• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  * Copyright (c) 2020 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  */
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"
29 #include "src/core/common/StdTypes.h"
30 #include "src/core/common/Validate.h"
31 
32 #include <arm_neon.h>
33 #include <cmath>
34 #include <cstddef>
35 
36 namespace arm_compute
37 {
38 namespace cpu
39 {
40 namespace
41 {
42 #ifndef __aarch64__
mask_float_vector(const float32x4_t & in,const uint32x4_t & mask)43 inline float32x4_t mask_float_vector(const float32x4_t &in, const uint32x4_t &mask)
44 {
45     auto int_in = vreinterpretq_u32_f32(in);
46     return vreinterpretq_f32_u32(wrapper::vand(int_in, mask));
47 }
48 #endif /* __arch64__ */
49 } // namespace
50 
fp32_neon_activation(const ITensor * src,ITensor * dst,const ActivationLayerInfo & act_info,const Window & window)51 void fp32_neon_activation(const ITensor *src, ITensor *dst, const ActivationLayerInfo &act_info, const Window &window)
52 {
53     /** NEON vector tag type. */
54     using ExactTagType = typename arm_compute::wrapper::traits::neon_bitvector_tag_t<float, wrapper::traits::BitWidth::W128>;
55 
56     constexpr int                                 window_step_x  = 4;
57     const auto                                    window_start_x = static_cast<int>(window.x().start());
58     const auto                                    window_end_x   = static_cast<int>(window.x().end());
59     const ActivationLayerInfo::ActivationFunction act            = act_info.activation();
60 
61     Window win_collapsed = window.collapse_if_possible(window, Window::DimZ);
62     win_collapsed.set(Window::DimX, Window::Dimension(0, 1, 1));
63 
64     Iterator input(src, win_collapsed);
65     Iterator output(dst, win_collapsed);
66 
67     // In case of non-aarch64, a small delta value is added to the input
68     // to prevent NAN values caused by zeros in inputs to SQRT.
69     // In case of aarh64, we call vsqrt directly, so we don't use delta.
70 #ifndef __aarch64__
71     const auto delta = wrapper::vdup_n(static_cast<float>(1e-24), ExactTagType {});
72 #endif /* __aarch64 */
73     const auto const_1     = wrapper::vdup_n(static_cast<float>(1.f), ExactTagType {});
74     const auto const_0     = wrapper::vdup_n(static_cast<float>(0.f), ExactTagType{});
75     const auto const_6     = wrapper::vdup_n(static_cast<float>(6.f), ExactTagType{});
76     const auto const_3     = wrapper::vdup_n(static_cast<float>(3.f), ExactTagType{});
77     const auto const_inv_6 = wrapper::vdup_n(static_cast<float>(0.166666667f), ExactTagType{});
78 
79     const auto va = wrapper::vdup_n(static_cast<float>(act_info.a()), ExactTagType{});
80     const auto vb = wrapper::vdup_n(static_cast<float>(act_info.b()), ExactTagType{});
81     const auto a  = static_cast<float>(act_info.a());
82     const auto b  = static_cast<float>(act_info.b());
83     execute_window_loop(win_collapsed, [&](const Coordinates &)
84     {
85         const auto input_ptr  = reinterpret_cast<const float *>(input.ptr());
86         const auto output_ptr = reinterpret_cast<float *>(output.ptr());
87 
88         wrapper::traits::neon_bitvector_t<float, wrapper::traits::BitWidth::W128> tmp;
89 
90         // Compute S elements per iteration
91         int x = window_start_x;
92         for(; x <= (window_end_x - window_step_x); x += window_step_x)
93         {
94             const auto vin = wrapper::vloadq(input_ptr + x);
95             switch(act)
96             {
97                 case ActivationLayerInfo::ActivationFunction::ABS:
98                     tmp = wrapper::vabs(vin);
99                     break;
100                 case ActivationLayerInfo::ActivationFunction::LINEAR:
101                     tmp = wrapper::vmla(vb, va, vin);
102                     break;
103                 case ActivationLayerInfo::ActivationFunction::LOGISTIC:
104                     tmp = wrapper::vinv(wrapper::vadd(const_1, wrapper::vexpq(wrapper::vneg(vin))));
105                     break;
106                 case ActivationLayerInfo::ActivationFunction::RELU:
107                     tmp = wrapper::vmax(const_0, vin);
108                     break;
109                 case ActivationLayerInfo::ActivationFunction::BOUNDED_RELU:
110                     tmp = wrapper::vmin(va, wrapper::vmax(const_0, vin));
111                     break;
112                 case ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU:
113                     tmp = wrapper::vmin(va, wrapper::vmax(vb, vin));
114                     break;
115                 case ActivationLayerInfo::ActivationFunction::LEAKY_RELU:
116                     tmp = wrapper::vbsl(wrapper::vcgt(vin, const_0), vin, wrapper::vmul(va, vin));
117                     break;
118                 case ActivationLayerInfo::ActivationFunction::SOFT_RELU:
119                     tmp = wrapper::vlog(wrapper::vadd(const_1, wrapper::vexpq(vin)));
120                     break;
121                 case ActivationLayerInfo::ActivationFunction::ELU:
122                     tmp = wrapper::vbsl(wrapper::vcge(vin, const_0), vin, wrapper::vmul(va, wrapper::vsub(wrapper::vexpq(vin), const_1)));
123                     break;
124                 case ActivationLayerInfo::ActivationFunction::SQRT:
125 #ifdef __aarch64__
126                     tmp = wrapper::vsqrt(vin);
127 #else  /* aarch64 */
128                     {
129                         const auto bitmask = wrapper::vceq(vin, wrapper::vdup_n(0.f, ExactTagType{}));
130                         tmp                 = wrapper::vinv(wrapper::vinvsqrt(wrapper::vadd(vin, mask_float_vector(delta, bitmask))));
131                         tmp                 = mask_float_vector(tmp, wrapper::vnot(bitmask));
132                     }
133 #endif /* aarch64 */
134                     break;
135                 case ActivationLayerInfo::ActivationFunction::SQUARE:
136                     tmp = wrapper::vmul(vin, vin);
137                     break;
138                 case ActivationLayerInfo::ActivationFunction::TANH:
139                     tmp = wrapper::vmul(va, wrapper::vtanh(wrapper::vmul(vb, vin)));
140                     break;
141                 case ActivationLayerInfo::ActivationFunction::IDENTITY:
142                     tmp = vin;
143                     break;
144                 case ActivationLayerInfo::ActivationFunction::HARD_SWISH:
145                     tmp = wrapper::vmul(vin, wrapper::vmul(const_inv_6, wrapper::vmin(const_6, wrapper::vmax(const_0, wrapper::vadd(vin, const_3)))));
146                     break;
147                 default:
148                     ARM_COMPUTE_ERROR("Unsupported activation function");
149             }
150             wrapper::vstore(output_ptr + x, tmp);
151         }
152 
153         // Compute left-over elements
154         for(; x < window_end_x; ++x)
155         {
156             const float in = *(reinterpret_cast<const float *>(input_ptr + x));
157             float       tmp;
158             switch(act)
159             {
160                 case ActivationLayerInfo::ActivationFunction::ABS:
161                     tmp = std::abs(in);
162                     break;
163                 case ActivationLayerInfo::ActivationFunction::LINEAR:
164                     tmp = a * in + b;
165                     break;
166                 case ActivationLayerInfo::ActivationFunction::LOGISTIC:
167                     tmp = static_cast<float>(1) / (static_cast<float>(1) + std::exp(-in));
168                     break;
169                 case ActivationLayerInfo::ActivationFunction::RELU:
170                     tmp = std::max<float>(static_cast<float>(0), in);
171                     break;
172                 case ActivationLayerInfo::ActivationFunction::BOUNDED_RELU:
173                     tmp = std::min<float>(a, std::max(static_cast<float>(0), in));
174                     break;
175                 case ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU:
176                     tmp = std::min<float>(a, std::max<float>(b, in));
177                     break;
178                 case ActivationLayerInfo::ActivationFunction::LEAKY_RELU:
179                     tmp = (in > 0) ? in : a * in;
180                     break;
181                 case ActivationLayerInfo::ActivationFunction::SOFT_RELU:
182                     tmp = std::log(static_cast<float>(1) + std::exp(in));
183                     break;
184                 case ActivationLayerInfo::ActivationFunction::ELU:
185                     tmp = (in >= 0) ? in : a * (std::exp(in) - 1);
186                     break;
187                 case ActivationLayerInfo::ActivationFunction::SQRT:
188                     tmp = std::sqrt(in);
189                     break;
190                 case ActivationLayerInfo::ActivationFunction::SQUARE:
191                     tmp = in * in;
192                     break;
193                 case ActivationLayerInfo::ActivationFunction::TANH:
194                     tmp = a * std::tanh(b * in);
195                     break;
196                 case ActivationLayerInfo::ActivationFunction::IDENTITY:
197                     tmp = in;
198                     break;
199                 case ActivationLayerInfo::ActivationFunction::HARD_SWISH:
200                     tmp = in * ((std::min(std::max((in + 3), 0.0f), 6.0f)) * 0.166666667f);
201                     break;
202                 default:
203                     ARM_COMPUTE_ERROR("Unsupported activation function");
204             }
205             *(output_ptr + x) = tmp;
206         }
207     },
208     input, output);
209 }
210 } // namespace cpu
211 } // namespace arm_compute
212