1 /*
2 * Copyright (c) 2016-2020 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/runtime/NEON/functions/NEScale.h"
25
26 #include "arm_compute/core/Coordinates.h"
27 #include "arm_compute/core/Error.h"
28 #include "arm_compute/core/Helpers.h"
29 #include "arm_compute/core/ITensor.h"
30 #include "arm_compute/core/PixelValue.h"
31 #include "arm_compute/core/TensorInfo.h"
32 #include "arm_compute/core/Window.h"
33 #include "arm_compute/runtime/NEON/NEScheduler.h"
34 #include "arm_compute/runtime/TensorAllocator.h"
35 #include "src/core/NEON/kernels/NEScaleKernel.h"
36
37 #include "src/core/utils/ScaleUtils.h"
38
39 #include "support/MemorySupport.h"
40 #include "support/Rounding.h"
41
42 #include <cmath>
43 #include <cstddef>
44 #include <utility>
45
46 namespace arm_compute
47 {
48 namespace
49 {
precompute_dx_dy_offsets(ITensor * dx,ITensor * dy,ITensor * offsets,float wr,float hr,SamplingPolicy sampling_policy,bool align_corners)50 void precompute_dx_dy_offsets(ITensor *dx, ITensor *dy, ITensor *offsets, float wr, float hr, SamplingPolicy sampling_policy, bool align_corners)
51 {
52 ARM_COMPUTE_ERROR_ON(nullptr == offsets);
53 ARM_COMPUTE_UNUSED(sampling_policy);
54 float sampling_offset = 0.0f;
55 if(sampling_policy == SamplingPolicy::CENTER)
56 {
57 sampling_offset = 0.5f;
58 }
59
60 Window win;
61 win.set(Window::DimX, Window::Dimension(0, offsets->info()->dimension(0), 1));
62 win.set(Window::DimY, Window::Dimension(0, offsets->info()->dimension(1), 1));
63
64 if(dx != nullptr && dy != nullptr)
65 {
66 // Pre-compute the offset and pixel's distance for BILINEAR interpolation
67 Iterator offsets_it(offsets, win);
68 Iterator dx_it(dx, win);
69 Iterator dy_it(dy, win);
70
71 execute_window_loop(win, [&](const Coordinates & id)
72 {
73 const float in_x = (id.x() + sampling_offset) * wr - sampling_offset;
74 const float in_y = (id.y() + sampling_offset) * hr - sampling_offset;
75 const int in_xi = std::floor(in_x);
76 const int in_yi = std::floor(in_y);
77
78 *reinterpret_cast<int32_t *>(offsets_it.ptr()) = in_xi;
79 *reinterpret_cast<float *>(dx_it.ptr()) = in_x - in_xi;
80 *reinterpret_cast<float *>(dy_it.ptr()) = in_y - in_yi;
81 },
82 offsets_it, dx_it, dy_it);
83 }
84 else
85 {
86 // Pre-compute the offset for NEAREST interpolation
87 Iterator offsets_it(offsets, win);
88
89 execute_window_loop(win, [&](const Coordinates & id)
90 {
91 const float float_in_xi = (id.x() + sampling_offset) * wr;
92 const auto in_xi = static_cast<size_t>(align_corners ? arm_compute::utils::rounding::round_half_away_from_zero(float_in_xi) : std::floor(float_in_xi));
93 *reinterpret_cast<int32_t *>(offsets_it.ptr()) = in_xi;
94 },
95 offsets_it);
96 }
97 }
98 } // namespace
99
NEScale()100 NEScale::NEScale()
101 : _offsets(), _dx(), _dy()
102 {
103 }
104
configure(ITensor * input,ITensor * output,const ScaleKernelInfo & info)105 void NEScale::configure(ITensor *input, ITensor *output, const ScaleKernelInfo &info)
106 {
107 ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
108 ARM_COMPUTE_ERROR_THROW_ON(NEScale::validate(input->info(), output->info(), info));
109
110 const bool is_align_corners_used = info.align_corners && arm_compute::scale_utils::is_align_corners_allowed_sampling_policy(info.sampling_policy);
111
112 // Get data layout and width/height indices
113 const DataLayout data_layout = input->info()->data_layout();
114 const int idx_width = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH);
115 const int idx_height = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT);
116
117 // Get the tensor shape
118 TensorShape shape(output->info()->dimension(idx_width));
119 shape.set(1, output->info()->dimension(idx_height), false);
120
121 // Compute the ratio between source width/height and destination width/height
122 const auto wr = arm_compute::scale_utils::calculate_resize_ratio(input->info()->dimension(idx_width), output->info()->dimension(idx_width), is_align_corners_used);
123 const auto hr = arm_compute::scale_utils::calculate_resize_ratio(input->info()->dimension(idx_height), output->info()->dimension(idx_height), is_align_corners_used);
124
125 // Area interpolation behaves as Nearest Neighbour in case of up-sampling
126 const auto policy_to_use = (info.interpolation_policy == InterpolationPolicy::AREA && wr <= 1.f && hr <= 1.f) ? InterpolationPolicy::NEAREST_NEIGHBOR : info.interpolation_policy;
127
128 auto scale_kernel = arm_compute::support::cpp14::make_unique<NEScaleKernel>();
129 switch(policy_to_use)
130 {
131 case InterpolationPolicy::NEAREST_NEIGHBOR:
132 {
133 TensorInfo tensor_info_offsets(shape, Format::S32);
134 _offsets.allocator()->init(tensor_info_offsets);
135
136 scale_kernel->configure(input, nullptr, nullptr, &_offsets, output, info);
137
138 // Allocate once the configure methods have been called
139 _offsets.allocator()->allocate();
140
141 // Pre-compute offsets for nearest interpolation
142 precompute_dx_dy_offsets(nullptr, nullptr, &_offsets, wr, hr, info.sampling_policy, is_align_corners_used);
143 break;
144 }
145 case InterpolationPolicy::BILINEAR:
146 {
147 TensorInfo tensor_info_offsets(shape, Format::S32);
148 TensorInfo tensor_info_dxdy(shape, Format::F32);
149
150 _offsets.allocator()->init(tensor_info_offsets);
151 _dx.allocator()->init(tensor_info_dxdy);
152 _dy.allocator()->init(tensor_info_dxdy);
153
154 scale_kernel->configure(input, &_dx, &_dy, &_offsets, output, info);
155
156 // Allocate once the configure methods have been called
157 _offsets.allocator()->allocate();
158 _dx.allocator()->allocate();
159 _dy.allocator()->allocate();
160
161 // Pre-compute dx, dy and offsets for bilinear interpolation
162 precompute_dx_dy_offsets(&_dx, &_dy, &_offsets, wr, hr, info.sampling_policy, is_align_corners_used);
163 break;
164 }
165 case InterpolationPolicy::AREA:
166 {
167 scale_kernel->configure(input, nullptr, nullptr, nullptr, output, info);
168 break;
169 }
170 default:
171 ARM_COMPUTE_ERROR("Unsupported interpolation mode");
172 }
173 _kernel = std::move(scale_kernel);
174 }
175
validate(const ITensorInfo * input,const ITensorInfo * output,const ScaleKernelInfo & info)176 Status NEScale::validate(const ITensorInfo *input, const ITensorInfo *output, const ScaleKernelInfo &info)
177 {
178 ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output);
179 ARM_COMPUTE_RETURN_ERROR_ON(info.sampling_policy != SamplingPolicy::CENTER && info.sampling_policy != SamplingPolicy::TOP_LEFT);
180
181 ITensorInfo *offsets = nullptr;
182 ITensorInfo *dx = nullptr;
183 ITensorInfo *dy = nullptr;
184
185 // Get data layout and width/height indices
186 const DataLayout data_layout = input->data_layout();
187 const int idx_width = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH);
188 const int idx_height = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT);
189
190 // Get the tensor shape of auxilary buffers
191 const TensorShape shape(output->dimension(idx_width), output->dimension(idx_height));
192
193 TensorInfo tensor_info_offsets(shape, Format::S32);
194 TensorInfo tensor_info_dx(shape, Format::F32);
195 TensorInfo tensor_info_dy(shape, Format::F32);
196
197 switch(info.interpolation_policy)
198 {
199 case InterpolationPolicy::NEAREST_NEIGHBOR:
200 offsets = &tensor_info_offsets;
201 break;
202 case InterpolationPolicy::BILINEAR:
203 offsets = &tensor_info_offsets;
204 dx = &tensor_info_dx;
205 dy = &tensor_info_dy;
206 break;
207 default:
208 break;
209 }
210
211 ARM_COMPUTE_RETURN_ON_ERROR(NEScaleKernel::validate(input->clone().get(), dx, dy, offsets, output->clone().get(), info));
212 return Status{};
213 }
214 } // namespace arm_compute
215