giuros01 | fc1da13 | 2019-02-18 16:48:35 +0000 | [diff] [blame] | 1 | /* |
| 2 | * Copyright (c) 2019 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 "arm_compute/core/NEON/kernels/NEBatchToSpaceLayerKernel.h" |
| 25 | |
| 26 | #include "arm_compute/core/Helpers.h" |
| 27 | #include "arm_compute/core/ITensor.h" |
| 28 | #include "arm_compute/core/NEON/wrapper/wrapper.h" |
| 29 | #include "arm_compute/core/Types.h" |
| 30 | #include "arm_compute/core/Validate.h" |
| 31 | #include "arm_compute/core/utils/misc/ShapeCalculator.h" |
| 32 | #include <arm_neon.h> |
| 33 | #include <cstdint> |
| 34 | |
| 35 | using namespace arm_compute::misc::shape_calculator; |
| 36 | |
| 37 | namespace arm_compute |
| 38 | { |
| 39 | namespace |
| 40 | { |
| 41 | Status validate_arguments(const ITensorInfo *input, const ITensorInfo *block_info, const ITensorInfo *output) |
| 42 | { |
| 43 | ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, block_info, output); |
| 44 | ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(block_info, 1, DataType::S32); |
| 45 | ARM_COMPUTE_RETURN_ERROR_ON(input->num_dimensions() > 4); |
| 46 | |
| 47 | // Validate output if initialized |
| 48 | if(output->total_size() != 0) |
| 49 | { |
| 50 | ARM_COMPUTE_RETURN_ERROR_ON(output->num_dimensions() > 4); |
| 51 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); |
| 52 | } |
| 53 | |
| 54 | return Status{}; |
| 55 | } |
| 56 | Status validate_arguments_static(const ITensorInfo *input, const int block_shape_x, const int block_shape_y, const ITensorInfo *output) |
| 57 | { |
| 58 | ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output); |
| 59 | ARM_COMPUTE_RETURN_ERROR_ON(input->num_dimensions() > 4); |
| 60 | ARM_COMPUTE_RETURN_ERROR_ON(block_shape_x <= 0); |
| 61 | ARM_COMPUTE_RETURN_ERROR_ON(block_shape_y <= 0); |
| 62 | |
| 63 | const DataLayout data_layout = input->data_layout(); |
| 64 | const int idx_batch = get_data_layout_dimension_index(data_layout, DataLayoutDimension::BATCHES); |
| 65 | ARM_COMPUTE_RETURN_ERROR_ON(input->tensor_shape()[idx_batch] % (block_shape_x * block_shape_y) != 0); |
giuros01 | fc1da13 | 2019-02-18 16:48:35 +0000 | [diff] [blame] | 66 | // Validate output if initialized |
| 67 | if(output->total_size() != 0) |
| 68 | { |
| 69 | const int idx_width = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH); |
| 70 | const int idx_height = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT); |
| 71 | const int idx_channel = get_data_layout_dimension_index(data_layout, DataLayoutDimension::CHANNEL); |
| 72 | ARM_COMPUTE_RETURN_ERROR_ON(output->tensor_shape()[idx_width] != (block_shape_x * input->tensor_shape()[idx_width])); |
Pablo Tello | 0a33a21 | 2019-04-23 12:06:19 +0100 | [diff] [blame] | 73 | ARM_COMPUTE_RETURN_ERROR_ON(output->tensor_shape()[idx_height] != (block_shape_y * input->tensor_shape()[idx_height])); |
giuros01 | fc1da13 | 2019-02-18 16:48:35 +0000 | [diff] [blame] | 74 | ARM_COMPUTE_RETURN_ERROR_ON(output->tensor_shape()[idx_channel] != input->tensor_shape()[idx_channel]); |
| 75 | ARM_COMPUTE_RETURN_ERROR_ON(output->num_dimensions() > 4); |
| 76 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); |
| 77 | } |
| 78 | |
| 79 | return Status{}; |
| 80 | } |
| 81 | } // namespace |
| 82 | |
| 83 | NEBatchToSpaceLayerKernel::NEBatchToSpaceLayerKernel() |
| 84 | : _input(nullptr), _block_shape(nullptr), _output(nullptr), _block_shape_x(), _block_shape_y() |
| 85 | { |
| 86 | } |
| 87 | |
| 88 | void NEBatchToSpaceLayerKernel::configure(const ITensor *input, const ITensor *block_shape, ITensor *output) |
| 89 | { |
| 90 | ARM_COMPUTE_ERROR_ON_NULLPTR(input, output); |
| 91 | ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), block_shape->info(), output->info())); |
| 92 | |
| 93 | _input = input; |
| 94 | _block_shape = block_shape; |
| 95 | _output = output; |
| 96 | |
| 97 | // Configure kernel window |
| 98 | Window win = calculate_max_window(*input->info(), Steps()); |
| 99 | ICPPKernel::configure(win); |
| 100 | } |
| 101 | |
| 102 | void NEBatchToSpaceLayerKernel::configure(const ITensor *input, const int32_t block_shape_x, const int32_t block_shape_y, ITensor *output) |
| 103 | { |
| 104 | ARM_COMPUTE_ERROR_ON_NULLPTR(input, output); |
| 105 | TensorShape output_shape = compute_batch_to_space_shape(input->info(), block_shape_x, block_shape_y); |
| 106 | // Output auto inizialitation if not yet initialized |
| 107 | auto_init_if_empty(*output->info(), input->info()->clone()->set_tensor_shape(output_shape)); |
| 108 | |
| 109 | // Perform validation step |
| 110 | ARM_COMPUTE_ERROR_THROW_ON(validate_arguments_static(input->info(), block_shape_x, block_shape_y, output->info())); |
| 111 | |
| 112 | _input = input; |
| 113 | _output = output; |
| 114 | _block_shape_x = block_shape_x; |
| 115 | _block_shape_y = block_shape_y; |
| 116 | |
| 117 | // Configure kernel window |
| 118 | Window win = calculate_max_window(*input->info(), Steps()); |
| 119 | ICPPKernel::configure(win); |
| 120 | } |
| 121 | |
| 122 | Status NEBatchToSpaceLayerKernel::validate(const ITensorInfo *input, const ITensorInfo *block_shape, const ITensorInfo *output) |
| 123 | { |
| 124 | ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, block_shape, output); |
| 125 | ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, block_shape, output)); |
| 126 | return Status{}; |
| 127 | } |
| 128 | |
| 129 | Status NEBatchToSpaceLayerKernel::validate(const ITensorInfo *input, const int32_t block_shape_x, const int32_t block_shape_y, const ITensorInfo *output) |
| 130 | { |
| 131 | ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output); |
| 132 | ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments_static(input, block_shape_x, block_shape_y, output)); |
| 133 | return Status{}; |
| 134 | } |
| 135 | |
| 136 | void NEBatchToSpaceLayerKernel::run(const Window &window, const ThreadInfo &info) |
| 137 | { |
| 138 | ARM_COMPUTE_UNUSED(info); |
| 139 | ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); |
| 140 | ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICPPKernel::window(), window); |
| 141 | |
| 142 | if(_block_shape != nullptr) |
| 143 | { |
| 144 | // Retrieve the block shapes dynamically |
| 145 | _block_shape_x = *(reinterpret_cast<const int *>(_block_shape->ptr_to_element(0))); |
| 146 | _block_shape_y = *(reinterpret_cast<const int *>(_block_shape->ptr_to_element(1))); |
| 147 | } |
| 148 | |
| 149 | const int batch_size = _input->info()->dimension(3); |
| 150 | const int r = (batch_size / (_block_shape_x * _block_shape_y)); |
| 151 | const int element_size = _input->info()->element_size(); |
| 152 | |
| 153 | Window slice_in = window.first_slice_window_3D(); |
| 154 | Window slice_out = window.first_slice_window_4D(); |
| 155 | |
| 156 | // The slice_out slice does not move |
| 157 | slice_out.set(Window::DimX, Window::Dimension(0, 0, 0)); |
| 158 | slice_out.set(Window::DimY, Window::Dimension(0, 0, 0)); |
| 159 | slice_out.set(Window::DimZ, Window::Dimension(0, 0, 0)); |
| 160 | slice_out.set(3, Window::Dimension(0, 0, 0)); |
| 161 | |
| 162 | int batch_id = 0; |
| 163 | // Main loop for NCHW and NHWC |
| 164 | if(_input->info()->data_layout() == DataLayout::NCHW) |
| 165 | { |
| 166 | do |
| 167 | { |
| 168 | Iterator in(_input, slice_in); |
| 169 | execute_window_loop(slice_in, [&](const Coordinates & id) |
| 170 | { |
| 171 | |
| 172 | const int x = id.x(); |
| 173 | const int y = id.y(); |
| 174 | const int z = id.z(); |
| 175 | |
| 176 | const int w = batch_id % r; |
| 177 | const int out_x = x * _block_shape_x + (batch_id / r) % _block_shape_x; |
| 178 | const int out_y = y * _block_shape_y + (batch_id / r) / _block_shape_x; |
| 179 | Coordinates output_coords{ out_x, out_y, z, w }; |
| 180 | memcpy(_output->ptr_to_element(output_coords), in.ptr(), element_size); |
| 181 | }, |
| 182 | in); |
| 183 | ++batch_id; |
| 184 | } |
| 185 | while(window.slide_window_slice_3D(slice_in)); |
| 186 | } |
| 187 | else |
| 188 | { |
| 189 | do |
| 190 | { |
| 191 | Iterator in(_input, slice_in); |
| 192 | execute_window_loop(slice_in, [&](const Coordinates & id) |
| 193 | { |
| 194 | |
| 195 | const int z = id.x(); |
| 196 | const int x = id.y(); |
| 197 | const int y = id.z(); |
| 198 | |
| 199 | const int w = batch_id % r; |
| 200 | const int out_x = x * _block_shape_x + (batch_id / r) % _block_shape_x; |
| 201 | const int out_y = y * _block_shape_y + (batch_id / r) / _block_shape_x; |
| 202 | Coordinates output_coords{ z, out_x, out_y, w }; |
| 203 | memcpy(_output->ptr_to_element(output_coords), in.ptr(), element_size); |
| 204 | }, |
| 205 | in); |
| 206 | ++batch_id; |
| 207 | } |
| 208 | while(window.slide_window_slice_3D(slice_in)); |
| 209 | } |
| 210 | } |
| 211 | } // namespace arm_compute |