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
25 #include "src/cpu/kernels/add/generic/sve/impl.h"
26 #include "arm_compute/core/Helpers.h"
27 #include "arm_compute/core/utils/misc/Traits.h"
28 #include "src/core/NEON/SVEMath.h"
29 #include "src/core/NEON/wrapper/intrinsics/intrinsics.h"
30 #include <arm_sve.h>
31 namespace arm_compute
32 {
33 namespace cpu
34 {
35 template <typename ScalarType>
add_same_sve(const ITensor * src0,const ITensor * src1,ITensor * dst,const ConvertPolicy & policy,const Window & window)36 void add_same_sve(const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window)
37 {
38 const auto all_true_pg = wrapper::svptrue<ScalarType>();
39 const auto window_start_x = static_cast<int>(window.x().start());
40 const auto window_end_x = static_cast<int>(window.x().end());
41 const bool is_broadcast_across_x = src0->info()->tensor_shape().x() != src1->info()->tensor_shape().x();
42 const bool is_sat = (policy == ConvertPolicy::SATURATE);
43
44 // Clear X Dimension on execution window as we handle manually
45 Window win = window;
46 win.set(Window::DimX, Window::Dimension(0, 1, 1));
47
48 // Create input windows
49 Window input1_win = window.broadcast_if_dimension_le_one(src0->info()->tensor_shape());
50 Window input2_win = window.broadcast_if_dimension_le_one(src1->info()->tensor_shape());
51
52 Iterator input1(src0, window.broadcast_if_dimension_le_one(src0->info()->tensor_shape()));
53 Iterator input2(src1, window.broadcast_if_dimension_le_one(src1->info()->tensor_shape()));
54 Iterator output(dst, window);
55
56 if(is_broadcast_across_x)
57 {
58 const bool is_broadcast_input_2 = input2_win.x().step() == 0;
59 Window broadcast_win = is_broadcast_input_2 ? input2_win : input1_win;
60 Window non_broadcast_win = !is_broadcast_input_2 ? input2_win : input1_win;
61 const ITensor *broadcast_tensor = is_broadcast_input_2 ? src1 : src0;
62 const ITensor *non_broadcast_tensor = !is_broadcast_input_2 ? src1 : src0;
63
64 // Clear X Dimension on execution window as we handle manually
65 non_broadcast_win.set(Window::DimX, Window::Dimension(0, 1, 1));
66
67 Iterator broadcast_input(broadcast_tensor, broadcast_win);
68 Iterator non_broadcast_input(non_broadcast_tensor, non_broadcast_win);
69 Iterator output(dst, win);
70
71 execute_window_loop(win, [&](const Coordinates &)
72 {
73 const auto non_broadcast_input_ptr = reinterpret_cast<const ScalarType *>(non_broadcast_input.ptr());
74 const auto output_ptr = reinterpret_cast<ScalarType *>(output.ptr());
75
76 const ScalarType broadcast_value = *reinterpret_cast<const ScalarType *>(broadcast_input.ptr());
77 const auto broadcast_value_vec = wrapper::svdup_n(broadcast_value);
78
79 int x = window_start_x;
80 svbool_t pg = wrapper::svwhilelt<ScalarType>(x, window_end_x);
81 do
82 {
83 const auto non_broadcast_v = svld1(pg, non_broadcast_input_ptr + x);
84 auto res = is_sat ? wrapper::svqadd(broadcast_value_vec, non_broadcast_v) : svadd_z(pg, broadcast_value_vec, non_broadcast_v);
85 svst1(pg, output_ptr + x, res);
86
87 x += wrapper::svcnt<ScalarType>();
88 pg = wrapper::svwhilelt<ScalarType>(x, window_end_x);
89 }
90 while(svptest_any(all_true_pg, pg));
91 },
92 broadcast_input, non_broadcast_input, output);
93 }
94 else
95 {
96 // Clear X Dimension on execution window as we handle manually
97 input1_win.set(Window::DimX, Window::Dimension(0, 1, 1));
98 input2_win.set(Window::DimX, Window::Dimension(0, 1, 1));
99
100 Iterator input1(src0, input1_win);
101 Iterator input2(src1, input2_win);
102 Iterator output(dst, win);
103
104 execute_window_loop(win, [&](const Coordinates &)
105 {
106 const auto input1_ptr = reinterpret_cast<const ScalarType *>(input1.ptr());
107 const auto input2_ptr = reinterpret_cast<const ScalarType *>(input2.ptr());
108 const auto output_ptr = reinterpret_cast<ScalarType *>(output.ptr());
109
110 int x = window_start_x;
111 svbool_t pg = wrapper::svwhilelt<ScalarType>(x, window_end_x);
112 do
113 {
114 const auto val1 = svld1(pg, input1_ptr + x);
115 const auto val2 = svld1(pg, input2_ptr + x);
116 const auto res = is_sat ? wrapper::svqadd(val1, val2) : svadd_z(pg, val1, val2);
117 svst1(pg, output_ptr + x, res);
118
119 x += wrapper::svcnt<ScalarType>();
120 pg = wrapper::svwhilelt<ScalarType>(x, window_end_x);
121 }
122 while(svptest_any(all_true_pg, pg));
123 },
124 input1, input2, output);
125 }
126 }
127 template void add_same_sve<float>(const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window);
128 template void add_same_sve<uint8_t>(const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window);
129 template void add_same_sve<int16_t>(const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window);
130 template void add_same_sve<int32_t>(const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window);
131 #if defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && defined(ENABLE_FP16_KERNELS)
132 template void add_same_sve<float16_t>(const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window);
133 #endif /* (__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && defined(ENABLE_FP16_KERNELS) */
134 } // namespace cpu
135 } // namespace arm_compute
136