1 /*
2 * Copyright (c) 2018-2022 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 "src/cpu/kernels/fuse_batch_normalization/nhwc/neon/impl.h"
25
26 namespace arm_compute
27 {
28 namespace cpu
29 {
30 template <typename T>
fused_batch_normalization_dwc_nhwc(const ITensor * dwc_weights,const ITensor * dwc_bias,ITensor * fused_weights,ITensor * fused_bias,const ITensor * bn_mean,const ITensor * bn_var,const ITensor * bn_beta,const ITensor * bn_gamma,float epsilon,const Window & window)31 void fused_batch_normalization_dwc_nhwc(const ITensor *dwc_weights, const ITensor *dwc_bias, ITensor *fused_weights, ITensor *fused_bias,
32 const ITensor *bn_mean, const ITensor *bn_var, const ITensor *bn_beta, const ITensor *bn_gamma, float epsilon, const Window &window)
33 {
34 using ScalarType = T;
35 const int size = 16 / dwc_weights->info()->element_size();
36 using ExactTagType = typename wrapper::traits::neon_bitvector_tag_t<T, wrapper::traits::BitWidth::W128>;
37
38 const bool run_in_place_weights = (fused_weights == nullptr) || (fused_weights == dwc_weights);
39 const bool run_in_place_bias = (fused_bias == nullptr) || (dwc_bias != nullptr && fused_bias == dwc_bias);
40
41 // Set build options
42 Window win = window;
43 win.set(Window::DimX, Window::Dimension(0, 1, 1));
44
45 const int window_step_x = size;
46 const auto window_start_x = static_cast<int>(window.x().start());
47 const auto window_end_x = static_cast<int>(window.x().end());
48
49 Iterator dwc_w_in(dwc_weights, win);
50 Iterator dwc_w_out(run_in_place_weights ? dwc_weights : fused_weights, win);
51
52 const auto dwc_bias_in = (dwc_bias != nullptr ? reinterpret_cast<ScalarType *>(dwc_bias->ptr_to_element(Coordinates(0, 0))) : nullptr);
53 auto dwc_bias_out = (run_in_place_bias ? dwc_bias_in : reinterpret_cast<ScalarType *>(fused_bias->ptr_to_element(Coordinates(0, 0))));
54
55 const auto input_mean = reinterpret_cast<const ScalarType *>(bn_mean->ptr_to_element(Coordinates(0, 0)));
56 const auto input_var = reinterpret_cast<const ScalarType *>(bn_var->ptr_to_element(Coordinates(0, 0)));
57 const auto input_gamma = (bn_gamma != nullptr) ? reinterpret_cast<const ScalarType *>(bn_gamma->ptr_to_element(Coordinates(0, 0))) : nullptr;
58 const auto input_beta = (bn_beta != nullptr) ? reinterpret_cast<const ScalarType *>(bn_beta->ptr_to_element(Coordinates(0, 0))) : nullptr;
59
60 auto mean_vec = wrapper::vdup_n(ScalarType(0), ExactTagType{});
61 auto var_vec = wrapper::vdup_n(ScalarType(0), ExactTagType{});
62 auto gamma_vec = wrapper::vdup_n(ScalarType(1), ExactTagType{});
63 auto beta_vec = wrapper::vdup_n(ScalarType(0), ExactTagType{});
64 auto rvar_vec = wrapper::vdup_n(ScalarType(0), ExactTagType{});
65 auto dwc_bias_vec = wrapper::vdup_n(ScalarType(0), ExactTagType{});
66 const auto epsilon_vec = wrapper::vdup_n(ScalarType(epsilon), ExactTagType{});
67
68 auto gamma = ScalarType(1.0);
69 auto beta = ScalarType(0.0);
70 auto dwc_bias_in_scalar = ScalarType(0);
71
72 execute_window_loop(win, [&](const Coordinates & id)
73 {
74 int x = window_start_x;
75 for(; x <= (window_end_x - window_step_x); x += window_step_x)
76 {
77 var_vec = wrapper::vloadq(input_var + x);
78 if(input_gamma != nullptr)
79 {
80 gamma_vec = wrapper::vloadq(input_gamma + x);
81 }
82
83 if((id[2] == 0) && (id[1] == 0))
84 {
85 mean_vec = wrapper::vloadq(input_mean + x);
86
87 // Construct vectors
88 if(input_beta != nullptr)
89 {
90 beta_vec = wrapper::vloadq(input_beta + x);
91 }
92
93 if(dwc_bias_in != nullptr)
94 {
95 dwc_bias_vec = wrapper::vloadq(dwc_bias_in + x);
96 }
97
98 auto dwc_bias_tmp_vec = wrapper::vmul(wrapper::vsub(dwc_bias_vec, mean_vec), wrapper::vinvsqrt(wrapper::vadd(var_vec, epsilon_vec)));
99 dwc_bias_tmp_vec = wrapper::vadd(wrapper::vmul(dwc_bias_tmp_vec, gamma_vec), beta_vec);
100 wrapper::vstore(dwc_bias_out + x, dwc_bias_tmp_vec);
101 }
102
103 auto dwc_w_in_ptr = reinterpret_cast<const ScalarType *>(dwc_w_in.ptr());
104 auto dwc_w_out_ptr = reinterpret_cast<ScalarType *>(dwc_w_out.ptr());
105
106 auto wn = wrapper::vloadq(dwc_w_in_ptr + x);
107 rvar_vec = wrapper::vinvsqrt(wrapper::vadd(var_vec, epsilon_vec));
108 wn = wrapper::vmul(wn, rvar_vec);
109 wn = wrapper::vmul(wn, gamma_vec);
110
111 // Store results
112 wrapper::vstore(dwc_w_out_ptr + x, wn);
113 }
114
115 // Compute left-over elements
116 for(; x < window_end_x; ++x)
117 {
118 auto var = input_var[x];
119 if(input_gamma != nullptr)
120 {
121 gamma = input_gamma[x];
122 }
123
124 if(id[2] == 0 && id[1] == 0)
125 {
126 auto mean = input_mean[x];
127 if(input_beta != nullptr)
128 {
129 beta = input_beta[x];
130 }
131 if(dwc_bias_in != nullptr)
132 {
133 dwc_bias_in_scalar = dwc_bias_in[x];
134 }
135
136 auto dwc_bias_tmp_scalar = (dwc_bias_in_scalar - mean) / std::sqrt(var + ScalarType(epsilon));
137 dwc_bias_out[x] = (dwc_bias_tmp_scalar * gamma) + beta;
138 }
139
140 const auto dwc_w_in_ptr = reinterpret_cast<const ScalarType *>(dwc_w_in.ptr());
141 auto dwc_w_out_ptr = reinterpret_cast<ScalarType *>(dwc_w_out.ptr());
142
143 *(dwc_w_out_ptr + x) = *(dwc_w_in_ptr + x) / std::sqrt(var + ScalarType(epsilon)) * gamma;
144 }
145 },
146 dwc_w_in, dwc_w_out);
147 }
148
149 template void fused_batch_normalization_dwc_nhwc<float32_t>(const ITensor *dwc_weights, const ITensor *dwc_bias, ITensor *fused_weights, ITensor *fused_bias,
150 const ITensor *bn_mean, const ITensor *bn_var, const ITensor *bn_beta, const ITensor *bn_gamma, float epsilon, const Window &window);
151
152 #if defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && defined(ENABLE_FP16_KERNELS)
153 template void fused_batch_normalization_dwc_nhwc<float16_t>(const ITensor *dwc_weights, const ITensor *dwc_bias, ITensor *fused_weights, ITensor *fused_bias,
154 const ITensor *bn_mean, const ITensor *bn_var, const ITensor *bn_beta, const ITensor *bn_gamma, float epsilon, const Window &window);
155 #endif /* defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && defined(ENABLE_FP16_KERNELS) */
156
157 } // namespace cpu
158 } // namespace arm_compute
159