Compute Library
 22.05
fp32.cpp
Go to the documentation of this file.
1 /*
2  * Copyright (c) 2020-2021 Arm Limited.
3  *
4  * SPDX-License-Identifier: MIT
5  *
6  * Permission is hereby granted, free of charge, to any person obtaining a copy
7  * of this software and associated documentation files (the "Software"), to
8  * deal in the Software without restriction, including without limitation the
9  * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
10  * sell copies of the Software, and to permit persons to whom the Software is
11  * furnished to do so, subject to the following conditions:
12  *
13  * The above copyright notice and this permission notice shall be included in all
14  * copies or substantial portions of the Software.
15  *
16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
19  * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
21  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
22  * SOFTWARE.
23  */
27 #include "src/core/NEON/NEMath.h"
30 
31 #include <arm_neon.h>
32 #include <cmath>
33 #include <cstddef>
34 
35 namespace arm_compute
36 {
37 namespace
38 {
39 using BatchNomalizationPtr = void (*)(ITensor *src, ITensor *dst, const ITensor *mean, const ITensor *var, const ITensor *beta, const ITensor *gamma,
40  float epsilon, ActivationLayerInfo &act_info, const Window &window);
41 
42 template <typename T>
43 void batch_normalization(ITensor *src, ITensor *dst, const ITensor *mean, const ITensor *var, const ITensor *beta, const ITensor *gamma,
44  float epsilon, ActivationLayerInfo &act_info, const Window &window)
45 {
46  /** SIMD vector tag type. */
47  using ExactTagType = typename wrapper::traits::neon_bitvector_tag_t<float, wrapper::traits::BitWidth::W128>;
48 
49  const int window_step_x = 4;
50  const auto window_start_x = static_cast<int>(window.x().start());
51  const auto window_end_x = static_cast<int>(window.x().end());
52 
53  Window win_collapsed = window.collapse_if_possible(window, Window::DimZ);
54  win_collapsed.set(Window::DimX, Window::Dimension(0, 1, 1));
55 
56  Iterator input(src, win_collapsed);
57  Iterator output(dst, win_collapsed);
58 
59  const auto input_mean = reinterpret_cast<const float *>(mean->ptr_to_element(Coordinates(0, 0)));
60  const auto input_var = reinterpret_cast<const float *>(var->ptr_to_element(Coordinates(0, 0)));
61  const auto input_gamma = (gamma != nullptr) ? reinterpret_cast<const float *>(gamma->ptr_to_element(Coordinates(0, 0))) : nullptr;
62  const auto input_beta = (beta != nullptr) ? reinterpret_cast<const float *>(beta->ptr_to_element(Coordinates(0, 0))) : nullptr;
63 
64  T activation_functor(act_info);
65 
66  const auto epsilon_vec = wrapper::vdup_n(static_cast<float>(epsilon), ExactTagType{});
67  execute_window_loop(win_collapsed, [&](const Coordinates &)
68  {
69  const auto input_ptr = reinterpret_cast<const float *>(input.ptr());
70  const auto output_ptr = reinterpret_cast<float *>(output.ptr());
71 
72  // Perform core calculations using vector operations
73  int x = window_start_x;
74  for(; x <= (window_end_x - window_step_x); x += window_step_x)
75  {
76  // Conctruct vectors
77  const auto mean_vec = wrapper::vloadq(input_mean + x);
78  const auto var_vec = wrapper::vloadq(input_var + x);
79  const auto gamma_vec = (input_gamma != nullptr) ? wrapper::vloadq(input_gamma + x) : wrapper::vdup_n(static_cast<float>(1.f), ExactTagType{});
80  const auto beta_vec = (input_beta != nullptr) ? wrapper::vloadq(input_beta + x) : wrapper::vdup_n(static_cast<float>(0.f), ExactTagType{});
81 
82  // Calculate denominator
83  const auto denominator = wrapper::vinvsqrt(wrapper::vadd(var_vec, epsilon_vec));
84 
85  // Calculate x bar
86  const auto numerator = wrapper::vsub(wrapper::vloadq(input_ptr + x), mean_vec);
87  const auto x_bar = wrapper::vmul(numerator, denominator);
88  auto res = wrapper::vmla(beta_vec, x_bar, gamma_vec);
89 
90  // Perform fused activation
91  if(act_info.enabled())
92  {
93  activation_functor(res);
94  }
95 
96  // Store results
97  wrapper::vstore(output_ptr + x, res);
98  }
99 
100  // Compute left-over elements
101  for(; x < window_end_x; ++x)
102  {
103  // Conctruct vectors
104  const float gamma = (input_gamma != nullptr) ? input_gamma[x] : 1.f;
105  const float beta = (input_beta != nullptr) ? input_beta[x] : 0.f;
106 
107  const float denominator = sqrt(input_var[x] + epsilon);
108  const float numerator = input_ptr[x] - input_mean[x];
109  const float x_bar = numerator / denominator;
110  float res = beta + x_bar * gamma;
111 
112  // Perform fused activation
113  if(act_info.enabled())
114  {
115  activation_functor(res);
116  }
117 
118  // Store results
119  *reinterpret_cast<float *>(output_ptr + x) = res;
120  }
121  },
122  input, output);
123 }
124 
125 // Fused Batched Normalization with activation functions
126 static std::map<ActivationLayerInfo::ActivationFunction, BatchNomalizationPtr> fused_map =
127 {
128  { ActivationLayerInfo::ActivationFunction::RELU, &batch_normalization<detail::relu<float, 4>> },
129  { ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, &batch_normalization<detail::brelu<float, 4>> },
130  { ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, &batch_normalization<detail::lubrelu<float, 4>> }
131 };
132 }
133 namespace cpu
134 {
135 void fp32_neon_batch_normalization(ITensor *src, ITensor *dst, const ITensor *mean, const ITensor *var, const ITensor *beta, const ITensor *gamma,
136  float epsilon, ActivationLayerInfo &act_info, const Window &window)
137 {
138  if(act_info.enabled())
139  {
140  fused_map[act_info.activation()](src, dst, mean, var, beta, gamma, epsilon, act_info, window);
141  }
142  else
143  {
144  batch_normalization<detail::dummy<float, 4>>(src, dst, mean, var, beta, gamma, epsilon, act_info, window);
145  }
146 }
147 } // namespace cpu
148 } // namespace arm_compute
bool enabled() const
Check if initialised.
Definition: Types.h:1675
float32x2_t vinvsqrt(const float32x2_t &a)
Definition: invsqrt.h:47
uint8x16_t vloadq(const uint8_t *ptr)
Definition: load.h:58
uint8x8_t vadd(const uint8x8_t &a, const uint8x8_t &b)
Definition: add.h:39
uint8x8_t vsub(const uint8x8_t &a, const uint8x8_t &b)
Definition: sub.h:39
Activation Layer Information class.
Definition: Types.h:1625
Interface for CPU tensor.
Definition: ITensor.h:36
SimpleTensor< float > src
Definition: DFT.cpp:155
Copyright (c) 2017-2022 Arm Limited.
static constexpr size_t DimX
Alias for dimension 0 also known as X dimension.
Definition: Window.h:43
void fp32_neon_batch_normalization(ITensor *src, ITensor *dst, const ITensor *mean, const ITensor *var, const ITensor *beta, const ITensor *gamma, float epsilon, ActivationLayerInfo &act_info, const Window &window)
Definition: fp32.cpp:135
uint8x8_t vmul(const uint8x8_t &a, const uint8x8_t &b)
Definition: mul.h:39
static constexpr size_t DimZ
Alias for dimension 2 also known as Z dimension.
Definition: Window.h:47
void vstore(uint8_t *ptr, uint8x8_t val)
Definition: store.h:39
uint8x8_t vdup_n(uint8_t value, traits::vector_64_tag)
Definition: dup_n.h:41
void execute_window_loop(const Window &w, L &&lambda_function, Ts &&... iterators)
Iterate through the passed window, automatically adjusting the iterators and calling the lambda_funct...
Definition: Helpers.inl:77
ActivationFunction activation() const
Get the type of activation function.
Definition: Types.h:1660
Includes all wrapper headers at once.
uint8x8_t vmla(const uint8x8_t &a, const uint8x8_t &b, const uint8x8_t &c)
Definition: mla.h:46
Describe a multidimensional execution window.
Definition: Window.h:39