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 */
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/kernels/detail/NEActivationFunctionDetail.h"
29 #include "src/core/NEON/wrapper/wrapper.h"
30
31 #include <arm_neon.h>
32 #include <cmath>
33 #include <cstddef>
34
35 #if defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && defined(ENABLE_FP16_KERNELS)
36 namespace arm_compute
37 {
38 namespace
39 {
40 using BatchNomalizationPtr = void (*)(ITensor *src, ITensor *dst, const ITensor *mean, const ITensor *var, const ITensor *beta, const ITensor *gamma,
41 float epsilon, ActivationLayerInfo &act_info, const Window &window);
42
43 template <typename T>
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)44 void batch_normalization(ITensor *src, ITensor *dst, const ITensor *mean, const ITensor *var, const ITensor *beta, const ITensor *gamma,
45 float epsilon, ActivationLayerInfo &act_info, const Window &window)
46 {
47 /** SIMD vector tag type. */
48 using ExactTagType = typename wrapper::traits::neon_bitvector_tag_t<float16_t, wrapper::traits::BitWidth::W128>;
49
50 const int window_step_x = 8;
51 const auto window_start_x = static_cast<int>(window.x().start());
52 const auto window_end_x = static_cast<int>(window.x().end());
53
54 Window win_collapsed = window.collapse_if_possible(window, Window::DimZ);
55 win_collapsed.set(Window::DimX, Window::Dimension(0, 1, 1));
56
57 Iterator input(src, win_collapsed);
58 Iterator output(dst, win_collapsed);
59
60 const auto input_mean = reinterpret_cast<const float16_t *>(mean->ptr_to_element(Coordinates(0, 0)));
61 const auto input_var = reinterpret_cast<const float16_t *>(var->ptr_to_element(Coordinates(0, 0)));
62 const auto input_gamma = (gamma != nullptr) ? reinterpret_cast<const float16_t *>(gamma->ptr_to_element(Coordinates(0, 0))) : nullptr;
63 const auto input_beta = (beta != nullptr) ? reinterpret_cast<const float16_t *>(beta->ptr_to_element(Coordinates(0, 0))) : nullptr;
64
65 T activation_functor(act_info);
66
67 const auto epsilon_vec = wrapper::vdup_n(static_cast<float16_t>(epsilon), ExactTagType{});
68 execute_window_loop(win_collapsed, [&](const Coordinates &)
69 {
70 const auto input_ptr = reinterpret_cast<const float16_t *>(input.ptr());
71 const auto output_ptr = reinterpret_cast<float16_t *>(output.ptr());
72
73 // Perform core calculations using vector operations
74 int x = window_start_x;
75 for(; x <= (window_end_x - window_step_x); x += window_step_x)
76 {
77 // Conctruct vectors
78 const auto mean_vec = wrapper::vloadq(input_mean + x);
79 const auto var_vec = wrapper::vloadq(input_var + x);
80 const auto gamma_vec = (input_gamma != nullptr) ? wrapper::vloadq(input_gamma + x) : wrapper::vdup_n(static_cast<float16_t>(1.f), ExactTagType{});
81 const auto beta_vec = (input_beta != nullptr) ? wrapper::vloadq(input_beta + x) : wrapper::vdup_n(static_cast<float16_t>(0.f), ExactTagType{});
82
83 // Calculate denominator
84 const auto denominator = wrapper::vinvsqrt(wrapper::vadd(var_vec, epsilon_vec));
85
86 // Calculate x bar
87 const auto numerator = wrapper::vsub(wrapper::vloadq(input_ptr + x), mean_vec);
88 const auto x_bar = wrapper::vmul(numerator, denominator);
89 auto res = wrapper::vmla(beta_vec, x_bar, gamma_vec);
90
91 // Perform fused activation
92 if(act_info.enabled())
93 {
94 activation_functor(res);
95 }
96
97 // Store results
98 wrapper::vstore(output_ptr + x, res);
99 }
100
101 // Compute left-over elements
102 for(; x < window_end_x; ++x)
103 {
104 // Conctruct vectors
105 const float16_t gamma = (input_gamma != nullptr) ? input_gamma[x] : 1.f;
106 const float16_t beta = (input_beta != nullptr) ? input_beta[x] : 0.f;
107
108 const float16_t denominator = sqrt(input_var[x] + epsilon);
109 const float16_t numerator = input_ptr[x] - input_mean[x];
110 const float16_t x_bar = numerator / denominator;
111 float16_t res = beta + x_bar * gamma;
112
113 // Perform fused activation
114 if(act_info.enabled())
115 {
116 activation_functor(res);
117 }
118
119 // Store results
120 *reinterpret_cast<float16_t *>(output_ptr + x) = res;
121 }
122 },
123 input, output);
124 }
125
126 // Fused Batched Normalization with activation functions
127 static std::map<ActivationLayerInfo::ActivationFunction, BatchNomalizationPtr> fused_map =
128 {
129 { ActivationLayerInfo::ActivationFunction::RELU, &batch_normalization<detail::relu<float16_t, 8>> },
130 { ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, &batch_normalization<detail::brelu<float16_t, 8>> },
131 { ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, &batch_normalization<detail::lubrelu<float16_t, 8>> }
132 };
133 }
134 namespace cpu
135 {
fp16_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)136 void fp16_neon_batch_normalization(ITensor *src, ITensor *dst, const ITensor *mean, const ITensor *var, const ITensor *beta, const ITensor *gamma,
137 float epsilon, ActivationLayerInfo &act_info, const Window &window)
138 {
139 if(act_info.enabled())
140 {
141 fused_map[act_info.activation()](src, dst, mean, var, beta, gamma, epsilon, act_info, window);
142 }
143 else
144 {
145 batch_normalization<detail::dummy<float16_t, 8>>(src, dst, mean, var, beta, gamma, epsilon, act_info, window);
146 }
147 }
148 } // namespace cpu
149 } // namespace arm_compute
150
151 #endif /* defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && defined(ENABLE_FP16_KERNELS) */
152