Compute Library
 21.05
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"
29 
30 #include <arm_neon.h>
31 #include <cmath>
32 #include <cstddef>
33 
34 namespace arm_compute
35 {
36 namespace cpu
37 {
38 namespace
39 {
40 #ifndef __aarch64__
41 inline float32x4_t mask_float_vector(const float32x4_t &in, const uint32x4_t &mask)
42 {
43  auto int_in = vreinterpretq_u32_f32(in);
44  return vreinterpretq_f32_u32(wrapper::vand(int_in, mask));
45 }
46 #endif /* __aarch64__ */
47 } // namespace
48 
49 void fp32_neon_activation(const ITensor *src, ITensor *dst, const ActivationLayerInfo &act_info, const Window &window)
50 {
51  /** SIMD vector tag type. */
53 
54  constexpr int window_step_x = 4;
55  const auto window_start_x = static_cast<int>(window.x().start());
56  const auto window_end_x = static_cast<int>(window.x().end());
58 
59  Window win_collapsed = window.collapse_if_possible(window, Window::DimZ);
60  win_collapsed.set(Window::DimX, Window::Dimension(0, 1, 1));
61 
62  Iterator input(src, win_collapsed);
63  Iterator output(dst, win_collapsed);
64 
65  // In case of non-aarch64, a small delta value is added to the input
66  // to prevent NAN values caused by zeros in inputs to SQRT.
67  // In case of aarh64, we call vsqrt directly, so we don't use delta.
68 #ifndef __aarch64__
69  const auto delta = wrapper::vdup_n(static_cast<float>(1e-24), ExactTagType {});
70 #endif /* __aarch64__ */
71  const auto const_1 = wrapper::vdup_n(static_cast<float>(1.f), ExactTagType {});
72  const auto const_0 = wrapper::vdup_n(static_cast<float>(0.f), ExactTagType{});
73  const auto const_6 = wrapper::vdup_n(static_cast<float>(6.f), ExactTagType{});
74  const auto const_3 = wrapper::vdup_n(static_cast<float>(3.f), ExactTagType{});
75  const auto const_inv_6 = wrapper::vdup_n(static_cast<float>(0.166666667f), ExactTagType{});
76 
77  constexpr float soft_relu_thresh = 12.f;
78  const auto vsoft_relu_thresh = wrapper::vdup_n(static_cast<float>(soft_relu_thresh), ExactTagType{});
79 
80  const auto va = wrapper::vdup_n(static_cast<float>(act_info.a()), ExactTagType{});
81  const auto vb = wrapper::vdup_n(static_cast<float>(act_info.b()), ExactTagType{});
82  const auto a = static_cast<float>(act_info.a());
83  const auto b = static_cast<float>(act_info.b());
84  execute_window_loop(win_collapsed, [&](const Coordinates &)
85  {
86  const auto input_ptr = reinterpret_cast<const float *>(input.ptr());
87  const auto output_ptr = reinterpret_cast<float *>(output.ptr());
88 
90 
91  // Compute S elements per iteration
92  int x = window_start_x;
93  for(; x <= (window_end_x - window_step_x); x += window_step_x)
94  {
95  const auto vin = wrapper::vloadq(input_ptr + x);
96  switch(act)
97  {
99  tmp = wrapper::vabs(vin);
100  break;
102  tmp = wrapper::vmla(vb, va, vin);
103  break;
106  break;
108  tmp = wrapper::vmax(const_0, vin);
109  break;
111  tmp = wrapper::vmin(va, wrapper::vmax(const_0, vin));
112  break;
114  tmp = wrapper::vmin(va, wrapper::vmax(vb, vin));
115  break;
117  tmp = wrapper::vbsl(wrapper::vcgt(vin, const_0), vin, wrapper::vmul(va, vin));
118  break;
120  tmp = wrapper::vbsl(wrapper::vcgt(vin, vsoft_relu_thresh), vin, wrapper::vlog(wrapper::vadd(const_1, wrapper::vexpq(vin))));
121  break;
123  tmp = wrapper::vbsl(wrapper::vcge(vin, const_0), vin, wrapper::vmul(va, wrapper::vsub(wrapper::vexpq(vin), const_1)));
124  break;
126 #ifdef __aarch64__
127  tmp = wrapper::vsqrt(vin);
128 #else /* __aarch64__ */
129  {
130  const auto bitmask = wrapper::vceq(vin, wrapper::vdup_n(0.f, ExactTagType{}));
131  tmp = wrapper::vinv(wrapper::vinvsqrt(wrapper::vadd(vin, mask_float_vector(delta, bitmask))));
132  tmp = mask_float_vector(tmp, wrapper::vnot(bitmask));
133  }
134 #endif /* __aarch64__ */
135  break;
137  tmp = wrapper::vmul(vin, vin);
138  break;
140  tmp = wrapper::vmul(va, wrapper::vtanh(wrapper::vmul(vb, vin)));
141  break;
143  tmp = vin;
144  break;
146  tmp = wrapper::vmul(vin, wrapper::vmul(const_inv_6, wrapper::vmin(const_6, wrapper::vmax(const_0, wrapper::vadd(vin, const_3)))));
147  break;
148  default:
149  ARM_COMPUTE_ERROR("Unsupported activation function");
150  }
151  wrapper::vstore(output_ptr + x, tmp);
152  }
153 
154  // Compute left-over elements
155  for(; x < window_end_x; ++x)
156  {
157  const float in = *(reinterpret_cast<const float *>(input_ptr + x));
158  float tmp;
159  switch(act)
160  {
162  tmp = std::abs(in);
163  break;
165  tmp = a * in + b;
166  break;
168  tmp = static_cast<float>(1) / (static_cast<float>(1) + std::exp(-in));
169  break;
171  tmp = std::max<float>(static_cast<float>(0), in);
172  break;
174  tmp = std::min<float>(a, std::max(static_cast<float>(0), in));
175  break;
177  tmp = std::min<float>(a, std::max<float>(b, in));
178  break;
180  tmp = (in > 0) ? in : a * in;
181  break;
183  tmp = (in > soft_relu_thresh) ? in : std::log(static_cast<float>(1) + std::exp(in));
184  break;
186  tmp = (in >= 0) ? in : a * (std::exp(in) - 1);
187  break;
189  tmp = std::sqrt(in);
190  break;
192  tmp = in * in;
193  break;
195  tmp = a * std::tanh(b * in);
196  break;
198  tmp = in;
199  break;
201  tmp = in * ((std::min(std::max((in + 3), 0.0f), 6.0f)) * 0.166666667f);
202  break;
203  default:
204  ARM_COMPUTE_ERROR("Unsupported activation function");
205  }
206  *(output_ptr + x) = tmp;
207  }
208  },
209  input, output);
210 }
211 } // namespace cpu
212 } // namespace arm_compute
float32x4_t vlog(const float32x4_t &a)
Definition: log.h:47
float32x4_t vtanh(const float32x4_t &a)
Definition: tanh.h:40
SimpleTensor< float > b
Definition: DFT.cpp:157
#define ARM_COMPUTE_ERROR(msg)
Print the given message then throw an std::runtime_error.
Definition: Error.h:352
float a() const
Get the alpha value.
Definition: Types.h:1518
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
float32x2_t vinv(const float32x2_t &a)
Definition: inv.h:47
int8x8_t vabs(const int8x8_t &a)
Definition: abs.h:46
uint8x8_t vsub(const uint8x8_t &a, const uint8x8_t &b)
Definition: sub.h:39
Describe one of the image's dimensions with a start, end and step.
Definition: Window.h:77
Activation Layer Information class.
Definition: Types.h:1478
Interface for CPU tensor.
Definition: ITensor.h:36
SimpleTensor< float > src
Definition: DFT.cpp:155
Copyright (c) 2017-2021 Arm Limited.
ActivationFunction
Available activation functions.
Definition: Types.h:1482
typename neon_bitvector< T, BW >::tag_type neon_bitvector_tag_t
Helper type template to get the tag type of a neon vector.
Definition: traits.h:132
static constexpr size_t DimX
Alias for dimension 0 also known as X dimension.
Definition: Window.h:43
uint8x8_t vnot(const uint8x8_t &a)
Definition: not.h:39
Window collapse_if_possible(const Window &full_window, size_t first, size_t last, bool *has_collapsed=nullptr) const
Collapse the dimensions between first and last if possible.
Definition: Window.inl:68
uint8x8_t vmin(const uint8x8_t &a, const uint8x8_t &b)
Definition: min.h:39
Coordinates of an item.
Definition: Coordinates.h:37
uint8x8_t vand(const uint8x8_t &a, const uint8x8_t &b)
Definition: and.h:39
void fp32_neon_activation(const ITensor *src, ITensor *dst, const ActivationLayerInfo &act_info, const Window &window)
Definition: fp32.cpp:49
constexpr uint8_t * ptr() const
Return a pointer to the current pixel.
Definition: Helpers.inl:139
void set(size_t dimension, const Dimension &dim)
Set the values of a given dimension.
Definition: Window.inl:49
int8x8_t vneg(const int8x8_t &a)
Definition: neg.h:39
uint8x8_t vcgt(const uint8x8_t &a, const uint8x8_t &b)
Definition: cgt.h:39
uint8x8_t vmul(const uint8x8_t &a, const uint8x8_t &b)
Definition: mul.h:39
uint8x8_t vbsl(const uint8x8_t &a, const uint8x8_t &b, const uint8x8_t &c)
Definition: bsl.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
typename neon_bitvector< T, BW >::type neon_bitvector_t
Helper type template to get the type of a neon vector.
Definition: traits.h:130
ActivationFunction activation() const
Get the type of activation function.
Definition: Types.h:1513
float b() const
Get the beta value.
Definition: Types.h:1523
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
constexpr int end() const
Return the end of the dimension.
Definition: Window.h:99
uint8x8_t vcge(const uint8x8_t &a, const uint8x8_t &b)
Definition: cge.h:39
Iterator updated by execute_window_loop for each window element.
Definition: Helpers.h:46
uint8x8_t vmax(const uint8x8_t &a, const uint8x8_t &b)
Definition: max.h:39
constexpr int start() const
Return the start of the dimension.
Definition: Window.h:94
float32x4_t vexpq(const float32x4_t &a)
Definition: exp.h:47
Describe a multidimensional execution window.
Definition: Window.h:39
uint8x8_t vceq(const uint8x8_t &a, const uint8x8_t &b)
Definition: ceq.h:39
constexpr const Dimension & x() const
Alias to access the first dimension of the window.
Definition: Window.h:145