1 /*
2 * Copyright (c) 2019-2022 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 #include "src/core/NEON/kernels/NEGatherKernel.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/TensorInfo.h"
30 #include "arm_compute/core/Validate.h"
31 #include "arm_compute/core/Window.h"
32 #include "arm_compute/core/utils/misc/ShapeCalculator.h"
33 #include "src/core/CPP/Validate.h"
34 #include "src/core/helpers/AutoConfiguration.h"
35 #include "src/core/helpers/WindowHelpers.h"
36
37 namespace arm_compute
38 {
39 namespace
40 {
41 /** Validate the indices
42 *
43 * Validate that indices are not negative
44 *
45 * @param[in] indices Indices tensor info.
46 */
47
48 template <typename U>
validate_indices(const ITensor * indices)49 void validate_indices(const ITensor *indices)
50 {
51 Window window;
52 window.use_tensor_dimensions(indices->info()->tensor_shape());
53 execute_window_loop(window, [&](const Coordinates & id)
54 {
55 const auto i = *(reinterpret_cast<int32_t *>(indices->ptr_to_element(id)));
56 ARM_COMPUTE_UNUSED(i);
57 ARM_COMPUTE_ERROR_ON(i < 0);
58 });
59 }
60
validate_arguments(const ITensorInfo * input,const ITensorInfo * indices,const ITensorInfo * output,int axis)61 Status validate_arguments(const ITensorInfo *input, const ITensorInfo *indices, const ITensorInfo *output, int axis)
62 {
63 ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, indices, output);
64 ARM_COMPUTE_RETURN_ERROR_ON(input->num_dimensions() > 4);
65
66 if(axis < 0)
67 {
68 axis += input->num_dimensions();
69 }
70
71 ARM_COMPUTE_RETURN_ERROR_ON(0 > axis || axis >= static_cast<int32_t>(input->num_dimensions()));
72 ARM_COMPUTE_RETURN_ERROR_ON(axis != 1 && indices->num_dimensions() > 1);
73 ARM_COMPUTE_RETURN_ERROR_ON(input->data_type() == DataType::UNKNOWN);
74
75 if(output->total_size() != 0)
76 {
77 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
78 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_QUANTIZATION_INFO(input, output);
79 TensorShape output_shape = arm_compute::misc::shape_calculator::compute_gather_shape(input->tensor_shape(), indices->tensor_shape(), axis);
80 ARM_COMPUTE_RETURN_ERROR_ON(output_shape.total_size() != output->tensor_shape().total_size());
81 }
82
83 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(indices, 1, DataType::U32, DataType::S32);
84
85 return Status{};
86 }
87 } // namespace
88
NEGatherKernel()89 NEGatherKernel::NEGatherKernel()
90 : _input{}, _indices{}, _axis{}, _output{}, _func{}
91 {
92 }
93
94 template <typename U>
gather_multiindices_1_axis(const Window & window,const ThreadInfo & info)95 inline void NEGatherKernel::gather_multiindices_1_axis(const Window &window, const ThreadInfo &info)
96 {
97 ARM_COMPUTE_UNUSED(info);
98 ARM_COMPUTE_ERROR_ON(_indices->info()->num_dimensions() < 2 || _indices->info()->num_dimensions() > 3);
99 validate_indices<U>(_indices);
100 Window win = window;
101 win.set(Window::DimX, Window::Dimension(0, 1, 1));
102 execute_window_loop(win, [&](const Coordinates & id)
103 {
104 auto *dst_ptr = _output->ptr_to_element(id);
105 Coordinates index_offset;
106 for(uint32_t k = 0; k < _indices->info()->num_dimensions(); ++k)
107 {
108 index_offset.set(k, id[k + 1]);
109 }
110 const uint32_t row = *(reinterpret_cast<uint32_t *>(_indices->ptr_to_element(index_offset)));
111 Coordinates src_offset;
112 // Set up input coords to read the row specified by the current index
113 src_offset.set(0, 0);
114 src_offset.set(1, row);
115 for(uint32_t j = 2; j < _input->info()->num_dimensions(); ++j)
116 {
117 src_offset.set(j, id[1 + _indices->info()->num_dimensions() + (j - 2)]);
118 }
119 const auto in_ptr_row = _input->ptr_to_element(src_offset);
120 // Copy a row from input to output
121 memcpy(dst_ptr, in_ptr_row, _input->info()->tensor_shape()[0] * _input->info()->element_size());
122 });
123 }
124
125 template <typename U>
gather_0_axis(const Window & window,const ThreadInfo & info)126 inline void NEGatherKernel::gather_0_axis(const Window &window, const ThreadInfo &info)
127 {
128 ARM_COMPUTE_UNUSED(info);
129
130 // Validate that the indices are not negative
131 validate_indices<U>(_indices);
132
133 Iterator output_it(_output, window);
134 execute_window_loop(window, [&](const Coordinates & id)
135 {
136 Coordinates gather_id(id);
137
138 auto new_index = *(reinterpret_cast<U *>(_indices->ptr_to_element(Coordinates(id[0]))));
139 gather_id.set(0, new_index);
140
141 std::copy_n(_input->ptr_to_element(gather_id), _output->info()->element_size(), output_it.ptr());
142 },
143 output_it);
144 }
145
146 template <typename U>
gather_n_axis(const Window & window,const ThreadInfo & info)147 void NEGatherKernel::gather_n_axis(const Window &window, const ThreadInfo &info)
148 {
149 ARM_COMPUTE_UNUSED(info);
150
151 // Validate that the indices are not negative
152 validate_indices<U>(_indices);
153
154 Window output_window{ window };
155 output_window.set(Window::DimX, Window::Dimension(0, 1, 1));
156
157 Iterator output_it(_output, output_window);
158 execute_window_loop(output_window, [&](const Coordinates & id)
159 {
160 Coordinates gather_id(id);
161
162 auto new_index = *(reinterpret_cast<U *>(_indices->ptr_to_element(Coordinates(id[_axis]))));
163 gather_id.set(_axis, new_index);
164
165 std::copy_n(_input->ptr_to_element(gather_id), _input->info()->dimension(0) * _output->info()->element_size(), output_it.ptr());
166 },
167 output_it);
168 }
169
configure(const ITensor * input,const ITensor * indices,ITensor * output,int axis)170 void NEGatherKernel::configure(const ITensor *input, const ITensor *indices, ITensor *output, int axis)
171 {
172 ARM_COMPUTE_ERROR_ON_NULLPTR(input, output, indices);
173 ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), indices->info(), output->info(), axis));
174
175 _input = input;
176 _indices = indices;
177 _output = output;
178 _axis = axis;
179
180 if(_axis < 0)
181 {
182 _axis += input->info()->num_dimensions();
183 }
184 ARM_COMPUTE_ERROR_ON(0 > _axis || _axis >= static_cast<int32_t>(input->info()->num_dimensions()));
185
186 if(indices->info()->num_dimensions() == 1u)
187 {
188 if(_axis == 0)
189 {
190 switch(_indices->info()->data_type())
191 {
192 case DataType::U32:
193 _func = &NEGatherKernel::gather_0_axis<uint32_t>;
194 break;
195 case DataType::S32:
196 _func = &NEGatherKernel::gather_0_axis<int32_t>;
197 break;
198 default:
199 ARM_COMPUTE_ERROR("Not supported");
200 break;
201 }
202 }
203 else
204 {
205 switch(_indices->info()->data_type())
206 {
207 case DataType::U32:
208 _func = &NEGatherKernel::gather_n_axis<uint32_t>;
209 break;
210 case DataType::S32:
211 _func = &NEGatherKernel::gather_n_axis<int32_t>;
212 break;
213 default:
214 ARM_COMPUTE_ERROR("Not supported");
215 break;
216 }
217 }
218 }
219 else
220 {
221 if(_axis == 1)
222 {
223 switch(_indices->info()->data_type())
224 {
225 case DataType::U32:
226 _func = &NEGatherKernel::gather_multiindices_1_axis<uint32_t>;
227 break;
228 case DataType::S32:
229 _func = &NEGatherKernel::gather_multiindices_1_axis<int32_t>;
230 break;
231 default:
232 ARM_COMPUTE_ERROR("Not supported");
233 break;
234 }
235 }
236 else
237 {
238 ARM_COMPUTE_ERROR("Not supported");
239 }
240 }
241
242 // Output auto initialization if not yet initialized
243 const TensorShape output_shape = arm_compute::misc::shape_calculator::compute_gather_shape(input->info()->tensor_shape(), indices->info()->tensor_shape(), _axis);
244 auto_init_if_empty(*output->info(), input->info()->clone()->set_tensor_shape(output_shape));
245
246 // Create window
247 Window win = calculate_max_window(*output->info(), Steps());
248
249 INEKernel::configure(win);
250 }
251
validate(const ITensorInfo * input,const ITensorInfo * indices,const ITensorInfo * output,int axis)252 Status NEGatherKernel::validate(const ITensorInfo *input, const ITensorInfo *indices, const ITensorInfo *output, int axis)
253 {
254 ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, indices, output, axis));
255 return Status{};
256 }
257
run(const Window & window,const ThreadInfo & info)258 void NEGatherKernel::run(const Window &window, const ThreadInfo &info)
259 {
260 ARM_COMPUTE_UNUSED(info);
261 ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
262 ARM_COMPUTE_ERROR_ON(_func == nullptr);
263
264 (this->*_func)(window, info);
265 }
266
267 } // namespace arm_compute
268