1 /*
2 * Copyright (c) 2021-2022 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 #ifndef SRC_CORE_NEON_KERNELS_SUB_LIST_H
25 #define SRC_CORE_NEON_KERNELS_SUB_LIST_H
26
27 #include "arm_compute/core/Types.h"
28 #include "arm_compute/core/utils/misc/Traits.h"
29 #include "src/core/NEON/wrapper/wrapper.h"
30
31 namespace arm_compute
32 {
33 namespace cpu
34 {
35 #define DECLARE_SUB_KERNEL(func_name) \
36 void func_name(const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window)
37
38 DECLARE_SUB_KERNEL(sub_qasymm8_neon_fixedpoint);
39 DECLARE_SUB_KERNEL(sub_qasymm8_signed_neon_fixedpoint);
40 DECLARE_SUB_KERNEL(sub_qasymm8_neon);
41 DECLARE_SUB_KERNEL(sub_qasymm8_signed_neon);
42 DECLARE_SUB_KERNEL(sub_qsymm16_neon);
43
44 #undef DECLARE_SUB_KERNEL
45
46 template <typename T>
sub_same_neon(const ITensor * src0,const ITensor * src1,ITensor * dst,const ConvertPolicy & policy,const Window & window)47 void sub_same_neon(const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window)
48 {
49 /** SIMD vector tag type. */
50 using ExactTagType = typename wrapper::traits::neon_bitvector_tag_t<T, wrapper::traits::BitWidth::W128>;
51
52 bool is_sat = policy == ConvertPolicy::SATURATE;
53
54 // Create input windows
55 Window input1_win = window.broadcast_if_dimension_le_one(src0->info()->tensor_shape());
56 Window input2_win = window.broadcast_if_dimension_le_one(src1->info()->tensor_shape());
57
58 // Clear X Dimension on execution window as we handle manually
59 Window win = window;
60 win.set(Window::DimX, Window::Dimension(0, 1, 1));
61
62 constexpr int window_step_x = 16 / sizeof(T);
63 const auto window_start_x = static_cast<int>(window.x().start());
64 const auto window_end_x = static_cast<int>(window.x().end());
65 const bool is_broadcast_across_x = src0->info()->tensor_shape().x() != src1->info()->tensor_shape().x();
66
67 Iterator input1(src0, window.broadcast_if_dimension_le_one(src0->info()->tensor_shape()));
68 Iterator input2(src1, window.broadcast_if_dimension_le_one(src1->info()->tensor_shape()));
69 Iterator output(dst, window);
70
71 if(is_broadcast_across_x)
72 {
73 const bool is_broadcast_input_2 = input2_win.x().step() == 0;
74 Window broadcast_win = is_broadcast_input_2 ? input2_win : input1_win;
75 Window non_broadcast_win = !is_broadcast_input_2 ? input2_win : input1_win;
76 const ITensor *broadcast_tensor = is_broadcast_input_2 ? src1 : src0;
77 const ITensor *non_broadcast_tensor = !is_broadcast_input_2 ? src1 : src0;
78
79 // Clear X Dimension on execution window as we handle manually
80 non_broadcast_win.set(Window::DimX, Window::Dimension(0, 1, 1));
81
82 Iterator broadcast_input(broadcast_tensor, broadcast_win);
83 Iterator non_broadcast_input(non_broadcast_tensor, non_broadcast_win);
84 Iterator output(dst, win);
85
86 execute_window_loop(
87 win, [&](const Coordinates &)
88 {
89 const auto non_broadcast_input_ptr = reinterpret_cast<const T *>(non_broadcast_input.ptr());
90 const auto output_ptr = reinterpret_cast<T *>(output.ptr());
91
92 const T broadcast_value = *reinterpret_cast<const T *>(broadcast_input.ptr());
93 const auto broadcast_value_vec = wrapper::vdup_n(broadcast_value, ExactTagType{});
94
95 // Compute S elements per iteration
96 int x = window_start_x;
97 for(; x <= (window_end_x - window_step_x); x += window_step_x)
98 {
99 const auto non_broadcast_v = wrapper::vloadq(non_broadcast_input_ptr + x);
100 auto res = is_sat ? wrapper::vqsub(broadcast_value_vec, non_broadcast_v) : wrapper::vsub(broadcast_value_vec, non_broadcast_v);
101 if(is_broadcast_input_2)
102 {
103 res = wrapper::vmul(res, wrapper::vdup_n(static_cast<T>(-1), ExactTagType{}));
104 }
105 wrapper::vstore(output_ptr + x, res);
106 }
107
108 // Compute left-over elements
109 for(; x < window_end_x; ++x)
110 {
111 const auto non_broadcast_v = *(non_broadcast_input_ptr + x);
112 auto res = is_sat ? wrapper::sub_sat(broadcast_value, non_broadcast_v) : broadcast_value - non_broadcast_v;
113 if(is_broadcast_input_2)
114 {
115 res = static_cast<T>(-1) * res;
116 }
117
118 *(output_ptr + x) = res;
119 }
120 },
121 broadcast_input, non_broadcast_input, output);
122 }
123 else
124 {
125 // Clear X Dimension on execution window as we handle manually
126 input1_win.set(Window::DimX, Window::Dimension(0, 1, 1));
127 input2_win.set(Window::DimX, Window::Dimension(0, 1, 1));
128
129 Iterator input1(src0, input1_win);
130 Iterator input2(src1, input2_win);
131 Iterator output(dst, win);
132
133 execute_window_loop(
134 win, [&](const Coordinates &)
135 {
136 const auto input1_ptr = reinterpret_cast<const T *>(input1.ptr());
137 const auto input2_ptr = reinterpret_cast<const T *>(input2.ptr());
138 const auto output_ptr = reinterpret_cast<T *>(output.ptr());
139
140 // Compute S elements per iteration
141 int x = window_start_x;
142 for(; x <= (window_end_x - window_step_x); x += window_step_x)
143 {
144 const auto val1 = wrapper::vloadq(input1_ptr + x);
145 const auto val2 = wrapper::vloadq(input2_ptr + x);
146 const auto res = is_sat ? wrapper::vqsub(val1, val2) : wrapper::vsub(val1, val2);
147 wrapper::vstore(output_ptr + x, res);
148 }
149
150 // Compute left-over elements
151 for(; x < window_end_x; ++x)
152 {
153 const auto val1 = *(input1_ptr + x);
154 const auto val2 = *(input2_ptr + x);
155 *(output_ptr + x) = is_sat ? wrapper::sub_sat(val1, val2) : val1 - val2;
156 }
157 },
158 input1, input2, output);
159 }
160 }
161 } // namespace cpu
162 } // namespace arm_compute
163 #endif // SRC_CORE_NEON_KERNELS_SUB_LIST_H
164