeps {
  kNoDeps = 0x00,  // Do not need any forward input/output
  kDepX = 0x01,    // Only need forward input X
  kDepOut = 0x02,  // Only need forward output Out
};

C
chengduo 已提交
48 49 50 51 52 53
/* The following operator can be used to process SelectedRows, because the
 * output of those operator for zero is zero too.
 */
static std::unordered_set<std::string> CanBeUsedBySelectedRows = {
    "abs", "abs_grad", "square", "square_grad", "sqrt", "sqrt_grad"};

54 55 56 57 58
inline void ExtractActivationTensor(const framework::ExecutionContext& context,
                                    const framework::Tensor** X,
                                    framework::Tensor** Out) {
  auto x_var = context.InputVar("X");
  auto out_var = context.OutputVar("Out");
59 60 61 62 63 64 65 66
  PADDLE_ENFORCE_NOT_NULL(x_var,
                          platform::errors::NotFound(
                              "Cannot get input Variable X, variable name = %s",
                              context.InputName("X")));
  PADDLE_ENFORCE_NOT_NULL(
      out_var, platform::errors::NotFound(
                   "Cannot get output Variable Out, variable name = %s",
                   context.OutputName("Out")));
H
hong 已提交
67
  if (CanBeUsedBySelectedRows.count(context.Type())) {
68 69 70 71 72 73 74 75
    *X = paddle::framework::GetLoDTensorOrSelectedRowsValueFromVar(*x_var);
    *Out = paddle::framework::GetMutableLoDTensorOrSelectedRowsValueFromVar(
        out_var);
  } else {
    *X = context.Input<framework::Tensor>("X");
    *Out = context.Output<framework::Tensor>("Out");
  }

76 77 78 79
  PADDLE_ENFORCE_NOT_NULL(*Out, platform::errors::NotFound(
                                    "Cannot get the tensor from the Variable "
                                    "Output(Out), variable name = %s",
                                    context.OutputName("Out")));
80 81
}

82
template <ActBwdOpFwdDeps kDepValue>
83 84 85 86 87 88
inline void ExtractActivationGradTensor(
    const framework::ExecutionContext& context, const framework::Tensor** X,
    const framework::Tensor** Out, const framework::Tensor** dOut,
    framework::Tensor** dX) {
  auto out_grad_var = context.InputVar(framework::GradVarName("Out"));
  auto x_grad_var = context.OutputVar(framework::GradVarName("X"));
89 90 91 92
  const framework::Variable* out_var = nullptr;

  if (static_cast<int>(kDepValue) & static_cast<int>(kDepOut)) {
    out_var = context.InputVar("Out");
93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108
    PADDLE_ENFORCE_NOT_NULL(
        out_var, platform::errors::NotFound(
                     "Cannot get input Variable Out, variable name = %s",
                     context.InputName("Out")));
  }

  PADDLE_ENFORCE_NOT_NULL(
      out_grad_var, platform::errors::NotFound(
                        "Cannot get input Variable %s, variable name = %s",
                        framework::GradVarName("Out"),
                        context.InputName(framework::GradVarName("Out"))));
  PADDLE_ENFORCE_NOT_NULL(
      x_grad_var, platform::errors::NotFound(
                      "Cannot get output Variable %s, variable name = %s",
                      framework::GradVarName("X"),
                      context.OutputName(framework::GradVarName("X"))));
109

H
hong 已提交
110
  if (CanBeUsedBySelectedRows.count(context.Type())) {
111 112 113 114
    *dOut = paddle::framework::GetLoDTensorOrSelectedRowsValueFromVar(
        *out_grad_var);
    *dX = paddle::framework::GetMutableLoDTensorOrSelectedRowsValueFromVar(
        x_grad_var);
115 116 117 118 119 120 121 122

    if (out_var) {
      *Out =
          paddle::framework::GetLoDTensorOrSelectedRowsValueFromVar(*out_var);
    } else {
      *Out = *dOut;  // fake out
    }

123 124 125 126
  } else {
    *Out = context.Input<framework::Tensor>("Out");
    *dOut = context.Input<framework::Tensor>(framework::GradVarName("Out"));
    *dX = context.Output<framework::Tensor>(framework::GradVarName("X"));
127 128 129 130 131 132

    if (out_var) {
      *Out = &(out_var->Get<framework::LoDTensor>());
    } else {
      *Out = *dOut;  // fake out
    }
133
  }
134

135 136 137 138 139
  PADDLE_ENFORCE_NOT_NULL(*dX,
                          platform::errors::NotFound(
                              "Cannot get the tensor from the Variable "
                              "Output(Out), variable name = %s",
                              context.OutputName(framework::GradVarName("X"))));
140

141
  if (static_cast<int>(kDepValue) & static_cast<int>(kDepX)) {
C
chengduo 已提交
142
    auto x_var = context.InputVar("X");
143 144 145 146
    PADDLE_ENFORCE_NOT_NULL(x_var, platform::errors::NotFound(
                                       "Cannot get the tensor from the "
                                       "Variable Input(X), variable name = %s",
                                       context.InputName("X")));
H
hong 已提交
147
    if (CanBeUsedBySelectedRows.count(context.Type())) {
148
      *X = paddle::framework::GetLoDTensorOrSelectedRowsValueFromVar(*x_var);
C
chengduo 已提交
149
    } else {
150
      *X = context.Input<framework::Tensor>("X");
C
chengduo 已提交
151
    }
152
  } else {
H
hong 已提交
153
    VLOG(10) << " Inplace activation of Op : " << context.Type();
154 155 156
    *X = *dX;
  }
}
C
chengduo 已提交
157

158 159 160 161 162
template <typename DeviceContext, typename Functor>
class ActivationKernel
    : public framework::OpKernel<typename Functor::ELEMENT_TYPE> {
 public:
  using T = typename Functor::ELEMENT_TYPE;
C
chengduo 已提交
163

164 165 166 167
  void Compute(const framework::ExecutionContext& context) const override {
    const framework::Tensor* X = nullptr;
    framework::Tensor* Out = nullptr;
    ExtractActivationTensor(context, &X, &Out);
C
chengduo 已提交
168
    Out->mutable_data<T>(context.GetPlace());
169

170 171 172 173
    auto x = framework::EigenVector<T>::Flatten(
        GET_DATA_SAFELY(X, "Input", "X", "Activation"));
    auto out = framework::EigenVector<T>::Flatten(
        GET_DATA_SAFELY(Out, "Output", "Out", "Activation"));
Q
QI JUN 已提交
174 175
    auto* place =
        context.template device_context<DeviceContext>().eigen_device();
Q
qijun 已提交
176
    Functor functor;
177 178 179 180 181

    auto attrs = functor.GetAttrs();
    for (auto& attr : attrs) {
      *attr.second = context.Attr<float>(attr.first);
    }
182 183 184 185 186 187 188 189
    // use 32bit index to speed up computation
    bool use_32bit_index = out.size() < Eigen::NumTraits<int>::highest();
    bool is_gpu_place = platform::is_gpu_place(context.GetPlace());
    if (use_32bit_index && is_gpu_place) {
      functor(*place, To32BitIndex(x), To32BitIndex(out));
    } else {
      functor(*place, x, out);
    }
Q
qijun 已提交
190 191 192
  }
};

Q
QI JUN 已提交
193
template <typename DeviceContext, typename Functor>
194 195
class ActivationGradKernel
    : public framework::OpKernel<typename Functor::ELEMENT_TYPE> {
Q
qijun 已提交
196
 public:
197
  using T = typename Functor::ELEMENT_TYPE;
Q
qijun 已提交
198
  void Compute(const framework::ExecutionContext& context) const override {
199 200 201
    const framework::Tensor *X, *Out, *dOut;
    framework::Tensor* dX = nullptr;
    X = Out = dOut = nullptr;
202 203
    ExtractActivationGradTensor<Functor::FwdDeps()>(context, &X, &Out, &dOut,
                                                    &dX);
Q
qijun 已提交
204
    dX->mutable_data<T>(context.GetPlace());
205 206 207 208 209 210 211 212
    auto dout = framework::EigenVector<T>::Flatten(
        GET_DATA_SAFELY(dOut, "Input", "Out@GRAD", "ActivationGrad"));
    auto out = framework::EigenVector<T>::Flatten(
        GET_DATA_SAFELY(Out, "Input", "Out", "ActivationGrad"));
    auto dx = framework::EigenVector<T>::Flatten(
        GET_DATA_SAFELY(dX, "Input", "X@GRAD", "ActivationGrad"));
    auto x = framework::EigenVector<T>::Flatten(
        GET_DATA_SAFELY(X, "Input", "X", "ActivationGrad"));
Q
QI JUN 已提交
213 214
    auto* place =
        context.template device_context<DeviceContext>().eigen_device();
Q
qijun 已提交
215
    Functor functor;
216 217 218 219
    auto attrs = functor.GetAttrs();
    for (auto& attr : attrs) {
      *attr.second = context.Attr<float>(attr.first);
    }
220 221 222 223 224 225 226 227 228
    // use 32bit index to speed up computation
    bool use_32bit_index = out.size() < Eigen::NumTraits<int>::highest();
    bool is_gpu_place = platform::is_gpu_place(context.GetPlace());
    if (use_32bit_index && is_gpu_place) {
      functor(*place, To32BitIndex(x), To32BitIndex(out), To32BitIndex(dout),
              To32BitIndex(dx));
    } else {
      functor(*place, x, out, dout, dx);
    }
Q
qijun 已提交
229 230 231
  }
};

232 233 234 235 236 237 238 239 240
template <typename T>
struct BaseActivationFunctor {
  using ELEMENT_TYPE = T;

  using AttrPair = std::vector<std::pair<const char*, float*>>;

  AttrPair GetAttrs() { return AttrPair(); }
};

241
// sigmoid(x) = 1 / (1 + exp(-x))
Q
qijun 已提交
242
template <typename T>
243
struct SigmoidFunctor : public BaseActivationFunctor<T> {
F
fengjiayi 已提交
244 245 246
  template <typename Device, typename X, typename Out>
  void operator()(Device d, X x, Out out) const {
    out.device(d) = static_cast<T>(1) / (static_cast<T>(1) + (-x).exp());
Q
qijun 已提交
247 248 249
  }
};

250
template <typename T>
251
struct SigmoidGradFunctor : public BaseActivationFunctor<T> {
F
fengjiayi 已提交
252 253 254 255
  template <typename Device, typename X, typename Out, typename dOut,
            typename dX>
  void operator()(Device d, X x, Out out, dOut dout, dX dx) const {
    dx.device(d) = dout * out * (static_cast<T>(1) - out);
Q
qijun 已提交
256
  }
257 258

  static constexpr ActBwdOpFwdDeps FwdDeps() { return kDepOut; }
Q
qijun 已提交
259 260
};

261 262 263 264
// Originally: logsigmoid(x) = -log (1 + exp(-x))
// For numerical stability, we can use the log-sum-exp trick:
// https://hips.seas.harvard.edu/blog/2013/01/09/computing-log-sum-exp/
// We can rewrite the above equation as:
F
fengjiayi 已提交
265
// out = -log( exp(0) + exp(-x)) [since exp(0) = 1]
266 267 268 269 270 271 272 273 274 275
//   = -log( exp(max(-x, 0) - max(-x, 0)) + exp(-x + max(-x, 0) - max(-x, 0)))
//   = -log( exp(max(-x, 0)) * exp(-max(-x, 0)) - exp(max(-x, 0)) * exp(-x -
//           max(-x, 0)))
//   = -log( exp(max(-x, 0)) * (exp(-max(-x, 0)) + exp(-x - max(-x, 0))))
//   = -log( exp(max(-x, 0)) - log(exp(-max(-x, 0)) + exp(-x - max(-x, 0)))
//
// Hence, logsigmoid(x) = - (max(-x, 0) + log(exp(-max(-x, 0))
// + exp(-x - max(-x, 0))))
template <typename T>
struct LogSigmoidFunctor : public BaseActivationFunctor<T> {
F
fengjiayi 已提交
276 277
  template <typename Device, typename X, typename Out>
  void operator()(Device d, X x, Out out) const {
278
    auto temp = (-x).cwiseMax(static_cast<T>(0));  // temp = max(-x, 0)
F
fengjiayi 已提交
279
    out.device(d) = -temp - (((-temp).exp() + (-x - temp).exp()).log());
280 281 282 283 284 285 286 287
  }
};

// Originally: f' = exp(-x) / (1 + exp(-x))
// For numerical stability: f' = exp(-x - max(-x, 0)) / (exp(-max(-x, 0)) +
// exp(-x - max(-x, 0)))
template <typename T>
struct LogSigmoidGradFunctor : public BaseActivationFunctor<T> {
F
fengjiayi 已提交
288 289 290
  template <typename Device, typename X, typename Out, typename dOut,
            typename dX>
  void operator()(Device d, X x, Out out, dOut dout, dX dx) const {
291 292
    auto temp = (-x).cwiseMax(static_cast<T>(0));  // temp = max(-x, 0)
    dx.device(d) =
F
fengjiayi 已提交
293
        dout * ((-x - temp).exp() / ((-temp).exp() + (-x - temp).exp()));
294
  }
295 296

  static constexpr ActBwdOpFwdDeps FwdDeps() { return kDepX; }
297 298
};

Q
qijun 已提交
299
// exp(x) = e^x
300 301
template <typename T>
struct ExpFunctor : public BaseActivationFunctor<T> {
F
fengjiayi 已提交
302 303 304
  template <typename Device, typename X, typename Out>
  void operator()(Device d, X x, Out out) const {
    out.device(d) = x.exp();
Q
qijun 已提交
305 306 307
  }
};

308 309
template <typename T>
struct ExpGradFunctor : public BaseActivationFunctor<T> {
F
fengjiayi 已提交
310 311 312 313
  template <typename Device, typename X, typename Out, typename dOut,
            typename dX>
  void operator()(Device d, X x, Out out, dOut dout, dX dx) const {
    dx.device(d) = dout * out;
Q
qijun 已提交
314
  }
315 316

  static constexpr ActBwdOpFwdDeps FwdDeps() { return kDepOut; }
Q
qijun 已提交
317 318
};

Q
qijun 已提交
319
// relu(x) = max(x, 0)
Q
qijun 已提交
320
template <typename T>
321
struct ReluFunctor : public BaseActivationFunctor<T> {
F
fengjiayi 已提交
322 323 324
  template <typename Device, typename X, typename Out>
  void operator()(Device d, X x, Out out) const {
    out.device(d) = x.cwiseMax(static_cast<T>(0));
Q
qijun 已提交
325 326
  }
};
Q
qijun 已提交
327

Q
qijun 已提交
328
template <typename T>
329
struct ReluGradFunctor : public BaseActivationFunctor<T> {
F
fengjiayi 已提交
330 331 332
  template <typename Device, typename X, typename Out, typename dOut,
            typename dX>
  void operator()(Device d, X x, Out out, dOut dout, dX dx) const {
D
dzhwinter 已提交
333
    dx.device(d) = dout * (out > static_cast<T>(0)).template cast<T>();
Q
qijun 已提交
334
  }
335 336

  static constexpr ActBwdOpFwdDeps FwdDeps() { return kDepOut; }
Q
qijun 已提交
337
};
Q
qijun 已提交
338

339
// tanh(x) = (exp(x) - exp(-x)) / (exp(x) + exp(-x))
340 341
template <typename T>
struct TanhFunctor : public BaseActivationFunctor<T> {
F
fengjiayi 已提交
342 343 344
  template <typename Device, typename X, typename Out>
  void operator()(Device d, X x, Out out) const {
    out.device(d) = x.tanh();
Q
qijun 已提交
345 346 347 348
  }
};

template <typename T>
349
struct TanhGradFunctor : public BaseActivationFunctor<T> {
F
fengjiayi 已提交
350 351 352 353
  template <typename Device, typename X, typename Out, typename dOut,
            typename dX>
  void operator()(Device d, X x, Out out, dOut dout, dX dx) const {
    dx.device(d) = dout * (static_cast<T>(1) - out * out);
Q
qijun 已提交
354
  }
355 356

  static constexpr ActBwdOpFwdDeps FwdDeps() { return kDepOut; }
Q
qijun 已提交
357 358
};

K
Kavya Srinet 已提交
359 360 361 362
// tanhshrink(x) = x - tanh(x)
// where tanh(x) = (exp(x) - exp(-x)) / (exp(x) + exp(-x))
template <typename T>
struct TanhShrinkFunctor : public BaseActivationFunctor<T> {
F
fengjiayi 已提交
363 364 365
  template <typename Device, typename X, typename Out>
  void operator()(Device d, X x, Out out) const {
    out.device(d) = x - x.tanh();
K
Kavya Srinet 已提交
366 367 368 369 370
  }
};

template <typename T>
struct TanhShrinkGradFunctor : public BaseActivationFunctor<T> {
F
fengjiayi 已提交
371 372 373 374
  template <typename Device, typename X, typename Out, typename dOut,
            typename dX>
  void operator()(Device d, X x, Out out, dOut dout, dX dx) const {
    dx.device(d) = dout * (x.tanh() * x.tanh());
K
Kavya Srinet 已提交
375
  }
376 377

  static constexpr ActBwdOpFwdDeps FwdDeps() { return kDepX; }
K
Kavya Srinet 已提交
378 379
};

380 381 382 383 384 385 386 387 388
// tanhshrink(x) = x - tanh(x)
// where tanh(x) = (exp(x) - exp(-x)) / (exp(x) + exp(-x))
template <typename T>
struct HardShrinkFunctor : public BaseActivationFunctor<T> {
  float threshold;

  typename BaseActivationFunctor<T>::AttrPair GetAttrs() {
    return {{"threshold", &threshold}};
  }
F
fengjiayi 已提交
389 390
  template <typename Device, typename X, typename Out>
  void operator()(Device d, X x, Out out) const {
391 392 393
    auto temp1 = x < static_cast<T>(threshold * -1.f);
    auto temp2 = x > static_cast<T>(threshold);
    out.device(d) = x * (temp1 + temp2 > 0).template cast<T>();
394 395 396 397 398 399 400 401 402 403 404
  }
};

template <typename T>
struct HardShrinkGradFunctor : public BaseActivationFunctor<T> {
  float threshold;

  typename BaseActivationFunctor<T>::AttrPair GetAttrs() {
    return {{"threshold", &threshold}};
  }

F
fengjiayi 已提交
405 406 407
  template <typename Device, typename X, typename Out, typename dOut,
            typename dX>
  void operator()(Device d, X x, Out out, dOut dout, dX dx) const {
408 409 410
    auto temp1 = x < static_cast<T>(threshold * -1.f);
    auto temp2 = x > static_cast<T>(threshold);
    dx.device(d) = dout * (temp1 + temp2 > 0).template cast<T>();
411
  }
412 413

  static constexpr ActBwdOpFwdDeps FwdDeps() { return kDepX; }
414 415
};

K
Kexin Zhao 已提交
416
// softshrink(x) = x - lambda, if x > lambda; x + lambda, if x < -lambda; 0
417 418 419 420 421 422 423 424
// otherwise
template <typename T>
struct SoftShrinkFunctor : public BaseActivationFunctor<T> {
  float lambda;
  typename BaseActivationFunctor<T>::AttrPair GetAttrs() {
    return {{"lambda", &lambda}};
  }

F
fengjiayi 已提交
425 426
  template <typename Device, typename X, typename Out>
  void operator()(Device d, X x, Out out) const {
Y
Yu Yang 已提交
427
    auto lambdaT = static_cast<T>(lambda);
Z
Zeng Jinle 已提交
428 429
    auto temp1 = (x > lambdaT).template cast<T>();
    auto temp2 = (x < -lambdaT).template cast<T>();
F
fengjiayi 已提交
430
    out.device(d) = temp1 * (x - lambdaT) + temp2 * (x + lambdaT);
431 432 433 434 435 436 437 438 439
  }
};

template <typename T>
struct SoftShrinkGradFunctor : public BaseActivationFunctor<T> {
  float lambda;
  typename BaseActivationFunctor<T>::AttrPair GetAttrs() {
    return {{"lambda", &lambda}};
  }
F
fengjiayi 已提交
440 441 442
  template <typename Device, typename X, typename Out, typename dOut,
            typename dX>
  void operator()(Device d, X x, Out out, dOut dout, dX dx) const {
Y
Yu Yang 已提交
443
    auto lambdaT = static_cast<T>(lambda);
Z
Zeng Jinle 已提交
444 445
    auto temp1 = (x > lambdaT).template cast<T>();
    auto temp2 = (x < -lambdaT).template cast<T>();
F
fengjiayi 已提交
446
    dx.device(d) = dout * (temp1 + temp2).template cast<T>();
447
  }
448 449

  static constexpr ActBwdOpFwdDeps FwdDeps() { return kDepX; }
450 451
};

Q
qijun 已提交
452
// sqrt(x) = x^(1/2)
453 454
template <typename T>
struct SqrtFunctor : public BaseActivationFunctor<T> {
F
fengjiayi 已提交
455 456 457
  template <typename Device, typename X, typename Out>
  void operator()(Device d, X x, Out out) const {
    out.device(d) = x.sqrt();
Q
qijun 已提交
458 459 460 461
  }
};

template <typename T>
462
struct SqrtGradFunctor : public BaseActivationFunctor<T> {
F
fengjiayi 已提交
463 464 465
  template <typename Device, typename X, typename Out, typename dOut,
            typename dX>
  void operator()(Device d, X x, Out out, dOut dout, dX dx) const {
C
chengduo 已提交
466
    dx.device(d) = static_cast<T>(0.5) * dout / out;
Q
qijun 已提交
467
  }
468 469

  static constexpr ActBwdOpFwdDeps FwdDeps() { return kDepOut; }
Q
qijun 已提交
470 471
};

Z
zhoukunsheng 已提交
472 473 474 475 476 477 478 479 480 481 482 483 484 485
// rsqrt(x) = x^(-1/2)
template <typename T>
struct RsqrtFunctor : public BaseActivationFunctor<T> {
  template <typename Device, typename X, typename Out>
  void operator()(Device d, X x, Out out) const {
    out.device(d) = x.rsqrt();
  }
};

template <typename T>
struct RsqrtGradFunctor : public BaseActivationFunctor<T> {
  template <typename Device, typename X, typename Out, typename dOut,
            typename dX>
  void operator()(Device d, X x, Out out, dOut dout, dX dx) const {
486
    dx.device(d) = static_cast<T>(-0.5) * dout * out * out * out;
Z
zhoukunsheng 已提交
487
  }
Z
zhoukunsheng 已提交
488 489

  static constexpr ActBwdOpFwdDeps FwdDeps() { return kDepOut; }
Z
zhoukunsheng 已提交
490 491
};

D
dzhwinter 已提交
492 493 494
// ceil(x) = ceiling(x)
template <typename T>
struct CeilFunctor : public BaseActivationFunctor<T> {
F
fengjiayi 已提交
495 496 497
  template <typename Device, typename X, typename Out>
  void operator()(Device d, X x, Out out) const {
    out.device(d) = x.ceil();
D
dzhwinter 已提交
498 499 500 501 502
  }
};

template <typename T>
struct ZeroGradFunctor : public BaseActivationFunctor<T> {
F
fengjiayi 已提交
503 504 505
  template <typename Device, typename X, typename Out, typename dOut,
            typename dX>
  void operator()(Device d, X x, Out out, dOut dout, dX dx) const {
Z
Zeng Jinle 已提交
506
    dx.device(d) = static_cast<T>(0) * out;
D
dzhwinter 已提交
507
  }
508 509

  static constexpr ActBwdOpFwdDeps FwdDeps() { return kNoDeps; }
D
dzhwinter 已提交
510 511 512 513 514
};

// floor(x) = flooring(x)
template <typename T>
struct FloorFunctor : public BaseActivationFunctor<T> {
F
fengjiayi 已提交
515 516
  template <typename Device, typename X, typename Out>
  void operator()(Device d, X x, Out out) const {
Q
Qiao Longfei 已提交
517
    out.device(d) = x.floor();
D
dzhwinter 已提交
518 519 520
  }
};

C
add cos  
chengduoZH 已提交
521 522 523 524 525
template <typename T>
struct Sine {
  HOSTDEVICE T operator()(const T& val) const { return sin(val); }
};

526 527 528 529 530 531 532
template <>
struct Sine<platform::float16> {
  HOSTDEVICE platform::float16 operator()(const platform::float16& val) const {
    return platform::float16(sin(static_cast<float>(val)));
  }
};

C
add cos  
chengduoZH 已提交
533 534 535 536 537
template <typename T>
struct Cosine {
  HOSTDEVICE T operator()(const T& val) const { return cos(val); }
};

538 539 540 541 542 543 544
template <>
struct Cosine<platform::float16> {
  HOSTDEVICE platform::float16 operator()(const platform::float16& val) const {
    return platform::float16(cos(static_cast<float>(val)));
  }
};

C
add cos  
chengduoZH 已提交
545 546 547 548 549 550 551 552
// cosine'(x) = -sin(x)
template <typename T>
struct CosGradFunctor : public BaseActivationFunctor<T> {
  template <typename Device, typename X, typename Out, typename dOut,
            typename dX>
  void operator()(Device d, X x, Out out, dOut dout, dX dx) const {
    dx.device(d) = -dout * x.unaryExpr(Sine<T>());
  }
553 554

  static constexpr ActBwdOpFwdDeps FwdDeps() { return kDepX; }
C
add cos  
chengduoZH 已提交
555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573
};

// cosine(x) = cos(x)
template <typename T>
struct CosFunctor : public BaseActivationFunctor<T> {
  template <typename Device, typename X, typename Out>
  void operator()(Device d, X x, Out out) const {
    out.device(d) = x.unaryExpr(Cosine<T>());
  }
};

// sine'(x) = cos(x)
template <typename T>
struct SinGradFunctor : public BaseActivationFunctor<T> {
  template <typename Device, typename X, typename Out, typename dOut,
            typename dX>
  void operator()(Device d, X x, Out out, dOut dout, dX dx) const {
    dx.device(d) = dout * x.unaryExpr(Cosine<T>());
  }
574 575

  static constexpr ActBwdOpFwdDeps FwdDeps() { return kDepX; }
C
add cos  
chengduoZH 已提交
576 577 578 579 580 581 582 583 584 585 586
};

// sine(x) = sin(x)
template <typename T>
struct SinFunctor : public BaseActivationFunctor<T> {
  template <typename Device, typename X, typename Out>
  void operator()(Device d, X x, Out out) const {
    out.device(d) = x.unaryExpr(Sine<T>());
  }
};

587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652
template <typename T>
struct Sinh {
  HOSTDEVICE T operator()(const T& val) const { return sinh(val); }
};

template <>
struct Sinh<platform::float16> {
  HOSTDEVICE platform::float16 operator()(const platform::float16& val) const {
    return platform::float16(sinhf(static_cast<float>(val)));
  }
};

template <typename T>
struct Cosh {
  HOSTDEVICE T operator()(const T& val) const { return cosh(val); }
};

template <>
struct Cosh<platform::float16> {
  HOSTDEVICE platform::float16 operator()(const platform::float16& val) const {
    return platform::float16(coshf(static_cast<float>(val)));
  }
};

// sinh(x) = sinh(x)
template <typename T>
struct SinhFunctor : public BaseActivationFunctor<T> {
  template <typename Device, typename X, typename Out>
  void operator()(Device d, X x, Out out) const {
    out.device(d) = x.unaryExpr(Sinh<T>());
  }
};

// cosh(x) = cosh(x)
template <typename T>
struct CoshFunctor : public BaseActivationFunctor<T> {
  template <typename Device, typename X, typename Out>
  void operator()(Device d, X x, Out out) const {
    out.device(d) = x.unaryExpr(Cosh<T>());
  }
};

// sinh'(x) = cosh(x)
template <typename T>
struct SinhGradFunctor : public BaseActivationFunctor<T> {
  template <typename Device, typename X, typename Out, typename dOut,
            typename dX>
  void operator()(Device d, X x, Out out, dOut dout, dX dx) const {
    dx.device(d) = dout * x.unaryExpr(Cosh<T>());
  }

  static constexpr ActBwdOpFwdDeps FwdDeps() { return kDepX; }
};

// cosh'(x) = sinh(x)
template <typename T>
struct CoshGradFunctor : public BaseActivationFunctor<T> {
  template <typename Device, typename X, typename Out, typename dOut,
            typename dX>
  void operator()(Device d, X x, Out out, dOut dout, dX dx) const {
    dx.device(d) = dout * x.unaryExpr(Sinh<T>());
  }

  static constexpr ActBwdOpFwdDeps FwdDeps() { return kDepX; }
};

653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682
template <typename T>
struct Acos {
  HOSTDEVICE T operator()(const T& val) const { return acos(val); }
};

template <>
struct Acos<platform::float16> {
  HOSTDEVICE platform::float16 operator()(const platform::float16& val) const {
    return platform::float16(acos(static_cast<float>(val)));
  }
};

// Acos(x) = acos(x)
template <typename T>
struct AcosFunctor : public BaseActivationFunctor<T> {
  template <typename Device, typename X, typename Out>
  void operator()(Device d, X x, Out out) const {
    out.device(d) = x.unaryExpr(Acos<T>());
  }
};

// acos'(x) = -1/sqrt(1-x^2)
template <typename T>
struct AcosGradFunctor : public BaseActivationFunctor<T> {
  template <typename Device, typename X, typename Out, typename dOut,
            typename dX>
  void operator()(Device d, X x, Out out, dOut dout, dX dx) const {
    dx.device(d) =
        -dout * static_cast<T>(1) / (static_cast<T>(1) - x.square()).sqrt();
  }
683 684

  static constexpr ActBwdOpFwdDeps FwdDeps() { return kDepX; }
685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716
};

template <typename T>
struct Asin {
  HOSTDEVICE T operator()(const T& val) const { return asin(val); }
};

template <>
struct Asin<platform::float16> {
  HOSTDEVICE platform::float16 operator()(const platform::float16& val) const {
    return platform::float16(asin(static_cast<float>(val)));
  }
};

// Asin(x) = asin(x)
template <typename T>
struct AsinFunctor : public BaseActivationFunctor<T> {
  template <typename Device, typename X, typename Out>
  void operator()(Device d, X x, Out out) const {
    out.device(d) = x.unaryExpr(Asin<T>());
  }
};

// asin'(x) = 1/sqrt(1-x^2)
template <typename T>
struct AsinGradFunctor : public BaseActivationFunctor<T> {
  template <typename Device, typename X, typename Out, typename dOut,
            typename dX>
  void operator()(Device d, X x, Out out, dOut dout, dX dx) const {
    dx.device(d) =
        dout * static_cast<T>(1) / (static_cast<T>(1) - x.square()).sqrt();
  }
717 718

  static constexpr ActBwdOpFwdDeps FwdDeps() { return kDepX; }
719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749
};

template <typename T>
struct Atan {
  HOSTDEVICE T operator()(const T& val) const { return atan(val); }
};

template <>
struct Atan<platform::float16> {
  HOSTDEVICE platform::float16 operator()(const platform::float16& val) const {
    return platform::float16(atan(static_cast<float>(val)));
  }
};

// Atan(x) = atan(x)
template <typename T>
struct AtanFunctor : public BaseActivationFunctor<T> {
  template <typename Device, typename X, typename Out>
  void operator()(Device d, X x, Out out) const {
    out.device(d) = x.unaryExpr(Atan<T>());
  }
};

// atan'(x) =  1 / (1 + x^2)
template <typename T>
struct AtanGradFunctor : public BaseActivationFunctor<T> {
  template <typename Device, typename X, typename Out, typename dOut,
            typename dX>
  void operator()(Device d, X x, Out out, dOut dout, dX dx) const {
    dx.device(d) = dout * static_cast<T>(1) / (static_cast<T>(1) + x.square());
  }
750 751

  static constexpr ActBwdOpFwdDeps FwdDeps() { return kDepX; }
752 753
};

D
dzhwinter 已提交
754 755 756
// round(x) = [x]
template <typename T>
struct RoundFunctor : public BaseActivationFunctor<T> {
F
fengjiayi 已提交
757 758 759
  template <typename Device, typename X, typename Out>
  void operator()(Device d, X x, Out out) const {
    out.device(d) = x.round();
D
dzhwinter 已提交
760 761 762
  }
};

Q
qijun 已提交
763
// abs(x) = |x|
764 765
template <typename T>
struct AbsFunctor : public BaseActivationFunctor<T> {
F
fengjiayi 已提交
766 767 768
  template <typename Device, typename X, typename Out>
  void operator()(Device d, X x, Out out) const {
    out.device(d) = x.abs();
Q
qijun 已提交
769 770 771
  }
};

772 773
template <typename T>
struct AbsGradFunctor : public BaseActivationFunctor<T> {
F
fengjiayi 已提交
774 775 776 777
  template <typename Device, typename X, typename Out, typename dOut,
            typename dX>
  void operator()(Device d, X x, Out out, dOut dout, dX dx) const {
    dx.device(d) = dout * x.sign();
778
  }
779

780
  static constexpr ActBwdOpFwdDeps FwdDeps() { return kDepX; }
781 782
};

Q
qijun 已提交
783 784
// reciprocal(x) = 1 / x
template <typename T>
785
struct ReciprocalFunctor : public BaseActivationFunctor<T> {
F
fengjiayi 已提交
786 787 788
  template <typename Device, typename X, typename Out>
  void operator()(Device d, X x, Out out) const {
    out.device(d) = static_cast<T>(1) / x;
Q
qijun 已提交
789 790 791
  }
};

792
template <typename T>
793
struct ReciprocalGradFunctor : public BaseActivationFunctor<T> {
F
fengjiayi 已提交
794 795 796 797
  template <typename Device, typename X, typename Out, typename dOut,
            typename dX>
  void operator()(Device d, X x, Out out, dOut dout, dX dx) const {
    dx.device(d) = dout * static_cast<T>(-1) * out * out;
Q
qijun 已提交
798
  }
799 800

  static constexpr ActBwdOpFwdDeps FwdDeps() { return kDepOut; }
Q
qijun 已提交
801 802 803
};

// log(x) = natural logarithm of x
804 805
template <typename T>
struct LogFunctor : public BaseActivationFunctor<T> {
F
fengjiayi 已提交
806 807 808
  template <typename Device, typename X, typename Out>
  void operator()(Device d, X x, Out out) const {
    out.device(d) = x.log();
Q
qijun 已提交
809 810 811
  }
};

812
template <typename T>
813
struct LogGradFunctor : public BaseActivationFunctor<T> {
F
fengjiayi 已提交
814 815 816 817
  template <typename Device, typename X, typename Out, typename dOut,
            typename dX>
  void operator()(Device d, X x, Out out, dOut dout, dX dx) const {
    dx.device(d) = dout * (static_cast<T>(1) / x);
Q
qijun 已提交
818
  }
819 820

  static constexpr ActBwdOpFwdDeps FwdDeps() { return kDepX; }
Q
qijun 已提交
821 822
};

823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842
// log1p(x) = natural logarithm of x+1
template <typename T>
struct Log1pFunctor : public BaseActivationFunctor<T> {
  template <typename Device, typename X, typename Out>
  void operator()(Device d, X x, Out out) const {
    out.device(d) = (static_cast<T>(1) + x).log();
  }
};

template <typename T>
struct Log1pGradFunctor : public BaseActivationFunctor<T> {
  template <typename Device, typename X, typename Out, typename dOut,
            typename dX>
  void operator()(Device d, X x, Out out, dOut dout, dX dx) const {
    dx.device(d) = dout * (static_cast<T>(1) / (x + static_cast<T>(1)));
  }

  static constexpr ActBwdOpFwdDeps FwdDeps() { return kDepX; }
};

Q
qijun 已提交
843
// square(x) = x^2
844 845
template <typename T>
struct SquareFunctor : public BaseActivationFunctor<T> {
F
fengjiayi 已提交
846 847 848
  template <typename Device, typename X, typename Out>
  void operator()(Device d, X x, Out out) const {
    out.device(d) = x.square();
Q
qijun 已提交
849
  }
850
};
Q
qijun 已提交
851

852
template <typename T>
853
struct SquareGradFunctor : public BaseActivationFunctor<T> {
F
fengjiayi 已提交
854 855 856 857
  template <typename Device, typename X, typename Out, typename dOut,
            typename dX>
  void operator()(Device d, X x, Out out, dOut dout, dX dx) const {
    dx.device(d) = dout * static_cast<T>(2) * x;
858
  }
859 860

  static constexpr ActBwdOpFwdDeps FwdDeps() { return kDepX; }
861 862
};

863 864 865 866 867 868 869 870 871 872
template <typename T>
struct BReluFunctor : public BaseActivationFunctor<T> {
  float t_min;
  float t_max;

  // NOTE: Explicit hides the `BaseActivationFunctor<T>::GetAttrs`
  // not polymorphism for speed.
  typename BaseActivationFunctor<T>::AttrPair GetAttrs() {
    return {{"t_min", &t_min}, {"t_max", &t_max}};
  }
873

F
fengjiayi 已提交
874 875 876
  template <typename Device, typename X, typename Out>
  void operator()(Device d, X x, Out out) const {
    out.device(d) =
Y
Yu Yang 已提交
877
        x.cwiseMax(static_cast<T>(t_min)).cwiseMin(static_cast<T>(t_max));
878 879 880
  }
};

881 882 883 884 885 886 887
template <typename T>
struct BReluGradFunctor : public BaseActivationFunctor<T> {
  float t_min;
  float t_max;
  typename BaseActivationFunctor<T>::AttrPair GetAttrs() {
    return {{"t_min", &t_min}, {"t_max", &t_max}};
  }
F
fengjiayi 已提交
888 889 890 891
  template <typename Device, typename X, typename Out, typename dOut,
            typename dX>
  void operator()(Device d, X x, Out out, dOut dout, dX dx) const {
    dx.device(d) = dout *
Y
Yu Yang 已提交
892 893
                   ((x > static_cast<T>(t_min)) * (x < static_cast<T>(t_max)))
                       .template cast<T>();
894
  }
895 896

  static constexpr ActBwdOpFwdDeps FwdDeps() { return kDepX; }
897 898
};

899 900 901 902 903 904 905 906 907
// relu6(x) = min(max(0, x), 6)
template <typename T>
struct Relu6Functor : public BaseActivationFunctor<T> {
  float threshold;

  typename BaseActivationFunctor<T>::AttrPair GetAttrs() {
    return {{"threshold", &threshold}};
  }

F
fengjiayi 已提交
908 909 910
  template <typename Device, typename X, typename Out>
  void operator()(Device d, X x, Out out) const {
    out.device(d) =
Y
Yu Yang 已提交
911
        x.cwiseMax(static_cast<T>(0)).cwiseMin(static_cast<T>(threshold));
912 913 914 915 916 917 918 919 920
  }
};

template <typename T>
struct Relu6GradFunctor : public BaseActivationFunctor<T> {
  float threshold;
  typename BaseActivationFunctor<T>::AttrPair GetAttrs() {
    return {{"threshold", &threshold}};
  }
F
fengjiayi 已提交
921 922 923
  template <typename Device, typename X, typename Out, typename dOut,
            typename dX>
  void operator()(Device d, X x, Out out, dOut dout, dX dx) const {
D
dzhwinter 已提交
924 925 926 927
    dx.device(d) =
        dout *
        ((out > static_cast<T>(0)) * (out < static_cast<T>(threshold)))
            .template cast<T>();
928
  }
929 930

  static constexpr ActBwdOpFwdDeps FwdDeps() { return kDepOut; }
931 932
};

H
huangjun12 已提交
933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977
// HardSwish = min(max(0, x+3), 6) * x / 6
template <typename T>
struct HardSwishFunctor : public BaseActivationFunctor<T> {
  float threshold;
  float scale;
  float offset;

  typename BaseActivationFunctor<T>::AttrPair GetAttrs() {
    return {{"threshold", &threshold}, {"scale", &scale}, {"offset", &offset}};
  }

  template <typename Device, typename X, typename Out>
  void operator()(Device d, X x, Out out) const {
    out.device(d) = (x + static_cast<T>(offset))
                        .cwiseMax(static_cast<T>(0))
                        .cwiseMin(static_cast<T>(threshold)) *
                    x / static_cast<T>(scale);
  }
};

template <typename T>
struct HardSwishGradFunctor : public BaseActivationFunctor<T> {
  float threshold;
  float scale;
  float offset;

  typename BaseActivationFunctor<T>::AttrPair GetAttrs() {
    return {{"threshold", &threshold}, {"scale", &scale}, {"offset", &offset}};
  }
  template <typename Device, typename X, typename Out, typename dOut,
            typename dX>
  void operator()(Device d, X x, Out out, dOut dout, dX dx) const {
    auto tmp = ((x + static_cast<T>(offset)) < static_cast<T>(threshold))
                   .template cast<T>();
    dx.device(d) =
        dout *
        (((x + static_cast<T>(offset)) > static_cast<T>(0)).template cast<T>() *
             (static_cast<T>(2) * x + static_cast<T>(offset)) /
             static_cast<T>(scale) * tmp +
         static_cast<T>(1) * (static_cast<T>(1) - tmp));
  }

  static constexpr ActBwdOpFwdDeps FwdDeps() { return kDepX; }
};

978 979 980 981
// For numerical stability, using the following formula instead of softplus(x) =
// log(1 + exp(x))
// softplus(x) = log(1 + exp(beta * x)) / beta when beta * x <= threshold(beta =
// 1, threshold = 20 by default), otherwise x
K
kexinzhao 已提交
982 983
template <typename T>
struct SoftplusFunctor : public BaseActivationFunctor<T> {
984 985 986 987 988 989
  float beta;
  float threshold;
  typename BaseActivationFunctor<T>::AttrPair GetAttrs() {
    return {{"beta", &beta}, {"threshold", &threshold}};
  }

F
fengjiayi 已提交
990 991
  template <typename Device, typename X, typename Out>
  void operator()(Device d, X x, Out out) {
992 993 994 995
    auto x_beta = static_cast<T>(beta) * x;
    out.device(d) = (x_beta > static_cast<T>(threshold))
                        .select(x, (static_cast<T>(1) + x_beta.exp()).log() /
                                       static_cast<T>(beta));
K
kexinzhao 已提交
996 997 998
  }
};

999 1000 1001 1002
// For numerical stability, using the following formula instead of
// d(softplus(x))/dx = 1 / (1 + exp(-x))
// d(softplus(x))/dx = 1 / (1 + exp(-beta * x)) when beta * x <= threshold(beta
// = 1, threshold = 20 by default), otherwise x
K
kexinzhao 已提交
1003 1004
template <typename T>
struct SoftplusGradFunctor : public BaseActivationFunctor<T> {
1005 1006 1007 1008 1009 1010
  float beta;
  float threshold;
  typename BaseActivationFunctor<T>::AttrPair GetAttrs() {
    return {{"beta", &beta}, {"threshold", &threshold}};
  }

F
fengjiayi 已提交
1011 1012 1013
  template <typename Device, typename X, typename Out, typename dOut,
            typename dX>
  void operator()(Device d, X x, Out out, dOut dout, dX dx) {
1014
    auto x_beta = static_cast<T>(beta) * x;
F
fengjiayi 已提交
1015
    dx.device(d) =
1016 1017
        (x_beta > static_cast<T>(threshold))
            .select(dout, dout / (static_cast<T>(1) + (-x_beta).exp()));
K
kexinzhao 已提交
1018
  }
1019 1020

  static constexpr ActBwdOpFwdDeps FwdDeps() { return kDepX; }
K
kexinzhao 已提交
1021 1022
};

1023 1024
// softsign(x) = x / (1 + |x|)
template <typename T>
1025
struct SoftsignFunctor : public BaseActivationFunctor<T> {
F
fengjiayi 已提交
1026 1027 1028
  template <typename Device, typename X, typename Out>
  void operator()(Device d, X x, Out out) {
    out.device(d) = x / (static_cast<T>(1) + x.abs());
1029 1030 1031 1032 1033 1034
  }
};

// d(softsign(x))/dx = 1 / (1 + |x|)^2
// Taken from https://en.wikipedia.org/wiki/Activation_function
template <typename T>
1035
struct SoftsignGradFunctor : public BaseActivationFunctor<T> {
F
fengjiayi 已提交
1036 1037 1038
  template <typename Device, typename X, typename Out, typename dOut,
            typename dX>
  void operator()(Device d, X x, Out out, dOut dout, dX dx) {
1039
    dx.device(d) =
F
fengjiayi 已提交
1040
        dout * (static_cast<T>(1) / (static_cast<T>(1) + x.abs()).square());
1041
  }
1042 1043

  static constexpr ActBwdOpFwdDeps FwdDeps() { return kDepX; }
1044 1045
};

1046 1047 1048 1049 1050 1051
template <typename T>
struct SoftReluFunctor : public BaseActivationFunctor<T> {
  float threshold;
  typename BaseActivationFunctor<T>::AttrPair GetAttrs() {
    return {{"threshold", &threshold}};
  }
1052

F
fengjiayi 已提交
1053 1054
  template <typename Device, typename X, typename Out>
  void operator()(Device d, X x, Out out) const {
Y
Yu Yang 已提交
1055 1056
    auto tmp = static_cast<T>(threshold);
    auto temp = x.cwiseMax(-tmp).cwiseMin(tmp);
F
fengjiayi 已提交
1057
    out.device(d) = (static_cast<T>(1) + temp.exp()).log();
1058 1059 1060
  }
};

1061 1062 1063 1064 1065 1066
template <typename T>
struct SoftReluGradFunctor : public BaseActivationFunctor<T> {
  float threshold;
  typename BaseActivationFunctor<T>::AttrPair GetAttrs() {
    return {{"threshold", &threshold}};
  }
F
fengjiayi 已提交
1067 1068 1069
  template <typename Device, typename X, typename Out, typename dOut,
            typename dX>
  void operator()(Device d, X x, Out out, dOut dout, dX dx) const {
Y
Yu Yang 已提交
1070
    auto tmp = static_cast<T>(threshold);
Z
Zeng Jinle 已提交
1071
    auto temp = ((out > -tmp) * (out < tmp)).template cast<T>();
F
fengjiayi 已提交
1072
    dx.device(d) = dout * (static_cast<T>(1) - (-out).exp()) * temp;
1073
  }
1074 1075

  static constexpr ActBwdOpFwdDeps FwdDeps() { return kDepOut; }
1076 1077
};

K
Kavya Srinet 已提交
1078 1079 1080 1081 1082 1083
template <typename T>
struct LeakyReluFunctor : public BaseActivationFunctor<T> {
  float alpha;
  typename BaseActivationFunctor<T>::AttrPair GetAttrs() {
    return {{"alpha", &alpha}};
  }
1084

F
fengjiayi 已提交
1085 1086
  template <typename Device, typename X, typename Out>
  void operator()(Device d, X x, Out out) const {
1087 1088 1089 1090 1091
    if (alpha < 1.f) {
      out.device(d) = x.cwiseMax(static_cast<T>(alpha) * x);
    } else {
      out.device(d) = x.cwiseMin(static_cast<T>(alpha) * x);
    }
1092 1093 1094
  }
};

K
Kavya Srinet 已提交
1095 1096 1097 1098 1099 1100
template <typename T>
struct LeakyReluGradFunctor : public BaseActivationFunctor<T> {
  float alpha;
  typename BaseActivationFunctor<T>::AttrPair GetAttrs() {
    return {{"alpha", &alpha}};
  }
F
fengjiayi 已提交
1101 1102 1103
  template <typename Device, typename X, typename Out, typename dOut,
            typename dX>
  void operator()(Device d, X x, Out out, dOut dout, dX dx) const {
Z
Zeng Jinle 已提交
1104
    auto temp1 =
1105 1106
        static_cast<T>(alpha) * (x < static_cast<T>(0)).template cast<T>();
    auto temp2 = (x >= static_cast<T>(0)).template cast<T>();
F
fengjiayi 已提交
1107
    dx.device(d) = dout * (temp1 + temp2).template cast<T>();
1108
  }
1109

1110
  static constexpr ActBwdOpFwdDeps FwdDeps() { return kDepX; }
1111 1112
};

1113 1114 1115 1116 1117 1118
template <typename T>
struct ELUFunctor : public BaseActivationFunctor<T> {
  float alpha;
  typename BaseActivationFunctor<T>::AttrPair GetAttrs() {
    return {{"alpha", &alpha}};
  }
1119

F
fengjiayi 已提交
1120 1121 1122 1123 1124
  template <typename Device, typename X, typename Out>
  void operator()(Device d, X x, Out out) const {
    out.device(d) = x.cwiseMax(static_cast<T>(0)) +
                    (static_cast<T>(alpha) * (x.exp() - static_cast<T>(1)))
                        .cwiseMin(static_cast<T>(0));
1125 1126 1127
  }
};

1128 1129 1130 1131 1132 1133
template <typename T>
struct ELUGradFunctor : public BaseActivationFunctor<T> {
  float alpha;
  typename BaseActivationFunctor<T>::AttrPair GetAttrs() {
    return {{"alpha", &alpha}};
  }
F
fengjiayi 已提交
1134 1135 1136
  template <typename Device, typename X, typename Out, typename dOut,
            typename dX>
  void operator()(Device d, X x, Out out, dOut dout, dX dx) const {
1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150
    auto temp_a_pos = static_cast<T>(alpha > 0);
    auto temp_a_neg = static_cast<T>(alpha <= 0);
    auto temp_x_pos = (x > static_cast<T>(0)).template cast<T>();
    auto temp_x_neg = (x <= static_cast<T>(0)).template cast<T>();

    // dx = dout, if alpha > 0 and x > 0
    // dx = dout * alpha * x.exp(), if alpha > 0 and x <= 0
    // dx = dout * (1 + alpha * x.exp()), if alpha <= 0 and x > 0
    // dx = 0, if alpha <= 0 and x <=0
    dx.device(d) =
        dout * temp_a_pos * temp_x_pos +
        dout * static_cast<T>(alpha) * x.exp() * temp_a_pos * temp_x_neg +
        dout * (static_cast<T>(1) + static_cast<T>(alpha) * x.exp()) *
            temp_a_neg * temp_x_pos;
1151
  }
1152 1153

  static constexpr ActBwdOpFwdDeps FwdDeps() { return kDepX; }
1154 1155
};

Q
QI JUN 已提交
1156
// FIXME(qijun) https://github.com/PaddlePaddle/Paddle/issues/5198
1157 1158 1159 1160 1161 1162
template <typename T>
struct PowFunctor : public BaseActivationFunctor<T> {
  float factor;
  typename BaseActivationFunctor<T>::AttrPair GetAttrs() {
    return {{"factor", &factor}};
  }
F
fengjiayi 已提交
1163 1164 1165
  template <typename Device, typename X, typename Out>
  void operator()(Device d, X x, Out out) const {
    out.device(d) = x.pow(static_cast<T>(factor));
1166 1167 1168
  }
};

1169 1170 1171 1172 1173 1174
template <typename T>
struct PowGradFunctor : public BaseActivationFunctor<T> {
  float factor;
  typename BaseActivationFunctor<T>::AttrPair GetAttrs() {
    return {{"factor", &factor}};
  }
F
fengjiayi 已提交
1175 1176 1177 1178
  template <typename Device, typename X, typename Out, typename dOut,
            typename dX>
  void operator()(Device d, X x, Out out, dOut dout, dX dx) const {
    dx.device(d) = dout * static_cast<T>(factor) *
C
chengduo 已提交
1179
                   x.pow(static_cast<T>(factor) - static_cast<T>(1));
1180
  }
1181 1182

  static constexpr ActBwdOpFwdDeps FwdDeps() { return kDepX; }
1183 1184
};

1185 1186 1187 1188 1189 1190 1191
template <typename T>
struct STanhFunctor : public BaseActivationFunctor<T> {
  float scale_a;
  float scale_b;
  typename BaseActivationFunctor<T>::AttrPair GetAttrs() {
    return {{"scale_a", &scale_a}, {"scale_b", &scale_b}};
  }
1192

F
fengjiayi 已提交
1193 1194 1195
  template <typename Device, typename X, typename Out>
  void operator()(Device d, X x, Out out) const {
    out.device(d) =
Y
Yu Yang 已提交
1196
        static_cast<T>(scale_b) * (static_cast<T>(scale_a) * x).tanh();
1197 1198 1199
  }
};

1200 1201 1202 1203 1204 1205 1206
template <typename T>
struct STanhGradFunctor : public BaseActivationFunctor<T> {
  float scale_a;
  float scale_b;
  typename BaseActivationFunctor<T>::AttrPair GetAttrs() {
    return {{"scale_a", &scale_a}, {"scale_b", &scale_b}};
  }
1207

F
fengjiayi 已提交
1208 1209 1210
  template <typename Device, typename X, typename Out, typename dOut,
            typename dX>
  void operator()(Device d, X x, Out out, dOut dout, dX dx) const {
Y
Yu Yang 已提交
1211 1212 1213
    auto a = static_cast<T>(scale_a);
    auto b = static_cast<T>(scale_b);
    auto temp = (a * x).tanh() * (a * x).tanh();
F
fengjiayi 已提交
1214
    dx.device(d) = dout * a * b * (static_cast<T>(1) - temp);
Q
qijun 已提交
1215
  }
1216 1217

  static constexpr ActBwdOpFwdDeps FwdDeps() { return kDepX; }
Q
qijun 已提交
1218 1219
};

1220 1221 1222 1223 1224 1225 1226
template <typename T>
struct ThresholdedReluFunctor : public BaseActivationFunctor<T> {
  float threshold;
  typename BaseActivationFunctor<T>::AttrPair GetAttrs() {
    return {{"threshold", &threshold}};
  }

F
fengjiayi 已提交
1227 1228
  template <typename Device, typename X, typename Out>
  void operator()(Device d, X x, Out out) const {
Y
Yu Yang 已提交
1229
    auto th = static_cast<T>(threshold);
F
fengjiayi 已提交
1230
    out.device(d) = (x > th).template cast<T>() * x;
1231 1232 1233 1234 1235 1236 1237 1238 1239 1240
  }
};

template <typename T>
struct ThresholdedReluGradFunctor : public BaseActivationFunctor<T> {
  float threshold;
  typename BaseActivationFunctor<T>::AttrPair GetAttrs() {
    return {{"threshold", &threshold}};
  }

F
fengjiayi 已提交
1241 1242 1243
  template <typename Device, typename X, typename Out, typename dOut,
            typename dX>
  void operator()(Device d, X x, Out out, dOut dout, dX dx) const {
Y
Yu Yang 已提交
1244
    auto th = static_cast<T>(threshold);
F
fengjiayi 已提交
1245
    dx.device(d) = dout * (x > th).template cast<T>();
1246
  }
1247 1248

  static constexpr ActBwdOpFwdDeps FwdDeps() { return kDepX; }
1249 1250
};

1251 1252 1253 1254 1255 1256 1257 1258
template <typename T>
struct HardSigmoidFunctor : public BaseActivationFunctor<T> {
  float slope;
  float offset;
  typename BaseActivationFunctor<T>::AttrPair GetAttrs() {
    return {{"slope", &slope}, {"offset", &offset}};
  }

F
fengjiayi 已提交
1259 1260
  template <typename Device, typename X, typename Out>
  void operator()(Device d, X x, Out out) const {
1261
    auto temp = x * static_cast<T>(slope) + static_cast<T>(offset);
F
fengjiayi 已提交
1262 1263
    out.device(d) =
        temp.cwiseMax(static_cast<T>(0)).cwiseMin(static_cast<T>(1));
1264 1265 1266 1267 1268 1269 1270 1271 1272 1273
  }
};

template <typename T>
struct HardSigmoidGradFunctor : public BaseActivationFunctor<T> {
  float slope;
  float offset;
  typename BaseActivationFunctor<T>::AttrPair GetAttrs() {
    return {{"slope", &slope}, {"offset", &offset}};
  }
F
fengjiayi 已提交
1274 1275 1276 1277 1278 1279 1280
  template <typename Device, typename X, typename Out, typename dOut,
            typename dX>
  void operator()(Device d, X x, Out out, dOut dout, dX dx) const {
    dx.device(d) = dout *
                   ((out > static_cast<T>(0)) * (out < static_cast<T>(1)))
                       .template cast<T>() *
                   static_cast<T>(slope);
1281
  }
1282 1283

  static constexpr ActBwdOpFwdDeps FwdDeps() { return kDepOut; }
1284 1285
};

A
Abhinav Arora 已提交
1286 1287 1288 1289 1290 1291 1292
template <typename T>
struct SwishFunctor : public BaseActivationFunctor<T> {
  float beta;
  typename BaseActivationFunctor<T>::AttrPair GetAttrs() {
    return {{"beta", &beta}};
  }

F
fengjiayi 已提交
1293 1294 1295
  template <typename Device, typename X, typename Out>
  void operator()(Device d, X x, Out out) const {
    out.device(d) = x / (static_cast<T>(1) + (static_cast<T>(-beta) * x).exp());
A
Abhinav Arora 已提交
1296 1297 1298 1299 1300 1301 1302 1303 1304 1305
  }
};

template <typename T>
struct SwishGradFunctor : public BaseActivationFunctor<T> {
  float beta;
  typename BaseActivationFunctor<T>::AttrPair GetAttrs() {
    return {{"beta", &beta}};
  }

F
fengjiayi 已提交
1306 1307
  template <typename Device, typename X, typename Out, typename dOut,
            typename dX>
1308
  void operator()(Device d, X x, Out fake_out, dOut dout, dX dx) const {
A
Abhinav Arora 已提交
1309
    auto temp1 = static_cast<T>(1) /
1310
                 (static_cast<T>(1) + (static_cast<T>(-beta) * x).exp());
1311
    auto out = x * temp1;
D
dzhwinter 已提交
1312 1313
    auto temp2 = temp1 * (static_cast<T>(1) - (static_cast<T>(beta) * out));
    dx.device(d) = dout * ((static_cast<T>(beta) * out) + temp2);
A
Abhinav Arora 已提交
1314
  }
1315 1316

  static constexpr ActBwdOpFwdDeps FwdDeps() { return kDepX; }
A
Abhinav Arora 已提交
1317 1318
};

1319 1320 1321 1322 1323 1324 1325 1326 1327 1328 1329 1330
/*
 * in arguments: x, out, ddx
 * out arguments: ddout, dout, dx
 */
template <ActBwdOpFwdDeps kDepValue>
inline void ExtractActivationDoubleGradTensor(
    const framework::ExecutionContext& ctx, const framework::Tensor** X,
    const framework::Tensor** Out, const framework::Tensor** ddX,
    framework::Tensor** dX, framework::Tensor** dOut,
    framework::Tensor** ddOut) {
  auto ddx_var = ctx.InputVar("DDX");
  auto ddo_var = ctx.OutputVar("DDOut");
1331 1332 1333 1334
  PADDLE_ENFORCE_NOT_NULL(
      ddx_var, platform::errors::NotFound(
                   "Cannot get input Variable Out, variable name = %s",
                   ctx.InputName("DDX")));
H
hong 已提交
1335
  if (CanBeUsedBySelectedRows.count(ctx.Type())) {
1336 1337 1338 1339 1340 1341 1342 1343 1344 1345 1346
    *ddX = paddle::framework::GetLoDTensorOrSelectedRowsValueFromVar(*ddx_var);
    if (ddo_var) {
      *ddOut = paddle::framework::GetMutableLoDTensorOrSelectedRowsValueFromVar(
          ddo_var);
    }
  } else {
    *ddX = ctx.Input<framework::Tensor>("DDX");
    if (ddo_var) {
      *ddOut = ctx.Output<framework::Tensor>("DDOut");
    }
  }
1347 1348 1349 1350 1351
  PADDLE_ENFORCE_NOT_NULL(
      *ddX,
      platform::errors::NotFound(
          "Cannot get the tensor from the Variable Output, variable name = %s",
          ctx.OutputName("DDX")));
1352 1353 1354

  if (static_cast<int>(kDepValue) & static_cast<int>(kDepX)) {
    auto x_var = ctx.InputVar("X");
1355 1356
    PADDLE_ENFORCE_NOT_NULL(
        x_var, platform::errors::NotFound(
1357
                   "Cannot get input Variable Out, variable name = %s",
1358
                   ctx.InputName("X")));
1359
    auto dx_var = ctx.OutputVar("DX");
H
hong 已提交
1360
    if (CanBeUsedBySelectedRows.count(ctx.Type())) {
1361 1362 1363 1364 1365 1366 1367 1368 1369 1370 1371 1372
      *X = paddle::framework::GetLoDTensorOrSelectedRowsValueFromVar(*x_var);
      if (dx_var) {
        *dX = paddle::framework::GetMutableLoDTensorOrSelectedRowsValueFromVar(
            dx_var);
      }
    } else {
      *X = ctx.Input<framework::Tensor>("X");
      if (dx_var) {
        *dX = ctx.Output<framework::Tensor>("DX");
      }
    }
  } else {
H
hong 已提交
1373
    VLOG(10) << "Inplace activation of Op: " << ctx.Type();
1374 1375
    *X = *ddX;
  }
1376 1377
  if (static_cast<int>(kDepValue) & static_cast<int>(kDepOut)) {
    auto out_var = ctx.InputVar("Out");
1378 1379 1380 1381 1382
    PADDLE_ENFORCE_NOT_NULL(
        out_var,
        platform::errors::NotFound(
            "Cannot get the tensor from the Variable Out, variable name = %s",
            ctx.InputName("Out")));
1383
    auto dout_var = ctx.OutputVar("DOut");
H
hong 已提交
1384
    if (CanBeUsedBySelectedRows.count(ctx.Type())) {
1385 1386 1387 1388 1389 1390 1391 1392 1393 1394 1395 1396 1397 1398
      *Out =
          paddle::framework::GetLoDTensorOrSelectedRowsValueFromVar(*out_var);
      if (dout_var) {
        *dOut =
            paddle::framework::GetMutableLoDTensorOrSelectedRowsValueFromVar(
                dout_var);
      }
    } else {
      *Out = ctx.Input<framework::Tensor>("Out");
      if (dout_var) {
        *dOut = ctx.Output<framework::Tensor>("DOut");
      }
    }
  } else {
H
hong 已提交
1399
    VLOG(10) << "Inplace activation of Op: " << ctx.Type();
1400 1401
    *Out = *ddX;
  }
1402 1403 1404 1405 1406 1407 1408 1409 1410 1411 1412 1413 1414 1415 1416 1417 1418 1419 1420 1421 1422 1423 1424 1425 1426 1427 1428 1429 1430 1431 1432
}

template <typename DeviceContext, typename Functor>
class ActivationDoubleGradKernel
    : public framework::OpKernel<typename Functor::ELEMENT_TYPE> {
 public:
  using T = typename Functor::ELEMENT_TYPE;
  void Compute(const framework::ExecutionContext& ctx) const override {
    const framework::Tensor *X, *Out, *ddX;
    X = Out = ddX = nullptr;
    framework::Tensor *ddOut, *dOut, *dX;
    ddOut = dOut = dX = nullptr;

    ExtractActivationDoubleGradTensor<Functor::FwdDeps()>(ctx, &X, &Out, &ddX,
                                                          &dX, &dOut, &ddOut);

    if (ddOut) ddOut->mutable_data<T>(ctx.GetPlace());
    if (dOut) dOut->mutable_data<T>(ctx.GetPlace());
    if (dX) dX->mutable_data<T>(Out->dims(), ctx.GetPlace());

    auto& place = ctx.template device_context<DeviceContext>();

    Functor functor;
    auto attrs = functor.GetAttrs();
    for (auto& attr : attrs) {
      *attr.second = ctx.Attr<float>(attr.first);
    }
    functor(place, X, Out, ddX, ddOut, dOut, dX);
  }
};

Z
Zhong Hui 已提交
1433 1434 1435 1436 1437 1438 1439 1440 1441 1442 1443 1444 1445 1446 1447 1448 1449 1450 1451 1452 1453
template <typename T>
struct AbsGradGradFunctor : public BaseActivationFunctor<T> {
  template <typename Device>
  void operator()(const Device& dev, const framework::Tensor* X,
                  const framework::Tensor* Out, const framework::Tensor* ddX,
                  framework::Tensor* ddOut, framework::Tensor* dOut,
                  framework::Tensor* dX) const {
    auto* d = dev.eigen_device();
    auto ddx = framework::EigenVector<T>::Flatten(
        GET_DATA_SAFELY(ddX, "Input", "DDX", "AbsGradGrad"));
    auto x = framework::EigenVector<T>::Flatten(
        GET_DATA_SAFELY(X, "Input", "X", "AbsGradGrad"));
    if (ddOut) {
      auto ddout = framework::EigenVector<T>::Flatten(
          GET_DATA_SAFELY(ddOut, "Output", "DDOut", "AbsGradGrad"));
      ddout.device(*d) = ddx * x.sign();
    }
  }
  static constexpr ActBwdOpFwdDeps FwdDeps() { return kDepX; }
};

1454 1455 1456 1457 1458 1459 1460 1461
template <typename T>
struct ReluGradGradFunctor : public BaseActivationFunctor<T> {
  template <typename Device>
  void operator()(const Device& dev, const framework::Tensor* X,
                  const framework::Tensor* Out, const framework::Tensor* ddX,
                  framework::Tensor* ddOut, framework::Tensor* dOut,
                  framework::Tensor* dX) const {
    auto* d = dev.eigen_device();
1462 1463 1464 1465
    auto ddx = framework::EigenVector<T>::Flatten(
        GET_DATA_SAFELY(ddX, "Input", "DDX", "ReluGradGrad"));
    auto out = framework::EigenVector<T>::Flatten(
        GET_DATA_SAFELY(Out, "Output", "Out", "ReluGradGrad"));
1466
    if (ddOut) {
1467 1468
      auto ddout = framework::EigenVector<T>::Flatten(
          GET_DATA_SAFELY(ddOut, "Output", "DDOut", "ReluGradGrad"));
1469 1470 1471 1472 1473 1474
      ddout.device(*d) = ddx * (out > static_cast<T>(0)).template cast<T>();
    }
  }
  static constexpr ActBwdOpFwdDeps FwdDeps() { return kDepOut; }
};

1475 1476 1477 1478 1479 1480 1481 1482 1483 1484 1485 1486
template <typename T>
struct LeakyReluGradGradFunctor : public BaseActivationFunctor<T> {
  float alpha;
  typename BaseActivationFunctor<T>::AttrPair GetAttrs() {
    return {{"alpha", &alpha}};
  }
  template <typename Device>
  void operator()(const Device& dev, const framework::Tensor* X,
                  const framework::Tensor* Out, const framework::Tensor* ddX,
                  framework::Tensor* ddOut, framework::Tensor* dOut,
                  framework::Tensor* dX) const {
    if (ddOut) {
Z
Zeng Jinle 已提交
1487
      auto* d = dev.eigen_device();
1488 1489
      auto ddx = framework::EigenVector<T>::Flatten(
          GET_DATA_SAFELY(ddX, "Input", "DDX", "LeakyReluGradGrad"));
1490 1491
      auto x = framework::EigenVector<T>::Flatten(
          GET_DATA_SAFELY(X, "Input", "X", "LeakyReluGradGrad"));
1492 1493
      auto ddout = framework::EigenVector<T>::Flatten(
          GET_DATA_SAFELY(ddOut, "Output", "DOut", "LeakyReluGradGrad"));
1494 1495 1496 1497 1498
      ddout.device(*d) =
          ddx *
          ((x > static_cast<T>(0)).template cast<T>() +
           static_cast<T>(alpha) * (x <= static_cast<T>(0)).template cast<T>())
              .template cast<T>();
1499 1500
    }
  }
1501
  static constexpr ActBwdOpFwdDeps FwdDeps() { return kDepX; }
1502 1503
};

D
Double_V 已提交
1504 1505 1506 1507 1508 1509 1510 1511 1512 1513 1514
template <typename T>
struct ELUGradGradFunctor : public BaseActivationFunctor<T> {
  float alpha;
  typename BaseActivationFunctor<T>::AttrPair GetAttrs() {
    return {{"alpha", &alpha}};
  }
  template <typename Device>
  void operator()(const Device& dev, const framework::Tensor* X,
                  const framework::Tensor* ddX, framework::Tensor* ddOut,
                  const framework::Tensor* dOut, framework::Tensor* dX) const {
    auto* d = dev.eigen_device();
1515 1516 1517 1518
    auto ddx = framework::EigenVector<T>::Flatten(
        GET_DATA_SAFELY(ddX, "Input", "DDX", "ELUGradGrad"));
    auto x = framework::EigenVector<T>::Flatten(
        GET_DATA_SAFELY(X, "Input", "X", "ELUGradGrad"));
D
Double_V 已提交
1519 1520

    if (dX) {
1521 1522 1523 1524
      auto dx = framework::EigenVector<T>::Flatten(
          GET_DATA_SAFELY(dX, "Output", "DX", "ELUGradGrad"));
      auto dout = framework::EigenVector<T>::Flatten(
          GET_DATA_SAFELY(dOut, "Output", "DOut", "ELUGradGrad"));
D
Double_V 已提交
1525 1526 1527 1528 1529
      dx.device(*d) = ddx * dout * static_cast<T>(alpha) * x.exp() *
                      (x < static_cast<T>(0)).template cast<T>();
    }

    if (ddOut) {
1530 1531
      auto ddout = framework::EigenVector<T>::Flatten(
          GET_DATA_SAFELY(ddOut, "Output", "DDOut", "ELUGradGrad"));
D
Double_V 已提交
1532 1533 1534 1535 1536 1537 1538 1539 1540 1541
      ddout.device(*d) = ddx *
                         ((x > static_cast<T>(0)).template cast<T>() +
                          static_cast<T>(alpha) * x.exp() *
                              (x <= static_cast<T>(0)).template cast<T>())
                             .template cast<T>();
    }
  }
  static constexpr ActBwdOpFwdDeps FwdDeps() { return kDepX; }
};

L
lvmengsi 已提交
1542 1543 1544 1545 1546 1547 1548
template <typename T>
struct SqrtGradGradFunctor : public BaseActivationFunctor<T> {
  template <typename Device>
  void operator()(const Device& dev, const framework::Tensor* Out,
                  const framework::Tensor* ddX, framework::Tensor* ddOut,
                  framework::Tensor* dOut, const framework::Tensor* dX) const {
    auto* d = dev.eigen_device();
1549 1550 1551 1552
    auto ddx = framework::EigenVector<T>::Flatten(
        GET_DATA_SAFELY(ddX, "Input", "DDX", "SqrtGradGrad"));
    auto out = framework::EigenVector<T>::Flatten(
        GET_DATA_SAFELY(Out, "Output", "Out", "SqrtGradGrad"));
1553 1554
    // sqrt GradGrad: ddy = 0.5 * ddx / y, dy = -1 * dx * ddx
    // calculate dy first, so ddy can inplace ddx
L
lvmengsi 已提交
1555
    if (dOut) {
1556 1557 1558 1559
      auto dx = framework::EigenVector<T>::Flatten(
          GET_DATA_SAFELY(dX, "Output", "DX", "SqrtGradGrad"));
      auto dout = framework::EigenVector<T>::Flatten(
          GET_DATA_SAFELY(dOut, "Output", "DOut", "SqrtGradGrad"));
L
lvmengsi 已提交
1560 1561
      dout.device(*d) = dx * ddx * static_cast<T>(-1) / out;
    }
1562
    if (ddOut) {
1563 1564
      auto ddout = framework::EigenVector<T>::Flatten(
          GET_DATA_SAFELY(ddOut, "Output", "DDOut", "SqrtGradGrad"));
1565 1566
      ddout.device(*d) = ddx * static_cast<T>(0.5) / out;
    }
L
lvmengsi 已提交
1567 1568 1569 1570
  }
  static constexpr ActBwdOpFwdDeps FwdDeps() { return kDepOut; }
};

1571 1572 1573 1574 1575 1576 1577
template <typename T>
struct SquareGradGradFunctor : public BaseActivationFunctor<T> {
  template <typename Device>
  void operator()(const Device& dev, const framework::Tensor* X,
                  const framework::Tensor* ddX, framework::Tensor* ddOut,
                  const framework::Tensor* dOut, framework::Tensor* dX) const {
    auto* d = dev.eigen_device();
1578 1579 1580 1581
    auto ddx = framework::EigenVector<T>::Flatten(
        GET_DATA_SAFELY(ddX, "Input", "DDX", "SquareGradGrad"));
    auto x = framework::EigenVector<T>::Flatten(
        GET_DATA_SAFELY(X, "Input", "X", "SquareGradGrad"));
1582 1583
    // square GradGrad: ddy=2x*ddx, dx=2dy*ddx
    // calculate dx first, so ddy can inplace ddx
1584
    if (dX) {
1585 1586 1587 1588
      auto dx = framework::EigenVector<T>::Flatten(
          GET_DATA_SAFELY(dX, "Output", "DX", "SquareGradGrad"));
      auto dout = framework::EigenVector<T>::Flatten(
          GET_DATA_SAFELY(dOut, "Output", "DOut", "SquareGradGrad"));
1589 1590
      dx.device(*d) = ddx * static_cast<T>(2) * dout;
    }
1591
    if (ddOut) {
1592 1593
      auto ddout = framework::EigenVector<T>::Flatten(
          GET_DATA_SAFELY(ddOut, "Output", "DDOut", "SquareGradGrad"));
1594 1595
      ddout.device(*d) = ddx * static_cast<T>(2) * x;
    }
1596 1597 1598 1599 1600 1601 1602 1603 1604 1605 1606 1607 1608 1609
  }
  static constexpr ActBwdOpFwdDeps FwdDeps() { return kDepX; }
};

// TODO(dengkaipeng): double gradient calculation for Square/Sqrt need
// DOut(dy) as input(not output), tensor extraction is different from
// others. Impliment extraction kernel seperately here.
inline void ExtractDoubleGradTensorWithInputDOut(
    const framework::ExecutionContext& ctx, const framework::Tensor** X,
    const framework::Tensor** ddX, framework::Tensor** dX,
    const framework::Tensor** dOut, framework::Tensor** ddOut) {
  // extract ddX(output), ddOut(input)
  auto ddx_var = ctx.InputVar("DDX");
  auto ddo_var = ctx.OutputVar("DDOut");
1610 1611 1612 1613
  PADDLE_ENFORCE_NOT_NULL(
      ddx_var, platform::errors::NotFound(
                   "Cannot get input Variable Out, variable name = %s",
                   ctx.InputName("DDX")));
1614 1615 1616 1617
  *ddX = ctx.Input<framework::Tensor>("DDX");
  if (ddo_var) {
    *ddOut = ctx.Output<framework::Tensor>("DDOut");
  }
1618 1619 1620 1621 1622
  PADDLE_ENFORCE_NOT_NULL(
      ddX,
      platform::errors::NotFound(
          "Cannot get the tensor from the Variable DDX, variable name = %s",
          ctx.OutputName("DDX")));
1623 1624 1625

  // extract x(input), dx(output)
  auto x_var = ctx.InputVar("X");
1626 1627
  PADDLE_ENFORCE_NOT_NULL(
      x_var, platform::errors::NotFound(
1628
                 "Cannot get input Variable Out, variable name = %s",
1629
                 ctx.InputName("X")));
1630 1631 1632 1633 1634 1635 1636 1637 1638 1639 1640 1641 1642 1643 1644 1645 1646 1647 1648 1649 1650 1651 1652 1653 1654 1655
  auto dx_var = ctx.OutputVar("DX");
  *X = ctx.Input<framework::Tensor>("X");
  if (dx_var) {
    *dX = ctx.Output<framework::Tensor>("DX");
  }

  // extract dOut(input)
  auto dout_var = ctx.InputVar("DOut");
  if (dout_var) {
    *dOut = ctx.Input<framework::Tensor>("DOut");
  }
}

template <typename DeviceContext, typename Functor>
class SquareDoubleGradKernel
    : public framework::OpKernel<typename Functor::ELEMENT_TYPE> {
 public:
  using T = typename Functor::ELEMENT_TYPE;
  void Compute(const framework::ExecutionContext& ctx) const override {
    const framework::Tensor *X, *ddX, *dOut;
    X = ddX = dOut = nullptr;
    framework::Tensor *dX, *ddOut;
    dX = ddOut = nullptr;

    ExtractDoubleGradTensorWithInputDOut(ctx, &X, &ddX, &dX, &dOut, &ddOut);

L
lvmengsi 已提交
1656 1657
    if (dX) dX->mutable_data<T>(X->dims(), ctx.GetPlace());
    if (ddOut) ddOut->mutable_data<T>(ctx.GetPlace());
1658 1659 1660 1661 1662 1663 1664 1665

    auto& place = ctx.template device_context<DeviceContext>();

    Functor functor;
    functor(place, X, ddX, ddOut, dOut, dX);
  }
};

D
Double_V 已提交
1666 1667 1668 1669 1670 1671 1672 1673 1674 1675 1676 1677 1678 1679 1680 1681 1682 1683 1684 1685 1686 1687 1688 1689 1690 1691 1692
template <typename DeviceContext, typename Functor>
class ELUDoubleGradKernel
    : public framework::OpKernel<typename Functor::ELEMENT_TYPE> {
 public:
  using T = typename Functor::ELEMENT_TYPE;
  void Compute(const framework::ExecutionContext& ctx) const override {
    const framework::Tensor *X, *ddX, *dOut;
    X = ddX = dOut = nullptr;
    framework::Tensor *dX, *ddOut;
    dX = ddOut = nullptr;

    ExtractDoubleGradTensorWithInputDOut(ctx, &X, &ddX, &dX, &dOut, &ddOut);

    if (dX) dX->mutable_data<T>(X->dims(), ctx.GetPlace());
    if (ddOut) ddOut->mutable_data<T>(ctx.GetPlace());

    auto& place = ctx.template device_context<DeviceContext>();

    Functor functor;
    auto attrs = functor.GetAttrs();
    for (auto& attr : attrs) {
      *attr.second = ctx.Attr<float>(attr.first);
    }
    functor(place, X, ddX, ddOut, dOut, dX);
  }
};

L
lvmengsi 已提交
1693 1694 1695 1696 1697 1698 1699 1700 1701 1702 1703 1704 1705 1706
template <typename DeviceContext, typename Functor>
class SqrtDoubleGradKernel
    : public framework::OpKernel<typename Functor::ELEMENT_TYPE> {
 public:
  using T = typename Functor::ELEMENT_TYPE;
  void Compute(const framework::ExecutionContext& ctx) const override {
    const framework::Tensor *Out, *dX, *ddX;
    Out = dX = ddX = nullptr;
    framework::Tensor *ddOut, *dOut;
    ddOut = dOut = nullptr;

    // extract ddx(input), ddout(output)
    auto ddx_var = ctx.InputVar("DDX");
    auto ddo_var = ctx.OutputVar("DDOut");
1707 1708 1709 1710
    PADDLE_ENFORCE_NOT_NULL(
        ddx_var, platform::errors::NotFound(
                     "Cannot get input Variable DDX, variable name = %s",
                     ctx.InputName("DDX")));
L
lvmengsi 已提交
1711 1712 1713 1714
    ddX = ctx.Input<framework::Tensor>("DDX");
    if (ddo_var) {
      ddOut = ctx.Output<framework::Tensor>("DDOut");
    }
1715 1716 1717 1718
    PADDLE_ENFORCE_NOT_NULL(
        ddX, platform::errors::NotFound(
                 "Cannot get input Variable DDX, variable name = %s",
                 ctx.InputName("DDX")));
L
lvmengsi 已提交
1719 1720 1721

    // extract out(input), dout(output)
    auto out_var = ctx.InputVar("Out");
1722 1723 1724 1725
    PADDLE_ENFORCE_NOT_NULL(
        out_var, platform::errors::NotFound(
                     "Cannot get input Variable Out, variable name = %s",
                     ctx.InputName("Out")));
L
lvmengsi 已提交
1726 1727 1728 1729 1730 1731 1732 1733
    auto dout_var = ctx.OutputVar("DOut");
    Out = ctx.Input<framework::Tensor>("Out");
    if (dout_var) {
      dOut = ctx.Output<framework::Tensor>("DOut");
    }

    // extract dx(input)
    auto dx_var = ctx.InputVar("DX");
1734 1735 1736 1737
    PADDLE_ENFORCE_NOT_NULL(
        dx_var, platform::errors::NotFound(
                    "Cannot get input Variable DX, variable name = %s",
                    ctx.InputName("DX")));
L
lvmengsi 已提交
1738 1739 1740 1741 1742 1743 1744 1745 1746 1747 1748 1749 1750 1751
    if (dx_var) {
      dX = ctx.Input<framework::Tensor>("DX");
    }

    if (dOut) dOut->mutable_data<T>(Out->dims(), ctx.GetPlace());
    if (ddOut) ddOut->mutable_data<T>(Out->dims(), ctx.GetPlace());

    auto& place = ctx.template device_context<DeviceContext>();

    Functor functor;
    functor(place, Out, ddX, ddOut, dOut, dX);
  }
};

1752 1753 1754 1755 1756 1757 1758 1759 1760 1761 1762
template <typename DeviceContext, typename Functor>
class PowKernel : public framework::OpKernel<typename Functor::ELEMENT_TYPE> {
 public:
  using T = typename Functor::ELEMENT_TYPE;

  void Compute(const framework::ExecutionContext& context) const override {
    const framework::Tensor* X = nullptr;
    framework::Tensor* Out = nullptr;
    ExtractActivationTensor(context, &X, &Out);
    Out->mutable_data<T>(context.GetPlace());

1763 1764 1765 1766
    auto x = framework::EigenVector<T>::Flatten(
        GET_DATA_SAFELY(X, "Input", "X", "Pow"));
    auto out = framework::EigenVector<T>::Flatten(
        GET_DATA_SAFELY(Out, "Output", "Out", "Pow"));
1767 1768 1769 1770 1771 1772 1773 1774 1775 1776 1777 1778 1779 1780 1781 1782 1783 1784 1785 1786 1787
    auto* place =
        context.template device_context<DeviceContext>().eigen_device();
    Functor functor;
    auto attrs = functor.GetAttrs();
    for (auto& attr : attrs) {
      *attr.second = context.Attr<float>(attr.first);
    }
    // get FactorTensor
    auto* factor_tensor = context.HasInput("FactorTensor")
                              ? context.Input<framework::Tensor>("FactorTensor")
                              : nullptr;
    if (factor_tensor) {
      auto* factor_data = factor_tensor->data<float>();
      framework::Tensor cpu_factor_tensor;
      if (platform::is_gpu_place(factor_tensor->place())) {
        TensorCopySync(*factor_tensor, platform::CPUPlace(),
                       &cpu_factor_tensor);
        factor_data = cpu_factor_tensor.data<float>();
      }
      auto factor =
          std::vector<float>(factor_data, factor_data + factor_tensor->numel());
1788 1789 1790 1791 1792
      PADDLE_ENFORCE_EQ(
          factor.size(), 1,
          platform::errors::InvalidArgument(
              "The shape of factor(tensor) must be [1] rather than %d",
              factor.size()));
1793 1794 1795 1796 1797 1798 1799 1800 1801 1802 1803 1804 1805 1806 1807 1808 1809 1810 1811 1812
      for (auto& attr : attrs) {
        *attr.second = factor[0];
      }
    }
    functor(*place, x, out);
  }
};

template <typename DeviceContext, typename Functor>
class PowGradKernel
    : public framework::OpKernel<typename Functor::ELEMENT_TYPE> {
 public:
  using T = typename Functor::ELEMENT_TYPE;
  void Compute(const framework::ExecutionContext& context) const override {
    const framework::Tensor *X, *Out, *dOut;
    framework::Tensor* dX = nullptr;
    X = Out = dOut = nullptr;
    ExtractActivationGradTensor<Functor::FwdDeps()>(context, &X, &Out, &dOut,
                                                    &dX);
    dX->mutable_data<T>(context.GetPlace());
1813 1814 1815 1816 1817 1818 1819 1820
    auto dout = framework::EigenVector<T>::Flatten(
        GET_DATA_SAFELY(dOut, "Input", "Out@GRAD", "PowGrad"));
    auto out = framework::EigenVector<T>::Flatten(
        GET_DATA_SAFELY(Out, "Input", "Out", "PowGrad"));
    auto dx = framework::EigenVector<T>::Flatten(
        GET_DATA_SAFELY(dX, "Output", "X@GRAD", "PowGrad"));
    auto x = framework::EigenVector<T>::Flatten(
        GET_DATA_SAFELY(X, "Input", "X", "PowGrad"));
1821 1822 1823 1824 1825 1826 1827 1828 1829 1830 1831 1832 1833 1834 1835 1836 1837 1838 1839 1840 1841 1842
    auto* place =
        context.template device_context<DeviceContext>().eigen_device();
    Functor functor;
    auto attrs = functor.GetAttrs();
    for (auto& attr : attrs) {
      *attr.second = context.Attr<float>(attr.first);
    }
    // get FactorTensor
    auto* factor_tensor =
        context.HasInput("FactorTensor")
            ? context.Input<framework::LoDTensor>("FactorTensor")
            : nullptr;
    if (factor_tensor) {
      auto* factor_data = factor_tensor->data<float>();
      framework::Tensor cpu_factor_tensor;
      if (platform::is_gpu_place(factor_tensor->place())) {
        TensorCopySync(*factor_tensor, platform::CPUPlace(),
                       &cpu_factor_tensor);
        factor_data = cpu_factor_tensor.data<float>();
      }
      auto factor =
          std::vector<float>(factor_data, factor_data + factor_tensor->numel());
1843 1844 1845 1846 1847
      PADDLE_ENFORCE_EQ(
          factor.size(), 1,
          platform::errors::InvalidArgument(
              "The shape of factor(tensor) must be [1] rather than %d",
              factor.size()));
1848 1849 1850 1851 1852 1853 1854
      for (auto& attr : attrs) {
        *attr.second = factor[0];
      }
    }
    functor(*place, x, out, dout, dx);
  }
};
Q
qijun 已提交
1855 1856
}  // namespace operators
}  // namespace paddle
1857

1858 1859 1860 1861 1862 1863
#define FOR_EACH_ACTIVATION_OP(__macro)                                       \
  __macro(sigmoid, Sigmoid, SigmoidFunctor, SigmoidGradFunctor);              \
  __macro(logsigmoid, LogSigmoid, LogSigmoidFunctor, LogSigmoidGradFunctor);  \
  __macro(tanh, Tanh, TanhFunctor, TanhGradFunctor);                          \
  __macro(atan, Atan, AtanFunctor, AtanGradFunctor);                          \
  __macro(softshrink, SoftShrink, SoftShrinkFunctor, SoftShrinkGradFunctor);  \
Z
zhoukunsheng 已提交
1864
  __macro(rsqrt, Rsqrt, RsqrtFunctor, RsqrtGradFunctor);                      \
1865 1866 1867 1868 1869 1870
  __macro(ceil, Ceil, CeilFunctor, ZeroGradFunctor);                          \
  __macro(floor, Floor, FloorFunctor, ZeroGradFunctor);                       \
  __macro(cos, Cos, CosFunctor, CosGradFunctor);                              \
  __macro(acos, Acos, AcosFunctor, AcosGradFunctor);                          \
  __macro(sin, Sin, SinFunctor, SinGradFunctor);                              \
  __macro(asin, Asin, AsinFunctor, AsinGradFunctor);                          \
1871 1872
  __macro(sinh, Sinh, SinhFunctor, SinhGradFunctor);                          \
  __macro(cosh, Cosh, CoshFunctor, CoshGradFunctor);                          \
1873 1874 1875
  __macro(round, Round, RoundFunctor, ZeroGradFunctor);                       \
  __macro(reciprocal, Reciprocal, ReciprocalFunctor, ReciprocalGradFunctor);  \
  __macro(log, Log, LogFunctor, LogGradFunctor);                              \
1876
  __macro(log1p, Log1p, Log1pFunctor, Log1pGradFunctor);                      \
1877 1878 1879 1880 1881 1882 1883 1884 1885 1886 1887 1888
  __macro(brelu, BRelu, BReluFunctor, BReluGradFunctor);                      \
  __macro(soft_relu, SoftRelu, SoftReluFunctor, SoftReluGradFunctor);         \
  __macro(stanh, STanh, STanhFunctor, STanhGradFunctor);                      \
  __macro(softplus, Softplus, SoftplusFunctor, SoftplusGradFunctor);          \
  __macro(softsign, Softsign, SoftsignFunctor, SoftsignGradFunctor);          \
  __macro(relu6, Relu6, Relu6Functor, Relu6GradFunctor);                      \
  __macro(tanh_shrink, TanhShrink, TanhShrinkFunctor, TanhShrinkGradFunctor); \
  __macro(hard_shrink, HardShrink, HardShrinkFunctor, HardShrinkGradFunctor); \
  __macro(hard_sigmoid, HardSigmoid, HardSigmoidFunctor,                      \
          HardSigmoidGradFunctor);                                            \
  __macro(swish, Swish, SwishFunctor, SwishGradFunctor);                      \
  __macro(thresholded_relu, ThresholdedRelu, ThresholdedReluFunctor,          \
H
huangjun12 已提交
1889 1890
          ThresholdedReluGradFunctor);                                        \
  __macro(hard_swish, HardSwish, HardSwishFunctor, HardSwishGradFunctor);