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/generic/impl.h"
25
26 namespace arm_compute
27 {
28 namespace cpu
29 {
30 template <typename T>
fused_batch_normalization_conv(const ITensor * conv_weights,const ITensor * conv_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_conv(const ITensor *conv_weights, const ITensor *conv_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 / conv_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 == conv_weights);
39 const bool run_in_place_bias = (fused_bias == nullptr) || (conv_bias != nullptr && fused_bias == conv_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 conv_w_in(conv_weights, win);
50 Iterator conv_w_out(run_in_place_weights ? conv_weights : fused_weights, win);
51
52 const auto conv_bias_in = (conv_bias != nullptr ? reinterpret_cast<ScalarType *>(conv_bias->ptr_to_element(Coordinates(0, 0))) : nullptr);
53 auto conv_bias_out = (run_in_place_bias ? conv_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 const auto epsilon_vec = wrapper::vdup_n(ScalarType(epsilon), ExactTagType{});
66
67 auto mean = ScalarType(0.0);
68 auto var = ScalarType(0.0);
69 auto gamma = ScalarType(1.0);
70 auto beta = ScalarType(0.0);
71 auto conv_bias_in_scalar = ScalarType(0.0);
72 execute_window_loop(win, [&](const Coordinates & id)
73 {
74 var = input_var[id[3]];
75 if(input_gamma != nullptr)
76 {
77 gamma = input_gamma[id[3]];
78 }
79
80 if((id[0] == 0) && (id[1] == 0) && (id[2] == 0))
81 {
82 if(input_beta != nullptr)
83 {
84 beta = input_beta[id[3]];
85 beta_vec = wrapper::vdup_n(beta, ExactTagType{});
86 }
87
88 // Construct vectors
89 mean = input_mean[id[3]];
90 mean_vec = wrapper::vdup_n(mean, ExactTagType{});
91
92 if(conv_bias_in != nullptr)
93 {
94 conv_bias_in_scalar = conv_bias_in[id[3]];
95 }
96 auto conv_bias_tmp_scalar = (conv_bias_in_scalar - mean) / std::sqrt(var + ScalarType(epsilon));
97 conv_bias_out[id[3]] = (conv_bias_tmp_scalar * gamma) + beta;
98 }
99
100 int x = window_start_x;
101 auto conv_w_in_ptr = reinterpret_cast<const ScalarType *>(conv_w_in.ptr());
102 auto conv_w_out_ptr = reinterpret_cast<ScalarType *>(conv_w_out.ptr());
103 var_vec = wrapper::vdup_n(var, ExactTagType{});
104 gamma_vec = wrapper::vdup_n(gamma, ExactTagType{});
105 rvar_vec = wrapper::vinvsqrt(wrapper::vadd(var_vec, epsilon_vec));
106
107 for(; x <= (window_end_x - window_step_x); x += window_step_x)
108 {
109 auto wn = wrapper::vloadq(conv_w_in_ptr + x);
110 wn = wrapper::vmul(wn, rvar_vec);
111 wn = wrapper::vmul(wn, gamma_vec);
112
113 // Store results
114 wrapper::vstore(conv_w_out_ptr + x, wn);
115 }
116
117 // Compute left-over elements
118 for(; x < window_end_x; ++x)
119 {
120 *(conv_w_out_ptr + x) = *(conv_w_in_ptr + x) / std::sqrt(var + ScalarType(epsilon)) * gamma;
121 }
122 },
123 conv_w_in, conv_w_out);
124 }
125
126 template void fused_batch_normalization_conv<float32_t>(const ITensor *conv_weights, const ITensor *conv_bias, ITensor *fused_weights, ITensor *fused_bias,
127 const ITensor *bn_mean, const ITensor *bn_var, const ITensor *bn_beta, const ITensor *bn_gamma, float epsilon, const Window &window);
128
129 #if defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && defined(ENABLE_FP16_KERNELS)
130 template void fused_batch_normalization_conv<float16_t>(const ITensor *conv_weights, const ITensor *conv_bias, ITensor *fused_weights, ITensor *fused_bias,
131 const ITensor *bn_mean, const ITensor *bn_var, const ITensor *bn_beta, const ITensor *bn_gamma, float epsilon, const Window &window);
132 #endif /* defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && defined(ENABLE_FP16_KERNELS) */
133
134 } // namespace cpu
135 } // namespace arm_compute
136