activation_kernel.cc 12.0 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
/* 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_kernel.h"
16

17 18
#include "paddle/phi/backends/cpu/cpu_context.h"
#include "paddle/phi/core/kernel_registry.h"
19
#include "paddle/phi/kernels/funcs/activation_functor.h"
20 21 22 23
#include "paddle/phi/kernels/impl/activation_impl.h"

namespace phi {

Y
YuanRisheng 已提交
24 25 26 27 28
#define DEFINE_CPU_ACTIVATION_KERNEL(name, functor_class)               \
  template <typename T, typename Context>                               \
  void name##Kernel(                                                    \
      const Context& dev_ctx, const DenseTensor& x, DenseTensor* out) { \
    funcs::functor_class<T> functor;                                    \
29
    ActivationImpl<T, T, Context, funcs::functor_class<T>>(             \
Y
YuanRisheng 已提交
30
        dev_ctx, x, out, functor);                                      \
31 32
  }

33 34 35 36 37 38 39 40 41 42 43 44
#define DEFINE_CPU_ACTIVATION_KERNEL_WITH_INT_IN_FLOAT_OUT(name,           \
                                                           functor_class)  \
  template <typename T, typename Context>                                  \
  void name##Kernel(                                                       \
      const Context& dev_ctx, const DenseTensor& x, DenseTensor* out) {    \
    funcs::functor_class<T> functor;                                       \
    using U =                                                              \
        typename std::conditional_t<std::is_integral<T>::value, float, T>; \
    ActivationImpl<T, U, Context, funcs::functor_class<T>>(                \
        dev_ctx, x, out, functor);                                         \
  }

Y
YuanRisheng 已提交
45 46 47 48 49 50 51 52 53
#define DEFINE_CPU_ACT_KERNEL_WITH_ONE_ATTRS(name, functor_class, attr) \
  template <typename T, typename Context>                               \
  void name##Kernel(const Context& dev_ctx,                             \
                    const DenseTensor& x,                               \
                    float attr,                                         \
                    DenseTensor* out) {                                 \
    funcs::functor_class<T> functor;                                    \
    auto attrs = functor.GetAttrs();                                    \
    *(attrs[0].second) = attr;                                          \
54
    ActivationImpl<T, T, Context, funcs::functor_class<T>>(             \
Y
YuanRisheng 已提交
55
        dev_ctx, x, out, functor);                                      \
56 57
  }

58 59 60 61 62 63 64 65 66 67 68 69 70 71
#define DEFINE_CPU_ACT_KERNEL_WITH_TWO_ATTRS(               \
    name, functor_class, attr1, attr2)                      \
  template <typename T, typename Context>                   \
  void name##Kernel(const Context& dev_ctx,                 \
                    const DenseTensor& x,                   \
                    float attr1,                            \
                    float attr2,                            \
                    DenseTensor* out) {                     \
    funcs::functor_class<T> functor;                        \
    auto attrs = functor.GetAttrs();                        \
    *(attrs[0].second) = attr1;                             \
    *(attrs[1].second) = attr2;                             \
    ActivationImpl<T, T, Context, funcs::functor_class<T>>( \
        dev_ctx, x, out, functor);                          \
72 73
  }

Y
YuanRisheng 已提交
74 75 76 77 78 79 80 81 82 83 84 85 86 87 88
DEFINE_CPU_ACTIVATION_KERNEL(Sin, SinFunctor)
DEFINE_CPU_ACTIVATION_KERNEL(Cos, CosFunctor)
DEFINE_CPU_ACTIVATION_KERNEL(Tan, TanFunctor)
DEFINE_CPU_ACTIVATION_KERNEL(Asin, AsinFunctor)
DEFINE_CPU_ACTIVATION_KERNEL(Atan, AtanFunctor)
DEFINE_CPU_ACTIVATION_KERNEL(Acos, AcosFunctor)
DEFINE_CPU_ACTIVATION_KERNEL(Sinh, SinhFunctor)
DEFINE_CPU_ACTIVATION_KERNEL(Cosh, CoshFunctor)
DEFINE_CPU_ACTIVATION_KERNEL(Asinh, AsinhFunctor)
DEFINE_CPU_ACTIVATION_KERNEL(Acosh, AcoshFunctor)
DEFINE_CPU_ACTIVATION_KERNEL(Atanh, AtanhFunctor)
DEFINE_CPU_ACTIVATION_KERNEL(Relu, ReluCPUFunctor)
DEFINE_CPU_ACTIVATION_KERNEL(Tanh, TanhFunctor)
DEFINE_CPU_ACTIVATION_KERNEL(TanhShrink, TanhShrinkFunctor)
DEFINE_CPU_ACTIVATION_KERNEL(Silu, SiluFunctor)
89 90 91 92
DEFINE_CPU_ACTIVATION_KERNEL(Reciprocal, ReciprocalFunctor)
DEFINE_CPU_ACTIVATION_KERNEL(Square, SquareFunctor)
DEFINE_CPU_ACTIVATION_KERNEL(Sqrt, SqrtFunctor)
DEFINE_CPU_ACTIVATION_KERNEL(Rsqrt, RsqrtFunctor)
93
DEFINE_CPU_ACTIVATION_KERNEL(Softsign, SoftsignFunctor)
Y
YuanRisheng 已提交
94 95
DEFINE_CPU_ACTIVATION_KERNEL(Sigmoid, SigmoidFunctor)
DEFINE_CPU_ACTIVATION_KERNEL(LogSigmoid, LogSigmoidFunctor)
Y
YuanRisheng 已提交
96 97 98
DEFINE_CPU_ACTIVATION_KERNEL(Round, RoundFunctor)
DEFINE_CPU_ACTIVATION_KERNEL(Floor, FloorFunctor)
DEFINE_CPU_ACTIVATION_KERNEL(Ceil, CeilFunctor)
99
DEFINE_CPU_ACTIVATION_KERNEL(Negative, NegativeFunctor)
Y
YuanRisheng 已提交
100

101 102 103 104
DEFINE_CPU_ACTIVATION_KERNEL_WITH_INT_IN_FLOAT_OUT(Log, LogFunctor)
DEFINE_CPU_ACTIVATION_KERNEL_WITH_INT_IN_FLOAT_OUT(Log2, Log2Functor)
DEFINE_CPU_ACTIVATION_KERNEL_WITH_INT_IN_FLOAT_OUT(Log10, Log10Functor)
DEFINE_CPU_ACTIVATION_KERNEL_WITH_INT_IN_FLOAT_OUT(Log1p, Log1pFunctor)
105 106
DEFINE_CPU_ACTIVATION_KERNEL_WITH_INT_IN_FLOAT_OUT(Exp, ExpFunctor)
DEFINE_CPU_ACTIVATION_KERNEL_WITH_INT_IN_FLOAT_OUT(Expm1, Expm1Functor)
107

Y
YuanRisheng 已提交
108
DEFINE_CPU_ACT_KERNEL_WITH_ONE_ATTRS(LeakyRelu, LeakyReluFunctor, alpha)
109
DEFINE_CPU_ACT_KERNEL_WITH_ONE_ATTRS(ThresholdedRelu,
Y
YuanRisheng 已提交
110
                                     ThresholdedReluFunctor,
111
                                     threshold)
112
DEFINE_CPU_ACT_KERNEL_WITH_ONE_ATTRS(Mish, MishFunctor, threshold)
Y
YuanRisheng 已提交
113 114 115
DEFINE_CPU_ACT_KERNEL_WITH_ONE_ATTRS(HardShrink, HardShrinkFunctor, threshold)
DEFINE_CPU_ACT_KERNEL_WITH_ONE_ATTRS(SoftShrink, SoftShrinkFunctor, lambda)
DEFINE_CPU_ACT_KERNEL_WITH_ONE_ATTRS(Elu, ELUFunctor, alpha)
Y
YuanRisheng 已提交
116
DEFINE_CPU_ACT_KERNEL_WITH_ONE_ATTRS(Celu, CELUFunctor, alpha)
Y
YuanRisheng 已提交
117

118
DEFINE_CPU_ACT_KERNEL_WITH_TWO_ATTRS(HardTanh, HardTanhFunctor, t_min, t_max)
Y
YuanRisheng 已提交
119 120
DEFINE_CPU_ACT_KERNEL_WITH_TWO_ATTRS(STanh, STanhFunctor, scale_a, scale_b)
DEFINE_CPU_ACT_KERNEL_WITH_TWO_ATTRS(Softplus, SoftplusFunctor, beta, threshold)
Y
YuanRisheng 已提交
121 122 123 124
DEFINE_CPU_ACT_KERNEL_WITH_TWO_ATTRS(HardSigmoid,
                                     HardSigmoidFunctor,
                                     slope,
                                     offset)
125

Y
YuanRisheng 已提交
126
template <typename T, typename Context>
127 128 129
void HardSwishKernel(const Context& dev_ctx,
                     const DenseTensor& x,
                     DenseTensor* out) {
Y
YuanRisheng 已提交
130
  funcs::HardSwishFunctor<T> functor;
131 132 133
  float threshold = 6;
  float scale = 6;
  float offset = 3;
Y
YuanRisheng 已提交
134 135 136 137
  auto attrs = functor.GetAttrs();
  *(attrs[0].second) = threshold;
  *(attrs[1].second) = scale;
  *(attrs[2].second) = offset;
138
  ActivationImpl<T, T, Context, funcs::HardSwishFunctor<T>>(
Y
YuanRisheng 已提交
139 140 141
      dev_ctx, x, out, functor);
}

Z
zhangyuqin1998 已提交
142 143 144 145 146 147 148 149 150 151
template <typename T, typename Context>
void SwishKernel(const Context& dev_ctx,
                 const DenseTensor& x,
                 DenseTensor* out) {
  funcs::SwishFunctor<T> functor;
  auto attrs = functor.GetAttrs();
  *(attrs[0].second) = 1.0;
  ActivationImpl<T, T, Context, funcs::SwishFunctor<T>>(
      dev_ctx, x, out, functor);
}
Z
zhangyuqin1998 已提交
152 153 154 155 156 157 158 159 160 161 162

template <typename T, typename Context>
void Relu6Kernel(const Context& dev_ctx,
                 const DenseTensor& x,
                 DenseTensor* out) {
  funcs::Relu6Functor<T> functor;
  auto attrs = functor.GetAttrs();
  *(attrs[0].second) = 6.0;
  ActivationImpl<T, T, Context, funcs::Relu6Functor<T>>(
      dev_ctx, x, out, functor);
}
163 164
}  // namespace phi
PD_REGISTER_KERNEL(relu, CPU, ALL_LAYOUT, phi::ReluKernel, float, double) {}
165 166

#define PD_REGISTER_ACTIVATION_KERNEL(name, func) \
Y
YuanRisheng 已提交
167
  PD_REGISTER_KERNEL(name, CPU, ALL_LAYOUT, phi::func, float, double) {}
168

Y
YuanRisheng 已提交
169 170 171 172 173 174 175 176 177 178 179 180
PD_REGISTER_ACTIVATION_KERNEL(sin, SinKernel)
PD_REGISTER_ACTIVATION_KERNEL(cos, CosKernel)
PD_REGISTER_ACTIVATION_KERNEL(tan, TanKernel)
PD_REGISTER_ACTIVATION_KERNEL(acos, AcosKernel)
PD_REGISTER_ACTIVATION_KERNEL(asin, AsinKernel)
PD_REGISTER_ACTIVATION_KERNEL(atan, AtanKernel)
PD_REGISTER_ACTIVATION_KERNEL(sinh, SinhKernel)
PD_REGISTER_ACTIVATION_KERNEL(cosh, CoshKernel)
PD_REGISTER_ACTIVATION_KERNEL(asinh, AsinhKernel)
PD_REGISTER_ACTIVATION_KERNEL(acosh, AcoshKernel)
PD_REGISTER_ACTIVATION_KERNEL(atanh, AtanhKernel)
PD_REGISTER_ACTIVATION_KERNEL(tanh, TanhKernel)
Z
zyfncg 已提交
181
PD_REGISTER_ACTIVATION_KERNEL(hardtanh, HardTanhKernel)
Y
YuanRisheng 已提交
182 183 184
PD_REGISTER_ACTIVATION_KERNEL(leaky_relu, LeakyReluKernel)
PD_REGISTER_ACTIVATION_KERNEL(thresholded_relu, ThresholdedReluKernel)
PD_REGISTER_ACTIVATION_KERNEL(hard_shrink, HardShrinkKernel)
185
PD_REGISTER_ACTIVATION_KERNEL(softshrink, SoftShrinkKernel)
Y
YuanRisheng 已提交
186 187 188
PD_REGISTER_ACTIVATION_KERNEL(tanh_shrink, TanhShrinkKernel)
PD_REGISTER_ACTIVATION_KERNEL(elu, EluKernel)
PD_REGISTER_ACTIVATION_KERNEL(silu, SiluKernel)
189 190 191 192 193 194 195
PD_REGISTER_ACTIVATION_KERNEL(mish, MishKernel)
PD_REGISTER_ACTIVATION_KERNEL(stanh, STanhKernel)
PD_REGISTER_ACTIVATION_KERNEL(reciprocal, ReciprocalKernel)
PD_REGISTER_ACTIVATION_KERNEL(sqrt, SqrtKernel)
PD_REGISTER_ACTIVATION_KERNEL(rsqrt, RsqrtKernel)
PD_REGISTER_ACTIVATION_KERNEL(softplus, SoftplusKernel)

196 197 198 199 200 201 202 203 204 205
PD_REGISTER_KERNEL(exp,
                   CPU,
                   ALL_LAYOUT,
                   phi::ExpKernel,
                   float,
                   double,
                   int,
                   int64_t,
                   phi::dtype::float16) {}

206 207 208 209 210 211
PD_REGISTER_KERNEL(expm1,
                   CPU,
                   ALL_LAYOUT,
                   phi::Expm1Kernel,
                   float,
                   double,
212 213
                   int,
                   int64_t,
214
                   phi::dtype::float16) {}
215

216 217 218
PD_REGISTER_KERNEL(logit, CPU, ALL_LAYOUT, phi::LogitKernel, float, double) {}
PD_REGISTER_KERNEL(
    square, CPU, ALL_LAYOUT, phi::SquareKernel, float, double, int, int64_t) {}
219
PD_REGISTER_ACTIVATION_KERNEL(softsign, SoftsignKernel)
Y
YuanRisheng 已提交
220 221 222
PD_REGISTER_ACTIVATION_KERNEL(sigmoid, SigmoidKernel)
PD_REGISTER_ACTIVATION_KERNEL(logsigmoid, LogSigmoidKernel)
PD_REGISTER_ACTIVATION_KERNEL(hard_sigmoid, HardSigmoidKernel)
Z
zhangyuqin1998 已提交
223
PD_REGISTER_ACTIVATION_KERNEL(swish, SwishKernel)
Z
zhangyuqin1998 已提交
224
PD_REGISTER_ACTIVATION_KERNEL(relu6, Relu6Kernel)
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

PD_REGISTER_KERNEL(log,
                   CPU,
                   ALL_LAYOUT,
                   phi::LogKernel,
                   float,
                   double,
                   int,
                   int64_t,
                   phi::dtype::float16,
                   phi::dtype::bfloat16) {}
PD_REGISTER_KERNEL(log2,
                   CPU,
                   ALL_LAYOUT,
                   phi::Log2Kernel,
                   float,
                   double,
                   int,
                   int64_t,
                   phi::dtype::float16,
                   phi::dtype::bfloat16) {}
PD_REGISTER_KERNEL(log10,
                   CPU,
                   ALL_LAYOUT,
                   phi::Log10Kernel,
                   float,
                   double,
                   int,
                   int64_t,
                   phi::dtype::float16,
                   phi::dtype::bfloat16) {}
PD_REGISTER_KERNEL(log1p,
                   CPU,
                   ALL_LAYOUT,
                   phi::Log1pKernel,
                   float,
                   double,
                   int,
                   int64_t,
                   phi::dtype::float16,
                   phi::dtype::bfloat16) {}

267
PD_REGISTER_ACTIVATION_KERNEL(hardswish, HardSwishKernel)
Y
YuanRisheng 已提交
268 269 270
PD_REGISTER_ACTIVATION_KERNEL(round, RoundKernel)
PD_REGISTER_ACTIVATION_KERNEL(floor, FloorKernel)
PD_REGISTER_ACTIVATION_KERNEL(ceil, CeilKernel)
271 272 273 274 275 276 277 278 279
PD_REGISTER_KERNEL(negative,
                   CPU,
                   ALL_LAYOUT,
                   phi::NegativeKernel,
                   float,
                   double,
                   int16_t,
                   int,
                   int64_t) {}
Y
YuanRisheng 已提交
280
PD_REGISTER_ACTIVATION_KERNEL(celu, CeluKernel)
Y
YuanRisheng 已提交
281 282
PD_REGISTER_KERNEL(
    pow, CPU, ALL_LAYOUT, phi::PowKernel, float, double, int, int64_t) {}