Kaizen | 8938bd3 | 2017-09-28 14:38:23 +0100 | [diff] [blame^] | 1 | /* |
| 2 | * Copyright (c) 2017 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/graph/Graph.h" |
| 25 | |
| 26 | #include "arm_compute/graph/CL/CLMap.h" |
| 27 | #include "arm_compute/graph/CL/CLUnmap.h" |
| 28 | #include "arm_compute/graph/INode.h" |
| 29 | #include "arm_compute/graph/Tensor.h" |
| 30 | #include "arm_compute/runtime/CL/CLTensor.h" |
| 31 | #include "arm_compute/runtime/Tensor.h" |
| 32 | |
| 33 | using namespace arm_compute::graph; |
| 34 | |
| 35 | struct Stage |
| 36 | { |
| 37 | Tensor *_input; |
| 38 | Tensor *_output; |
| 39 | std::unique_ptr<arm_compute::IFunction> _function; |
| 40 | }; |
| 41 | |
| 42 | struct Graph::Private |
| 43 | { |
| 44 | public: |
| 45 | /** Finalizes the current node's configuration |
| 46 | * |
| 47 | * @param _next_hint Device execution hint |
| 48 | */ |
| 49 | void configure(Hint _next_hint); |
| 50 | |
| 51 | /** Sets whether to enable information print out |
| 52 | * |
| 53 | * @param[in] is_enabled Set to true if need info printed out |
| 54 | */ |
| 55 | void set_info_enablement(bool is_enabled); |
| 56 | |
| 57 | std::vector<Stage> _pipeline{}; |
| 58 | std::vector<std::unique_ptr<Tensor>> _tensors{}; |
| 59 | std::vector<std::unique_ptr<INode>> _nodes{}; |
| 60 | Hint _current_hint{ Hint::DONT_CARE }; |
| 61 | Hint _next_hint{ Hint::DONT_CARE }; |
| 62 | std::unique_ptr<Tensor> _graph_input{ nullptr }; |
| 63 | std::unique_ptr<Tensor> _graph_output{ nullptr }; |
| 64 | std::unique_ptr<INode> _current_node{ nullptr }; |
| 65 | Tensor *_current_output{ nullptr }; |
| 66 | bool _info_enabled{ false }; |
| 67 | |
| 68 | private: |
| 69 | Tensor *_current_input{ nullptr }; |
| 70 | Hint _previous_hint{ Hint::DONT_CARE }; |
| 71 | }; |
| 72 | |
| 73 | Graph::~Graph() //NOLINT |
| 74 | { |
| 75 | //Can't use =default because the destructor must be defined after Graph::Private's definition |
| 76 | } |
| 77 | |
| 78 | Graph::Graph() |
| 79 | : _pimpl{ new Private() } |
| 80 | { |
| 81 | } |
| 82 | |
| 83 | void Graph::run() |
| 84 | { |
| 85 | while(true) |
| 86 | { |
| 87 | if(!_pimpl->_graph_input->call_accessor()) |
| 88 | { |
| 89 | return; |
| 90 | } |
| 91 | |
| 92 | for(auto &stage : _pimpl->_pipeline) |
| 93 | { |
| 94 | stage._function->run(); |
| 95 | } |
| 96 | |
| 97 | if(!_pimpl->_graph_output->call_accessor()) |
| 98 | { |
| 99 | return; |
| 100 | } |
| 101 | } |
| 102 | } |
| 103 | |
| 104 | //Finalize current node's configuration |
| 105 | void Graph::Private::configure(Hint _next_hint) |
| 106 | { |
| 107 | ARM_COMPUTE_ERROR_ON(_current_node == nullptr); |
| 108 | ARM_COMPUTE_ERROR_ON(_graph_input == nullptr); |
| 109 | |
| 110 | // Is it the first node of the graph ? |
| 111 | if(_current_input == nullptr) |
| 112 | { |
| 113 | _graph_input->set_target(_current_hint); |
| 114 | _current_input = _graph_input.get(); |
| 115 | _previous_hint = _current_hint; // For the first node just assume the previous node was of the same type as this one |
| 116 | } |
| 117 | |
| 118 | //Automatic output configuration ? |
| 119 | if(_current_output == nullptr) |
| 120 | { |
| 121 | _tensors.push_back(arm_compute::support::cpp14::make_unique<Tensor>(TensorInfo())); |
| 122 | _current_output = _tensors.back().get(); |
| 123 | } |
| 124 | |
| 125 | // If either the writer or reader node needs OpenCL then use OpenCL memory: |
| 126 | if((_next_hint == Hint::OPENCL || _current_hint == Hint::OPENCL)) |
| 127 | { |
| 128 | _current_output->set_target(Hint::OPENCL); |
| 129 | } |
| 130 | else |
| 131 | { |
| 132 | _current_output->set_target(Hint::NEON); |
| 133 | } |
| 134 | |
| 135 | // Map input if needed |
| 136 | std::unique_ptr<arm_compute::IFunction> func = _current_node->instantiate_node(_current_hint, _current_input->tensor(), _current_output->tensor()); |
| 137 | _current_input->allocate(); |
| 138 | |
| 139 | if(_current_input->target() == Hint::OPENCL) |
| 140 | { |
| 141 | if(_previous_hint == Hint::NEON) |
| 142 | { |
| 143 | ARM_COMPUTE_ERROR_ON(_current_hint == Hint::NEON); |
| 144 | _pipeline.push_back({ _current_input, _current_input, arm_compute::support::cpp14::make_unique<CLUnmap>(_current_input) }); |
| 145 | } |
| 146 | if(_current_hint == Hint::NEON) |
| 147 | { |
| 148 | ARM_COMPUTE_ERROR_ON(_previous_hint == Hint::NEON); |
| 149 | _pipeline.push_back({ _current_input, _current_input, arm_compute::support::cpp14::make_unique<CLMap>(_current_input, true) }); |
| 150 | } |
| 151 | } |
| 152 | |
| 153 | _pipeline.push_back({ _current_input, _current_output, std::move(func) }); |
| 154 | |
| 155 | _current_input = _current_output; |
| 156 | _current_output = nullptr; |
| 157 | _previous_hint = _current_hint; |
| 158 | _current_hint = _next_hint; |
| 159 | } |
| 160 | |
| 161 | void Graph::Private::set_info_enablement(bool is_enabled) |
| 162 | { |
| 163 | _info_enabled = is_enabled; |
| 164 | } |
| 165 | |
| 166 | void Graph::add_node(std::unique_ptr<INode> node) |
| 167 | { |
| 168 | ARM_COMPUTE_ERROR_ON_MSG(_pimpl->_graph_input == nullptr, "The graph's input must be set before the first node is added"); |
| 169 | ARM_COMPUTE_ERROR_ON_MSG(_pimpl->_graph_output != nullptr, "Nothing can be added after the output tensor"); |
| 170 | //Trigger the creation of the current Node: |
| 171 | |
| 172 | Hint _next_hint = node->override_hint(_pimpl->_next_hint); |
| 173 | ARM_COMPUTE_ERROR_ON(_next_hint == Hint::DONT_CARE); |
| 174 | if(_pimpl->_current_node) |
| 175 | { |
| 176 | //Finalize the previous Node: |
| 177 | _pimpl->configure(_pimpl->_next_hint); |
| 178 | |
| 179 | if(_pimpl->_info_enabled) |
| 180 | { |
| 181 | _pimpl->_current_node->print_info(); |
| 182 | } |
| 183 | } |
| 184 | else |
| 185 | { |
| 186 | // If that's the first node then use the same Hint before and after the node. |
| 187 | _pimpl->_current_hint = _next_hint; |
| 188 | } |
| 189 | if(_pimpl->_current_node) |
| 190 | { |
| 191 | _pimpl->_nodes.push_back(std::move(_pimpl->_current_node)); |
| 192 | } |
| 193 | _pimpl->_current_node = std::move(node); |
| 194 | } |
| 195 | void Graph::set_hint(Hint hint) |
| 196 | { |
| 197 | _pimpl->_next_hint = hint; |
| 198 | } |
| 199 | |
| 200 | void Graph::set_info_enablement(bool is_enabled) |
| 201 | { |
| 202 | _pimpl->set_info_enablement(is_enabled); |
| 203 | } |
| 204 | |
| 205 | //Add a tensor with an Accessor (i.e either the input or output of the graph) |
| 206 | void Graph::add_tensor(std::unique_ptr<Tensor> tensor) |
| 207 | { |
| 208 | // If it's the first Tensor added then it will be the input of the Graph. |
| 209 | if(_pimpl->_graph_input == nullptr) |
| 210 | { |
| 211 | ARM_COMPUTE_ERROR_ON(_pimpl->_graph_output != nullptr); |
| 212 | ARM_COMPUTE_ERROR_ON(_pimpl->_current_node != nullptr); |
| 213 | _pimpl->_graph_input = std::move(tensor); |
| 214 | } |
| 215 | else |
| 216 | { |
| 217 | // Else it will be the output of the Graph |
| 218 | ARM_COMPUTE_ERROR_ON(_pimpl->_graph_output != nullptr); |
| 219 | ARM_COMPUTE_ERROR_ON(_pimpl->_current_node == nullptr); |
| 220 | _pimpl->_graph_output = std::move(tensor); |
| 221 | _pimpl->_current_output = _pimpl->_graph_output.get(); |
| 222 | |
| 223 | // Finalize the graph by configuring the last Node of the graph: |
| 224 | _pimpl->configure(_pimpl->_current_hint); // Ignore _next_hint as this is the last node, and just use the same hint as before this node. |
| 225 | _pimpl->_graph_output->allocate(); |
| 226 | } |
| 227 | } |
| 228 | |
| 229 | void Graph::set_temp(TensorInfo &&tmp) |
| 230 | { |
| 231 | ARM_COMPUTE_ERROR_ON(_pimpl->_graph_input == nullptr); |
| 232 | ARM_COMPUTE_ERROR_ON(_pimpl->_graph_output != nullptr); |
| 233 | ARM_COMPUTE_ERROR_ON_MSG(_pimpl->_current_output != nullptr, "TensorInfo for temporary tensor already set"); |
| 234 | |
| 235 | _pimpl->_tensors.push_back(arm_compute::support::cpp14::make_unique<Tensor>(std::move(tmp))); |
| 236 | _pimpl->_current_output = _pimpl->_tensors.back().get(); |
| 237 | } |
| 238 | |
| 239 | Graph &arm_compute::graph::operator<<(Graph &graph, TensorInfo &&info) |
| 240 | { |
| 241 | graph.set_temp(std::move(info)); |
| 242 | return graph; |
| 243 | } |
| 244 | |
| 245 | Graph &arm_compute::graph::operator<<(Graph &graph, Tensor &&tensor) |
| 246 | { |
| 247 | graph.add_tensor(arm_compute::support::cpp14::make_unique<Tensor>(std::move(tensor))); |
| 248 | return graph; |
| 249 | } |
| 250 | |
| 251 | Graph &arm_compute::graph::operator<<(Graph &graph, Hint hint) |
| 252 | { |
| 253 | graph.set_hint(hint); |
| 254 | return graph; |
| 255 | } |