avx_functions.cc 2.6 KB
Newer Older
D
dangqingqing 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
/* 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. */

#include <immintrin.h>
16
#include "paddle/operators/math/detail/activation_functions.h"
17 18
// TODO(qingqing) refine this dependence
#include "paddle/cuda/src/avx_mathfun.h"
D
dangqingqing 已提交
19

20 21 22 23
namespace paddle {
namespace operators {
namespace math {
namespace detail {
D
dangqingqing 已提交
24

25
__m256 exp(__m256 a) { return exp256_ps(a); }
D
dangqingqing 已提交
26

27 28
namespace forward {
namespace avx {
D
dangqingqing 已提交
29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57
__m256 relu(const __m256 a) {
  __m256 tmp = _mm256_set1_ps(0.0f);
  return _mm256_max_ps(a, tmp);
}

__m256 sigmoid(const __m256 a) {
  __m256 max = _mm256_set1_ps(SIGMOID_THRESHOLD_MAX);
  __m256 min = _mm256_set1_ps(SIGMOID_THRESHOLD_MIN);
  __m256 tmp = _mm256_max_ps(a, min);
  tmp = _mm256_min_ps(tmp, max);
  tmp = _mm256_sub_ps(_mm256_set1_ps(0.0f), tmp);
  tmp = exp(tmp);
  tmp = _mm256_add_ps(_mm256_set1_ps(1.0f), tmp);
  tmp = _mm256_div_ps(_mm256_set1_ps(1.0f), tmp);
  return tmp;
}

__m256 tanh(const __m256 a) {
  __m256 max = _mm256_set1_ps(EXP_MAX_INPUT);
  __m256 tmp = _mm256_mul_ps(_mm256_set1_ps(-2.0f), a);
  tmp = _mm256_min_ps(tmp, max);
  tmp = exp(tmp);
  return _mm256_sub_ps(_mm256_div_ps(_mm256_set1_ps(2.0f),
                                     _mm256_add_ps(_mm256_set1_ps(1.0f), tmp)),
                       _mm256_set1_ps(1.0f));
}

__m256 linear(const __m256 a) { return a; }

58 59 60 61 62
}  // namespace avx
}  // namespace forward

namespace backward {
namespace avx {
D
dangqingqing 已提交
63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79
__m256 relu(const __m256 a, const __m256 b) {
  return _mm256_mul_ps(
      a, _mm256_and_ps(_mm256_cmp_ps(b, _mm256_set1_ps(0.0f), _CMP_GT_OS),
                       _mm256_set1_ps(1.0f)));
}

__m256 sigmoid(const __m256 a, const __m256 b) {
  return _mm256_mul_ps(_mm256_mul_ps(a, b),
                       _mm256_sub_ps(_mm256_set1_ps(1.0f), b));
}

__m256 tanh(const __m256 a, const __m256 b) {
  return _mm256_mul_ps(
      a, _mm256_sub_ps(_mm256_set1_ps(1.0f), _mm256_mul_ps(b, b)));
}

__m256 linear(const __m256 a, const __m256 b) { return a; }
80 81 82 83 84 85 86
}  // namespace avx
}  // namespace backward

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