xref: /aosp_15_r20/external/ComputeLibrary/src/cpu/kernels/activation/generic/sve2/qsymm16.cpp (revision c217d954acce2dbc11938adb493fc0abd69584f3)
1 /*
2  * Copyright (c) 2020-2022 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 "arm_compute/core/Helpers.h"
25 #include "arm_compute/core/ITensorPack.h"
26 #include "arm_compute/core/Window.h"
27 #include "arm_compute/core/experimental/Types.h"
28 
29 #include <cmath>
30 #include <cstddef>
31 
32 #include "src/core/NEON/SVEMath.h"
33 #include "src/core/NEON/SVESymm.h"
34 #include <arm_sve.h>
35 
36 namespace arm_compute
37 {
38 namespace cpu
39 {
sve2_qsymm16_activation(const ITensor * src,ITensor * dst,const ActivationLayerInfo & act_info,const Window & window)40 void sve2_qsymm16_activation(const ITensor *src, ITensor *dst, const 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     const ActivationLayerInfo::ActivationFunction act            = act_info.activation();
45 
46     Window win_collapsed = window.collapse_if_possible(window, Window::DimZ);
47     win_collapsed.set(Window::DimX, Window::Dimension(0, 1, 1));
48 
49     Iterator input(src, win_collapsed);
50     Iterator output(dst, win_collapsed);
51 
52     const UniformQuantizationInfo qi_in    = src->info()->quantization_info().uniform();
53     const UniformQuantizationInfo qi_out   = dst->info()->quantization_info().uniform();
54     const auto                    vconst_1 = svdup_n_f32(1.f);
55     const auto                    va_f32   = svdup_n_f32(act_info.a());
56     const auto                    vb_f32   = svdup_n_f32(act_info.b());
57 
58     execute_window_loop(win_collapsed, [&](const Coordinates &)
59     {
60         const auto input_ptr  = reinterpret_cast<const int16_t *>(input.ptr());
61         const auto output_ptr = reinterpret_cast<int16_t *>(output.ptr());
62 
63         svint16_t tmp;
64 
65         int      x  = window_start_x;
66         svbool_t pg = svwhilelt_b16(x, window_end_x);
67         do
68         {
69             const auto vin = svld1_s16(pg, input_ptr + x);
70             if(act == ActivationLayerInfo::ActivationFunction::LOGISTIC)
71             {
72                 // De-quantize
73                 auto vin_deq = svdequantize_qsymm16_z(pg, vin, qi_in.scale);
74                 // Perform activation
75                 const svfloat32x2_t tmp_dep = svcreate2_f32(svdiv_f32_z(pg, vconst_1, svadd_f32_z(pg, vconst_1, svexp_f32_z(pg, svneg_f32_z(pg, svget2_f32(vin_deq, 0))))),
76                                                             svdiv_f32_z(pg, vconst_1, svadd_f32_z(pg, vconst_1, svexp_f32_z(pg, svneg_f32_z(pg, svget2_f32(vin_deq, 1))))));
77                 // Re-quantize to new output space
78                 tmp = svquantize_qsymm16_z(pg, tmp_dep, qi_out.scale);
79             }
80             else if(act == ActivationLayerInfo::ActivationFunction::TANH)
81             {
82                 // De-quantize
83                 auto vin_deq = svdequantize_qsymm16_z(pg, vin, qi_in.scale);
84                 // Perform activation
85                 const svfloat32x2_t tmp_dep = svcreate2_f32(svmul_f32_z(pg, va_f32, svtanh_f32_z(pg, svmul_f32_z(pg, svget2_f32(vin_deq, 0), vb_f32))),
86                                                             svmul_f32_z(pg, va_f32, svtanh_f32_z(pg, svmul_f32_z(pg, svget2_f32(vin_deq, 1), vb_f32))));
87                 // Re-quantize to new output space
88                 tmp = svquantize_qsymm16_z(pg, tmp_dep, qi_out.scale);
89             }
90             else if(act == ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU)
91             {
92                 // De-quantize
93                 auto vin_deq = svdequantize_qsymm16_z(pg, vin, qi_in.scale);
94                 // Perform activation
95                 const svfloat32x2_t tmp_dep = svcreate2_f32(svmin_f32_z(pg, va_f32, svmax_f32_z(pg, vb_f32, svget2_f32(vin_deq, 0))),
96                                                             svmin_f32_z(pg, va_f32, svmax_f32_z(pg, vb_f32, svget2_f32(vin_deq, 1))));
97                 // Re-quantize to new output space
98                 tmp = svquantize_qsymm16_z(pg, tmp_dep, qi_out.scale);
99             }
100             else
101             {
102                 ARM_COMPUTE_ERROR("Unsupported activation function");
103             }
104 
105             svst1_s16(pg, output_ptr + x, tmp);
106 
107             x += svcnth();
108             pg = svwhilelt_b16(x, window_end_x);
109 
110         }
111         while(svptest_any(svptrue_b16(), pg));
112     },
113     input, output);
114 }
115 } // namespace cpu
116 } // namespace arm_compute
117