xref: /aosp_15_r20/external/ComputeLibrary/src/core/AccessWindowTranspose.cpp (revision c217d954acce2dbc11938adb493fc0abd69584f3)
1 /*
2  * Copyright (c) 2017-2020 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 "src/core/AccessWindowTranspose.h"
25 
26 #include "arm_compute/core/Helpers.h"
27 #include "arm_compute/core/ITensorInfo.h"
28 #include "arm_compute/core/Window.h"
29 
30 using namespace arm_compute;
31 
compute_valid_region(const Window & window,ValidRegion input_valid_region,bool border_undefined,BorderSize border_size) const32 ValidRegion AccessWindowTranspose::compute_valid_region(const Window &window, ValidRegion input_valid_region, bool border_undefined, BorderSize border_size) const
33 {
34     if(_info == nullptr)
35     {
36         return input_valid_region;
37     }
38 
39     Coordinates &anchor = input_valid_region.anchor;
40     TensorShape &shape  = input_valid_region.shape;
41     Coordinates  old_anchor(anchor);
42     TensorShape  old_shape(shape);
43 
44     if(!border_undefined)
45     {
46         border_size = BorderSize(0);
47     }
48 
49     // Start of the valid region is equal to the start of the window. But it
50     // cannot be less than the start of the input's valid region plus the border
51     // size required by this kernel (if undefined).
52     // Additionally the valid region is shifted by the offset that is used by
53     // the kernel to write back output values.
54     // As the relation between input and output is transposed window.y() is
55     // used for x anchor and window.x() for y anchor.
56     if(_info->dimension(0) > 1)
57     {
58         anchor.set(0, std::max<int>(window.y().start() * _scale_x, anchor[1] + border_size.top) + _x);
59     }
60     anchor.set(1, std::max<int>(window.x().start() * _scale_y, anchor[0] + border_size.left) + _y);
61 
62     // End of the valid region is equal to the start of the last write of the
63     // kernel plus the number of written elements. (This assumes that all
64     // written elements are valid). Nevertheless the end cannot be larger than
65     // the end of the input's valid region minus the border size.
66     // Note: not the end points of the region are stored but its size. Thus the
67     // old size is first converted into end points to compared against the
68     // execution window. Afterwards the new end points are converted back into
69     // a size of the region.
70     // As the relation between input and output is transposed window.y() is
71     // used for x shape and window.x() for y shape.
72     if(_info->dimension(0) > 1)
73     {
74         shape.set(0, std::min<int>((old_anchor[1] + old_shape[0]) * _scale_x - border_size.right, (window.y().end() - window.y().step()) * _scale_x + _width) - anchor[0]);
75     }
76     shape.set(1, std::min<int>((old_anchor[0] + old_shape[1]) * _scale_y - border_size.bottom, (window.x().end() - window.x().step()) * _scale_y + _height) - anchor[1]);
77 
78     // For higher dimensions use the intersection of the window size and the
79     // valid region of the input
80     for(size_t d = 2; d < _info->num_dimensions(); ++d)
81     {
82         anchor.set(d, std::max(window[d].start(), input_valid_region.anchor[d]));
83         shape.set(d, std::min<int>(window[d].end(), input_valid_region.shape[d]) - anchor[d]);
84     }
85 
86     return input_valid_region;
87 }
88 
update_window_if_needed(Window & window) const89 bool AccessWindowTranspose::update_window_if_needed(Window &window) const
90 {
91     // Only update the window size if we can't use padding
92     if(_info == nullptr || _info->is_resizable())
93     {
94         return false;
95     }
96 
97     const TensorShape &shape                = _info->tensor_shape();
98     const Strides     &strides              = _info->strides_in_bytes();
99     const size_t       offset_first_element = _info->offset_first_element_in_bytes();
100 
101     bool window_modified = false;
102 
103     int front_pad_y = 0;
104 
105     // Transpose and scale
106     const int min_y = window.x().start() * _scale_y + _y;
107     const int max_y = window.x().end() * _scale_y + _y;
108 
109     // Adjust window start for output's Y dimension (so X in (input) window)
110     if(min_y < 0)
111     {
112         // Calculate rows available above the tensor
113         const int front_pad_y_available = -offset_first_element / strides[1];
114 
115         if(min_y < front_pad_y_available)
116         {
117             // Not enough padding available, need to shrink the window
118             const int start = adjust_up(min_y, front_pad_y_available, window.x().step() * _scale_y) - _y;
119 
120             window.set(0, Window::Dimension(start / _scale_y, window.x().end(), window.x().step()));
121             window_modified = true;
122         }
123 
124         // Update front padding with reconstructed value
125         front_pad_y = std::max(0, static_cast<int>(std::floor(-window.x().start() * _scale_y)) - _y);
126     }
127 
128     // Adjust window end for Y dimension
129     if(max_y > static_cast<int>(shape[1]))
130     {
131         const int stride_z = _info->num_dimensions() > 2 ? strides[2] : _info->total_size();
132 
133         // Calculate rows available below the tensor
134         const int tail_pad_y_available = (stride_z / strides[1]) - shape[1] - front_pad_y;
135 
136         if(static_cast<int>(shape[1]) + tail_pad_y_available < max_y)
137         {
138             // Not enough padding available, need to shrink the window
139             const int end = adjust_down(max_y, shape[1] + tail_pad_y_available, window.x().step() * _scale_y) + window.x().step() * _scale_y - _y - _height;
140             window.set(0, Window::Dimension(window.x().start(), end / _scale_y, window.x().step()));
141             window_modified = true;
142         }
143     }
144 
145     int front_pad_x = 0;
146 
147     // Transpose and scale
148     const int min_x = window.y().start() * _scale_x + _x;
149     const int max_x = window.y().end() * _scale_x + _x;
150 
151     const int stride_y = _info->num_dimensions() > 1 ? strides[1] : _info->total_size();
152 
153     // Adjust window start for X dimension
154     if(min_x < 0)
155     {
156         const int front_pad_x_available = -std::min<int>(static_cast<int>(offset_first_element) - front_pad_y * strides[1], stride_y - shape[0] * strides[0]) / static_cast<int>(strides[0]);
157 
158         if(min_x < front_pad_x_available)
159         {
160             // Not enough padding available, need to shrink the window
161             const int start = adjust_up(min_x, front_pad_x_available, window.y().step() * _scale_x) - _x;
162             window.set(1, Window::Dimension(start / _scale_x, window.y().end(), window.y().step()));
163             window_modified = true;
164         }
165 
166         // Update front padding with reconstructed value
167         front_pad_x = std::max(0, static_cast<int>(std::floor(-window.y().start() * _scale_x)) - _x);
168     }
169 
170     // Adjust window end for X dimension
171     if(max_x > static_cast<int>(shape[0]))
172     {
173         const int tail_pad_x_available = (stride_y / strides[0]) - shape[0] - front_pad_x;
174 
175         if(static_cast<int>(shape[0]) + tail_pad_x_available < max_x)
176         {
177             // Not enough padding available, need to shrink the window
178             const int end = adjust_down(max_x, shape[0] + tail_pad_x_available, window.y().step() * _scale_x) + window.y().step() * _scale_x - _x - _width;
179             window.set(1, Window::Dimension(window.y().start(), end / _scale_x, window.y().step()));
180             window_modified = true;
181         }
182     }
183 
184     window.validate();
185 
186     return window_modified;
187 }
188 
update_padding_if_needed(const Window & window)189 bool AccessWindowTranspose::update_padding_if_needed(const Window &window)
190 {
191     // Only update the padding if the tensor allows it
192     if(_info == nullptr || !_info->is_resizable())
193     {
194         return false;
195     }
196 
197     ARM_COMPUTE_ERROR_ON(window.y().step() == 0);
198     ARM_COMPUTE_ERROR_ON(window.x().step() == 0);
199 
200     const int min_x = window.y().start() * _scale_x + _x;
201     const int max_x = (window.y().end() - window.y().step()) * _scale_x + _x + _width;
202     const int min_y = window.x().start() * _scale_y + _y;
203     const int max_y = (window.x().end() - window.x().step()) * _scale_y + _y + _height;
204 
205     const TensorShape &shape = _info->tensor_shape();
206 
207     PaddingSize padding;
208     padding.left   = std::max(0, -min_x);
209     padding.right  = std::max<int>(0, max_x - shape[0]);
210     padding.top    = std::max(0, -min_y);
211     padding.bottom = std::max<int>(0, max_y - shape[1]);
212 
213     // Update strides in tensor info
214     return _info->extend_padding(padding);
215 }
216