“2e08de2b45bdf99c320176c64f29a11452ab1647”上不存在“src/client/git@gitcode.net:qq_37101384/tdengine.git”
lstm_kernel.h 6.2 KB
Newer Older
D
dangqingqing 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.

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. */

Y
Yi Wang 已提交
15 16
#include "paddle/fluid/operators/math/detail/activation_functions.h"
#include "paddle/fluid/platform/hostdevice.h"
D
dangqingqing 已提交
17

Y
Yu Yang 已提交
18
#include <type_traits>
D
dangqingqing 已提交
19 20 21 22 23 24 25 26 27 28 29

namespace paddle {
namespace operators {
namespace math {
namespace detail {

namespace forward {

template <class T>
class lstm {
 public:
30 31
  HOSTDEVICE void operator()(T &value_in, T &value_ig, T &value_fg, T &value_og,
                             T &prev_state, T &state, T &state_atv, T &output,
32
                             T &checkI, T &checkF, T &checkO,
33 34 35
                             ActivationType active_node,
                             ActivationType active_gate,
                             ActivationType active_state) {
36 37 38 39 40 41 42
    value_in = activation(value_in, active_node);
    value_ig = activation(value_ig + prev_state * checkI, active_gate);
    value_fg = activation(value_fg + prev_state * checkF, active_gate);
    state = value_in * value_ig + prev_state * value_fg;
    value_og = activation(value_og + state * checkO, active_gate);
    state_atv = activation(state, active_state);
    output = value_og * state_atv;
D
dangqingqing 已提交
43 44
  }
#ifndef __NVCC__
Y
Yu Yang 已提交
45
#ifndef __AVX__  // If not compiled with AVX instructs. Disable AVX by default
D
dangqingqing 已提交
46 47
  static const bool avx = false;
#else
Y
Yu Yang 已提交
48 49 50
  // Only float support AVX optimization
  static const bool avx = std::is_same<T, float>::value;

51 52 53 54
  HOSTDEVICE void operator()(__m256 &value_in, __m256 &value_ig,
                             __m256 &value_fg, __m256 &value_og,
                             __m256 &prev_state, __m256 &state,
                             __m256 &state_atv, __m256 &output, __m256 &checkI,
Y
Yu Yang 已提交
55
                             __m256 &checkF, __m256 &checkO,
56 57 58
                             ActivationType active_node,
                             ActivationType active_gate,
                             ActivationType active_state) {
59 60 61 62 63 64 65 66 67 68 69 70 71
    value_in = activation(value_in, active_node);
    value_ig =
        activation(_mm256_add_ps(value_ig, _mm256_mul_ps(prev_state, checkI)),
                   active_gate);
    value_fg =
        activation(_mm256_add_ps(value_fg, _mm256_mul_ps(prev_state, checkF)),
                   active_gate);
    state = _mm256_add_ps(_mm256_mul_ps(value_in, value_ig),
                          _mm256_mul_ps(prev_state, value_fg));
    value_og = activation(_mm256_add_ps(value_og, _mm256_mul_ps(state, checkO)),
                          active_gate);
    state_atv = activation(state, active_state);
    output = _mm256_mul_ps(value_og, state_atv);
D
dangqingqing 已提交
72 73 74 75 76 77 78 79 80 81 82 83
  }
#endif
#endif
};

}  // namespace forward

namespace backward {

template <class T>
class lstm {
 public:
84 85 86 87
  HOSTDEVICE void operator()(T &value_in, T &value_ig, T &value_fg, T &value_og,
                             T &grad_in, T &grad_ig, T &grad_fg, T &grad_og,
                             T &prev_state, T &prev_state_grad, T &state,
                             T &state_grad, T &state_atv, T &output_grad,
Y
Yu Yang 已提交
88
                             T &checkI, T &checkF, T &checkO, T &checkIGrad,
89
                             T &checkFGrad, T &checkOGrad,
90 91 92
                             ActivationType active_node,
                             ActivationType active_gate,
                             ActivationType active_state) {
93 94 95 96 97 98 99 100 101 102 103
    grad_og = activation(output_grad * state_atv, value_og, active_gate);
    state_grad += activation(output_grad * value_og, state_atv, active_state) +
                  grad_og * checkO;
    grad_in = activation(state_grad * value_ig, value_in, active_node);
    grad_ig = activation(state_grad * value_in, value_ig, active_gate);
    grad_fg = activation(state_grad * prev_state, value_fg, active_gate);
    prev_state_grad =
        grad_ig * checkI + grad_fg * checkF + state_grad * value_fg;
    checkIGrad = grad_ig * prev_state;
    checkFGrad = grad_fg * prev_state;
    checkOGrad = grad_og * state;
D
dangqingqing 已提交
104 105
  }
#ifndef __NVCC__
Y
Yu Yang 已提交
106
#ifndef __AVX__  // If not compiled with AVX instructs. Disable AVX by default
D
dangqingqing 已提交
107 108
  static const bool avx = false;
#else
Y
Yu Yang 已提交
109 110
  // Only float support AVX optimization
  static const bool avx = std::is_same<T, float>::value;
111
  HOSTDEVICE void operator()(
112 113 114 115 116
      __m256 &value_in, __m256 &value_ig, __m256 &value_fg, __m256 &value_og,
      __m256 &grad_in, __m256 &grad_ig, __m256 &grad_fg, __m256 &grad_og,
      __m256 &prev_state, __m256 &prev_state_grad, __m256 &state,
      __m256 &state_grad, __m256 &state_atv, __m256 &output_grad,
      __m256 &checkI, __m256 &checkF, __m256 &checkO, __m256 &checkIGrad,
117 118
      __m256 &checkFGrad, __m256 &checkOGrad, ActivationType active_node,
      ActivationType active_gate, ActivationType active_state) {
119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137
    grad_og = activation(_mm256_mul_ps(output_grad, state_atv), value_og,
                         active_gate);
    state_grad = _mm256_add_ps(activation(_mm256_mul_ps(output_grad, value_og),
                                          state_atv, active_state),
                               state_grad);
    state_grad = _mm256_add_ps(_mm256_mul_ps(grad_og, checkO), state_grad);
    grad_in =
        activation(_mm256_mul_ps(state_grad, value_ig), value_in, active_node);
    grad_ig =
        activation(_mm256_mul_ps(state_grad, value_in), value_ig, active_gate);
    grad_fg = activation(_mm256_mul_ps(state_grad, prev_state), value_fg,
                         active_gate);
    prev_state_grad = _mm256_add_ps(_mm256_mul_ps(grad_ig, checkI),
                                    _mm256_mul_ps(grad_fg, checkF));
    prev_state_grad =
        _mm256_add_ps(_mm256_mul_ps(state_grad, value_fg), prev_state_grad);
    checkIGrad = _mm256_mul_ps(grad_ig, prev_state);
    checkFGrad = _mm256_mul_ps(grad_fg, prev_state);
    checkOGrad = _mm256_mul_ps(grad_og, state);
D
dangqingqing 已提交
138 139 140 141 142 143 144 145 146 147 148
  }
#endif
#endif
};

}  // namespace backward

}  // namespace detail
}  // namespace math
}  // namespace operators
}  // namespace paddle