Gitiles
Code Review Sign In
review.mlplatform.org / ml / ComputeLibrary / 33f41fabd30fb444aaa0cf3e65b61794d498d151 / . / src / core / cpu / kernels / activation / NEON / fp32.cpp
blob: 54301d45ad5bd4ab731b2bc5cc3fc152432e9b04 [file] [log] [blame]
Michalis Spyrouc4d45552020-10-19 12:41:30 +0100[diff] [blame]1/*
Georgios Pinitas70eb53b2021-01-06 19:42:21 +0000[diff] [blame]2 * Copyright (c) 2020-2021 Arm Limited.
Michalis Spyrouc4d45552020-10-19 12:41:30 +0100[diff] [blame]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 */
24#include "arm_compute/core/Helpers.h"
25#include "arm_compute/core/ITensorPack.h"
26#include "arm_compute/core/Window.h"
27#include "src/core/NEON/NEMath.h"
28#include "src/core/NEON/wrapper/wrapper.h"
Michalis Spyrouc4d45552020-10-19 12:41:30 +0100[diff] [blame]29
30#include <arm_neon.h>
31#include <cmath>
32#include <cstddef>
33
34namespace arm_compute
35{
36namespace cpu
37{
38namespace
39{
40#ifndef __aarch64__
41inline 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}
Michele Di Giorgio6c25aad2021-01-14 16:17:48 +0000[diff] [blame]46#endif /* __aarch64__ */
Michalis Spyrouc4d45552020-10-19 12:41:30 +0100[diff] [blame]47} // namespace
48
49void fp32_neon_activation(const ITensor *src, ITensor *dst, const ActivationLayerInfo &act_info, const Window &window)
50{
Michele Di Giorgio33f41fa2021-03-09 14:09:08 +0000[diff] [blame^]51 /** SIMD vector tag type. */
Michalis Spyrouc4d45552020-10-19 12:41:30 +0100[diff] [blame]52 using ExactTagType = typename arm_compute::wrapper::traits::neon_bitvector_tag_t<float, wrapper::traits::BitWidth::W128>;
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());
57 const ActivationLayerInfo::ActivationFunction act = act_info.activation();
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 {});
Michele Di Giorgio6c25aad2021-01-14 16:17:48 +0000[diff] [blame]70#endif /* __aarch64__ */
Michalis Spyrouc4d45552020-10-19 12:41:30 +0100[diff] [blame]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
Michele Di Giorgio6c25aad2021-01-14 16:17:48 +0000[diff] [blame]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
Michalis Spyrouc4d45552020-10-19 12:41:30 +0100[diff] [blame]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
89 wrapper::traits::neon_bitvector_t<float, wrapper::traits::BitWidth::W128> tmp;
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 {
98 case ActivationLayerInfo::ActivationFunction::ABS:
99 tmp = wrapper::vabs(vin);
100 break;
101 case ActivationLayerInfo::ActivationFunction::LINEAR:
102 tmp = wrapper::vmla(vb, va, vin);
103 break;
104 case ActivationLayerInfo::ActivationFunction::LOGISTIC:
105 tmp = wrapper::vinv(wrapper::vadd(const_1, wrapper::vexpq(wrapper::vneg(vin))));
106 break;
107 case ActivationLayerInfo::ActivationFunction::RELU:
108 tmp = wrapper::vmax(const_0, vin);
109 break;
110 case ActivationLayerInfo::ActivationFunction::BOUNDED_RELU:
111 tmp = wrapper::vmin(va, wrapper::vmax(const_0, vin));
112 break;
113 case ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU:
114 tmp = wrapper::vmin(va, wrapper::vmax(vb, vin));
115 break;
116 case ActivationLayerInfo::ActivationFunction::LEAKY_RELU:
117 tmp = wrapper::vbsl(wrapper::vcgt(vin, const_0), vin, wrapper::vmul(va, vin));
118 break;
119 case ActivationLayerInfo::ActivationFunction::SOFT_RELU:
Michele Di Giorgio6c25aad2021-01-14 16:17:48 +0000[diff] [blame]120 tmp = wrapper::vbsl(wrapper::vcgt(vin, vsoft_relu_thresh), vin, wrapper::vlog(wrapper::vadd(const_1, wrapper::vexpq(vin))));
Michalis Spyrouc4d45552020-10-19 12:41:30 +0100[diff] [blame]121 break;
122 case ActivationLayerInfo::ActivationFunction::ELU:
123 tmp = wrapper::vbsl(wrapper::vcge(vin, const_0), vin, wrapper::vmul(va, wrapper::vsub(wrapper::vexpq(vin), const_1)));
124 break;
125 case ActivationLayerInfo::ActivationFunction::SQRT:
126#ifdef __aarch64__
127 tmp = wrapper::vsqrt(vin);
Michele Di Giorgio6c25aad2021-01-14 16:17:48 +0000[diff] [blame]128#else /* __aarch64__ */
Michalis Spyrouc4d45552020-10-19 12:41:30 +0100[diff] [blame]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 }
Michele Di Giorgio6c25aad2021-01-14 16:17:48 +0000[diff] [blame]134#endif /* __aarch64__ */
Michalis Spyrouc4d45552020-10-19 12:41:30 +0100[diff] [blame]135 break;
136 case ActivationLayerInfo::ActivationFunction::SQUARE:
137 tmp = wrapper::vmul(vin, vin);
138 break;
139 case ActivationLayerInfo::ActivationFunction::TANH:
140 tmp = wrapper::vmul(va, wrapper::vtanh(wrapper::vmul(vb, vin)));
141 break;
142 case ActivationLayerInfo::ActivationFunction::IDENTITY:
143 tmp = vin;
144 break;
145 case ActivationLayerInfo::ActivationFunction::HARD_SWISH:
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 {
161 case ActivationLayerInfo::ActivationFunction::ABS:
162 tmp = std::abs(in);
163 break;
164 case ActivationLayerInfo::ActivationFunction::LINEAR:
165 tmp = a * in + b;
166 break;
167 case ActivationLayerInfo::ActivationFunction::LOGISTIC:
168 tmp = static_cast<float>(1) / (static_cast<float>(1) + std::exp(-in));
169 break;
170 case ActivationLayerInfo::ActivationFunction::RELU:
171 tmp = std::max<float>(static_cast<float>(0), in);
172 break;
173 case ActivationLayerInfo::ActivationFunction::BOUNDED_RELU:
174 tmp = std::min<float>(a, std::max(static_cast<float>(0), in));
175 break;
176 case ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU:
177 tmp = std::min<float>(a, std::max<float>(b, in));
178 break;
179 case ActivationLayerInfo::ActivationFunction::LEAKY_RELU:
180 tmp = (in > 0) ? in : a * in;
181 break;
182 case ActivationLayerInfo::ActivationFunction::SOFT_RELU:
Michele Di Giorgio6c25aad2021-01-14 16:17:48 +0000[diff] [blame]183 tmp = (in > soft_relu_thresh) ? in : std::log(static_cast<float>(1) + std::exp(in));
Michalis Spyrouc4d45552020-10-19 12:41:30 +0100[diff] [blame]184 break;
185 case ActivationLayerInfo::ActivationFunction::ELU:
186 tmp = (in >= 0) ? in : a * (std::exp(in) - 1);
187 break;
188 case ActivationLayerInfo::ActivationFunction::SQRT:
189 tmp = std::sqrt(in);
190 break;
191 case ActivationLayerInfo::ActivationFunction::SQUARE:
192 tmp = in * in;
193 break;
194 case ActivationLayerInfo::ActivationFunction::TANH:
195 tmp = a * std::tanh(b * in);
196 break;
197 case ActivationLayerInfo::ActivationFunction::IDENTITY:
198 tmp = in;
199 break;
200 case ActivationLayerInfo::ActivationFunction::HARD_SWISH:
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