David Beck | 1b61be5 | 2018-11-08 09:19:14 +0000 | [diff] [blame] | 1 | // |
| 2 | // Copyright © 2017 Arm Ltd. All rights reserved. |
| 3 | // SPDX-License-Identifier: MIT |
| 4 | // |
| 5 | |
| 6 | #include "ClBackendContext.hpp" |
Derek Lamberti | 836b27b | 2019-11-20 10:51:57 +0000 | [diff] [blame] | 7 | #include "ClContextControl.hpp" |
Derek Lamberti | 0844697 | 2019-11-26 16:38:31 +0000 | [diff] [blame] | 8 | |
| 9 | #include <armnn/Logging.hpp> |
| 10 | |
David Beck | 1b61be5 | 2018-11-08 09:19:14 +0000 | [diff] [blame] | 11 | #include <arm_compute/core/CL/OpenCL.h> |
| 12 | #include <arm_compute/core/CL/CLKernelLibrary.h> |
| 13 | #include <arm_compute/runtime/CL/CLScheduler.h> |
Derek Lamberti | 836b27b | 2019-11-20 10:51:57 +0000 | [diff] [blame] | 14 | #include <arm_compute/runtime/CL/CLTunerTypes.h> |
| 15 | |
| 16 | #include <boost/polymorphic_cast.hpp> |
Aron Virginas-Tar | d46e647 | 2018-11-16 10:26:23 +0000 | [diff] [blame] | 17 | |
David Beck | 1b61be5 | 2018-11-08 09:19:14 +0000 | [diff] [blame] | 18 | namespace armnn |
| 19 | { |
| 20 | |
| 21 | struct ClBackendContext::ClContextControlWrapper |
| 22 | { |
Derek Lamberti | 836b27b | 2019-11-20 10:51:57 +0000 | [diff] [blame] | 23 | ClContextControlWrapper(arm_compute::CLTuner* tuner, |
David Beck | 1b61be5 | 2018-11-08 09:19:14 +0000 | [diff] [blame] | 24 | bool profilingEnabled) |
Derek Lamberti | 836b27b | 2019-11-20 10:51:57 +0000 | [diff] [blame] | 25 | : m_ClContextControl(tuner, profilingEnabled) |
David Beck | 1b61be5 | 2018-11-08 09:19:14 +0000 | [diff] [blame] | 26 | {} |
| 27 | |
| 28 | bool Sync() |
| 29 | { |
David Beck | 1b61be5 | 2018-11-08 09:19:14 +0000 | [diff] [blame] | 30 | if (arm_compute::CLScheduler::get().context()() != NULL) |
| 31 | { |
| 32 | // Waits for all queued CL requests to finish before unloading the network they may be using. |
| 33 | try |
| 34 | { |
| 35 | // Coverity fix: arm_compute::CLScheduler::sync() may throw an exception of type cl::Error. |
| 36 | arm_compute::CLScheduler::get().sync(); |
| 37 | } |
| 38 | catch (const cl::Error&) |
| 39 | { |
Derek Lamberti | 0844697 | 2019-11-26 16:38:31 +0000 | [diff] [blame] | 40 | ARMNN_LOG(warning) << "Runtime::UnloadNetwork(): an error occurred while waiting for " |
| 41 | "the queued CL requests to finish"; |
David Beck | 1b61be5 | 2018-11-08 09:19:14 +0000 | [diff] [blame] | 42 | return false; |
| 43 | } |
| 44 | } |
Aron Virginas-Tar | d46e647 | 2018-11-16 10:26:23 +0000 | [diff] [blame] | 45 | |
David Beck | 1b61be5 | 2018-11-08 09:19:14 +0000 | [diff] [blame] | 46 | return true; |
| 47 | } |
| 48 | |
| 49 | void ClearClCache() |
| 50 | { |
David Beck | 1b61be5 | 2018-11-08 09:19:14 +0000 | [diff] [blame] | 51 | if (arm_compute::CLScheduler::get().context()() != NULL) |
| 52 | { |
| 53 | // There are no loaded networks left, so clear the CL cache to free up memory |
| 54 | m_ClContextControl.ClearClCache(); |
| 55 | } |
David Beck | 1b61be5 | 2018-11-08 09:19:14 +0000 | [diff] [blame] | 56 | } |
| 57 | |
David Beck | 1b61be5 | 2018-11-08 09:19:14 +0000 | [diff] [blame] | 58 | ClContextControl m_ClContextControl; |
| 59 | }; |
| 60 | |
Derek Lamberti | 836b27b | 2019-11-20 10:51:57 +0000 | [diff] [blame] | 61 | std::string LowerString(std::string value) |
| 62 | { |
| 63 | std::transform(value.begin(), value.end(), value.begin(), |
| 64 | [](unsigned char c){ return std::tolower(c); }); |
| 65 | |
| 66 | return value; |
| 67 | } |
| 68 | |
| 69 | enum class TuningLevel |
| 70 | { |
| 71 | None, |
| 72 | Rapid, |
| 73 | Normal, |
| 74 | Exhaustive |
| 75 | }; |
| 76 | |
| 77 | |
| 78 | TuningLevel ParseTuningLevel(const BackendOptions::Var& value, TuningLevel defaultValue) |
| 79 | { |
| 80 | if (value.IsInt()) |
| 81 | { |
| 82 | int v = value.IsInt(); |
| 83 | if (v > static_cast<int>(TuningLevel::Exhaustive) || |
| 84 | v < static_cast<int>(TuningLevel::None)) |
| 85 | { |
| 86 | ARMNN_LOG(warning) << "Invalid GpuAcc tuning level ("<< v << ") selected. " |
| 87 | "Using default(" << static_cast<int>(defaultValue) << ")"; |
| 88 | } else |
| 89 | { |
| 90 | return static_cast<TuningLevel>(v); |
| 91 | } |
| 92 | } |
| 93 | return defaultValue; |
| 94 | } |
| 95 | |
| 96 | bool ParseBoolean(const BackendOptions::Var& value, bool defaultValue) |
| 97 | { |
| 98 | if (value.IsBool()) |
| 99 | { |
| 100 | return value.AsBool(); |
| 101 | } |
| 102 | |
| 103 | return defaultValue; |
| 104 | } |
| 105 | |
| 106 | std::string ParseFile(const BackendOptions::Var& value, std::string defaultValue) |
| 107 | { |
| 108 | if (value.IsString()) |
| 109 | { |
| 110 | return value.AsString(); |
| 111 | } |
| 112 | return defaultValue; |
| 113 | } |
| 114 | |
| 115 | template <typename F> |
| 116 | void ParseOptions(const std::vector<BackendOptions>& options, BackendId backend, F f) |
| 117 | { |
| 118 | for (auto optionsGroup : options) |
| 119 | { |
| 120 | if (optionsGroup.GetBackendId() == backend) |
| 121 | { |
| 122 | for (size_t i=0; i < optionsGroup.GetOptionCount(); i++) |
| 123 | { |
| 124 | const BackendOptions::BackendOption option = optionsGroup.GetOption(i); |
| 125 | f(option.GetName(), option.GetValue()); |
| 126 | } |
| 127 | } |
| 128 | } |
| 129 | } |
David Beck | 1b61be5 | 2018-11-08 09:19:14 +0000 | [diff] [blame] | 130 | |
| 131 | ClBackendContext::ClBackendContext(const IRuntime::CreationOptions& options) |
| 132 | : IBackendContext(options) |
David Beck | 1b61be5 | 2018-11-08 09:19:14 +0000 | [diff] [blame] | 133 | { |
Derek Lamberti | 836b27b | 2019-11-20 10:51:57 +0000 | [diff] [blame] | 134 | bool kernelProfiling = options.m_EnableGpuProfiling; |
| 135 | const TuningLevel defaultTuningLevel = TuningLevel::None; |
| 136 | auto tuningLevel = defaultTuningLevel; |
| 137 | m_TuningFile = ""; |
| 138 | |
| 139 | |
| 140 | arm_compute::CLTuner* tuner = nullptr; |
| 141 | if (m_TuningFile.empty() == false) |
| 142 | { |
| 143 | bool useLegacyTunerAPI = options.m_GpuAccTunedParameters.get() != nullptr; |
| 144 | if (useLegacyTunerAPI) |
| 145 | { |
| 146 | auto clTunerParams = boost::polymorphic_downcast<ClTunedParameters*>( |
| 147 | options.m_GpuAccTunedParameters.get()); |
| 148 | auto clTuner = &clTunerParams->m_Tuner; |
| 149 | |
| 150 | if (clTuner) |
| 151 | { |
| 152 | auto ConvertTuningLevel = [](IGpuAccTunedParameters::TuningLevel level) |
| 153 | { |
| 154 | switch(level) |
| 155 | { |
| 156 | case IGpuAccTunedParameters::TuningLevel::Rapid: |
| 157 | return arm_compute::CLTunerMode::RAPID; |
| 158 | case IGpuAccTunedParameters::TuningLevel::Normal: |
| 159 | return arm_compute::CLTunerMode::NORMAL; |
| 160 | case IGpuAccTunedParameters::TuningLevel::Exhaustive: |
| 161 | return arm_compute::CLTunerMode::EXHAUSTIVE; |
| 162 | default: |
| 163 | { |
| 164 | BOOST_ASSERT_MSG(false, "Tuning level not recognised."); |
| 165 | return arm_compute::CLTunerMode::NORMAL; |
| 166 | } |
| 167 | } |
| 168 | }; |
| 169 | |
| 170 | clTuner->set_tuner_mode(ConvertTuningLevel(clTunerParams->m_TuningLevel)); |
| 171 | clTuner->set_tune_new_kernels( |
| 172 | clTunerParams->m_Mode == armnn::IGpuAccTunedParameters::Mode::UpdateTunedParameters); |
| 173 | } |
| 174 | } |
| 175 | else //New backend options API |
| 176 | { |
| 177 | ParseOptions(options.m_BackendOptions, "GpuAcc", [&](std::string name, const BackendOptions::Var& value) |
| 178 | { |
| 179 | if (name == "KernelProfilingEnabled") |
| 180 | { |
| 181 | kernelProfiling |= ParseBoolean(value, false); |
| 182 | } else if (name == "TuningFile") |
| 183 | { |
| 184 | m_TuningFile = ParseFile(value, ""); |
| 185 | } else if (name == "TuningLevel") |
| 186 | { |
| 187 | tuningLevel = ParseTuningLevel(value, defaultTuningLevel); |
| 188 | } |
| 189 | }); |
| 190 | |
| 191 | // Create the tuner, in tuning mode initially. |
| 192 | m_Tuner = std::make_unique<arm_compute::CLTuner>(true); |
| 193 | |
| 194 | switch (tuningLevel) |
| 195 | { |
| 196 | case TuningLevel::Rapid: |
| 197 | m_Tuner->set_tuner_mode(arm_compute::CLTunerMode::RAPID); |
| 198 | break; |
| 199 | case TuningLevel::Normal: |
| 200 | m_Tuner->set_tuner_mode(arm_compute::CLTunerMode::NORMAL); |
| 201 | break; |
| 202 | case TuningLevel::Exhaustive: |
| 203 | m_Tuner->set_tuner_mode(arm_compute::CLTunerMode::EXHAUSTIVE); |
| 204 | break; |
| 205 | case TuningLevel::None: |
| 206 | default: |
| 207 | m_Tuner->set_tune_new_kernels(false); // Turn of tuning. Set to "use" only mode. |
| 208 | break; |
| 209 | } |
| 210 | |
| 211 | try |
| 212 | { |
| 213 | m_Tuner->load_from_file(m_TuningFile.c_str()); |
| 214 | } catch (const std::exception& e) |
| 215 | { |
| 216 | ARMNN_LOG(warning) << "Could not load GpuAcc tuner data file."; |
| 217 | } |
| 218 | |
| 219 | tuner = m_Tuner.get(); |
| 220 | } |
| 221 | } |
| 222 | |
| 223 | m_ClContextControlWrapper = std::make_unique<ClContextControlWrapper>( |
| 224 | tuner, |
| 225 | kernelProfiling |
| 226 | ); |
David Beck | 1b61be5 | 2018-11-08 09:19:14 +0000 | [diff] [blame] | 227 | } |
| 228 | |
| 229 | bool ClBackendContext::BeforeLoadNetwork(NetworkId) |
| 230 | { |
| 231 | return true; |
| 232 | } |
| 233 | |
| 234 | bool ClBackendContext::AfterLoadNetwork(NetworkId networkId) |
| 235 | { |
| 236 | { |
| 237 | std::lock_guard<std::mutex> lockGuard(m_Mutex); |
| 238 | m_NetworkIds.insert(networkId); |
| 239 | } |
| 240 | return true; |
| 241 | } |
| 242 | |
| 243 | bool ClBackendContext::BeforeUnloadNetwork(NetworkId) |
| 244 | { |
| 245 | return m_ClContextControlWrapper->Sync(); |
| 246 | } |
| 247 | |
| 248 | bool ClBackendContext::AfterUnloadNetwork(NetworkId networkId) |
| 249 | { |
| 250 | bool clearCache = false; |
| 251 | { |
| 252 | std::lock_guard<std::mutex> lockGuard(m_Mutex); |
| 253 | m_NetworkIds.erase(networkId); |
| 254 | clearCache = m_NetworkIds.empty(); |
| 255 | } |
| 256 | |
| 257 | if (clearCache) |
| 258 | { |
| 259 | m_ClContextControlWrapper->ClearClCache(); |
| 260 | } |
| 261 | |
| 262 | return true; |
| 263 | } |
| 264 | |
| 265 | ClBackendContext::~ClBackendContext() |
| 266 | { |
Derek Lamberti | 836b27b | 2019-11-20 10:51:57 +0000 | [diff] [blame] | 267 | if (m_Tuner && !m_TuningFile.empty()) |
| 268 | { |
| 269 | try |
| 270 | { |
| 271 | m_Tuner->save_to_file(m_TuningFile.c_str()); |
| 272 | } |
| 273 | catch(const std::exception& e) |
| 274 | { |
| 275 | ARMNN_LOG(warning) << "Could not save GpuAcc tuner data to file " << m_TuningFile; |
| 276 | } |
| 277 | } |
David Beck | 1b61be5 | 2018-11-08 09:19:14 +0000 | [diff] [blame] | 278 | } |
| 279 | |
| 280 | } // namespace armnn |