David Monahan | 8a57046 | 2023-11-22 13:24:25 +0000 | [diff] [blame] | 1 | // |
| 2 | // Copyright © 2023 Arm Ltd and Contributors. All rights reserved. |
| 3 | // SPDX-License-Identifier: MIT |
| 4 | // |
| 5 | |
| 6 | #include "GpuFsaBackendContext.hpp" |
| 7 | #include "GpuFsaContextControl.hpp" |
| 8 | |
| 9 | #include <armnn/utility/Assert.hpp> |
| 10 | #include <armnn/utility/PolymorphicDowncast.hpp> |
| 11 | |
| 12 | #include <arm_compute/core/CL/OpenCL.h> |
| 13 | #include <arm_compute/core/CL/CLKernelLibrary.h> |
| 14 | #include <arm_compute/runtime/CL/CLScheduler.h> |
| 15 | #include <arm_compute/runtime/CL/CLTunerTypes.h> |
| 16 | |
| 17 | namespace armnn |
| 18 | { |
| 19 | |
| 20 | struct GpuFsaBackendContext::GpuFsaContextControlWrapper |
| 21 | { |
| 22 | GpuFsaContextControlWrapper(arm_compute::CLTuner* tuner, |
| 23 | arm_compute::CLGEMMHeuristicsHandle* heuristicsHandle, |
| 24 | bool profilingEnabled) |
| 25 | : m_GpuFsaContextControl(tuner, heuristicsHandle, profilingEnabled) |
| 26 | {} |
| 27 | |
| 28 | bool Sync() |
| 29 | { |
| 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& err) |
| 39 | { |
| 40 | ARMNN_LOG(warning) << "Runtime::UnloadNetwork(): an error occurred while waiting for " |
| 41 | "the queued CL requests to finish"; |
| 42 | throw err; |
| 43 | } |
| 44 | } |
| 45 | |
| 46 | return true; |
| 47 | } |
| 48 | |
| 49 | void ClearClCache() |
| 50 | { |
| 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_GpuFsaContextControl.ClearClCache(); |
| 55 | } |
| 56 | } |
| 57 | |
| 58 | GpuFsaContextControl m_GpuFsaContextControl; |
| 59 | }; |
| 60 | |
| 61 | GpuFsaBackendContext::GpuFsaBackendContext(const IRuntime::CreationOptions& options) |
| 62 | : IBackendContext(options) |
| 63 | , m_TuningFile() |
| 64 | { |
| 65 | bool kernelProfiling = options.m_EnableGpuProfiling; |
| 66 | |
| 67 | arm_compute::CLTuner* tuner = nullptr; |
| 68 | arm_compute::CLGEMMHeuristicsHandle* mlgoTuner = nullptr; |
| 69 | bool useLegacyTunerAPI = options.m_GpuAccTunedParameters.get() != nullptr; |
| 70 | if (useLegacyTunerAPI) |
| 71 | { |
| 72 | auto clTunerParams = PolymorphicDowncast<ClTunedParameters*>( |
| 73 | options.m_GpuAccTunedParameters.get()); |
| 74 | tuner = &clTunerParams->m_Tuner; |
| 75 | |
| 76 | if (tuner) |
| 77 | { |
| 78 | auto ConvertTuningLevel = [](IGpuAccTunedParameters::TuningLevel level, |
| 79 | armnn::IGpuAccTunedParameters::Mode mode) |
| 80 | { |
| 81 | if (mode == armnn::IGpuAccTunedParameters::Mode::UseTunedParameters) |
| 82 | { |
| 83 | return TuningLevel::None; |
| 84 | } |
| 85 | |
| 86 | switch(level) |
| 87 | { |
| 88 | case IGpuAccTunedParameters::TuningLevel::Rapid: |
| 89 | return TuningLevel::Rapid; |
| 90 | case IGpuAccTunedParameters::TuningLevel::Normal: |
| 91 | return TuningLevel::Normal; |
| 92 | case IGpuAccTunedParameters::TuningLevel::Exhaustive: |
| 93 | return TuningLevel::Exhaustive; |
| 94 | default: |
| 95 | { |
| 96 | ARMNN_LOG(warning) << "Tuning level not recognised."; |
| 97 | return TuningLevel::None; |
| 98 | } |
| 99 | } |
| 100 | }; |
| 101 | |
| 102 | TuningLevel tuningLevel = ConvertTuningLevel(clTunerParams->m_TuningLevel, clTunerParams->m_Mode); |
| 103 | ConfigureTuner(*tuner, tuningLevel); |
| 104 | } |
| 105 | } |
| 106 | else //New backend options API |
| 107 | { |
| 108 | const TuningLevel defaultTuningLevel = TuningLevel::None; |
| 109 | auto tuningLevel = defaultTuningLevel; |
| 110 | |
| 111 | ParseOptions(options.m_BackendOptions, "GpuFsa", [&](std::string name, const BackendOptions::Var& value) |
| 112 | { |
| 113 | if (name == "KernelProfilingEnabled") |
| 114 | { |
| 115 | kernelProfiling |= ParseBooleanBackendOption(value, false); |
| 116 | } else if (name == "TuningFile") |
| 117 | { |
| 118 | m_TuningFile = ParseStringBackendOption(value, ""); |
| 119 | } else if (name == "TuningLevel") |
| 120 | { |
| 121 | tuningLevel = ParseTuningLevel(value, defaultTuningLevel); |
| 122 | } |
| 123 | else if (name == "MLGOTuningFilePath") |
| 124 | { |
| 125 | m_MLGOTuningFile = ParseStringBackendOption(value, ""); |
| 126 | } |
| 127 | }); |
| 128 | |
| 129 | // Create the tuner, in tuning mode initially. |
| 130 | m_Tuner = std::make_unique<arm_compute::CLTuner>(true); |
| 131 | |
| 132 | ConfigureTuner(*(m_Tuner.get()), tuningLevel); |
| 133 | |
| 134 | if (!m_TuningFile.empty()) |
| 135 | { |
| 136 | try |
| 137 | { |
| 138 | ARMNN_LOG(info) << "Loading Gpu tuning data from file: " << m_TuningFile; |
| 139 | m_Tuner->load_from_file(m_TuningFile.c_str()); |
| 140 | } |
| 141 | catch (const std::exception& e) |
| 142 | { |
| 143 | // Warn if not tuning, otherwise tuning will generate new params |
| 144 | if (tuningLevel == TuningLevel::None) |
| 145 | { |
| 146 | ARMNN_LOG(warning) << "Could not load GpuFsa tuner data file."; |
| 147 | } |
| 148 | } |
| 149 | } |
| 150 | |
| 151 | if (!m_MLGOTuningFile.empty()) |
| 152 | { |
| 153 | try |
| 154 | { |
| 155 | ARMNN_LOG(info) << "Loading Gpu MLGO tuning data from file: " << m_TuningFile; |
| 156 | if(m_MLGOTuner.reload_from_file(m_MLGOTuningFile.c_str())) |
| 157 | { |
| 158 | mlgoTuner = &m_MLGOTuner; |
| 159 | } |
| 160 | } |
| 161 | catch (const std::exception& e) |
| 162 | { |
| 163 | ARMNN_LOG(warning) << "Could not load GpuFsa MLGO tuner data file."; |
| 164 | } |
| 165 | } |
| 166 | |
| 167 | tuner = m_Tuner.get(); |
| 168 | } |
| 169 | |
| 170 | m_GpuFsaContextControlWrapper = std::make_unique<GpuFsaContextControlWrapper>( |
| 171 | tuner, |
| 172 | mlgoTuner, |
| 173 | kernelProfiling |
| 174 | ); |
| 175 | } |
| 176 | |
| 177 | bool GpuFsaBackendContext::BeforeLoadNetwork(NetworkId) |
| 178 | { |
| 179 | return true; |
| 180 | } |
| 181 | |
| 182 | bool GpuFsaBackendContext::AfterLoadNetwork(NetworkId networkId) |
| 183 | { |
| 184 | { |
| 185 | std::lock_guard<std::mutex> lockGuard(m_Mutex); |
| 186 | m_NetworkIds.insert(networkId); |
| 187 | } |
| 188 | return true; |
| 189 | } |
| 190 | |
| 191 | bool GpuFsaBackendContext::BeforeUnloadNetwork(NetworkId) |
| 192 | { |
| 193 | return m_GpuFsaContextControlWrapper->Sync(); |
| 194 | } |
| 195 | |
| 196 | bool GpuFsaBackendContext::AfterUnloadNetwork(NetworkId networkId) |
| 197 | { |
| 198 | bool clearCache = false; |
| 199 | { |
| 200 | std::lock_guard<std::mutex> lockGuard(m_Mutex); |
| 201 | m_NetworkIds.erase(networkId); |
| 202 | clearCache = m_NetworkIds.empty(); |
| 203 | } |
| 204 | |
| 205 | if (clearCache) |
| 206 | { |
| 207 | m_GpuFsaContextControlWrapper->ClearClCache(); |
| 208 | } |
| 209 | |
| 210 | return true; |
| 211 | } |
| 212 | |
| 213 | bool GpuFsaBackendContext::AfterEnqueueWorkload(NetworkId) |
| 214 | { |
| 215 | return m_GpuFsaContextControlWrapper->Sync(); |
| 216 | } |
| 217 | |
| 218 | GpuFsaBackendContext::~GpuFsaBackendContext() |
| 219 | { |
| 220 | if (m_Tuner && !m_TuningFile.empty()) |
| 221 | { |
| 222 | try |
| 223 | { |
| 224 | m_Tuner->save_to_file(m_TuningFile.c_str()); |
| 225 | } |
| 226 | catch(const std::exception& e) |
| 227 | { |
| 228 | ARMNN_LOG(warning) << "Could not save GpuFsa tuner data to file " << m_TuningFile; |
| 229 | } |
| 230 | } |
| 231 | } |
| 232 | |
| 233 | } // namespace armnn |