/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
    http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */

#include "paddle/fluid/operators/activation_op.h"
#include "paddle/fluid/operators/amp/fp16_type_traits.h"
#include "paddle/fluid/operators/elementwise/elementwise_op_impl.cu.h"
#include "paddle/fluid/platform/bfloat16.h"
#include "paddle/fluid/platform/device/gpu/gpu_device_function.h"

#include "paddle/phi/kernels/funcs/activation_functor.h"

namespace paddle {
namespace operators {

template <typename T>
struct CudaSigmoidFunctor : public BaseActivationFunctor<T> {
  using MPType = typename details::MPTypeTrait<T>::Type;
  MPType one = static_cast<MPType>(1.0f);

  // sigmoid(x) = 1 / (1 + exp(-x))
  __device__ __forceinline__ T operator()(const T arg_x) const {
    MPType x = static_cast<MPType>(arg_x);
    return static_cast<T>(one / (one + exp(-x)));
  }
};

template <typename T>
struct CudaSigmoidGradFunctor : public BaseActivationFunctor<T> {
  T one = static_cast<T>(1.0f);

  // dx = dout * out * (1 - out)
  __device__ __forceinline__ T operator()(const T dout, const T out) const {
    return dout * out * (one - out);
  }

  static constexpr ActBwdOpFwdDeps FwdDeps() {
    return ActBwdOpFwdDeps::kDepOut;
  }
};

template <typename T>
struct CudaLogSigmoidFunctor : public BaseActivationFunctor<T> {
  using MPType = typename details::MPTypeTrait<T>::Type;
  MPType zero = static_cast<MPType>(0.0f);

  // logsigmoid(x) = log(1 / (1 + exp(-x)))
  // For numerical stability,
  // logsigmoid(x) =
  //          - (max(-x, 0) + log(exp(-max(-x, 0)) + exp(-x - max(-x, 0))))
  __device__ __forceinline__ T operator()(const T arg_x) const {
    MPType x = static_cast<MPType>(arg_x);
    MPType temp = x > zero ? zero : -x;
    return static_cast<T>(-temp - log(exp(-temp) + exp(-x - temp)));
  }
};

template <typename T>
struct CudaLogSigmoidGradFunctor : public BaseActivationFunctor<T> {
  using MPType = typename details::MPTypeTrait<T>::Type;
  MPType zero = static_cast<MPType>(0.0f);

  // dx = dout * exp(-x) / (1 + exp(-x))
  // For numerical stability:
  // dx = dout * exp(-x - max(-x, 0)) / (exp(-max(-x, 0)) + exp(-x - max(-x,
  // 0)))
  __device__ __forceinline__ T operator()(const T arg_dout,
                                          const T arg_x) const {
    MPType dout = static_cast<MPType>(arg_dout);
    MPType x = static_cast<MPType>(arg_x);
    MPType temp1 = x > zero ? zero : -x;
    MPType temp2 = exp(-x - temp1);
    return static_cast<T>(dout * (temp2 / (exp(-temp1) + temp2)));
  }

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

template <typename T>
struct CudaCeilFunctor : public BaseActivationFunctor<T> {
  using MPType = typename details::MPTypeTrait<T>::Type;

  // ceil(x) = ceil(x)
  __device__ __forceinline__ T operator()(const T arg_x) const {
    MPType x = static_cast<MPType>(arg_x);
    return static_cast<T>(ceil(x));
  }
};

template <typename T>
struct CudaFloorFunctor : public BaseActivationFunctor<T> {
  using MPType = typename details::MPTypeTrait<T>::Type;

  // floor(x) = floor(x)
  __device__ __forceinline__ T operator()(const T arg_x) const {
    MPType x = static_cast<MPType>(arg_x);
    return static_cast<T>(floor(x));
  }
};

template <typename T>
struct CudaRoundFunctor : public BaseActivationFunctor<T> {
  using MPType = typename details::MPTypeTrait<T>::Type;

  // round(x) = round(x)
  __device__ __forceinline__ T operator()(const T arg_x) const {
    MPType x = static_cast<MPType>(arg_x);
    return static_cast<T>(round(x));
  }
};

// GradFunctor for ceil, floor and round
template <typename T>
struct CudaZeroGradFunctor : public BaseActivationFunctor<T> {
  __device__ __forceinline__ T operator()(const T x) const {
    return static_cast<T>(0.0f);
  }

  static constexpr ActBwdOpFwdDeps FwdDeps() {
    return ActBwdOpFwdDeps::kNoDeps;
  }
};

template <typename T>
struct CudaReciprocalGradFunctor : public BaseActivationFunctor<T> {
  // dx = -dout * out^2
  __device__ __forceinline__ T operator()(const T dout, const T out) const {
    return -dout * out * out;
  }

  static constexpr ActBwdOpFwdDeps FwdDeps() {
    return ActBwdOpFwdDeps::kDepOut;
  }
};

template <typename T>
struct CudaLogFunctor : public BaseActivationFunctor<T> {
  using MPType = typename details::MPTypeTrait<T>::Type;

  // log(x) = log(x)
  __device__ __forceinline__ T operator()(const T arg_x) const {
    MPType x = static_cast<MPType>(arg_x);
    return static_cast<T>(log(x));
  }
};

template <typename T>
struct CudaLogGradFunctor : public BaseActivationFunctor<T> {
  // dx = dout / x
  __device__ __forceinline__ T operator()(const T dout, const T x) const {
    return dout / x;
  }

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

template <typename T>
struct CudaSquareGradFunctor : public BaseActivationFunctor<T> {
  T two = static_cast<T>(2.0f);

  // dx = dout * 2 * x
  __device__ __forceinline__ T operator()(const T dout, const T x) const {
    return dout * two * x;
  }

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

template <typename T>
struct CudaSqrtGradFunctor : public BaseActivationFunctor<T> {
  T one_half = static_cast<T>(0.5f);

  // dx = dout * 0.5 / out
  __device__ __forceinline__ T operator()(const T dout, const T out) const {
    return one_half * dout / out;
  }

  static constexpr ActBwdOpFwdDeps FwdDeps() {
    return ActBwdOpFwdDeps::kDepOut;
  }
};

template <typename T>
struct CudaRsqrtGradFunctor : public BaseActivationFunctor<T> {
  T minus_one_half = static_cast<T>(-0.5f);

  // dx = -0.5 * dout * out^3
  __device__ __forceinline__ T operator()(const T dout, const T out) const {
    return minus_one_half * dout * out * out * out;
  }

  static constexpr ActBwdOpFwdDeps FwdDeps() {
    return ActBwdOpFwdDeps::kDepOut;
  }
};

template <typename T>
struct CudaLog1pFunctor : public BaseActivationFunctor<T> {
  using MPType = typename details::MPTypeTrait<T>::Type;
  MPType one = static_cast<MPType>(1.0f);

  // log1p(x) = log(1 + x)
  __device__ __forceinline__ T operator()(const T arg_x) const {
    MPType x = static_cast<MPType>(arg_x);
    return static_cast<T>(log(one + x));
  }
};

template <typename T>
struct CudaLog1pGradFunctor : public BaseActivationFunctor<T> {
  T one = static_cast<T>(1.0f);

  // dx = dout / (1 + x)
  __device__ __forceinline__ T operator()(const T dout, const T x) const {
    return dout / (one + x);
  }

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

template <typename T>
struct CudaLog2Functor : public BaseActivationFunctor<T> {
  using MPType = typename details::MPTypeTrait<T>::Type;

  // log2(x) = log2(x)
  __device__ __forceinline__ T operator()(const T arg_x) const {
    MPType x = static_cast<MPType>(arg_x);
    return static_cast<T>(log2(x));
  }
};

template <typename T>
struct CudaLog2GradFunctor : public BaseActivationFunctor<T> {
  using MPType = typename details::MPTypeTrait<T>::Type;
  T log_two = static_cast<T>(log(static_cast<MPType>(2.0f)));

  // dx = dout / (x * log(2))
  __device__ __forceinline__ T operator()(const T dout, const T x) const {
    return dout / (x * log_two);
  }

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

template <typename T>
struct CudaLog10Functor : public BaseActivationFunctor<T> {
  using MPType = typename details::MPTypeTrait<T>::Type;

  // log10(x) = log10(x)
  __device__ __forceinline__ T operator()(const T arg_x) const {
    MPType x = static_cast<MPType>(arg_x);
    return static_cast<T>(log10(x));
  }
};

template <typename T>
struct CudaLog10GradFunctor : public BaseActivationFunctor<T> {
  using MPType = typename details::MPTypeTrait<T>::Type;
  T log_ten = static_cast<T>(log(static_cast<MPType>(10.0f)));

  // dx = dout / (x * log(10))
  __device__ __forceinline__ T operator()(const T dout, const T x) const {
    return dout / (x * log_ten);
  }

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

template <typename T>
struct CudaSoftReluFunctor : public BaseActivationFunctor<T> {
  using MPType = typename details::MPTypeTrait<T>::Type;
  MPType one = static_cast<MPType>(1.0f);
  float threshold;

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

  // soft_relu(x) = log(1 + exp(max(min(x, threshold), -threshold)))
  // threshold should not be negative
  __device__ __forceinline__ T operator()(const T arg_x) const {
    MPType x = static_cast<MPType>(arg_x);
    MPType t = static_cast<MPType>(threshold);
    MPType temp_min = x < t ? x : t;
    MPType temp_max = temp_min > -t ? temp_min : -t;
    return static_cast<T>(log(one + exp(temp_max)));
  }
};

template <typename T>
struct CudaSoftReluGradFunctor : public BaseActivationFunctor<T> {
  using MPType = typename details::MPTypeTrait<T>::Type;
  MPType one = static_cast<MPType>(1.0f);
  float threshold;

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

  // dx = (out > -threshold && out < threshold) ? dout * (1 - exp(-out)) : 0
  // threshold should not be negative
  __device__ __forceinline__ T operator()(const T arg_dout,
                                          const T arg_out) const {
    MPType dout = static_cast<MPType>(arg_dout);
    MPType out = static_cast<MPType>(arg_out);
    MPType t = static_cast<MPType>(threshold);
    return (out > -t && out < t) ? static_cast<T>(dout * (one - exp(-out)))
                                 : static_cast<T>(0.0f);
  }

  static constexpr ActBwdOpFwdDeps FwdDeps() {
    return ActBwdOpFwdDeps::kDepOut;
  }
};

template <typename T>
struct CudaSTanhGradFunctor : public BaseActivationFunctor<T> {
  using MPType = typename details::MPTypeTrait<T>::Type;
  MPType one = static_cast<MPType>(1.0f);
  float scale_a;
  float scale_b;

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

  // dx = dout * a * b * (1 - tanh(a * x) * tanh(a * x))
  __device__ __forceinline__ T operator()(const T arg_dout,
                                          const T arg_x) const {
    MPType dout = static_cast<MPType>(arg_dout);
    MPType x = static_cast<MPType>(arg_x);
    MPType a = static_cast<MPType>(scale_a);
    MPType b = static_cast<MPType>(scale_b);
    MPType temp = tanh(a * x);
    return static_cast<T>(dout * a * b * (one - temp * temp));
  }

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

template <typename T>
struct CudaSoftplusGradFunctor : public BaseActivationFunctor<T> {
  using MPType = typename details::MPTypeTrait<T>::Type;
  MPType one = static_cast<MPType>(1.0f);
  float beta;
  float threshold;

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

  // dx = x * beta > threshold ? dout : dout / (1 + exp(-beta * x))
  __device__ __forceinline__ T operator()(const T arg_dout,
                                          const T arg_x) const {
    MPType dout = static_cast<MPType>(arg_dout);
    MPType x = static_cast<MPType>(arg_x);
    MPType b = static_cast<MPType>(beta);
    MPType t = static_cast<MPType>(threshold);
    MPType x_beta = x * beta;
    return x_beta > t ? arg_dout : static_cast<T>(dout / (one + exp(-x_beta)));
  }

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

template <typename T>
struct CudaSoftsignGradFunctor : public BaseActivationFunctor<T> {
  T one = static_cast<T>(1.0f);

  // dx = dout / (1 + abs(x))^2
  __device__ __forceinline__ T operator()(const T dout, const T x) const {
    T temp = one + abs(x);
    return dout / (temp * temp);
  }

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

template <typename T>
struct CudaRelu6Functor : public BaseActivationFunctor<T> {
  T zero = static_cast<T>(0.0f);
  float threshold;

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

  // relu6(x) = min(max(0, x), 6)
  __device__ __forceinline__ T operator()(const T x) const {
    T t = static_cast<T>(threshold);
    return x <= zero ? zero : (x < t ? x : t);
  }
};

template <typename T>
struct CudaRelu6GradFunctor : public BaseActivationFunctor<T> {
  T zero = static_cast<T>(0.0f);
  float threshold;

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

  // dx = (out > 0 && out < t) ? dout : 0
  __device__ __forceinline__ T operator()(const T dout, const T out) const {
    T t = static_cast<T>(threshold);
    return (out > zero && out < t) ? dout : zero;
  }

  static constexpr ActBwdOpFwdDeps FwdDeps() {
    return ActBwdOpFwdDeps::kDepOut;
  }
};

template <typename T>
struct CudaHardSigmoidFunctor : public BaseActivationFunctor<T> {
  T zero = static_cast<T>(0.0f);
  T one = static_cast<T>(1.0f);
  float slope;
  float offset;

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

  // hard_sigmoid(x) = 0, when x <= -3
  //                   1, when x >= 3
  //                   x * slope + offset, otherwise
  __device__ __forceinline__ T operator()(const T x) const {
    T temp = x * static_cast<T>(slope) + static_cast<T>(offset);
    T temp_max = temp > zero ? temp : zero;
    T temp_min = temp_max < one ? temp_max : one;
    return temp_min;
  }
};

template <typename T>
struct CudaHardSigmoidGradFunctor : public BaseActivationFunctor<T> {
  T zero = static_cast<T>(0.0f);
  T one = static_cast<T>(1.0f);
  float slope;
  float offset;

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

  // dx = (out > 0 && out < 1) ? dout * slope : 0
  __device__ __forceinline__ T operator()(const T dout, const T out) const {
    return (out > zero && out < one) ? dout * static_cast<T>(slope) : zero;
  }

  static constexpr ActBwdOpFwdDeps FwdDeps() {
    return ActBwdOpFwdDeps::kDepOut;
  }
};

template <typename T>
struct CudaSwishFunctor : public BaseActivationFunctor<T> {
  using MPType = typename details::MPTypeTrait<T>::Type;
  MPType one = static_cast<MPType>(1.0f);
  float beta;

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

  // swish(x) = x / (1 + exp(-beta * x))
  __device__ __forceinline__ T operator()(const T arg_x) const {
    MPType x = static_cast<MPType>(arg_x);
    MPType b = static_cast<MPType>(beta);
    return static_cast<T>(x / (one + exp(-b * x)));
  }
};

template <typename T>
struct CudaSwishGradFunctor : public BaseActivationFunctor<T> {
  using MPType = typename details::MPTypeTrait<T>::Type;
  MPType one = static_cast<MPType>(1.0f);
  float beta;

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

  // dx = dout * (1 + exp(-b * x) + b * x * exp(-b * x) / (1 + exp(-b * x))^2)
  __device__ __forceinline__ T operator()(const T arg_dout,
                                          const T arg_x) const {
    MPType dout = static_cast<MPType>(arg_dout);
    MPType x = static_cast<MPType>(arg_x);
    MPType b = static_cast<MPType>(beta);
    MPType temp1 = one / (one + exp(-b * x));
    MPType out = x * temp1;
    MPType temp2 = b * out;
    MPType temp3 = temp1 * (one - temp2);
    return static_cast<T>(dout * (temp2 + temp3));
  }

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

template <typename T>
struct CudaMishGradFunctor : public BaseActivationFunctor<T> {
  using MPType = typename details::MPTypeTrait<T>::Type;
  MPType one = static_cast<MPType>(1.0f);
  float threshold;

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

  // dx = dout * (tanh(sp) + x * (1 - tanh(sp) ** 2) * (1 - exp(-sp)))
  // sp = softplus(x)
  // Inputs: args[0], the input dout
  //         args[1], the input x
  __device__ __forceinline__ T operator()(const T arg_dout,
                                          const T arg_x) const {
    MPType dout = static_cast<MPType>(arg_dout);
    MPType x = static_cast<MPType>(arg_x);
    MPType sp = (x > static_cast<MPType>(threshold)) ? x : log(one + exp(x));
    MPType gsp =
        (x > static_cast<MPType>(threshold)) ? one : one / (one + exp(-x));
    MPType tsp = tanh(sp);
    return static_cast<T>(dout * (tsp + x * (one - tsp * tsp) * gsp));
  }

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

template <typename T>
struct CudaHardSwishFunctor : public BaseActivationFunctor<T> {
  T zero = static_cast<T>(0.0f);
  float threshold;
  float scale;
  float offset;

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

  // hard_swish(x) = 0, when x <= -offset
  //                 x , when x >= threshold - offset
  //                 x * (x + offset) / scale, otherwise
  // threshold = scale = 6, offset = 3 by default
  __device__ __forceinline__ T operator()(const T x) const {
    T t = static_cast<T>(threshold);
    T temp = x + static_cast<T>(offset);
    T temp_max = temp > zero ? temp : zero;
    T temp_min = temp_max < t ? temp_max : t;
    return temp_min * x / static_cast<T>(scale);
  }
};

template <typename T>
struct CudaHardSwishGradFunctor : public BaseActivationFunctor<T> {
  T zero = static_cast<T>(0.0f);
  T one = static_cast<T>(1.0f);
  T two = static_cast<T>(2.0f);
  float threshold;
  float scale;
  float offset;

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

  // dx = 0, when x <= -offset
  //      dout , when x >= threshold - offset
  //      dout * (2 * x / scale + offset / scale), otherwise
  // threshold = scale = 6, offset = 3 by default
  __device__ __forceinline__ T operator()(const T dout, const T x) const {
    T o = static_cast<T>(offset);
    T s = static_cast<T>(scale);
    T temp1 = static_cast<T>(x + o > zero);
    T temp2 = static_cast<T>(x + o < static_cast<T>(threshold));
    return dout * (temp1 * temp2 * (two * x + o) / s + one - temp2);
  }

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

template <typename T>
struct CudaCELUFunctor : public BaseActivationFunctor<T> {
  using CT = typename details::MPTypeTrait<T>::Type;
  CT zero = static_cast<CT>(0.0f);
  CT one = static_cast<CT>(1.0f);
  float alpha;

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

  // celu(x) = max(0, x) + min(0, alpha * (exp(x/alpha) - 1))
  __device__ __forceinline__ T operator()(const T arg_x) const {
    CT x = static_cast<CT>(arg_x);
    CT temp = static_cast<CT>(alpha) * (exp(x / static_cast<CT>(alpha)) - one);
    CT res = (x > zero ? x : zero) + (temp > zero ? zero : temp);
    return static_cast<T>(res);
  }
};

template <typename T>
struct CudaCELUGradFunctor : public BaseActivationFunctor<T> {
  using MPType = typename details::MPTypeTrait<T>::Type;
  MPType zero = static_cast<MPType>(0.0f);
  MPType one = static_cast<MPType>(1.0f);
  float alpha;

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

  // dx = dout, if alpha > 0 and x > 0
  // dx = dout * (x/alpha).exp(), if alpha > 0 and x <= 0
  // dx = dout , if alpha < 0 and x > 0
  // dx = dout * (x/alpha).exp(), if alpha < 0 and x <=0
  __device__ __forceinline__ T operator()(const T arg_dout,
                                          const T arg_x) const {
    MPType dout = static_cast<MPType>(arg_dout);
    MPType x = static_cast<MPType>(arg_x);
    MPType a = static_cast<MPType>(alpha);
    MPType temp_a_pos = static_cast<MPType>(alpha > 0.0f);
    MPType temp_a_neg = static_cast<MPType>(alpha <= 0.0f);
    MPType temp_x_pos = static_cast<MPType>(x > zero);
    MPType temp_x_neg = static_cast<MPType>(x <= zero);
    return static_cast<T>(
        dout *
        (temp_a_pos * temp_x_pos + temp_a_pos * temp_x_neg * exp(x / a) +
         temp_a_neg * temp_x_pos + exp(x / a) * temp_a_neg * temp_x_neg));
  }

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

template <typename DeviceContext, typename Functor>
class ActivationCudaKernel
    : 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 = nullptr;
    framework::Tensor* out = nullptr;
    ExtractActivationTensor(ctx, &x, &out);
    out->mutable_data<T>(ctx.GetPlace());
    auto& dev_ctx = ctx.template device_context<DeviceContext>();
    std::vector<const framework::Tensor*> ins = {x};
    std::vector<framework::Tensor*> outs = {out};
    auto functor = Functor();
    auto attrs = functor.GetAttrs();
    for (auto& attr : attrs) {
      *attr.second = ctx.Attr<float>(attr.first);
    }
    paddle::operators::LaunchSameDimsElementwiseCudaKernel<T>(dev_ctx, ins,
                                                              &outs, functor);
  }
};

template <typename DeviceContext, typename Functor>
class ActivationGradCudaKernel
    : 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, *d_out;
    framework::Tensor* d_x = nullptr;
    x = out = d_out = nullptr;
    ExtractActivationGradTensor<Functor::FwdDeps()>(ctx, &x, &out, &d_out,
                                                    &d_x);
    d_x->mutable_data<T>(ctx.GetPlace());
    auto& dev_ctx = ctx.template device_context<DeviceContext>();
    auto functor = Functor();
    auto attrs = functor.GetAttrs();
    for (auto& attr : attrs) {
      *attr.second = ctx.Attr<float>(attr.first);
    }

    std::vector<const framework::Tensor*> ins = {d_out};
    std::vector<framework::Tensor*> outs = {d_x};

    if (static_cast<int>(Functor::FwdDeps()) ==
        static_cast<int>(ActBwdOpFwdDeps::kDepOut)) {
      // Only need forward output Out
      ins.push_back(out);
      paddle::operators::LaunchSameDimsElementwiseCudaKernel<T>(dev_ctx, ins,
                                                                &outs, functor);
    } else if (static_cast<int>(Functor::FwdDeps()) ==
               static_cast<int>(ActBwdOpFwdDeps::kDepX)) {
      // Only need forward input X
      ins.push_back(x);
      paddle::operators::LaunchSameDimsElementwiseCudaKernel<T>(dev_ctx, ins,
                                                                &outs, functor);
    } else {
      paddle::operators::LaunchSameDimsElementwiseCudaKernel<T>(dev_ctx, ins,
                                                                &outs, functor);
    }
  }
};

USE_PHI_FUNCTOR(CudaCos)
USE_PHI_FUNCTOR(CudaTan)
USE_PHI_FUNCTOR(CudaAcos)
USE_PHI_FUNCTOR(CudaSin)
USE_PHI_FUNCTOR(CudaAsin)
USE_PHI_FUNCTOR(CudaAtan)
USE_PHI_FUNCTOR(CudaSinh)
USE_PHI_FUNCTOR(CudaCosh)
USE_PHI_FUNCTOR(CudaAsinh)
USE_PHI_FUNCTOR(CudaAcosh)
USE_PHI_FUNCTOR(CudaAtanh)
USE_PHI_FUNCTOR(CudaTanh)
USE_PHI_FUNCTOR(CudaBRelu)
USE_PHI_FUNCTOR(CudaLeakyRelu)
USE_PHI_FUNCTOR(CudaThresholdedRelu)
USE_PHI_FUNCTOR(CudaHardShrink)
USE_PHI_FUNCTOR(CudaSoftShrink)
USE_PHI_FUNCTOR(CudaTanhShrink)
USE_PHI_FUNCTOR(CudaSilu)
USE_PHI_FUNCTOR(CudaELU)

template <typename T>
using CudaELUGradNegativeAlphaFunctor =
    phi::funcs::CudaELUGradNegativeAlphaFunctor<T>;

}  // namespace operators
}  // namespace paddle

namespace ops = paddle::operators;
namespace plat = paddle::platform;

#define REGISTER_ACTIVATION_CUDA_KERNEL(act_type, op_name, functor,            \
                                        grad_functor)                          \
  REGISTER_OP_CUDA_KERNEL(                                                     \
      act_type, ops::ActivationCudaKernel<paddle::platform::CUDADeviceContext, \
                                          ops::functor<float>>,                \
      ops::ActivationCudaKernel<paddle::platform::CUDADeviceContext,           \
                                ops::functor<double>>,                         \
      ops::ActivationCudaKernel<plat::CUDADeviceContext,                       \
                                ops::functor<plat::float16>>,                  \
      ops::ActivationCudaKernel<plat::CUDADeviceContext,                       \
                                ops::functor<plat::bfloat16>>);                \
  REGISTER_OP_CUDA_KERNEL(                                                     \
      act_type##_grad,                                                         \
      ops::ActivationGradCudaKernel<plat::CUDADeviceContext,                   \
                                    ops::grad_functor<float>>,                 \
      ops::ActivationGradCudaKernel<plat::CUDADeviceContext,                   \
                                    ops::grad_functor<double>>,                \
      ops::ActivationGradCudaKernel<plat::CUDADeviceContext,                   \
                                    ops::grad_functor<plat::float16>>,         \
      ops::ActivationGradCudaKernel<plat::CUDADeviceContext,                   \
                                    ops::grad_functor<plat::bfloat16>>);

#define REGISTER_ACTIVATION_CUDA_KERNEL_INT(act_type, op_name, functor,        \
                                            grad_functor)                      \
  REGISTER_OP_CUDA_KERNEL(                                                     \
      act_type, ops::ActivationCudaKernel<paddle::platform::CUDADeviceContext, \
                                          ops::functor<float>>,                \
      ops::ActivationCudaKernel<paddle::platform::CUDADeviceContext,           \
                                ops::functor<double>>,                         \
      ops::ActivationCudaKernel<paddle::platform::CUDADeviceContext,           \
                                ops::functor<int>>,                            \
      ops::ActivationCudaKernel<paddle::platform::CUDADeviceContext,           \
                                ops::functor<int64_t>>,                        \
      ops::ActivationCudaKernel<plat::CUDADeviceContext,                       \
                                ops::functor<plat::float16>>,                  \
      ops::ActivationCudaKernel<plat::CUDADeviceContext,                       \
                                ops::functor<plat::bfloat16>>);                \
  REGISTER_OP_CUDA_KERNEL(                                                     \
      act_type##_grad,                                                         \
      ops::ActivationGradCudaKernel<plat::CUDADeviceContext,                   \
                                    ops::grad_functor<float>>,                 \
      ops::ActivationGradCudaKernel<plat::CUDADeviceContext,                   \
                                    ops::grad_functor<double>>,                \
      ops::ActivationGradCudaKernel<plat::CUDADeviceContext,                   \
                                    ops::grad_functor<int>>,                   \
      ops::ActivationGradCudaKernel<plat::CUDADeviceContext,                   \
                                    ops::grad_functor<int64_t>>,               \
      ops::ActivationGradCudaKernel<plat::CUDADeviceContext,                   \
                                    ops::grad_functor<plat::float16>>,         \
      ops::ActivationGradCudaKernel<plat::CUDADeviceContext,                   \
                                    ops::grad_functor<plat::bfloat16>>);

/* ========================================================================== */

/* ======================== celu register  ============================ */
REGISTER_ACTIVATION_CUDA_KERNEL(celu, CELU, CudaCELUFunctor,
                                CudaCELUGradFunctor);

REGISTER_OP_CUDA_KERNEL(
    celu_grad_grad, ops::CELUDoubleGradKernel<plat::CUDADeviceContext,
                                              ops::CELUGradGradFunctor<float>>,
    ops::CELUDoubleGradKernel<plat::CUDADeviceContext,
                              ops::CELUGradGradFunctor<double>>,
    ops::CELUDoubleGradKernel<plat::CUDADeviceContext,
                              ops::CELUGradGradFunctor<plat::float16>>);
/* ========================================================================== */

/* ===========================    sigmoid register  ============================
 */
REGISTER_ACTIVATION_CUDA_KERNEL(sigmoid, Sigmoid, CudaSigmoidFunctor,
                                CudaSigmoidGradFunctor);

REGISTER_OP_CUDA_KERNEL(
    sigmoid_grad_grad,
    ops::SigmoidDoubleGradKernel<paddle::platform::CUDADeviceContext,
                                 ops::SigmoidGradGradFunctor<float>>,
    ops::SigmoidDoubleGradKernel<paddle::platform::CUDADeviceContext,
                                 ops::SigmoidGradGradFunctor<double>>,
    ops::SigmoidDoubleGradKernel<plat::CUDADeviceContext,
                                 ops::SigmoidGradGradFunctor<plat::float16>>,
    ops::SigmoidDoubleGradKernel<plat::CUDADeviceContext,
                                 ops::SigmoidGradGradFunctor<plat::bfloat16>>);

REGISTER_OP_CUDA_KERNEL(
    sigmoid_triple_grad,
    ops::SigmoidTripleGradKernel<paddle::platform::CUDADeviceContext,
                                 ops::SigmoidTripleGradFunctor<float>>,
    ops::SigmoidTripleGradKernel<paddle::platform::CUDADeviceContext,
                                 ops::SigmoidTripleGradFunctor<double>>,
    ops::SigmoidTripleGradKernel<plat::CUDADeviceContext,
                                 ops::SigmoidTripleGradFunctor<plat::float16>>,
    ops::SigmoidTripleGradKernel<
        plat::CUDADeviceContext,
        ops::SigmoidTripleGradFunctor<plat::bfloat16>>);
/* ========================================================================== */

/* ===========================   sqrt register  ============================= */
REGISTER_ACTIVATION_CUDA_KERNEL(sqrt, Sqrt, CudaSqrtFunctor,
                                CudaSqrtGradFunctor);

REGISTER_OP_CUDA_KERNEL(
    sqrt_grad_grad,
    ops::SqrtDoubleGradKernel<paddle::platform::CUDADeviceContext,
                              ops::SqrtGradGradFunctor<float>>,
    ops::SqrtDoubleGradKernel<paddle::platform::CUDADeviceContext,
                              ops::SqrtGradGradFunctor<double>>,
    ops::SqrtDoubleGradKernel<paddle::platform::CUDADeviceContext,
                              ops::SqrtGradGradFunctor<plat::float16>>,
    ops::SqrtDoubleGradKernel<paddle::platform::CUDADeviceContext,
                              ops::SqrtGradGradFunctor<plat::bfloat16>>);
/* ========================================================================== */

/* ===========================   rsqrt register  =============================
 */
REGISTER_ACTIVATION_CUDA_KERNEL(rsqrt, Rsqrt, CudaRsqrtFunctor,
                                CudaRsqrtGradFunctor);

REGISTER_OP_CUDA_KERNEL(
    rsqrt_grad_grad,
    ops::RsqrtDoubleGradKernel<paddle::platform::CUDADeviceContext,
                               ops::RsqrtGradGradFunctor<float>>,
    ops::RsqrtDoubleGradKernel<paddle::platform::CUDADeviceContext,
                               ops::RsqrtGradGradFunctor<double>>,
    ops::RsqrtDoubleGradKernel<paddle::platform::CUDADeviceContext,
                               ops::RsqrtGradGradFunctor<plat::float16>>);
/* ========================================================================== */

/* ===========================  square register  ============================ */
REGISTER_ACTIVATION_CUDA_KERNEL_INT(square, Square, CudaSquareFunctor,
                                    CudaSquareGradFunctor);

REGISTER_OP_CUDA_KERNEL(
    square_grad_grad,
    ops::SquareDoubleGradKernel<paddle::platform::CUDADeviceContext,
                                ops::SquareGradGradFunctor<float>>,
    ops::SquareDoubleGradKernel<paddle::platform::CUDADeviceContext,
                                ops::SquareGradGradFunctor<double>>,
    ops::SquareDoubleGradKernel<plat::CUDADeviceContext,
                                ops::SquareGradGradFunctor<plat::float16>>,
    ops::SquareDoubleGradKernel<plat::CUDADeviceContext,
                                ops::SquareGradGradFunctor<plat::bfloat16>>,
    ops::SquareDoubleGradKernel<paddle::platform::CUDADeviceContext,
                                ops::SquareGradGradFunctor<int>>,
    ops::SquareDoubleGradKernel<paddle::platform::CUDADeviceContext,
                                ops::SquareGradGradFunctor<int64_t>>);
/* ========================================================================== */

/* ==========================   pow register  ============================ */
REGISTER_OP_CUDA_KERNEL(
    pow, ops::PowKernel<plat::CUDADeviceContext, ops::PowFunctor<float>>,
    ops::PowKernel<plat::CUDADeviceContext, ops::PowFunctor<double>>,
    ops::PowKernel<plat::CUDADeviceContext, ops::PowFunctor<int>>,
    ops::PowKernel<plat::CUDADeviceContext, ops::PowFunctor<int64_t>>,
    ops::PowKernel<plat::CUDADeviceContext, ops::PowFunctor<plat::float16>>);
REGISTER_OP_CUDA_KERNEL(
    pow_grad,
    ops::PowGradKernel<plat::CUDADeviceContext, ops::PowGradFunctor<float>>,
    ops::PowGradKernel<plat::CUDADeviceContext, ops::PowGradFunctor<double>>,
    ops::PowGradKernel<plat::CUDADeviceContext, ops::PowGradFunctor<int>>,
    ops::PowGradKernel<plat::CUDADeviceContext, ops::PowGradFunctor<int64_t>>,
    ops::PowGradKernel<plat::CUDADeviceContext,
                       ops::PowGradFunctor<plat::float16>>);
/* ========================================================================== */

/* ==========================   logit register  ============================ */
namespace ops = paddle::operators;
REGISTER_OP_CUDA_KERNEL(
    logit_grad,
    ops::LogitGradKernel<paddle::platform::CUDADeviceContext, float>,
    ops::LogitGradKernel<paddle::platform::CUDADeviceContext, double>,
    ops::LogitGradKernel<paddle::platform::CUDADeviceContext,
                         paddle::platform::float16>);
/* ========================================================================== */

/* ==========================   exp register  ============================ */
REGISTER_OP_CUDA_KERNEL(
    exp, ops::ActivationCudaKernel<plat::CUDADeviceContext,
                                   ops::CudaExpFunctor<float>>,
    ops::ActivationCudaKernel<plat::CUDADeviceContext,
                              ops::CudaExpFunctor<double>>,
    ops::ActivationKernel<plat::CUDADeviceContext, ops::ExpFunctor<int>>,
    ops::ActivationKernel<plat::CUDADeviceContext, ops::ExpFunctor<int64_t>>,
    ops::ActivationCudaKernel<plat::CUDADeviceContext,
                              ops::CudaExpFunctor<plat::float16>>);
REGISTER_OP_CUDA_KERNEL(
    exp_grad, ops::ActivationGradCudaKernel<plat::CUDADeviceContext,
                                            ops::CudaExpGradFunctor<float>>,
    ops::ActivationGradCudaKernel<plat::CUDADeviceContext,
                                  ops::CudaExpGradFunctor<double>>,
    ops::ActivationGradCudaKernel<plat::CUDADeviceContext,
                                  ops::CudaExpGradFunctor<int>>,
    ops::ActivationGradCudaKernel<plat::CUDADeviceContext,
                                  ops::CudaExpGradFunctor<int64_t>>,
    ops::ActivationGradCudaKernel<plat::CUDADeviceContext,
                                  ops::CudaExpGradFunctor<plat::float16>>);
/* ========================================================================== */

/* ==========================   expm1 register  ============================ */

REGISTER_OP_CUDA_KERNEL(
    expm1, ops::ActivationCudaKernel<plat::CUDADeviceContext,
                                     ops::CudaExpm1Functor<float>>,
    ops::ActivationCudaKernel<plat::CUDADeviceContext,
                              ops::CudaExpm1Functor<double>>,
    ops::ActivationCudaKernel<plat::CUDADeviceContext,
                              ops::CudaExpm1Functor<plat::float16>>);
REGISTER_OP_CUDA_KERNEL(
    expm1_grad, ops::ActivationGradCudaKernel<plat::CUDADeviceContext,
                                              ops::CudaExpm1GradFunctor<float>>,
    ops::ActivationGradCudaKernel<plat::CUDADeviceContext,
                                  ops::CudaExpm1GradFunctor<double>>,
    ops::ActivationGradCudaKernel<plat::CUDADeviceContext,
                                  ops::CudaExpm1GradFunctor<plat::float16>>);
/* ========================================================================== */

/* ==========================  Log register ==================================*/
REGISTER_ACTIVATION_CUDA_KERNEL(log, Log, CudaLogFunctor, CudaLogGradFunctor);

REGISTER_OP_CUDA_KERNEL(
    log_grad_grad, ops::LogDoubleGradKernel<plat::CUDADeviceContext,
                                            ops::LogGradGradFunctor<float>>,
    ops::LogDoubleGradKernel<plat::CUDADeviceContext,
                             ops::LogGradGradFunctor<double>>,
    ops::LogDoubleGradKernel<plat::CUDADeviceContext,
                             ops::LogGradGradFunctor<plat::float16>>);
/* ========================================================================== */

#define FOR_EACH_ACTIVATION_CUDA_OP(__macro)                                  \
  __macro(logsigmoid, LogSigmoid, CudaLogSigmoidFunctor,                      \
          CudaLogSigmoidGradFunctor);                                         \
  __macro(softshrink, SoftShrink, CudaSoftShrinkFunctor,                      \
          CudaSoftShrinkGradFunctor);                                         \
  __macro(ceil, Ceil, CudaCeilFunctor, CudaZeroGradFunctor);                  \
  __macro(floor, Floor, CudaFloorFunctor, CudaZeroGradFunctor);               \
  __macro(round, Round, CudaRoundFunctor, CudaZeroGradFunctor);               \
  __macro(reciprocal, Reciprocal, CudaReciprocalFunctor,                      \
          CudaReciprocalGradFunctor);                                         \
  __macro(log1p, Log1p, CudaLog1pFunctor, CudaLog1pGradFunctor);              \
  __macro(log2, Log2, CudaLog2Functor, CudaLog2GradFunctor);                  \
  __macro(log10, Log10, CudaLog10Functor, CudaLog10GradFunctor);              \
  __macro(soft_relu, SoftRelu, CudaSoftReluFunctor, CudaSoftReluGradFunctor); \
  __macro(stanh, STanh, CudaSTanhFunctor, CudaSTanhGradFunctor);              \
  __macro(softplus, Softplus, CudaSoftplusFunctor, CudaSoftplusGradFunctor);  \
  __macro(softsign, Softsign, CudaSoftsignFunctor, CudaSoftsignGradFunctor);  \
  __macro(relu6, Relu6, CudaRelu6Functor, CudaRelu6GradFunctor);              \
  __macro(tanh_shrink, TanhShrink, CudaTanhShrinkFunctor,                     \
          CudaTanhShrinkGradFunctor);                                         \
  __macro(hard_shrink, HardShrink, CudaHardShrinkFunctor,                     \
          CudaHardShrinkGradFunctor);                                         \
  __macro(hard_sigmoid, HardSigmoid, CudaHardSigmoidFunctor,                  \
          CudaHardSigmoidGradFunctor);                                        \
  __macro(swish, Swish, CudaSwishFunctor, CudaSwishGradFunctor);              \
  __macro(mish, Mish, CudaMishFunctor, CudaMishGradFunctor);                  \
  __macro(hard_swish, HardSwish, CudaHardSwishFunctor,                        \
          CudaHardSwishGradFunctor);
FOR_EACH_ACTIVATION_CUDA_OP(REGISTER_ACTIVATION_CUDA_KERNEL)

#ifdef PADDLE_WITH_XPU_KP
REGISTER_OP_KERNEL(
    brelu, KP, plat::XPUPlace,
    ops::ActivationCudaKernel<paddle::platform::XPUDeviceContext,
                              phi::funcs::CudaBReluFunctor<float>>);
REGISTER_OP_KERNEL(
    brelu_grad, KP, plat::XPUPlace,
    ops::ActivationGradCudaKernel<paddle::platform::XPUDeviceContext,
                                  phi::funcs::CudaBReluGradFunctor<float>>);

REGISTER_OP_KERNEL(ceil, KP, plat::XPUPlace,
                   ops::ActivationCudaKernel<paddle::platform::XPUDeviceContext,
                                             ops::CudaCeilFunctor<float>>);
REGISTER_OP_KERNEL(
    ceil_grad, KP, plat::XPUPlace,
    ops::ActivationGradCudaKernel<paddle::platform::XPUDeviceContext,
                                  ops::CudaZeroGradFunctor<float>>);

REGISTER_OP_KERNEL(celu, KP, plat::XPUPlace,
                   ops::ActivationCudaKernel<paddle::platform::XPUDeviceContext,
                                             ops::CudaCELUFunctor<float>>);
REGISTER_OP_KERNEL(
    celu_grad, KP, plat::XPUPlace,
    ops::ActivationGradCudaKernel<paddle::platform::XPUDeviceContext,
                                  ops::CudaCELUGradFunctor<float>>);

REGISTER_OP_KERNEL(elu, KP, plat::XPUPlace,
                   ops::ActivationCudaKernel<paddle::platform::XPUDeviceContext,
                                             ops::CudaELUFunctor<float>>);
REGISTER_OP_KERNEL(
    elu_grad, KP, plat::XPUPlace,
    ops::ActivationGradCudaKernel<paddle::platform::XPUDeviceContext,
                                  ops::CudaELUGradFunctor<float>>);

REGISTER_OP_KERNEL(exp, KP, plat::XPUPlace,
                   ops::ActivationCudaKernel<paddle::platform::XPUDeviceContext,
                                             ops::CudaExpFunctor<float>>);
REGISTER_OP_KERNEL(
    exp_grad, KP, plat::XPUPlace,
    ops::ActivationGradCudaKernel<paddle::platform::XPUDeviceContext,
                                  ops::CudaExpGradFunctor<float>>);

REGISTER_OP_KERNEL(floor, KP, plat::XPUPlace,
                   ops::ActivationCudaKernel<paddle::platform::XPUDeviceContext,
                                             ops::CudaFloorFunctor<float>>);
REGISTER_OP_KERNEL(
    floor_grad, KP, plat::XPUPlace,
    ops::ActivationGradCudaKernel<paddle::platform::XPUDeviceContext,
                                  ops::CudaZeroGradFunctor<float>>);

REGISTER_OP_KERNEL(
    hard_shrink, KP, plat::XPUPlace,
    ops::ActivationCudaKernel<paddle::platform::XPUDeviceContext,
                              ops::CudaHardShrinkFunctor<float>>);
REGISTER_OP_KERNEL(
    hard_shrink_grad, KP, plat::XPUPlace,
    ops::ActivationGradCudaKernel<paddle::platform::XPUDeviceContext,
                                  ops::CudaHardShrinkGradFunctor<float>>);

REGISTER_OP_KERNEL(
    hard_sigmoid, KP, plat::XPUPlace,
    ops::ActivationCudaKernel<paddle::platform::XPUDeviceContext,
                              ops::CudaHardSigmoidFunctor<float>>);
REGISTER_OP_KERNEL(
    hard_sigmoid_grad, KP, plat::XPUPlace,
    ops::ActivationGradCudaKernel<paddle::platform::XPUDeviceContext,
                                  ops::CudaHardSigmoidGradFunctor<float>>);

REGISTER_OP_KERNEL(hard_swish, KP, plat::XPUPlace,
                   ops::ActivationCudaKernel<paddle::platform::XPUDeviceContext,
                                             ops::CudaHardSwishFunctor<float>>);
REGISTER_OP_KERNEL(
    hard_swish_grad, KP, plat::XPUPlace,
    ops::ActivationGradCudaKernel<paddle::platform::XPUDeviceContext,
                                  ops::CudaHardSwishGradFunctor<float>>);

REGISTER_OP_KERNEL(
    leaky_relu, KP, plat::XPUPlace,
    ops::ActivationCudaKernel<paddle::platform::XPUDeviceContext,
                              phi::funcs::CudaLeakyReluFunctor<float>>);
REGISTER_OP_KERNEL(
    leaky_relu_grad, KP, plat::XPUPlace,
    ops::ActivationGradCudaKernel<paddle::platform::XPUDeviceContext,
                                  phi::funcs::CudaLeakyReluGradFunctor<float>>);

REGISTER_OP_KERNEL(log, KP, plat::XPUPlace,
                   ops::ActivationCudaKernel<paddle::platform::XPUDeviceContext,
                                             ops::CudaLogFunctor<float>>);
REGISTER_OP_KERNEL(
    log_grad, KP, plat::XPUPlace,
    ops::ActivationGradCudaKernel<paddle::platform::XPUDeviceContext,
                                  ops::CudaLogGradFunctor<float>>);

REGISTER_OP_KERNEL(log1p, KP, plat::XPUPlace,
                   ops::ActivationCudaKernel<paddle::platform::XPUDeviceContext,
                                             ops::CudaLog1pFunctor<float>>);
REGISTER_OP_KERNEL(
    log1p_grad, KP, plat::XPUPlace,
    ops::ActivationGradCudaKernel<paddle::platform::XPUDeviceContext,
                                  ops::CudaLog1pGradFunctor<float>>);

REGISTER_OP_KERNEL(
    logsigmoid, KP, plat::XPUPlace,
    ops::ActivationCudaKernel<paddle::platform::XPUDeviceContext,
                              ops::CudaLogSigmoidFunctor<float>>);
REGISTER_OP_KERNEL(
    logsigmoid_grad, KP, plat::XPUPlace,
    ops::ActivationGradCudaKernel<paddle::platform::XPUDeviceContext,
                                  ops::CudaLogSigmoidGradFunctor<float>>);

REGISTER_OP_KERNEL(
    reciprocal, KP, plat::XPUPlace,
    ops::ActivationCudaKernel<paddle::platform::XPUDeviceContext,
                              ops::CudaReciprocalFunctor<float>>);
REGISTER_OP_KERNEL(
    reciprocal_grad, KP, plat::XPUPlace,
    ops::ActivationGradCudaKernel<paddle::platform::XPUDeviceContext,
                                  ops::CudaReciprocalGradFunctor<float>>);

REGISTER_OP_KERNEL(
    relu, KP, plat::XPUPlace,
    ops::ActivationCudaKernel<paddle::platform::XPUDeviceContext,
                              phi::funcs::CudaReluFunctor<float>>);
REGISTER_OP_KERNEL(
    relu_grad, KP, plat::XPUPlace,
    ops::ActivationGradCudaKernel<paddle::platform::XPUDeviceContext,
                                  phi::funcs::CudaReluGradFunctor<float>>);

REGISTER_OP_KERNEL(relu6, KP, plat::XPUPlace,
                   ops::ActivationCudaKernel<paddle::platform::XPUDeviceContext,
                                             ops::CudaRelu6Functor<float>>);
REGISTER_OP_KERNEL(
    relu6_grad, KP, plat::XPUPlace,
    ops::ActivationGradCudaKernel<paddle::platform::XPUDeviceContext,
                                  ops::CudaRelu6GradFunctor<float>>);

REGISTER_OP_KERNEL(sigmoid, KP, plat::XPUPlace,
                   ops::ActivationCudaKernel<paddle::platform::XPUDeviceContext,
                                             ops::CudaSigmoidFunctor<float>>);
REGISTER_OP_KERNEL(
    sigmoid_grad, KP, plat::XPUPlace,
    ops::ActivationGradCudaKernel<paddle::platform::XPUDeviceContext,
                                  ops::CudaSigmoidGradFunctor<float>>);

REGISTER_OP_KERNEL(silu, KP, plat::XPUPlace,
                   ops::ActivationCudaKernel<paddle::platform::XPUDeviceContext,
                                             ops::CudaSiluFunctor<float>>);
REGISTER_OP_KERNEL(
    silu_grad, KP, plat::XPUPlace,
    ops::ActivationGradCudaKernel<paddle::platform::XPUDeviceContext,
                                  ops::CudaSiluGradFunctor<float>>);

REGISTER_OP_KERNEL(soft_relu, KP, plat::XPUPlace,
                   ops::ActivationCudaKernel<paddle::platform::XPUDeviceContext,
                                             ops::CudaSoftReluFunctor<float>>);
REGISTER_OP_KERNEL(
    soft_relu_grad, KP, plat::XPUPlace,
    ops::ActivationGradCudaKernel<paddle::platform::XPUDeviceContext,
                                  ops::CudaSoftReluGradFunctor<float>>);

REGISTER_OP_KERNEL(softplus, KP, plat::XPUPlace,
                   ops::ActivationCudaKernel<paddle::platform::XPUDeviceContext,
                                             ops::CudaSoftplusFunctor<float>>);
REGISTER_OP_KERNEL(
    softplus_grad, KP, plat::XPUPlace,
    ops::ActivationGradCudaKernel<paddle::platform::XPUDeviceContext,
                                  ops::CudaSoftplusGradFunctor<float>>);

REGISTER_OP_KERNEL(
    softshrink, KP, plat::XPUPlace,
    ops::ActivationCudaKernel<paddle::platform::XPUDeviceContext,
                              ops::CudaSoftShrinkFunctor<float>>);
REGISTER_OP_KERNEL(
    softshrink_grad, KP, plat::XPUPlace,
    ops::ActivationGradCudaKernel<paddle::platform::XPUDeviceContext,
                                  ops::CudaSoftShrinkGradFunctor<float>>);

REGISTER_OP_KERNEL(softsign, KP, plat::XPUPlace,
                   ops::ActivationCudaKernel<paddle::platform::XPUDeviceContext,
                                             ops::CudaSoftsignFunctor<float>>);
REGISTER_OP_KERNEL(
    softsign_grad, KP, plat::XPUPlace,
    ops::ActivationGradCudaKernel<paddle::platform::XPUDeviceContext,
                                  ops::CudaSoftsignGradFunctor<float>>);

REGISTER_OP_KERNEL(sqrt, KP, plat::XPUPlace,
                   ops::ActivationCudaKernel<paddle::platform::XPUDeviceContext,
                                             ops::CudaSqrtFunctor<float>>);
REGISTER_OP_KERNEL(
    sqrt_grad, KP, plat::XPUPlace,
    ops::ActivationGradCudaKernel<paddle::platform::XPUDeviceContext,
                                  ops::CudaSqrtGradFunctor<float>>);

REGISTER_OP_KERNEL(square, KP, plat::XPUPlace,
                   ops::ActivationCudaKernel<paddle::platform::XPUDeviceContext,
                                             ops::CudaSquareFunctor<float>>);
REGISTER_OP_KERNEL(
    square_grad, KP, plat::XPUPlace,
    ops::ActivationGradCudaKernel<paddle::platform::XPUDeviceContext,
                                  ops::CudaSquareGradFunctor<float>>);

REGISTER_OP_KERNEL(swish, KP, plat::XPUPlace,
                   ops::ActivationCudaKernel<paddle::platform::XPUDeviceContext,
                                             ops::CudaSwishFunctor<float>>);
REGISTER_OP_KERNEL(
    swish_grad, KP, plat::XPUPlace,
    ops::ActivationGradCudaKernel<paddle::platform::XPUDeviceContext,
                                  ops::CudaSwishGradFunctor<float>>);

REGISTER_OP_KERNEL(
    thresholded_relu, KP, plat::XPUPlace,
    ops::ActivationCudaKernel<paddle::platform::XPUDeviceContext,
                              ops::CudaThresholdedReluFunctor<float>>);
REGISTER_OP_KERNEL(
    thresholded_relu_grad, KP, plat::XPUPlace,
    ops::ActivationGradCudaKernel<paddle::platform::XPUDeviceContext,
                                  ops::CudaThresholdedReluGradFunctor<float>>);

#endif  // PADDLE_WITH_XPU_KP