1 /*
2 * Copyright (c) 2018-2021 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/NEChannelShuffleLayerKernel.h"
25
26 #include "arm_compute/core/Error.h"
27 #include "arm_compute/core/Helpers.h"
28 #include "arm_compute/core/ITensor.h"
29 #include "arm_compute/core/TensorInfo.h"
30 #include "arm_compute/core/Utils.h"
31 #include "arm_compute/core/Validate.h"
32 #include "arm_compute/core/Window.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 {
validate_arguments(const ITensorInfo * input,const ITensorInfo * output,unsigned int num_groups)41 Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, unsigned int num_groups)
42 {
43 // Note: ARM_COMPUTE_RETURN_ERROR_ON_CPU_F16_UNSUPPORTED(input) is not needed here as this kernel doesn't use CPU FP16 instructions.
44 ARM_COMPUTE_RETURN_ERROR_ON(input->data_type() == DataType::UNKNOWN);
45 ARM_COMPUTE_RETURN_ERROR_ON_DATA_LAYOUT_NOT_IN(input, DataLayout::NCHW, DataLayout::NHWC);
46
47 const unsigned int channels = input->dimension(get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::CHANNEL));
48
49 ARM_COMPUTE_RETURN_ERROR_ON_MSG(num_groups < 2, "Channel shuffling with less than 2 groups would be inefficient");
50 ARM_COMPUTE_RETURN_ERROR_ON_MSG(num_groups == channels, "Channel shuffling with same number of groups as number of channels would be inefficient");
51 ARM_COMPUTE_RETURN_ERROR_ON(num_groups > channels); // There cannot be more groups than channels
52 ARM_COMPUTE_RETURN_ERROR_ON_MSG((channels % num_groups) != 0, "The number of channels must be a multiple of the number of groups");
53
54 // Checks performed when output is configured
55 if(output->total_size() != 0)
56 {
57 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input, output);
58 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
59 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_LAYOUT(input, output);
60 }
61
62 return Status{};
63 }
channel_shuffle_nhwc(const ITensor * input,ITensor * output,unsigned int num_groups,const Window & window)64 void channel_shuffle_nhwc(const ITensor *input, ITensor *output, unsigned int num_groups, const Window &window)
65 {
66 const DataLayout data_layout = input->info()->data_layout();
67 const unsigned int channel_idx = get_data_layout_dimension_index(data_layout, DataLayoutDimension::CHANNEL);
68
69 const size_t element_size = input->info()->element_size();
70 const unsigned int K = input->info()->dimension(channel_idx) / num_groups;
71 const double rK = 1.0 / K;
72
73 Iterator in(input, window);
74
75 execute_window_loop(window, [&](const Coordinates & id)
76 {
77 // Shuffle channel
78 const unsigned int curr_channel = id.x();
79 const unsigned int group_id = curr_channel * rK;
80 const unsigned int r = group_id * K;
81 const unsigned int channel_id = curr_channel - r;
82
83 // Calculate output coordinates
84 Coordinates out_coords = id;
85 out_coords.set(Window::DimX, channel_id * num_groups + group_id);
86 std::copy_n(in.ptr(), element_size, output->ptr_to_element(out_coords));
87 },
88 in);
89 }
channel_shuffle_nchw(const ITensor * input,ITensor * output,unsigned int num_groups,const Window & window)90 void channel_shuffle_nchw(const ITensor *input, ITensor *output, unsigned int num_groups, const Window &window)
91 {
92 Window win = window;
93 win.set(Window::DimX, Window::Dimension(0, 1, 1));
94 win.set(Window::DimY, Window::Dimension(0, 1, 1));
95
96 const DataLayout data_layout = input->info()->data_layout();
97 const unsigned int width_idx = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH);
98 const unsigned int channel_idx = get_data_layout_dimension_index(data_layout, DataLayoutDimension::CHANNEL);
99
100 const unsigned int height = input->info()->tensor_shape().y();
101 const size_t input_stride_y = input->info()->strides_in_bytes().y();
102 const size_t output_stride_y = output->info()->strides_in_bytes().y();
103 const size_t row_size = input->info()->dimension(width_idx) * input->info()->element_size();
104
105 const unsigned int K = input->info()->dimension(channel_idx) / num_groups;
106 const double rK = 1.0 / K;
107
108 Iterator in(input, win);
109
110 execute_window_loop(win, [&](const Coordinates & id)
111 {
112 // Shuffle channel
113 const unsigned int curr_channel = id.z();
114 const unsigned int group_id = curr_channel * rK;
115 const unsigned int r = group_id * K;
116 const unsigned int channel_id = curr_channel - r;
117
118 // Calculate output coordinates
119 Coordinates out_coords = id;
120 out_coords.set(Window::DimZ, channel_id * num_groups + group_id);
121 const uint8_t *input_ptr = in.ptr();
122 uint8_t *output_ptr = output->ptr_to_element(out_coords);
123
124 // Copy plane
125 for(unsigned int y = 0; y < height; ++y)
126 {
127 std::copy_n(input_ptr, row_size, output_ptr);
128 input_ptr += input_stride_y;
129 output_ptr += output_stride_y;
130 }
131 },
132 in);
133 }
134 } // namespace
135
NEChannelShuffleLayerKernel()136 NEChannelShuffleLayerKernel::NEChannelShuffleLayerKernel()
137 : _input(nullptr), _output(nullptr), _num_groups()
138 {
139 }
140
configure(const ITensor * input,ITensor * output,unsigned int num_groups)141 void NEChannelShuffleLayerKernel::configure(const ITensor *input, ITensor *output, unsigned int num_groups)
142 {
143 ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
144
145 // Output tensor auto initialization if not yet initialized
146 auto_init_if_empty(*output->info(), *input->info()->clone());
147
148 _input = input;
149 _output = output;
150 _num_groups = num_groups;
151
152 ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info(), num_groups));
153
154 // Configure kernel window
155 Window win = calculate_max_window(*input->info(), Steps());
156
157 // The NEChannelShuffleLayerKernel doesn't need padding so update_window_and_padding() can be skipped
158 INEKernel::configure(win);
159 }
160
validate(const ITensorInfo * input,const ITensorInfo * output,unsigned int num_groups)161 Status NEChannelShuffleLayerKernel::validate(const ITensorInfo *input, const ITensorInfo *output, unsigned int num_groups)
162 {
163 ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output, num_groups));
164 return Status{};
165 }
166
run(const Window & window,const ThreadInfo & info)167 void NEChannelShuffleLayerKernel::run(const Window &window, const ThreadInfo &info)
168 {
169 ARM_COMPUTE_UNUSED(info);
170 ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
171 ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(IKernel::window(), window);
172
173 switch(_input->info()->data_layout())
174 {
175 case DataLayout::NHWC:
176 channel_shuffle_nhwc(_input, _output, _num_groups, window);
177 break;
178 case DataLayout::NCHW:
179 channel_shuffle_nchw(_input, _output, _num_groups, window);
180 break;
181 default:
182 ARM_COMPUTE_ERROR("Unsupported data layout!");
183 break;
184 }
185 }
186 } // namespace arm_compute
187