xref: /aosp_15_r20/external/ComputeLibrary/src/core/NEON/kernels/batchnormalization/impl/SVE/fp16.cpp (revision c217d954acce2dbc11938adb493fc0abd69584f3)
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/SVEMath.h"
28 
29 #include <cmath>
30 #include <cstddef>
31 
32 #if defined(ARM_COMPUTE_ENABLE_SVE)
33 #include <arm_sve.h>
34 
35 namespace arm_compute
36 {
37 namespace cpu
38 {
fp16_sve_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)39 void fp16_sve_batch_normalization(ITensor *src, ITensor *dst, const ITensor *mean, const ITensor *var, const ITensor *beta, const ITensor *gamma,
40                                   float epsilon, ActivationLayerInfo &act_info, const Window &window)
41 {
42     const auto window_start_x = static_cast<int>(window.x().start());
43     const auto window_end_x   = static_cast<int>(window.x().end());
44 
45     Window win_collapsed = window.collapse_if_possible(window, Window::DimZ);
46     win_collapsed.set(Window::DimX, Window::Dimension(0, 1, 1));
47 
48     Iterator input(src, win_collapsed);
49     Iterator output(dst, win_collapsed);
50 
51     const auto input_mean  = reinterpret_cast<const float16_t *>(mean->ptr_to_element(Coordinates(0, 0)));
52     const auto input_var   = reinterpret_cast<const float16_t *>(var->ptr_to_element(Coordinates(0, 0)));
53     const auto input_gamma = (gamma != nullptr) ? reinterpret_cast<const float16_t *>(gamma->ptr_to_element(Coordinates(0, 0))) : nullptr;
54     const auto input_beta  = (beta != nullptr) ? reinterpret_cast<const float16_t *>(beta->ptr_to_element(Coordinates(0, 0))) : nullptr;
55 
56     const auto epsilon_vec = svdup_n_f16(epsilon);
57     const auto const_1     = svdup_n_f16(1.f);
58     const auto const_0     = svdup_n_f16(0.f);
59     const auto va          = svdup_n_f16(act_info.a());
60     const auto vb          = svdup_n_f16(act_info.b());
61     execute_window_loop(win_collapsed, [&](const Coordinates &)
62     {
63         const auto input_ptr  = reinterpret_cast<const float16_t *>(input.ptr());
64         const auto output_ptr = reinterpret_cast<float16_t *>(output.ptr());
65 
66         // Compute S elements per iteration
67         int      x  = window_start_x;
68         svbool_t pg = svwhilelt_b16(x, window_end_x);
69         do
70         {
71             // Conctruct vectors
72             const auto mean_vec  = svld1_f16(pg, input_mean + x);
73             const auto var_vec   = svld1_f16(pg, input_var + x);
74             const auto gamma_vec = (input_gamma != nullptr) ? svld1_f16(pg, input_gamma + x) : const_1;
75             const auto beta_vec  = (input_beta != nullptr) ? svld1_f16(pg, input_beta + x) : const_0;
76 
77             // Calculate denominator
78             const auto tmp         = svadd_f16_z(pg, var_vec, epsilon_vec);
79             auto       denominator = svrsqrte_f16(tmp);
80             denominator            = svmul_f16_z(pg, svrsqrts_f16(svmul_f16_z(pg, tmp, denominator), denominator), denominator);
81             denominator            = svmul_f16_z(pg, svrsqrts_f16(svmul_f16_z(pg, tmp, denominator), denominator), denominator);
82 
83             // Calculate x bar
84             const auto numerator = svsub_f16_z(pg, svld1_f16(pg, input_ptr + x), mean_vec);
85             const auto x_bar     = svmul_f16_z(pg, numerator, denominator);
86             auto       res       = svmla_f16_z(pg, beta_vec, x_bar, gamma_vec);
87 
88             // Perform fused activation
89             if(act_info.enabled())
90             {
91                 if(act_info.activation() == ActivationLayerInfo::ActivationFunction::RELU)
92                 {
93                     res = svmax_f16_z(pg, const_0, res);
94                 }
95                 else if(act_info.activation() == ActivationLayerInfo::ActivationFunction::BOUNDED_RELU)
96                 {
97                     res = svmin_f16_z(pg, va, svmax_f16_z(pg, const_0, res));
98                 }
99                 else if(act_info.activation() == ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU)
100                 {
101                     res = svmin_f16_z(pg, va, svmax_f16_z(pg, vb, res));
102                 }
103             }
104 
105             // Store results
106             svst1_f16(pg, output_ptr + x, res);
107 
108             x += svcntw();
109             pg = svwhilelt_b16(x, window_end_x);
110         }
111         while(svptest_any(svptrue_b16(), pg));
112     },
113     input, output);
114 }
115 } // namespace cpu
116 } // namespace arm_compute
117 #endif // ENABLE_SVE
118