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/NESpaceToDepthLayerKernel.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 * output,int32_t block_shape)44 Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, int32_t block_shape)
45 {
46 ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output);
47 ARM_COMPUTE_RETURN_ERROR_ON(input->data_type() == DataType::UNKNOWN);
48 ARM_COMPUTE_RETURN_ERROR_ON(input->num_dimensions() > 4);
49
50 ARM_COMPUTE_RETURN_ERROR_ON(block_shape < 1);
51
52 // Validate output if initialized
53 if(output->total_size() != 0)
54 {
55 const DataLayout data_layout = input->data_layout();
56 const int idx_width = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH);
57 const int idx_height = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT);
58 const int idx_channel = get_data_layout_dimension_index(data_layout, DataLayoutDimension::CHANNEL);
59 const int idx_batch = get_data_layout_dimension_index(data_layout, DataLayoutDimension::BATCHES);
60 ARM_COMPUTE_RETURN_ERROR_ON(input->tensor_shape()[idx_width] % block_shape != 0);
61 ARM_COMPUTE_RETURN_ERROR_ON(input->tensor_shape()[idx_height] % block_shape != 0);
62 ARM_COMPUTE_RETURN_ERROR_ON(input->tensor_shape()[idx_batch] != output->tensor_shape()[idx_batch]);
63 ARM_COMPUTE_RETURN_ERROR_ON(output->tensor_shape()[idx_channel] % (block_shape * block_shape) != 0);
64 ARM_COMPUTE_RETURN_ERROR_ON(input->tensor_shape().total_size() != output->tensor_shape().total_size());
65 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
66 }
67
68 return Status{};
69 }
70 } // namespace
71
NESpaceToDepthLayerKernel()72 NESpaceToDepthLayerKernel::NESpaceToDepthLayerKernel()
73 : _input(nullptr), _output(nullptr), _block_shape(), _data_layout(DataLayout::UNKNOWN)
74 {
75 }
76
configure(const ITensor * input,ITensor * output,int32_t block_shape)77 void NESpaceToDepthLayerKernel::configure(const ITensor *input, ITensor *output, int32_t block_shape)
78 {
79 ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
80
81 TensorShape output_shape = misc::shape_calculator::compute_space_to_depth_shape(input->info(), block_shape);
82 auto_init_if_empty(*output->info(), output_shape, 1, input->info()->data_type());
83
84 ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info(), block_shape));
85
86 _input = input;
87 _block_shape = block_shape;
88 _output = output;
89 _data_layout = input->info()->data_layout();
90
91 // Configure kernel window
92 Window win = calculate_max_window(*output->info(), Steps());
93 INEKernel::configure(win);
94 }
95
validate(const ITensorInfo * input,const ITensorInfo * output,int32_t block_shape)96 Status NESpaceToDepthLayerKernel::validate(const ITensorInfo *input, const ITensorInfo *output, int32_t block_shape)
97 {
98 ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output, block_shape));
99 return Status{};
100 }
101
run(const Window & window,const ThreadInfo & info)102 void NESpaceToDepthLayerKernel::run(const Window &window, const ThreadInfo &info)
103 {
104 ARM_COMPUTE_UNUSED(info);
105 ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
106 ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICPPKernel::window(), window);
107
108 const int channel_idx = get_data_layout_dimension_index(_data_layout, DataLayoutDimension::CHANNEL);
109 const int element_size = _input->info()->element_size();
110
111 const size_t channel_size = _input->info()->dimension(channel_idx);
112
113 Window slice_out = window.first_slice_window_3D();
114
115 int batch_id = 0;
116
117 // Main loop for NCHW and NHWC
118 if(_data_layout == DataLayout::NCHW)
119 {
120 do
121 {
122 Iterator out(_output, slice_out);
123 execute_window_loop(slice_out, [&](const Coordinates & id)
124 {
125 const size_t channel_id = id.z();
126 const size_t in_x = id.x() * _block_shape + (channel_id / channel_size) % _block_shape;
127 const size_t in_y = id.y() * _block_shape + (channel_id / channel_size) / _block_shape;
128 const int z = channel_id % channel_size;
129 Coordinates input_coords{ in_x, in_y, z, batch_id };
130 memcpy(out.ptr(), _input->ptr_to_element(input_coords), element_size);
131 },
132 out);
133 ++batch_id;
134 }
135 while(window.slide_window_slice_3D(slice_out));
136 }
137 else
138 {
139 do
140 {
141 Iterator out(_output, slice_out);
142 execute_window_loop(slice_out, [&](const Coordinates & id)
143 {
144 const size_t channel_id = id.x();
145 const size_t in_x = id.y() * _block_shape + (channel_id / channel_size) % _block_shape;
146 const size_t in_y = id.z() * _block_shape + (channel_id / channel_size) / _block_shape;
147 const int z = channel_id % channel_size;
148 Coordinates input_coords{ z, in_x, in_y, batch_id };
149 memcpy(out.ptr(), _input->ptr_to_element(input_coords), element_size);
150 },
151 out);
152 ++batch_id;
153 }
154 while(window.slide_window_slice_3D(slice_out));
155 }
156 }
157 } // namespace arm_compute
158