activation_kernel.cu 12.0 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14
/* 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. */

15 16
#include "paddle/phi/kernels/activation_kernel.h"

17
#include "paddle/fluid/platform/device/gpu/gpu_device_function.h"
18 19 20 21 22
#include "paddle/phi/backends/gpu/gpu_context.h"
#include "paddle/phi/common/bfloat16.h"
#include "paddle/phi/common/float16.h"
#include "paddle/phi/core/kernel_registry.h"
#include "paddle/phi/kernels/funcs/elementwise_base.h"
23
#include "paddle/phi/kernels/impl/activation_grad_impl.h"
24
#include "paddle/phi/kernels/impl/activation_impl.h"
25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40

namespace phi {

template <typename T, typename Context, typename Functor>
void ActivationGPUImpl(const Context& dev_ctx,
                       const DenseTensor& x,
                       DenseTensor* out,
                       const Functor& functor) {
  PADDLE_ENFORCE_NOT_NULL(out,
                          errors::NotFound("Output Out should not be nullptr"));
  dev_ctx.template Alloc<T>(out);
  std::vector<const DenseTensor*> ins = {&x};
  std::vector<DenseTensor*> outs = {out};
  funcs::ElementwiseKernel<T>(dev_ctx, ins, &outs, functor);
}

Y
YuanRisheng 已提交
41 42 43 44 45 46 47
#define DEFINE_GPU_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;                                    \
    ActivationGPUImpl<T, Context, funcs::functor_class<T>>(             \
        dev_ctx, x, out, functor);                                      \
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
#define DEFINE_GPU_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;                                          \
    ActivationGPUImpl<T, Context, funcs::functor_class<T>>(             \
        dev_ctx, x, out, functor);                                      \
  }

#define DEFINE_GPU_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;                             \
    ActivationGPUImpl<T, Context, funcs::functor_class<T>>( \
        dev_ctx, x, out, functor);                          \
  }

Y
YuanRisheng 已提交
79 80 81 82 83 84 85 86 87 88 89 90 91 92 93
DEFINE_GPU_ACTIVATION_KERNEL(Cos, CudaCosFunctor)
DEFINE_GPU_ACTIVATION_KERNEL(Tan, CudaTanFunctor)
DEFINE_GPU_ACTIVATION_KERNEL(Acos, CudaAcosFunctor)
DEFINE_GPU_ACTIVATION_KERNEL(Sin, CudaSinFunctor)
DEFINE_GPU_ACTIVATION_KERNEL(Asin, CudaAsinFunctor)
DEFINE_GPU_ACTIVATION_KERNEL(Atan, CudaAtanFunctor)
DEFINE_GPU_ACTIVATION_KERNEL(Sinh, CudaSinhFunctor)
DEFINE_GPU_ACTIVATION_KERNEL(Cosh, CudaCoshFunctor)
DEFINE_GPU_ACTIVATION_KERNEL(Asinh, CudaAsinhFunctor)
DEFINE_GPU_ACTIVATION_KERNEL(Acosh, CudaAcoshFunctor)
DEFINE_GPU_ACTIVATION_KERNEL(Atanh, CudaAtanhFunctor)
DEFINE_GPU_ACTIVATION_KERNEL(Relu, CudaReluFunctor)
DEFINE_GPU_ACTIVATION_KERNEL(Tanh, CudaTanhFunctor)
DEFINE_GPU_ACTIVATION_KERNEL(TanhShrink, CudaTanhShrinkFunctor)
DEFINE_GPU_ACTIVATION_KERNEL(Silu, CudaSiluFunctor)
94 95 96 97 98 99
DEFINE_GPU_ACTIVATION_KERNEL(Exp, CudaExpFunctor)
DEFINE_GPU_ACTIVATION_KERNEL(Expm1, CudaExpm1Functor)
DEFINE_GPU_ACTIVATION_KERNEL(Reciprocal, CudaReciprocalFunctor)
DEFINE_GPU_ACTIVATION_KERNEL(Square, CudaSquareFunctor)
DEFINE_GPU_ACTIVATION_KERNEL(Sqrt, CudaSqrtFunctor)
DEFINE_GPU_ACTIVATION_KERNEL(Rsqrt, CudaRsqrtFunctor)
100
DEFINE_GPU_ACTIVATION_KERNEL(Softsign, CudaSoftsignFunctor)
Y
YuanRisheng 已提交
101 102
DEFINE_GPU_ACTIVATION_KERNEL(Sigmoid, CudaSigmoidFunctor)
DEFINE_GPU_ACTIVATION_KERNEL(LogSigmoid, CudaLogSigmoidFunctor)
103 104 105 106
DEFINE_GPU_ACTIVATION_KERNEL(Log, CudaLogFunctor)
DEFINE_GPU_ACTIVATION_KERNEL(Log2, CudaLog2Functor)
DEFINE_GPU_ACTIVATION_KERNEL(Log10, CudaLog10Functor)
DEFINE_GPU_ACTIVATION_KERNEL(Log1p, CudaLog1pFunctor)
Y
YuanRisheng 已提交
107 108 109
DEFINE_GPU_ACTIVATION_KERNEL(Round, CudaRoundFunctor)
DEFINE_GPU_ACTIVATION_KERNEL(Floor, CudaFloorFunctor)
DEFINE_GPU_ACTIVATION_KERNEL(Ceil, CudaCeilFunctor)
110 111 112 113 114

DEFINE_GPU_ACT_KERNEL_WITH_ONE_ATTRS(LeakyRelu, CudaLeakyReluFunctor, alpha)
DEFINE_GPU_ACT_KERNEL_WITH_ONE_ATTRS(ThresholdedRelu,
                                     CudaThresholdedReluFunctor,
                                     threshold)
115
DEFINE_GPU_ACT_KERNEL_WITH_ONE_ATTRS(Relu6, CudaRelu6Functor, threshold)
Y
YuanRisheng 已提交
116 117 118 119 120
DEFINE_GPU_ACT_KERNEL_WITH_ONE_ATTRS(HardShrink,
                                     CudaHardShrinkFunctor,
                                     threshold)
DEFINE_GPU_ACT_KERNEL_WITH_ONE_ATTRS(SoftShrink, CudaSoftShrinkFunctor, lambda)
DEFINE_GPU_ACT_KERNEL_WITH_ONE_ATTRS(Elu, CudaELUFunctor, alpha)
Y
YuanRisheng 已提交
121
DEFINE_GPU_ACT_KERNEL_WITH_ONE_ATTRS(Swish, CudaSwishFunctor, beta)
122
DEFINE_GPU_ACT_KERNEL_WITH_ONE_ATTRS(Mish, CudaMishFunctor, threshold)
Y
YuanRisheng 已提交
123
DEFINE_GPU_ACT_KERNEL_WITH_ONE_ATTRS(Celu, CudaCELUFunctor, alpha)
124

125
DEFINE_GPU_ACT_KERNEL_WITH_TWO_ATTRS(BRelu, CudaBReluFunctor, t_min, t_max)
126 127 128 129 130
DEFINE_GPU_ACT_KERNEL_WITH_TWO_ATTRS(Stanh, CudaSTanhFunctor, scale_a, scale_b)
DEFINE_GPU_ACT_KERNEL_WITH_TWO_ATTRS(Softplus,
                                     CudaSoftplusFunctor,
                                     beta,
                                     threshold)
Y
YuanRisheng 已提交
131 132 133 134
DEFINE_GPU_ACT_KERNEL_WITH_TWO_ATTRS(HardSigmoid,
                                     CudaHardSigmoidFunctor,
                                     slope,
                                     offset)
135
DEFINE_GPU_ACT_KERNEL_WITH_TWO_ATTRS(Selu, CudaSeluFunctor, scale, alpha)
136

Y
YuanRisheng 已提交
137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152
template <typename T, typename Context>
void HardSwishKernel(const Context& dev_ctx,
                     const DenseTensor& x,
                     float threshold,
                     float scale,
                     float offset,
                     DenseTensor* out) {
  funcs::CudaHardSwishFunctor<T> functor;
  auto attrs = functor.GetAttrs();
  *(attrs[0].second) = threshold;
  *(attrs[1].second) = scale;
  *(attrs[2].second) = offset;
  ActivationGPUImpl<T, Context, funcs::CudaHardSwishFunctor<T>>(
      dev_ctx, x, out, functor);
}

153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172
}  // namespace phi

#ifdef PADDLE_WITH_HIP
PD_REGISTER_KERNEL(relu,
                   GPU,
                   ALL_LAYOUT,
                   phi::ReluKernel,
                   float,
                   double,
                   phi::dtype::float16) {}
#else
PD_REGISTER_KERNEL(relu,
                   GPU,
                   ALL_LAYOUT,
                   phi::ReluKernel,
                   float,
                   double,
                   phi::dtype::float16,
                   phi::dtype::bfloat16) {}
#endif
173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197

#define PD_REGISTER_ACTIVATION_KERNEL(name, func) \
  PD_REGISTER_KERNEL(name,                        \
                     GPU,                         \
                     ALL_LAYOUT,                  \
                     phi::func,                   \
                     float,                       \
                     double,                      \
                     phi::dtype::float16,         \
                     phi::dtype::bfloat16) {}

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)
PD_REGISTER_ACTIVATION_KERNEL(brelu, BReluKernel)
PD_REGISTER_ACTIVATION_KERNEL(thresholded_relu, ThresholdedReluKernel)
198
PD_REGISTER_ACTIVATION_KERNEL(relu6, Relu6Kernel)
199
PD_REGISTER_ACTIVATION_KERNEL(leaky_relu, LeakyReluKernel)
200 201 202 203 204 205 206 207 208 209 210 211 212 213 214
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)

PD_REGISTER_KERNEL(exp,
                   GPU,
                   ALL_LAYOUT,
                   phi::ExpKernel,
                   float,
                   double,
                   int,
                   int64_t,
215 216
                   phi::dtype::float16,
                   phi::dtype::bfloat16) {}
217 218 219 220 221 222
PD_REGISTER_KERNEL(expm1,
                   GPU,
                   ALL_LAYOUT,
                   phi::Expm1Kernel,
                   float,
                   double,
223 224
                   phi::dtype::float16,
                   phi::dtype::bfloat16) {}
225 226 227 228 229 230
PD_REGISTER_KERNEL(logit,
                   GPU,
                   ALL_LAYOUT,
                   phi::LogitKernel,
                   float,
                   double,
231 232
                   phi::dtype::float16,
                   phi::dtype::bfloat16) {}
233 234 235 236 237 238 239 240 241 242
PD_REGISTER_KERNEL(square,
                   GPU,
                   ALL_LAYOUT,
                   phi::SquareKernel,
                   float,
                   double,
                   int,
                   int64_t,
                   phi::dtype::float16,
                   phi::dtype::bfloat16) {}
Y
YuanRisheng 已提交
243

Y
YuanRisheng 已提交
244 245 246 247 248
PD_REGISTER_ACTIVATION_KERNEL(hard_shrink, HardShrinkKernel)
PD_REGISTER_ACTIVATION_KERNEL(soft_shrink, SoftShrinkKernel)
PD_REGISTER_ACTIVATION_KERNEL(tanh_shrink, TanhShrinkKernel)
PD_REGISTER_ACTIVATION_KERNEL(elu, EluKernel)
PD_REGISTER_ACTIVATION_KERNEL(silu, SiluKernel)
249
PD_REGISTER_ACTIVATION_KERNEL(softsign, SoftsignKernel)
Y
YuanRisheng 已提交
250 251 252
PD_REGISTER_ACTIVATION_KERNEL(sigmoid, SigmoidKernel)
PD_REGISTER_ACTIVATION_KERNEL(logsigmoid, LogSigmoidKernel)
PD_REGISTER_ACTIVATION_KERNEL(hard_sigmoid, HardSigmoidKernel)
253 254 255 256
PD_REGISTER_ACTIVATION_KERNEL(log, LogKernel)
PD_REGISTER_ACTIVATION_KERNEL(log2, Log2Kernel)
PD_REGISTER_ACTIVATION_KERNEL(log10, Log10Kernel)
PD_REGISTER_ACTIVATION_KERNEL(log1p, Log1pKernel)
Y
YuanRisheng 已提交
257 258 259 260 261
PD_REGISTER_ACTIVATION_KERNEL(hard_swish, HardSwishKernel)
PD_REGISTER_ACTIVATION_KERNEL(swish, SwishKernel)
PD_REGISTER_ACTIVATION_KERNEL(round, RoundKernel)
PD_REGISTER_ACTIVATION_KERNEL(floor, FloorKernel)
PD_REGISTER_ACTIVATION_KERNEL(ceil, CeilKernel)
Y
YuanRisheng 已提交
262
PD_REGISTER_ACTIVATION_KERNEL(celu, CeluKernel)
Y
YuanRisheng 已提交
263 264 265 266 267 268 269 270
PD_REGISTER_KERNEL(pow,
                   GPU,
                   ALL_LAYOUT,
                   phi::PowKernel,
                   float,
                   double,
                   int,
                   int64_t,
271 272 273 274 275 276 277 278 279
                   phi::dtype::float16,
                   phi::dtype::bfloat16) {}
PD_REGISTER_KERNEL(selu,
                   GPU,
                   ALL_LAYOUT,
                   phi::SeluKernel,
                   float,
                   double,
                   phi::dtype::bfloat16) {}