xref: /aosp_15_r20/external/ComputeLibrary/src/core/NEON/kernels/NESpaceToBatchLayerKernel.cpp (revision c217d954acce2dbc11938adb493fc0abd69584f3)
1 /*
2  * Copyright (c) 2019-2020 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/NEON/kernels/NESpaceToBatchLayerKernel.h"
25 
26 #include "arm_compute/core/Helpers.h"
27 #include "arm_compute/core/ITensor.h"
28 #include "arm_compute/core/Types.h"
29 #include "arm_compute/core/Validate.h"
30 #include "arm_compute/core/utils/misc/ShapeCalculator.h"
31 #include "src/core/NEON/wrapper/wrapper.h"
32 #include "src/core/helpers/AutoConfiguration.h"
33 #include "src/core/helpers/WindowHelpers.h"
34 
35 #include <arm_neon.h>
36 #include <cstdint>
37 
38 using namespace arm_compute::misc::shape_calculator;
39 
40 namespace arm_compute
41 {
42 namespace
43 {
validate_arguments(const ITensorInfo * input,const ITensorInfo * block_info,const ITensorInfo * paddings,const ITensorInfo * output)44 Status validate_arguments(const ITensorInfo *input, const ITensorInfo *block_info, const ITensorInfo *paddings, const ITensorInfo *output)
45 {
46     ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, block_info, paddings, output);
47     ARM_COMPUTE_RETURN_ERROR_ON(input->data_type() == DataType::UNKNOWN);
48     ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(block_info, 1, DataType::S32);
49     ARM_COMPUTE_RETURN_ERROR_ON(input->num_dimensions() > 4);
50     ARM_COMPUTE_RETURN_ERROR_ON(block_info->num_dimensions() > 1);
51     ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(block_info->tensor_shape(), TensorShape{ 2 });
52     ARM_COMPUTE_RETURN_ERROR_ON(paddings->num_dimensions() > 2);
53     ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(paddings->tensor_shape(), TensorShape{ 2, 2 });
54 
55     // Validate output if initialized
56     if(output->total_size() != 0)
57     {
58         const DataLayout data_layout = input->data_layout();
59         const int        idx_channel = get_data_layout_dimension_index(data_layout, DataLayoutDimension::CHANNEL);
60         ARM_COMPUTE_RETURN_ERROR_ON(input->tensor_shape()[idx_channel] != output->tensor_shape()[idx_channel]);
61         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
62         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_QUANTIZATION_INFO(input, output);
63     }
64 
65     return Status{};
66 }
validate_arguments_static(const ITensorInfo * input,const int block_shape_x,const int block_shape_y,const Size2D & padding_left,const Size2D & padding_right,const ITensorInfo * output)67 Status validate_arguments_static(const ITensorInfo *input, const int block_shape_x, const int block_shape_y, const Size2D &padding_left, const Size2D &padding_right,
68                                  const ITensorInfo *output)
69 {
70     ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output);
71     ARM_COMPUTE_RETURN_ERROR_ON(input->data_type() == DataType::UNKNOWN);
72     ARM_COMPUTE_RETURN_ERROR_ON(input->num_dimensions() > 4);
73     ARM_COMPUTE_RETURN_ERROR_ON(block_shape_x < 1 || block_shape_y < 1);
74 
75     // Validate output if initialized
76     if(output->total_size() != 0)
77     {
78         TensorShape expected_output_shape = misc::shape_calculator::compute_space_to_batch_shape(input, block_shape_x, block_shape_y, padding_left, padding_right);
79         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(output->tensor_shape(), expected_output_shape);
80         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
81         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_QUANTIZATION_INFO(input, output);
82     }
83 
84     return Status{};
85 }
86 } // namespace
87 
NESpaceToBatchLayerKernel()88 NESpaceToBatchLayerKernel::NESpaceToBatchLayerKernel()
89     : _input(nullptr), _block_shape(nullptr), _paddings(nullptr), _output(nullptr), _data_layout(DataLayout::UNKNOWN), _padding_left(), _block_shape_x(), _block_shape_y()
90 {
91 }
92 
configure(const ITensor * input,const ITensor * block_shape,const ITensor * paddings,ITensor * output)93 void NESpaceToBatchLayerKernel::configure(const ITensor *input, const ITensor *block_shape, const ITensor *paddings, ITensor *output)
94 {
95     ARM_COMPUTE_ERROR_ON_NULLPTR(input, block_shape, paddings, output);
96     ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), block_shape->info(), paddings->info(), output->info()));
97 
98     _input       = input;
99     _block_shape = block_shape;
100     _paddings    = paddings;
101     _output      = output;
102     _data_layout = input->info()->data_layout();
103 
104     // Configure kernel window
105     Window win = calculate_max_window(*output->info(), Steps());
106     ICPPKernel::configure(win);
107 }
108 
configure(const ITensor * input,const int block_shape_x,const int block_shape_y,const Size2D & padding_left,const Size2D & padding_right,ITensor * output)109 void NESpaceToBatchLayerKernel::configure(const ITensor *input, const int block_shape_x, const int block_shape_y, const Size2D &padding_left, const Size2D &padding_right,
110                                           ITensor *output)
111 {
112     ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
113 
114     TensorShape output_shape = misc::shape_calculator::compute_space_to_batch_shape(input->info(), block_shape_x, block_shape_y, padding_left, padding_right);
115     auto_init_if_empty(*output->info(), output_shape, 1, input->info()->data_type(), input->info()->quantization_info());
116 
117     ARM_COMPUTE_ERROR_THROW_ON(validate_arguments_static(input->info(), block_shape_x, block_shape_y, padding_left, padding_right, output->info()));
118 
119     _input         = input;
120     _output        = output;
121     _block_shape_x = block_shape_x;
122     _block_shape_y = block_shape_y;
123     _padding_left  = padding_left;
124     _data_layout   = input->info()->data_layout();
125 
126     // Configure kernel window
127     Window win = calculate_max_window(*output->info(), Steps());
128     INEKernel::configure(win);
129 }
130 
validate(const ITensorInfo * input,const ITensorInfo * block_shape,const ITensorInfo * paddings,const ITensorInfo * output)131 Status NESpaceToBatchLayerKernel::validate(const ITensorInfo *input, const ITensorInfo *block_shape, const ITensorInfo *paddings, const ITensorInfo *output)
132 {
133     ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, block_shape, paddings, output));
134     return Status{};
135 }
validate(const ITensorInfo * input,const int block_shape_x,const int block_shape_y,const Size2D & padding_left,const Size2D & padding_right,const ITensorInfo * output)136 Status NESpaceToBatchLayerKernel::validate(const ITensorInfo *input, const int block_shape_x, const int block_shape_y, const Size2D &padding_left, const Size2D &padding_right,
137                                            const ITensorInfo *output)
138 {
139     ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments_static(input, block_shape_x, block_shape_y, padding_left, padding_right, output));
140     return Status{};
141 }
142 
run(const Window & window,const ThreadInfo & info)143 void NESpaceToBatchLayerKernel::run(const Window &window, const ThreadInfo &info)
144 {
145     ARM_COMPUTE_UNUSED(info);
146     ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
147     ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICPPKernel::window(), window);
148 
149     if(_block_shape != nullptr)
150     {
151         // Retrieve the block shapes dynamically
152         _block_shape_x = *(reinterpret_cast<const int *>(_block_shape->ptr_to_element(0)));
153         _block_shape_y = *(reinterpret_cast<const int *>(_block_shape->ptr_to_element(1)));
154     }
155 
156     if(_paddings != nullptr)
157     {
158         const size_t pad_left_x = *reinterpret_cast<const size_t *>(_paddings->ptr_to_element({ 0, 0 }));
159         const size_t pad_left_y = *reinterpret_cast<const size_t *>(_paddings->ptr_to_element({ 1, 0 }));
160         _padding_left           = Size2D(pad_left_x, pad_left_y);
161     }
162     const int height_idx   = get_data_layout_dimension_index(_data_layout, DataLayoutDimension::HEIGHT);
163     const int width_idx    = get_data_layout_dimension_index(_data_layout, DataLayoutDimension::WIDTH);
164     const int batch_idx    = get_data_layout_dimension_index(_data_layout, DataLayoutDimension::BATCHES);
165     const int element_size = _input->info()->element_size();
166 
167     const size_t height     = _input->info()->dimension(height_idx);
168     const size_t width      = _input->info()->dimension(width_idx);
169     const size_t batch_size = _input->info()->dimension(batch_idx);
170 
171     Window slice_out = window.first_slice_window_3D();
172 
173     int batch_id = 0;
174 
175     // Main loop for NCHW and NHWC
176     if(_data_layout == DataLayout::NCHW)
177     {
178         do
179         {
180             Iterator out(_output, slice_out);
181             execute_window_loop(slice_out, [&](const Coordinates & id)
182             {
183                 const size_t out_x = id.x();
184                 const size_t out_y = id.y();
185                 const size_t z     = id.z();
186                 const size_t pos_x = out_x * _block_shape_x + (batch_id / batch_size) % _block_shape_x;
187                 const size_t pos_y = out_y * _block_shape_y + (batch_id / batch_size) / _block_shape_x;
188                 if(pos_y >= _padding_left.y() && pos_y < _padding_left.y() + height && pos_x >= _padding_left.x() && pos_x < _padding_left.x() + width)
189                 {
190                     const int   w    = batch_id % batch_size;
191                     const int   in_x = pos_x - _padding_left.x();
192                     const int   in_y = pos_y - _padding_left.y();
193                     Coordinates input_coords{ in_x, in_y, z, w };
194                     memcpy(out.ptr(), _input->ptr_to_element(input_coords), element_size);
195                 }
196             },
197             out);
198             ++batch_id;
199         }
200         while(window.slide_window_slice_3D(slice_out));
201     }
202     else
203     {
204         do
205         {
206             Iterator out(_output, slice_out);
207             execute_window_loop(slice_out, [&](const Coordinates & id)
208             {
209                 const size_t out_x = id.y();
210                 const size_t out_y = id.z();
211                 const size_t z     = id.x();
212                 const size_t pos_x = out_x * _block_shape_x + (batch_id / batch_size) % _block_shape_x;
213                 const size_t pos_y = out_y * _block_shape_y + (batch_id / batch_size) / _block_shape_x;
214                 if(pos_y >= _padding_left.y() && pos_y < _padding_left.y() + height && pos_x >= _padding_left.x() && pos_x < _padding_left.x() + width)
215                 {
216                     const int   w    = batch_id % batch_size;
217                     const int   in_x = pos_x - _padding_left.x();
218                     const int   in_y = pos_y - _padding_left.y();
219                     Coordinates input_coords{ z, in_x, in_y, w };
220                     memcpy(out.ptr(), _input->ptr_to_element(input_coords), element_size);
221                 }
222             },
223             out);
224             ++batch_id;
225         }
226         while(window.slide_window_slice_3D(slice_out));
227     }
228 }
229 } // namespace arm_compute
230