activation_grad_kernel.cc 23.2 KB
Newer Older
Y
ykkk2333 已提交
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
/* Copyright (c) 2022 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. */

#include "paddle/phi/kernels/activation_grad_kernel.h"

#include "paddle/phi/backends/xpu/enforce_xpu.h"
#include "paddle/phi/core/kernel_registry.h"
#include "paddle/phi/kernels/funcs/activation_functor.h"

namespace phi {

template <typename T, typename Context, typename Functor>
void ActivationGradXPUImpl(const Context& dev_ctx,
                           const DenseTensor* x,
                           const DenseTensor* out,
                           const DenseTensor* d_out,
                           DenseTensor* d_x,
                           const Functor& functor) {
  PADDLE_ENFORCE_NOT_NULL(
      d_out, errors::NotFound("The input DenseTensor dOut can not be nullptr"));
  PADDLE_ENFORCE_NOT_NULL(
      d_x, errors::NotFound("The output DenseTensor dX can not be nullptr"));
  if (!out) {
    out = d_out;  // fake out
  }
  dev_ctx.template Alloc<T>(d_x);
  functor(dev_ctx, x, out, d_out, d_x);
}

#define DEFINE_XPU_ACTIVATION_GRAD_KERNEL_DEPX(name, functor_class) \
  template <typename T, typename Context>                           \
  void name##GradKernel(const Context& dev_ctx,                     \
                        const DenseTensor& x,                       \
                        const DenseTensor& dout,                    \
                        DenseTensor* dx) {                          \
    functor_class<T> functor;                                       \
    ActivationGradXPUImpl<T, Context, functor_class<T>>(            \
        dev_ctx, &x, nullptr, &dout, dx, functor);                  \
  }

#define DEFINE_XPU_ACT_GRAD_KERNEL_WITH_ONE_ATTRS_DEPX(  \
    name, functor_class, attr)                           \
  template <typename T, typename Context>                \
  void name##GradKernel(const Context& dev_ctx,          \
                        const DenseTensor& x,            \
                        const DenseTensor& dout,         \
                        float attr,                      \
                        DenseTensor* dx) {               \
    functor_class<T> functor;                            \
    auto attrs = functor.GetAttrs();                     \
    *(attrs[0].second) = attr;                           \
    ActivationGradXPUImpl<T, Context, functor_class<T>>( \
        dev_ctx, &x, nullptr, &dout, dx, functor);       \
  }

#define DEFINE_XPU_ACT_GRAD_KERNEL_WITH_TWO_ATTRS_DEPX(  \
    name, functor_class, attr1, attr2)                   \
  template <typename T, typename Context>                \
  void name##GradKernel(const Context& dev_ctx,          \
                        const DenseTensor& x,            \
                        const DenseTensor& dout,         \
                        float attr1,                     \
                        float attr2,                     \
                        DenseTensor* dx) {               \
    functor_class<T> functor;                            \
    auto attrs = functor.GetAttrs();                     \
    *(attrs[0].second) = attr1;                          \
    *(attrs[1].second) = attr2;                          \
    ActivationGradXPUImpl<T, Context, functor_class<T>>( \
        dev_ctx, &x, nullptr, &dout, dx, functor);       \
  }

#define DEFINE_XPU_ACTIVATION_GRAD_KERNEL_DEPOUT(name, functor_class) \
  template <typename T, typename Context>                             \
  void name##GradKernel(const Context& dev_ctx,                       \
                        const DenseTensor& out,                       \
                        const DenseTensor& dout,                      \
                        DenseTensor* dx) {                            \
    functor_class<T> functor;                                         \
    ActivationGradXPUImpl<T, Context, functor_class<T>>(              \
        dev_ctx, nullptr, &out, &dout, dx, functor);                  \
  }

#define DEFINE_XPU_ACT_GRAD_KERNEL_WITH_ONE_ATTRS_DEPOUT( \
    name, functor_class, attr)                            \
  template <typename T, typename Context>                 \
  void name##GradKernel(const Context& dev_ctx,           \
                        const DenseTensor& out,           \
                        const DenseTensor& dout,          \
                        float attr,                       \
                        DenseTensor* dx) {                \
    functor_class<T> functor;                             \
    auto attrs = functor.GetAttrs();                      \
    *(attrs[0].second) = attr;                            \
    ActivationGradXPUImpl<T, Context, functor_class<T>>(  \
        dev_ctx, nullptr, &out, &dout, dx, functor);      \
  }

#define DEFINE_XPU_ACT_GRAD_KERNEL_WITH_TWO_ATTRS_DEPOUT( \
    name, functor_class, attr1, attr2)                    \
  template <typename T, typename Context>                 \
  void name##GradKernel(const Context& dev_ctx,           \
                        const DenseTensor& out,           \
                        const DenseTensor& dout,          \
                        float attr1,                      \
                        float attr2,                      \
                        DenseTensor* dx) {                \
    functor_class<T> functor;                             \
    auto attrs = functor.GetAttrs();                      \
    *(attrs[0].second) = attr1;                           \
    *(attrs[1].second) = attr2;                           \
    ActivationGradXPUImpl<T, Context, functor_class<T>>(  \
        dev_ctx, nullptr, &out, &dout, dx, functor);      \
  }

#define DEFINE_XPU_ACTIVATION_GRAD_KERNEL_NODEP(name, functor_class)      \
  template <typename T, typename Context>                                 \
  void name##GradKernel(                                                  \
      const Context& dev_ctx, const DenseTensor& dout, DenseTensor* dx) { \
    functor_class<T> functor;                                             \
    ActivationGradXPUImpl<T, Context, functor_class<T>>(                  \
        dev_ctx, nullptr, nullptr, &dout, dx, functor);                   \
  }

template <typename Context, typename T, typename XPUType>
int xpu_activation_backward(const Context& dev_ctx,
                            const DenseTensor* x,
                            const DenseTensor* out,
                            const DenseTensor* dout,
                            DenseTensor* dx,
                            std::function<int(xpu::Context*,
                                              const XPUType*,
                                              const XPUType*,
                                              const XPUType*,
                                              XPUType*,
                                              int)> func) {
  /* TODO: relu tanh sigmoid are inplace */
  const XPUType* x_data = nullptr;
  const XPUType* y_data = nullptr;
  const XPUType* y_grad = nullptr;
  if (x != nullptr) x_data = reinterpret_cast<const XPUType*>(x->data<T>());
  if (out != nullptr) y_data = reinterpret_cast<const XPUType*>(out->data<T>());
  if (dout != nullptr)
    y_grad = reinterpret_cast<const XPUType*>(dout->data<T>());
  XPUType* x_grad = reinterpret_cast<XPUType*>(dx->data<T>());

  int r =
      func(dev_ctx.x_context(), x_data, y_data, y_grad, x_grad, dx->numel());
  return r;
}

template <typename T>
struct XPULogGradFunctor : public funcs::BaseActivationFunctor<T> {
  template <typename Context>
  void operator()(const Context& dev_ctx,
                  const DenseTensor* x,
                  const DenseTensor* out,
                  const DenseTensor* dOut,
                  DenseTensor* dX) const {
    const T* x_data = nullptr;
172
    const T* dout_data = nullptr;
Y
ykkk2333 已提交
173
    if (x != nullptr) x_data = x->data<T>();
174 175 176 177 178
    if (dOut != nullptr) dout_data = dOut->data<T>();

    T* dx_data = dev_ctx.template Alloc<T>(dX);
    int r = xpu::constant<T>(
        dev_ctx.x_context(), dx_data, x->numel(), static_cast<T>(1.0));
Y
ykkk2333 已提交
179 180
    PADDLE_ENFORCE_XDNN_SUCCESS(r, "constant");

181 182 183 184 185 186 187
    auto x_dims = vectorize<int>(x->dims());

    // use [1] to replace [], because xpu not support []
    if (x_dims.size() == 0) {
      x_dims = std::vector<int>({1});
    }

Y
ykkk2333 已提交
188 189
    // dx.device(d) = dout * (static_cast<T>(1) / x);
    r = xpu::broadcast_div(dev_ctx.x_context(),
190
                           reinterpret_cast<const float*>(dx_data),
Y
ykkk2333 已提交
191
                           reinterpret_cast<const float*>(x_data),
192 193 194
                           reinterpret_cast<float*>(dx_data),
                           x_dims,
                           x_dims);
Y
ykkk2333 已提交
195 196 197
    PADDLE_ENFORCE_XDNN_SUCCESS(r, "broadcast_div");

    r = xpu::broadcast_mul(dev_ctx.x_context(),
198 199 200 201 202
                           reinterpret_cast<const float*>(dx_data),
                           reinterpret_cast<const float*>(dout_data),
                           reinterpret_cast<float*>(dx_data),
                           x_dims,
                           x_dims);
Y
ykkk2333 已提交
203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574
    PADDLE_ENFORCE_XDNN_SUCCESS(r, "broadcast_mul");
  }
};

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

  template <typename Context>
  void operator()(const Context& dev_ctx,
                  const DenseTensor* x,
                  const DenseTensor* out,
                  const DenseTensor* dout,
                  DenseTensor* dx) const {
    const T* x_data = nullptr;
    const T* y_grad = nullptr;
    if (x != nullptr) x_data = x->data<T>();
    if (dout != nullptr) y_grad = dout->data<T>();
    T* x_grad = dx->data<T>();
    auto xpu_context = dev_ctx.x_context();

    // The signs of x and y are the same,
    // y == nullptr here,
    // so we give 2 x to the api
    int r = xpu::leaky_relu_grad(xpu_context,
                                 reinterpret_cast<const float*>(x_data),
                                 reinterpret_cast<const float*>(x_data),
                                 reinterpret_cast<const float*>(y_grad),
                                 reinterpret_cast<float*>(x_grad),
                                 dx->numel(),
                                 alpha);
    PADDLE_ENFORCE_XDNN_SUCCESS(r, "leaky_relu_grad");
  }
};

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

  typename funcs::BaseActivationFunctor<T>::AttrPair GetAttrs() {
    return {{"threshold", &threshold}, {"scale", &scale}, {"offset", &offset}};
  }
  template <typename Context>
  void operator()(const Context& dev_ctx,
                  const DenseTensor* x,
                  const DenseTensor* out,
                  const DenseTensor* dout,
                  DenseTensor* dx) const {
    using XPUType = typename XPUTypeTrait<T>::Type;
    PADDLE_ENFORCE_EQ(
        threshold,
        6.0f,
        errors::External("Not support threshold [%f] in XPU", threshold));
    PADDLE_ENFORCE_EQ(
        scale, 6.0f, errors::External("Not support scale [%f] in XPU", scale));
    PADDLE_ENFORCE_EQ(
        offset,
        3.0f,
        errors::External("Not support offset [%f] in XPU", offset));
    int r = xpu_activation_backward<Context, T, XPUType>(
        dev_ctx, x, out, dout, dx, xpu::hard_swish_grad<XPUType>);
    PADDLE_ENFORCE_XDNN_SUCCESS(r, "hard_swish_grad");
  }
};

template <typename T>
struct XPUReciprocalGradFunctor : public funcs::BaseActivationFunctor<T> {
  using XPUType = typename XPUTypeTrait<T>::Type;
  template <typename Context>
  void operator()(const Context& dev_ctx,
                  const DenseTensor* x,
                  const DenseTensor* out,
                  const DenseTensor* dout,
                  DenseTensor* dx) const {
    int r = xpu_activation_backward<Context, T, XPUType>(
        dev_ctx, x, out, dout, dx, xpu::reciprocal_grad<XPUType>);
    PADDLE_ENFORCE_XDNN_SUCCESS(r, "reciprocal_grad");
  }
};

template <typename T>
struct XPUReluGradFunctor : public funcs::BaseActivationFunctor<T> {
  using XPUType = typename XPUTypeTrait<T>::Type;
  template <typename Context>
  void operator()(const Context& dev_ctx,
                  const DenseTensor* x,
                  const DenseTensor* out,
                  const DenseTensor* dout,
                  DenseTensor* dx) const {
    int r = xpu_activation_backward<Context, T, XPUType>(
        dev_ctx, x, out, dout, dx, xpu::relu_grad<XPUType>);
    PADDLE_ENFORCE_XDNN_SUCCESS(r, "relu_grad");
  }
};

template <typename T>
struct XPURelu6GradFunctor : public funcs::BaseActivationFunctor<T> {
  using XPUType = typename XPUTypeTrait<T>::Type;
  float threshold;
  typename funcs::BaseActivationFunctor<T>::AttrPair GetAttrs() {
    return {{"threshold", &threshold}};
  }
  template <typename Context>
  void operator()(const Context& dev_ctx,
                  const DenseTensor* x,
                  const DenseTensor* out,
                  const DenseTensor* dout,
                  DenseTensor* dx) const {
    int r = xpu_activation_backward<Context, T, XPUType>(
        dev_ctx, x, out, dout, dx, xpu::relu6_grad<XPUType>);
    PADDLE_ENFORCE_XDNN_SUCCESS(r, "relu6_grad");
  }
};

template <typename T>
struct XPUSigmoidGradFunctor : public funcs::BaseActivationFunctor<T> {
  using XPUType = typename XPUTypeTrait<T>::Type;
  template <typename Context>
  void operator()(const Context& dev_ctx,
                  const DenseTensor* x,
                  const DenseTensor* out,
                  const DenseTensor* dout,
                  DenseTensor* dx) const {
    int r = xpu_activation_backward<Context, T, XPUType>(
        dev_ctx, x, out, dout, dx, xpu::sigmoid_grad<XPUType>);
    PADDLE_ENFORCE_XDNN_SUCCESS(r, "sigmoid_grad");
  }
};

template <typename T>
struct XPUTanhGradFunctor : public funcs::BaseActivationFunctor<T> {
  using XPUType = typename XPUTypeTrait<T>::Type;
  template <typename Context>
  void operator()(const Context& dev_ctx,
                  const DenseTensor* x,
                  const DenseTensor* out,
                  const DenseTensor* dout,
                  DenseTensor* dx) const {
    int r = xpu_activation_backward<Context, T, XPUType>(
        dev_ctx, x, out, dout, dx, xpu::tanh_grad<XPUType>);
    PADDLE_ENFORCE_XDNN_SUCCESS(r, "tanh_grad");
  }
};

template <typename T>
struct XPUSquareGradFunctor : public funcs::BaseActivationFunctor<T> {
  using XPUType = typename XPUTypeTrait<T>::Type;
  template <typename Context>
  void operator()(const Context& dev_ctx,
                  const DenseTensor* x,
                  const DenseTensor* out,
                  const DenseTensor* dout,
                  DenseTensor* dx) const {
    int r = xpu_activation_backward<Context, T, XPUType>(
        dev_ctx, x, out, dout, dx, xpu::square_grad<XPUType>);
    PADDLE_ENFORCE_XDNN_SUCCESS(r, "square_grad");
  }
};

template <typename T>
struct XPUSqrtGradFunctor : public funcs::BaseActivationFunctor<T> {
  using XPUType = typename XPUTypeTrait<T>::Type;
  template <typename Context>
  void operator()(const Context& dev_ctx,
                  const DenseTensor* x,
                  const DenseTensor* out,
                  const DenseTensor* dout,
                  DenseTensor* dx) const {
    int r = xpu_activation_backward<Context, T, XPUType>(
        dev_ctx, x, out, dout, dx, xpu::sqrt_grad<XPUType>);
    PADDLE_ENFORCE_XDNN_SUCCESS(r, "sqrt_grad");
  }
};

template <typename T, typename Context>
void PowGradKernel(const Context& dev_ctx,
                   const DenseTensor& x,
                   const DenseTensor& dout,
                   const Scalar& factor,
                   DenseTensor* dx) {
  dev_ctx.template Alloc<T>(dx);
  const T* x_data = x.data<T>();
  const T* y_grad = dout.data<T>();
  T* x_grad = dx->data<T>();

  // check dims: all dims should equal
  auto x_dims = vectorize<int>(x.dims());
  auto dy_dims = vectorize<int>(dout.dims());
  auto dx_dims = vectorize<int>(dx->dims());
  PADDLE_ENFORCE_EQ(x_dims,
                    dy_dims,
                    errors::PreconditionNotMet("x_dims should match dy_dims."));
  PADDLE_ENFORCE_EQ(x_dims,
                    dx_dims,
                    errors::PreconditionNotMet("x_dims should match dx_dims."));
  float pow_factor = factor.to<float>();

  auto xpu_context = dev_ctx.x_context();
  // int pow_grad(Context* ctx, const T* x, const T* dy, T* dx, int len, float
  // factor);
  int r =
      xpu::pow_grad(xpu_context, x_data, y_grad, x_grad, x.numel(), pow_factor);
  PADDLE_ENFORCE_XDNN_SUCCESS(r, "pow_grad");
}

template <typename T>
struct XPUSwishGradFunctor : public funcs::BaseActivationFunctor<T> {
  using XPUType = typename XPUTypeTrait<T>::Type;
  float beta;
  typename funcs::BaseActivationFunctor<T>::AttrPair GetAttrs() {
    return {{"beta", &beta}};
  }

  template <typename Context>
  void operator()(const Context& dev_ctx,
                  const DenseTensor* x,
                  const DenseTensor* out,
                  const DenseTensor* dout,
                  DenseTensor* dx) const {
    const XPUType* x_data = reinterpret_cast<const XPUType*>(x->data<T>());
    const XPUType* y_grad = reinterpret_cast<const XPUType*>(dout->data<T>());
    XPUType* x_grad = reinterpret_cast<XPUType*>(dx->data<T>());

    auto xpu_context = dev_ctx.x_context();
    int r = xpu::swish_grad(xpu_context, x_data, y_grad, x_grad, dx->numel());
    PADDLE_ENFORCE_XDNN_SUCCESS(r, "swish_grad");
  }
};

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

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

  template <typename Context>
  void operator()(const Context& dev_ctx,
                  const DenseTensor* x,
                  const DenseTensor* out,
                  const DenseTensor* dout,
                  DenseTensor* dx) const {
    const T* x_data = x->data<T>();
    const T* y_grad = dout->data<T>();
    T* x_grad = dx->data<T>();

    auto xpu_context = dev_ctx.x_context();
    int r = xpu::mish_grad(
        xpu_context,
        reinterpret_cast<const float*>(x_data),
        reinterpret_cast<const float*>(x_data),  // mish_grad do not need y_data
        reinterpret_cast<const float*>(y_grad),
        reinterpret_cast<float*>(x_grad),
        dx->numel(),
        threshold);
    PADDLE_ENFORCE_XDNN_SUCCESS(r, "mish_grad");
  }
};

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

  template <typename Context>
  void operator()(const Context& dev_ctx,
                  const DenseTensor* x,
                  const DenseTensor* out,
                  const DenseTensor* dOut,
                  DenseTensor* dX) const {
    const T* x_data = x->data<T>();
    const T* y_grad = dOut->data<T>();
    T* x_grad = dX->data<T>();

    auto xpu_context = dev_ctx.x_context();
    int r = xpu::softplus_grad(xpu_context,
                               reinterpret_cast<const float*>(x_data),
                               reinterpret_cast<const float*>(
                                   x_data),  // softplus_grad do not need y_data
                               reinterpret_cast<const float*>(y_grad),
                               reinterpret_cast<float*>(x_grad),
                               dX->numel(),
                               beta,
                               threshold);
    PADDLE_ENFORCE_XDNN_SUCCESS(r, "softplus_grad");
  }
};

DEFINE_XPU_ACTIVATION_GRAD_KERNEL_DEPOUT(Reciprocal, XPUReciprocalGradFunctor);
DEFINE_XPU_ACTIVATION_GRAD_KERNEL_DEPOUT(Sigmoid, XPUSigmoidGradFunctor);
DEFINE_XPU_ACTIVATION_GRAD_KERNEL_DEPOUT(Sqrt, XPUSqrtGradFunctor);
DEFINE_XPU_ACTIVATION_GRAD_KERNEL_DEPOUT(Tanh, XPUTanhGradFunctor);
DEFINE_XPU_ACTIVATION_GRAD_KERNEL_DEPOUT(Relu, XPUReluGradFunctor);

DEFINE_XPU_ACTIVATION_GRAD_KERNEL_DEPX(Log, XPULogGradFunctor);
DEFINE_XPU_ACTIVATION_GRAD_KERNEL_DEPX(Square, XPUSquareGradFunctor);

DEFINE_XPU_ACT_GRAD_KERNEL_WITH_ONE_ATTRS_DEPX(Swish,
                                               XPUSwishGradFunctor,
                                               beta);
DEFINE_XPU_ACT_GRAD_KERNEL_WITH_ONE_ATTRS_DEPX(Mish,
                                               XPUMishGradFunctor,
                                               threshold);
DEFINE_XPU_ACT_GRAD_KERNEL_WITH_ONE_ATTRS_DEPX(LeakyRelu,
                                               XPULeakyReluGradFunctor,
                                               alpha);

DEFINE_XPU_ACT_GRAD_KERNEL_WITH_ONE_ATTRS_DEPOUT(Relu6,
                                                 XPURelu6GradFunctor,
                                                 threshold);

DEFINE_XPU_ACT_GRAD_KERNEL_WITH_TWO_ATTRS_DEPX(Softplus,
                                               XPUSoftPlusGradFunctor,
                                               beta,
                                               threshold)

template <typename T, typename Context>
void HardSwishGradKernel(const Context& dev_ctx,
                         const DenseTensor& x,
                         const DenseTensor& dout,
                         float threshold,
                         float scale,
                         float offset,
                         DenseTensor* dx) {
  XPUHardSwishGradFunctor<T> functor;
  auto attrs = functor.GetAttrs();
  *(attrs[0].second) = threshold;
  *(attrs[1].second) = scale;
  *(attrs[2].second) = offset;
  ActivationGradXPUImpl<T, Context, XPUHardSwishGradFunctor<T>>(
      dev_ctx, &x, nullptr, &dout, dx, functor);
}

}  // namespace phi

PD_REGISTER_KERNEL(relu_grad,
                   XPU,
                   ALL_LAYOUT,
                   phi::ReluGradKernel,
                   float,
                   phi::dtype::float16) {}

#define PD_REGISTER_ACTIVATION_GRAD_KERNEL(name, func) \
  PD_REGISTER_KERNEL(name, XPU, ALL_LAYOUT, phi::func, float) {}

PD_REGISTER_KERNEL(tanh_grad,
                   XPU,
                   ALL_LAYOUT,
                   phi::TanhGradKernel,
                   float,
                   phi::dtype::float16) {}
PD_REGISTER_ACTIVATION_GRAD_KERNEL(log_grad, LogGradKernel)
PD_REGISTER_ACTIVATION_GRAD_KERNEL(leaky_relu_grad, LeakyReluGradKernel)
PD_REGISTER_ACTIVATION_GRAD_KERNEL(hard_swish_grad, HardSwishGradKernel)
PD_REGISTER_ACTIVATION_GRAD_KERNEL(reciprocal_grad, ReciprocalGradKernel)
PD_REGISTER_ACTIVATION_GRAD_KERNEL(relu6_grad, Relu6GradKernel)
PD_REGISTER_ACTIVATION_GRAD_KERNEL(sigmoid_grad, SigmoidGradKernel)
PD_REGISTER_ACTIVATION_GRAD_KERNEL(sqrt_grad, SqrtGradKernel)
PD_REGISTER_ACTIVATION_GRAD_KERNEL(mish_grad, MishGradKernel)
PD_REGISTER_ACTIVATION_GRAD_KERNEL(swish_grad, SwishGradKernel)
PD_REGISTER_ACTIVATION_GRAD_KERNEL(softplus_grad, SoftplusGradKernel)
PD_REGISTER_ACTIVATION_GRAD_KERNEL(square_grad, SquareGradKernel)
PD_REGISTER_KERNEL(pow_grad, XPU, ALL_LAYOUT, phi::PowGradKernel, float) {}