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 "arm_compute/core/Helpers.h"
26 #include "arm_compute/core/ITensorPack.h"
27 #include "arm_compute/core/Window.h"
28 #include "src/core/NEON/NEMath.h"
29 #include "src/core/NEON/wrapper/wrapper.h"
30 #include "src/core/helpers/ScaleHelpers.h"
31 #include "src/core/utils/ScaleUtils.h"
32 #include "support/Rounding.h"
33
34 #include <arm_sve.h>
35 #include <cmath>
36 #include <cstddef>
37
38 namespace arm_compute
39 {
40 namespace
41 {
u8_sve_scale_nearest(const ITensor * src,ITensor * dst,const ITensor * offsets,float sampling_offset,bool align_corners,const Window & window)42 void u8_sve_scale_nearest(const ITensor *src, ITensor *dst, const ITensor *offsets,
43 float sampling_offset, bool align_corners, const Window &window)
44 {
45 const size_t in_stride_c = src->info()->dimension(0) + src->info()->padding().left + src->info()->padding().right;
46 const size_t in_stride_w = src->info()->dimension(1) + src->info()->padding().top + src->info()->padding().bottom;
47 const size_t in_stride_wc = in_stride_w * in_stride_c;
48 const size_t in_dim_h = src->info()->dimension(2);
49
50 // Compute the ratio between source height and destination height
51 const auto hr = scale_utils::calculate_resize_ratio(in_dim_h, dst->info()->dimension(2), align_corners);
52 const auto window_start_x = static_cast<int32_t>(window.x().start());
53 const auto window_end_x = static_cast<int32_t>(window.x().end());
54
55 Window win(window);
56 win.set(Window::DimX, Window::Dimension(0, 1, 1));
57 Iterator out(dst, win);
58
59 const uint8_t *in_ptr_start = src->buffer() + src->info()->offset_first_element_in_bytes();
60 const unsigned int in_stride_bytes_hwc = src->info()->strides_in_bytes()[3];
61
62 execute_window_loop(win, [&](const Coordinates & id)
63 {
64 const int32_t offset = *reinterpret_cast<const int32_t *>(offsets->ptr_to_element(Coordinates(id.y(), id.z()))) * in_stride_c;
65 const auto in_hi = static_cast<int>(align_corners ? utils::rounding::round_half_away_from_zero((id.z() + sampling_offset) * hr) : std::floor((id.z() + sampling_offset) * hr));
66 const int offset_row = in_hi * in_stride_wc;
67 const auto in_ptr = reinterpret_cast<const uint8_t *>(in_ptr_start + in_stride_bytes_hwc * id[3]);
68 const auto out_ptr = reinterpret_cast<uint8_t *>(out.ptr());
69
70 // Compute S elements per iteration
71 int x = window_start_x;
72 svbool_t pg = svwhilelt_b8(x, window_end_x);
73 do
74 {
75 // Store results
76 svst1_u8(pg, out_ptr + x, svld1_u8(pg, in_ptr + offset + offset_row + x));
77
78 x += svcntw();
79 pg = svwhilelt_b8(x, window_end_x);
80 }
81 while(svptest_any(svptrue_b8(), pg));
82 },
83 out);
84 }
85
s16_sve_scale_nearest(const ITensor * src,ITensor * dst,const ITensor * offsets,float sampling_offset,bool align_corners,const Window & window)86 void s16_sve_scale_nearest(const ITensor *src, ITensor *dst, const ITensor *offsets,
87 float sampling_offset, bool align_corners, const Window &window)
88 {
89 const size_t in_stride_c = src->info()->dimension(0) + src->info()->padding().left + src->info()->padding().right;
90 const size_t in_stride_w = src->info()->dimension(1) + src->info()->padding().top + src->info()->padding().bottom;
91 const size_t in_stride_wc = in_stride_w * in_stride_c;
92 const size_t in_dim_h = src->info()->dimension(2);
93
94 // Compute the ratio between source height and destination height
95 const auto hr = scale_utils::calculate_resize_ratio(in_dim_h, dst->info()->dimension(2), align_corners);
96 const auto window_start_x = static_cast<int32_t>(window.x().start());
97 const auto window_end_x = static_cast<int32_t>(window.x().end());
98
99 Window win(window);
100 win.set(Window::DimX, Window::Dimension(0, 1, 1));
101 Iterator out(dst, win);
102
103 const uint8_t *in_ptr_start = src->buffer() + src->info()->offset_first_element_in_bytes();
104 const unsigned int in_stride_bytes_hwc = src->info()->strides_in_bytes()[3];
105
106 execute_window_loop(win, [&](const Coordinates & id)
107 {
108 const int32_t offset = *reinterpret_cast<const int32_t *>(offsets->ptr_to_element(Coordinates(id.y(), id.z()))) * in_stride_c;
109 const auto in_hi = static_cast<int>(align_corners ? utils::rounding::round_half_away_from_zero((id.z() + sampling_offset) * hr) : std::floor((id.z() + sampling_offset) * hr));
110 const int offset_row = in_hi * in_stride_wc;
111 const auto in_ptr = reinterpret_cast<const int16_t *>(in_ptr_start + in_stride_bytes_hwc * id[3]);
112 const auto out_ptr = reinterpret_cast<int16_t *>(out.ptr());
113
114 // Compute S elements per iteration
115 int x = window_start_x;
116 svbool_t pg = svwhilelt_b16(x, window_end_x);
117 do
118 {
119 // Store results
120 svst1_s16(pg, out_ptr + x, svld1_s16(pg, in_ptr + offset + offset_row + x));
121
122 x += svcntw();
123 pg = svwhilelt_b16(x, window_end_x);
124 }
125 while(svptest_any(svptrue_b16(), pg));
126 },
127 out);
128 }
129 }
130 namespace cpu
131 {
u8_sve_scale(const ITensor * src,ITensor * dst,const ITensor * offsets,const ITensor * dx,const ITensor * dy,InterpolationPolicy policy,BorderMode border_mode,PixelValue constant_border_value,float sampling_offset,bool align_corners,const Window & window)132 void u8_sve_scale(const ITensor *src, ITensor *dst, const ITensor *offsets, const ITensor *dx, const ITensor *dy,
133 InterpolationPolicy policy, BorderMode border_mode, PixelValue constant_border_value, float sampling_offset,
134 bool align_corners, const Window &window)
135 {
136 ARM_COMPUTE_UNUSED(dx, dy, border_mode, constant_border_value);
137 if(policy == InterpolationPolicy::NEAREST_NEIGHBOR)
138 {
139 u8_sve_scale_nearest(src, dst, offsets, sampling_offset, align_corners, window);
140 }
141 else
142 {
143 ARM_COMPUTE_ERROR("Not Implemented");
144 }
145 }
146
s16_sve_scale(const ITensor * src,ITensor * dst,const ITensor * offsets,const ITensor * dx,const ITensor * dy,InterpolationPolicy policy,BorderMode border_mode,PixelValue constant_border_value,float sampling_offset,bool align_corners,const Window & window)147 void s16_sve_scale(const ITensor *src, ITensor *dst, const ITensor *offsets, const ITensor *dx, const ITensor *dy,
148 InterpolationPolicy policy, BorderMode border_mode, PixelValue constant_border_value, float sampling_offset,
149 bool align_corners, const Window &window)
150 {
151 ARM_COMPUTE_UNUSED(dx, dy, border_mode, constant_border_value);
152 if(policy == InterpolationPolicy::NEAREST_NEIGHBOR)
153 {
154 s16_sve_scale_nearest(src, dst, offsets, sampling_offset, align_corners, window);
155 }
156 else
157 {
158 ARM_COMPUTE_ERROR("Not Implemented");
159 }
160 }
161 } // namespace cpu
162 } // namespace arm_compute
163