cpu_lstm_compute.h 2.0 KB
Newer Older
T
tensor-tang 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13
/* Copyright (c) 2018 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. */

#pragma once
#include <string>
G
gongweibao 已提交
14 15 16 17 18
#include "paddle/fluid/operators/math/cpu_vec.h"
#include "paddle/fluid/platform/cpu_info.h"
#ifdef __AVX__
#include <immintrin.h>
#endif
T
tensor-tang 已提交
19 20 21 22 23 24 25

namespace paddle {
namespace operators {
namespace math {

// TODO(TJ): ugly workaround, clean me
template <typename T>
G
gongweibao 已提交
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
void lstm_compute_ctht(T* gates, const T* ct_1, T* ct, T* ht) {
  // gates: W_ch, W_ih, W_fh, W_oh
  vec_sigmoid<T, platform::jit::avx>(24, gates + 8, gates + 8);
  vec_tanh<T, platform::jit::avx>(8, gates, gates);
  const T *i = gates + 8, *f = gates + 16, *o = gates + 24;
  const T min = SIGMOID_THRESHOLD_MIN;
  const T max = SIGMOID_THRESHOLD_MAX;
  for (int d = 0; d < 8; ++d) {
    // C_t = C_t-1 * fgated + cand_gated * igated
    ct[d] = ct_1[d] * f[d] + gates[d] * i[d];
    // H_t = act_cell(C_t) * ogated
    T tmp = ct[d] * 2;
    tmp = static_cast<T>(0) - ((tmp < min) ? min : ((tmp > max) ? max : tmp));
    vec_exp<T>(1, &tmp, &tmp);
    tmp = static_cast<T>(2) / (static_cast<T>(1) + tmp) - static_cast<T>(1);
    ht[d] = tmp * o[d];
  }
}

#ifdef __AVX__
namespace detail {
namespace forward {
namespace avx {
__m256 Sigmoid(const __m256 a);
__m256 Tanh(const __m256 a);
51

G
gongweibao 已提交
52 53 54 55 56 57
}  // namespace avx
}  // namespace forward
}  // namespace detail

template <>
void lstm_compute_ctht<float>(float* gates, const float* ct_1, float* ct,
58
                              float* ht);
G
gongweibao 已提交
59 60

#endif
61

T
tensor-tang 已提交
62 63 64
}  // namespace math
}  // namespace operators
}  // namespace paddle