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/NEDepthToSpaceLayerKernel.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 ARM_COMPUTE_RETURN_ERROR_ON(block_shape < 2);
50
51 const DataLayout data_layout = input->data_layout();
52 const int idx_channel = get_data_layout_dimension_index(data_layout, DataLayoutDimension::CHANNEL);
53 ARM_COMPUTE_RETURN_ERROR_ON(input->tensor_shape()[idx_channel] % (block_shape * block_shape) != 0);
54 // Validate output if initialized
55 if(output->total_size() != 0)
56 {
57 const int idx_width = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH);
58 const int idx_height = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT);
59 ARM_COMPUTE_RETURN_ERROR_ON(output->tensor_shape()[idx_width] != (block_shape * input->tensor_shape()[idx_width]));
60 ARM_COMPUTE_RETURN_ERROR_ON(output->tensor_shape()[idx_height] != (block_shape * input->tensor_shape()[idx_height]));
61 ARM_COMPUTE_RETURN_ERROR_ON(output->num_dimensions() > 4);
62 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
63 }
64
65 return Status{};
66 }
67 } // namespace
68
NEDepthToSpaceLayerKernel()69 NEDepthToSpaceLayerKernel::NEDepthToSpaceLayerKernel()
70 : _input(nullptr), _output(nullptr), _block_shape(), _data_layout(DataLayout::UNKNOWN)
71 {
72 }
73
configure(const ITensor * input,ITensor * output,int32_t block_shape)74 void NEDepthToSpaceLayerKernel::configure(const ITensor *input, ITensor *output, int32_t block_shape)
75 {
76 ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
77 TensorShape output_shape = compute_depth_to_space_shape(input->info()->tensor_shape(), input->info()->data_layout(), block_shape);
78 // Output auto inizialitation if not yet initialized
79 auto_init_if_empty(*output->info(), input->info()->clone()->set_tensor_shape(output_shape));
80
81 // Perform validation step
82 ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info(), block_shape));
83
84 _input = input;
85 _output = output;
86 _block_shape = block_shape;
87 _data_layout = input->info()->data_layout();
88
89 // Configure kernel window
90 Window win = calculate_max_window(*input->info(), Steps());
91 ICPPKernel::configure(win);
92 }
93
validate(const ITensorInfo * input,const ITensorInfo * output,int32_t block_shape)94 Status NEDepthToSpaceLayerKernel::validate(const ITensorInfo *input, const ITensorInfo *output, int32_t block_shape)
95 {
96 ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output);
97 ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output, block_shape));
98 return Status{};
99 }
100
run(const Window & window,const ThreadInfo & info)101 void NEDepthToSpaceLayerKernel::run(const Window &window, const ThreadInfo &info)
102 {
103 ARM_COMPUTE_UNUSED(info);
104 ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
105 ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICPPKernel::window(), window);
106
107 const int idx_channel = get_data_layout_dimension_index(_data_layout, DataLayoutDimension::CHANNEL);
108 const int depth_size = _input->info()->dimension(idx_channel);
109 const int r = (depth_size / (_block_shape * _block_shape));
110 const int element_size = _input->info()->element_size();
111
112 Window slice_out = window.first_slice_window_3D();
113
114 // The slice_out slice does not move
115 slice_out.set(Window::DimX, Window::Dimension(0, 0, 0));
116 slice_out.set(Window::DimY, Window::Dimension(0, 0, 0));
117 slice_out.set(Window::DimZ, Window::Dimension(0, 0, 0));
118
119 // Main loop for NCHW and NHWC
120 if(_data_layout == DataLayout::NCHW)
121 {
122 Window slice_in = window.first_slice_window_2D();
123 do
124 {
125 Iterator in(_input, slice_in);
126 execute_window_loop(slice_in, [&](const Coordinates & id)
127 {
128 const int x = id.x();
129 const int y = id.y();
130
131 const int z = id.z() % r;
132 const int out_x = x * _block_shape + (id.z() / r) % _block_shape;
133 const int out_y = y * _block_shape + (id.z() / r) / _block_shape;
134 Coordinates output_coords{ out_x, out_y, z, id[3] };
135 memcpy(_output->ptr_to_element(output_coords), in.ptr(), element_size);
136 },
137 in);
138 }
139 while(window.slide_window_slice_2D(slice_in));
140 }
141 else
142 {
143 Window slice_in = window.first_slice_window_3D();
144 do
145 {
146 Iterator in(_input, slice_in);
147 execute_window_loop(slice_in, [&](const Coordinates & id)
148 {
149 const int x = id.y();
150 const int y = id.z();
151
152 const int z = id.x() % r;
153 const int out_x = x * _block_shape + (id.x() / r) % _block_shape;
154 const int out_y = y * _block_shape + (id.x() / r) / _block_shape;
155 Coordinates output_coords{ z, out_x, out_y, id[3] };
156 memcpy(_output->ptr_to_element(output_coords), in.ptr(), element_size);
157 },
158 in);
159 }
160 while(window.slide_window_slice_3D(slice_in));
161 }
162 }
163 } // namespace arm_compute
164