xref: /aosp_15_r20/external/ComputeLibrary/src/cpu/kernels/select/generic/neon/impl.cpp (revision c217d954acce2dbc11938adb493fc0abd69584f3)
1 /*
2  * Copyright (c) 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 
25 #include "arm_compute/core/TensorInfo.h"
26 #include "src/cpu/kernels/select/generic/neon/impl.h"
27 #include "src/core/NEON/NEAsymm.h"
28 
29 #include <arm_neon.h>
30 #include <map>
31 #include <string>
32 
33 namespace arm_compute
34 {
35 namespace cpu
36 {
37 template <typename ScalarType, typename VectorType>
select_op(const ITensor * cond,const ITensor * in1,const ITensor * in2,ITensor * out,const Window & window,const int window_step_x,const int window_start_x,const int window_end_x,const int limit,VectorType (* condition_conversion)(const uint8_t *))38 void select_op(const ITensor *cond, const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window,
39                const int window_step_x, const int window_start_x, const int window_end_x, const int limit, VectorType (*condition_conversion)(const uint8_t *))
40 {
41     Window win = window;
42     win.set(Window::DimX, Window::Dimension(0, 1, 1));
43 
44     Iterator condition(cond, win);
45     Iterator input1(in1, win);
46     Iterator input2(in2, win);
47     Iterator output(out, win);
48 
49     execute_window_loop(win, [&](const Coordinates &)
50     {
51         auto       output_ptr    = reinterpret_cast<ScalarType *>(output.ptr());
52         const auto condition_ptr = reinterpret_cast<const uint8_t *>(condition.ptr());
53         const auto input1_ptr    = reinterpret_cast<const ScalarType *>(input1.ptr());
54         const auto input2_ptr    = reinterpret_cast<const ScalarType *>(input2.ptr());
55 
56         int x = window_start_x;
57         for(; x <= limit; x += window_step_x)
58         {
59             const auto c = (*condition_conversion)(condition_ptr + x);
60             const auto a = wrapper::vloadq(input1_ptr + x);
61             const auto b = wrapper::vloadq(input2_ptr + x);
62             wrapper::vstore(output_ptr + x, wrapper::vbsl(c, a, b));
63         }
64         for(; x < window_end_x; ++x)
65         {
66             const auto c      = *(condition_ptr + x);
67             const auto a      = *(input1_ptr + x);
68             const auto b      = *(input2_ptr + x);
69             *(output_ptr + x) = static_cast<bool>(c) ? a : b;
70         }
71     },
72     condition, input1, input2, output);
73 }
74 
75 template <typename ScalarType, typename VectorType>
select_op_8(const ITensor * cond,const ITensor * in1,const ITensor * in2,ITensor * out,const Window & window)76 void select_op_8(const ITensor *cond, const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window)
77 {
78     const auto window_step_x  = 16 / sizeof(ScalarType);
79     const auto window_start_x = static_cast<int>(window.x().start());
80     const auto window_end_x   = static_cast<int>(window.x().end());
81 
82     select_op<ScalarType, VectorType>(cond, in1, in2, out, window, window_step_x, window_start_x, window_end_x, window_end_x - window_step_x, [](const uint8_t *condition_ptr) -> VectorType
83     {
84         static const auto zero = wrapper::vdup_n(static_cast<uint8_t>(0), arm_compute::wrapper::traits::vector_128_tag());
85         return wrapper::vcgt(wrapper::vloadq(condition_ptr), zero);
86     });
87 }
88 
89 template <typename ScalarType, typename VectorType>
select_op_16(const ITensor * cond,const ITensor * in1,const ITensor * in2,ITensor * out,const Window & window)90 void select_op_16(const ITensor *cond, const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window)
91 {
92     const auto window_step_x  = 16 / sizeof(ScalarType);
93     const auto window_start_x = static_cast<int>(window.x().start());
94     const auto window_end_x   = static_cast<int>(window.x().end());
95 
96     select_op<ScalarType, VectorType>(cond, in1, in2, out, window, window_step_x, window_start_x, window_end_x, window_end_x - window_step_x, [](const uint8_t *condition_ptr) -> VectorType
97     {
98         static const auto zero = wrapper::vdup_n(static_cast<uint16_t>(0), arm_compute::wrapper::traits::vector_128_tag());
99         return wrapper::vcgt(wrapper::vmovl(wrapper::vload(condition_ptr)), zero);
100     });
101 }
102 
103 template <typename ScalarType, typename VectorType>
select_op_32(const ITensor * cond,const ITensor * in1,const ITensor * in2,ITensor * out,const Window & window)104 void select_op_32(const ITensor *cond, const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window)
105 {
106     const auto window_step_x  = 16 / sizeof(ScalarType);
107     const auto window_start_x = static_cast<int>(window.x().start());
108     const auto window_end_x   = static_cast<int>(window.x().end());
109 
110     select_op<ScalarType, VectorType>(cond, in1, in2, out, window, window_step_x, window_start_x, window_end_x, window_end_x - window_step_x, [](const uint8_t *condition_ptr) -> VectorType
111     {
112         static const auto zero = wrapper::vdup_n(static_cast<uint32_t>(0), arm_compute::wrapper::traits::vector_128_tag());
113         return wrapper::vcgt(wrapper::vmovl(wrapper::vgetlow(wrapper::vmovl(wrapper::vload(condition_ptr)))), zero);
114     });
115 }
116 
117 template <typename ScalarType>
select_op_not_same_rank(const ITensor * cond,const ITensor * in1,const ITensor * in2,ITensor * out,const Window & window)118 void select_op_not_same_rank(const ITensor *cond, const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window)
119 {
120     ARM_COMPUTE_UNUSED(window);
121 
122     auto       output_ptr    = reinterpret_cast<ScalarType *>(out->buffer());
123     const auto condition_ptr = reinterpret_cast<const uint8_t *>(cond->buffer());
124     const auto input1_ptr    = reinterpret_cast<const ScalarType *>(in1->buffer());
125     const auto input2_ptr    = reinterpret_cast<const ScalarType *>(in2->buffer());
126 
127     const int outer_size = cond->info()->total_size() / cond->info()->element_size();
128     const int inner_size = (in1->info()->total_size() / in1->info()->element_size()) / outer_size;
129     int       offset     = 0;
130     const int step       = 16 / in1->info()->element_size();
131 
132     for(int i = 0; i < outer_size; ++i)
133     {
134         int        x         = offset;
135         const auto input_ptr = static_cast<bool>(*(condition_ptr + i)) ? input1_ptr : input2_ptr;
136         for(; x <= offset + inner_size - step; x += step)
137         {
138             wrapper::vstore(output_ptr + x, wrapper::vloadq(input_ptr + x));
139         }
140         if(x <= offset + inner_size - (step / 2))
141         {
142             wrapper::vstore(output_ptr + x, wrapper::vload(input_ptr + x));
143             x += step / 2;
144         }
145         for(; x < offset + inner_size; ++x)
146         {
147             *(output_ptr + x) = *(input_ptr + x);
148         }
149         offset += inner_size;
150     }
151 }
152 
153 template void select_op_32<float, uint32x4_t>(const ITensor *cond, const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window);
154 
155 template void select_op_not_same_rank<float>(const ITensor *cond, const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window);
156 
157 template void select_op_8<int8_t, uint8x16_t>(const ITensor *cond, const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window);
158 
159 template void select_op_16<int16_t, uint16x8_t>(const ITensor *cond, const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window);
160 
161 #if defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && defined(ENABLE_FP16_KERNELS)
162 template void select_op_16<float16_t, uint16x8_t>(const ITensor *cond, const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window);
163 #endif /* (__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && defined(ENABLE_FP16_KERNELS) */
164 
165 template void select_op_32<int32_t, uint32x4_t>(const ITensor *cond, const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window);
166 
167 template void select_op_not_same_rank<int8_t>(const ITensor *cond, const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window);
168 
169 template void select_op_not_same_rank<int16_t>(const ITensor *cond, const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window);
170 
171 #if defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && defined(ENABLE_FP16_KERNELS)
172 template void select_op_not_same_rank<float16_t>(const ITensor *cond, const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window);
173 #endif /* (__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && defined(ENABLE_FP16_KERNELS) */
174 
175 template void select_op_not_same_rank<int32_t>(const ITensor *cond, const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window);
176 
177 template void select_op_8<uint8_t, uint8x16_t>(const ITensor *cond, const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window);
178 
179 template void select_op_16<uint16_t, uint16x8_t>(const ITensor *cond, const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window);
180 
181 template void select_op_32<uint32_t, uint32x4_t>(const ITensor *cond, const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window);
182 
183 template void select_op_not_same_rank<uint8_t>(const ITensor *cond, const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window);
184 
185 template void select_op_not_same_rank<uint16_t>(const ITensor *cond, const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window);
186 
187 template void select_op_not_same_rank<uint32_t>(const ITensor *cond, const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window);
188 
189 } // namespace cpu
190 
191 } // namespace arm_compute
192