Mike Kelly | b5fdf38 | 2019-06-11 16:35:25 +0100 | [diff] [blame] | 1 | // |
| 2 | // Copyright © 2017 Arm Ltd. All rights reserved. |
| 3 | // SPDX-License-Identifier: MIT |
| 4 | // |
| 5 | |
| 6 | #include "ArmnnDriverImpl.hpp" |
| 7 | #include "../ArmnnPreparedModel_1_2.hpp" |
| 8 | #include "../ModelToINetworkConverter.hpp" |
| 9 | #include "../SystemPropertiesUtils.hpp" |
| 10 | |
| 11 | #include <log/log.h> |
| 12 | |
| 13 | namespace |
| 14 | { |
| 15 | |
| 16 | const char *g_RelaxedFloat32toFloat16PerformanceExecTime = "ArmNN.relaxedFloat32toFloat16Performance.execTime"; |
| 17 | void NotifyCallbackAndCheck(const sp<V1_2::IPreparedModelCallback>& callback, |
| 18 | ErrorStatus errorStatus, |
| 19 | const sp<V1_2::IPreparedModel>& preparedModelPtr) |
| 20 | { |
| 21 | Return<void> returned = callback->notify(errorStatus, preparedModelPtr); |
| 22 | // This check is required, if the callback fails and it isn't checked it will bring down the service |
| 23 | if (!returned.isOk()) |
| 24 | { |
| 25 | ALOGE("ArmnnDriverImpl::prepareModel: hidl callback failed to return properly: %s ", |
| 26 | returned.description().c_str()); |
| 27 | } |
| 28 | } |
| 29 | |
| 30 | Return<ErrorStatus> FailPrepareModel(ErrorStatus error, |
| 31 | const std::string& message, |
| 32 | const sp<V1_2::IPreparedModelCallback>& callback) |
| 33 | { |
| 34 | ALOGW("ArmnnDriverImpl::prepareModel: %s", message.c_str()); |
| 35 | NotifyCallbackAndCheck(callback, error, nullptr); |
| 36 | return error; |
| 37 | } |
| 38 | |
| 39 | } // anonymous namespace |
| 40 | |
| 41 | namespace armnn_driver |
| 42 | { |
| 43 | namespace hal_1_2 |
| 44 | { |
| 45 | |
| 46 | Return<ErrorStatus> ArmnnDriverImpl::prepareArmnnModel_1_2(const armnn::IRuntimePtr& runtime, |
| 47 | const armnn::IGpuAccTunedParametersPtr& clTunedParameters, |
| 48 | const DriverOptions& options, |
| 49 | const V1_2::Model& model, |
| 50 | const sp<V1_2::IPreparedModelCallback>& cb, |
| 51 | bool float32ToFloat16) |
| 52 | { |
| 53 | ALOGV("ArmnnDriverImpl::prepareModel()"); |
| 54 | |
| 55 | if (cb.get() == nullptr) |
| 56 | { |
| 57 | ALOGW("ArmnnDriverImpl::prepareModel: Invalid callback passed to prepareModel"); |
| 58 | return ErrorStatus::INVALID_ARGUMENT; |
| 59 | } |
| 60 | |
| 61 | if (!runtime) |
| 62 | { |
| 63 | return FailPrepareModel(ErrorStatus::DEVICE_UNAVAILABLE, "Device unavailable", cb); |
| 64 | } |
| 65 | |
| 66 | if (!android::nn::validateModel(model)) |
| 67 | { |
| 68 | return FailPrepareModel(ErrorStatus::INVALID_ARGUMENT, "Invalid model passed as input", cb); |
| 69 | } |
| 70 | |
| 71 | // Deliberately ignore any unsupported operations requested by the options - |
| 72 | // at this point we're being asked to prepare a model that we've already declared support for |
| 73 | // and the operation indices may be different to those in getSupportedOperations anyway. |
| 74 | std::set<unsigned int> unsupportedOperations; |
| 75 | ModelToINetworkConverter<HalPolicy> modelConverter(options.GetBackends(), |
| 76 | model, |
| 77 | unsupportedOperations); |
| 78 | |
| 79 | if (modelConverter.GetConversionResult() != ConversionResult::Success) |
| 80 | { |
| 81 | FailPrepareModel(ErrorStatus::GENERAL_FAILURE, "ModelToINetworkConverter failed", cb); |
| 82 | return ErrorStatus::NONE; |
| 83 | } |
| 84 | |
| 85 | // Optimize the network |
| 86 | armnn::IOptimizedNetworkPtr optNet(nullptr, nullptr); |
| 87 | armnn::OptimizerOptions OptOptions; |
| 88 | OptOptions.m_ReduceFp32ToFp16 = float32ToFloat16; |
| 89 | |
| 90 | std::vector<std::string> errMessages; |
| 91 | try |
| 92 | { |
| 93 | optNet = armnn::Optimize(*modelConverter.GetINetwork(), |
| 94 | options.GetBackends(), |
| 95 | runtime->GetDeviceSpec(), |
| 96 | OptOptions, |
| 97 | errMessages); |
| 98 | } |
| 99 | catch (armnn::Exception &e) |
| 100 | { |
| 101 | std::stringstream message; |
| 102 | message << "armnn::Exception (" << e.what() << ") caught from optimize."; |
| 103 | FailPrepareModel(ErrorStatus::GENERAL_FAILURE, message.str(), cb); |
| 104 | return ErrorStatus::NONE; |
| 105 | } |
| 106 | |
| 107 | // Check that the optimized network is valid. |
| 108 | if (!optNet) |
| 109 | { |
| 110 | std::stringstream message; |
| 111 | message << "Invalid optimized network"; |
| 112 | for (const std::string& msg : errMessages) |
| 113 | { |
| 114 | message << "\n" << msg; |
| 115 | } |
| 116 | FailPrepareModel(ErrorStatus::GENERAL_FAILURE, message.str(), cb); |
| 117 | return ErrorStatus::NONE; |
| 118 | } |
| 119 | |
| 120 | // Export the optimized network graph to a dot file if an output dump directory |
| 121 | // has been specified in the drivers' arguments. |
| 122 | ExportNetworkGraphToDotFile<hal_1_2::HalPolicy::Model>(*optNet, options.GetRequestInputsAndOutputsDumpDir(), |
| 123 | model); |
| 124 | |
| 125 | // Load it into the runtime. |
| 126 | armnn::NetworkId netId = 0; |
| 127 | try |
| 128 | { |
| 129 | if (runtime->LoadNetwork(netId, move(optNet)) != armnn::Status::Success) |
| 130 | { |
| 131 | return FailPrepareModel(ErrorStatus::GENERAL_FAILURE, "Network could not be loaded", cb); |
| 132 | } |
| 133 | } |
| 134 | catch (armnn::Exception& e) |
| 135 | { |
| 136 | std::stringstream message; |
| 137 | message << "armnn::Exception (" << e.what()<< ") caught from LoadNetwork."; |
| 138 | FailPrepareModel(ErrorStatus::GENERAL_FAILURE, message.str(), cb); |
| 139 | return ErrorStatus::NONE; |
| 140 | } |
| 141 | |
| 142 | std::unique_ptr<ArmnnPreparedModel_1_2<hal_1_2::HalPolicy>> preparedModel( |
| 143 | new ArmnnPreparedModel_1_2<hal_1_2::HalPolicy>( |
| 144 | netId, |
| 145 | runtime.get(), |
| 146 | model, |
| 147 | options.GetRequestInputsAndOutputsDumpDir(), |
| 148 | options.IsGpuProfilingEnabled())); |
| 149 | |
| 150 | // Run a single 'dummy' inference of the model. This means that CL kernels will get compiled (and tuned if |
| 151 | // this is enabled) before the first 'real' inference which removes the overhead of the first inference. |
| 152 | if (!preparedModel->ExecuteWithDummyInputs()) |
| 153 | { |
| 154 | return FailPrepareModel(ErrorStatus::GENERAL_FAILURE, "Network could not be executed", cb); |
| 155 | } |
| 156 | |
| 157 | if (clTunedParameters && |
| 158 | options.GetClTunedParametersMode() == armnn::IGpuAccTunedParameters::Mode::UpdateTunedParameters) |
| 159 | { |
| 160 | // Now that we've done one inference the CL kernel parameters will have been tuned, so save the updated file. |
| 161 | try |
| 162 | { |
| 163 | clTunedParameters->Save(options.GetClTunedParametersFile().c_str()); |
| 164 | } |
| 165 | catch (const armnn::Exception& error) |
| 166 | { |
| 167 | ALOGE("ArmnnDriverImpl::prepareModel: Failed to save CL tuned parameters file '%s': %s", |
| 168 | options.GetClTunedParametersFile().c_str(), error.what()); |
| 169 | } |
| 170 | } |
| 171 | |
| 172 | NotifyCallbackAndCheck(cb, ErrorStatus::NONE, preparedModel.release()); |
| 173 | |
| 174 | return ErrorStatus::NONE; |
| 175 | } |
| 176 | |
| 177 | Return<void> ArmnnDriverImpl::getCapabilities_1_2(const armnn::IRuntimePtr& runtime, |
| 178 | V1_2::IDevice::getCapabilities_1_2_cb cb) |
| 179 | { |
| 180 | ALOGV("hal_1_2::ArmnnDriverImpl::getCapabilities()"); |
| 181 | |
| 182 | V1_2::Capabilities capabilities; |
| 183 | |
| 184 | if (runtime) |
| 185 | { |
| 186 | capabilities.relaxedFloat32toFloat16PerformanceScalar.execTime = |
| 187 | ParseSystemProperty(g_RelaxedFloat32toFloat16PerformanceExecTime, .1f); |
| 188 | |
| 189 | capabilities.relaxedFloat32toFloat16PerformanceTensor.execTime = |
| 190 | ParseSystemProperty(g_RelaxedFloat32toFloat16PerformanceExecTime, .1f); |
| 191 | |
| 192 | cb(ErrorStatus::NONE, capabilities); |
| 193 | } |
| 194 | else |
| 195 | { |
| 196 | capabilities.relaxedFloat32toFloat16PerformanceScalar.execTime = 0; |
| 197 | capabilities.relaxedFloat32toFloat16PerformanceTensor.execTime = 0; |
| 198 | |
| 199 | cb(ErrorStatus::DEVICE_UNAVAILABLE, capabilities); |
| 200 | } |
| 201 | |
| 202 | return Void(); |
| 203 | } |
| 204 | |
| 205 | } // namespace hal_1_2 |
| 206 | } // namespace armnn_driver |