Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 1 | // |
| 2 | // Copyright © 2022 Arm Ltd and Contributors. All rights reserved. |
| 3 | // SPDX-License-Identifier: MIT |
| 4 | // |
| 5 | |
| 6 | #define LOG_TAG "arm-armnn-sl" |
| 7 | |
| 8 | #include "DriverOptions.hpp" |
| 9 | |
| 10 | #include "CanonicalUtils.hpp" |
| 11 | |
| 12 | #include <armnn/Version.hpp> |
| 13 | #include <log/log.h> |
| 14 | #include "SystemPropertiesUtils.hpp" |
| 15 | |
| 16 | #include <OperationsUtils.h> |
| 17 | |
| 18 | #include <cxxopts/cxxopts.hpp> |
| 19 | |
| 20 | #include <algorithm> |
| 21 | #include <cassert> |
| 22 | #include <functional> |
| 23 | #include <string> |
| 24 | #include <sstream> |
| 25 | |
| 26 | using namespace android; |
| 27 | using namespace std; |
| 28 | |
| 29 | namespace armnn_driver |
| 30 | { |
| 31 | |
| 32 | DriverOptions::DriverOptions(armnn::Compute computeDevice, bool fp16Enabled) |
| 33 | : m_Backends({computeDevice}) |
| 34 | , m_VerboseLogging(false) |
| 35 | , m_RequestInputsAndOutputsDumpDir(std::string("")) |
| 36 | , m_ServiceName(std::string("armnn_sl")) |
| 37 | , m_ForcedUnsupportedOperations({}) |
| 38 | , m_ClTunedParametersMode(armnn::IGpuAccTunedParameters::Mode::UseTunedParameters) |
| 39 | , m_ClTuningLevel(armnn::IGpuAccTunedParameters::TuningLevel::Rapid) |
| 40 | , m_EnableGpuProfiling(false) |
| 41 | , m_fp16Enabled(fp16Enabled) |
| 42 | , m_FastMathEnabled(false) |
| 43 | , m_ShouldExit(false) |
| 44 | , m_ExitCode(EXIT_SUCCESS) |
| 45 | , m_CachedNetworkFilePath(std::string("")) |
| 46 | , m_SaveCachedNetwork(false) |
| 47 | , m_NumberOfThreads(0) |
| 48 | , m_EnableAsyncModelExecution(false) |
| 49 | , m_ArmnnNumberOfThreads(1) |
| 50 | { |
| 51 | } |
| 52 | |
| 53 | DriverOptions::DriverOptions(const std::vector<armnn::BackendId>& backends, bool fp16Enabled) |
| 54 | : m_Backends(backends) |
| 55 | , m_VerboseLogging(false) |
| 56 | , m_RequestInputsAndOutputsDumpDir(std::string("")) |
| 57 | , m_ServiceName(std::string("armnn_sl")) |
| 58 | , m_ForcedUnsupportedOperations({}) |
| 59 | , m_ClTunedParametersMode(armnn::IGpuAccTunedParameters::Mode::UseTunedParameters) |
| 60 | , m_ClTuningLevel(armnn::IGpuAccTunedParameters::TuningLevel::Rapid) |
| 61 | , m_EnableGpuProfiling(false) |
| 62 | , m_fp16Enabled(fp16Enabled) |
| 63 | , m_FastMathEnabled(false) |
| 64 | , m_ShouldExit(false) |
| 65 | , m_ExitCode(EXIT_SUCCESS) |
| 66 | , m_CachedNetworkFilePath(std::string("")) |
| 67 | , m_SaveCachedNetwork(false) |
| 68 | , m_NumberOfThreads(0) |
| 69 | , m_EnableAsyncModelExecution(false) |
| 70 | , m_ArmnnNumberOfThreads(1) |
| 71 | { |
| 72 | } |
| 73 | |
| 74 | // This default constructor will example an environment variable called |
| 75 | // ARMNN_SL_OPTIONS. It will parse the parameters using the existing cxx |
| 76 | // opts mechanism. |
| 77 | DriverOptions::DriverOptions() |
| 78 | : m_VerboseLogging(false) |
| 79 | , m_RequestInputsAndOutputsDumpDir(std::string("")) |
| 80 | , m_ServiceName(std::string("armnn_sl")) |
| 81 | , m_ForcedUnsupportedOperations({}) |
| 82 | , m_ClTunedParametersMode(armnn::IGpuAccTunedParameters::Mode::UseTunedParameters) |
| 83 | , m_ClTuningLevel(armnn::IGpuAccTunedParameters::TuningLevel::Rapid) |
| 84 | , m_EnableGpuProfiling(false) |
| 85 | , m_fp16Enabled(false) |
| 86 | , m_FastMathEnabled(false) |
| 87 | , m_ShouldExit(false) |
| 88 | , m_SaveCachedNetwork(false) |
| 89 | , m_NumberOfThreads(0) |
| 90 | , m_EnableAsyncModelExecution(false) |
| 91 | , m_ArmnnNumberOfThreads(1) |
| 92 | { |
| 93 | std::string unsupportedOperationsAsString; |
| 94 | std::string clTunedParametersModeAsString; |
| 95 | std::string clTuningLevelAsString; |
| 96 | std::vector<std::string> backends; |
| 97 | bool showHelp = false; |
| 98 | bool showVersion = false; |
| 99 | |
| 100 | const char* rawEnv = std::getenv("ARMNN_SL_OPTIONS"); |
| 101 | // If the environment variable isn't set we'll continue as if it were an empty string. |
| 102 | if (!rawEnv) |
| 103 | { |
| 104 | rawEnv = ""; |
| 105 | } |
| 106 | string optionsAsString(rawEnv); |
| 107 | regex whiteSpaceRegex("\\s+"); |
| 108 | // Tokienize the string based on whitespace. |
| 109 | sregex_token_iterator iter(optionsAsString.begin(), optionsAsString.end(), whiteSpaceRegex, -1); |
| 110 | sregex_token_iterator end; |
| 111 | vector<string> cliAsVector(iter, end); |
| 112 | // As we're pretending to be a command line, argv[0] should be an executable name. |
| 113 | cliAsVector.insert(cliAsVector.begin(), "ARMNN_SL_OPTIONS"); |
| 114 | // Convert the vector of string to a vector of char* backed by the existing vector. |
| 115 | std::vector<char*> argVector; |
| 116 | for (const auto& arg : cliAsVector) |
| 117 | { |
| 118 | argVector.push_back((char*)arg.data()); |
| 119 | } |
| 120 | // Terminate the array. |
| 121 | argVector.push_back(nullptr); |
| 122 | // Create usable variables. |
| 123 | int argc = argVector.size() - 1; // Ignore the null pointer at the end. |
| 124 | char** argv = argVector.data(); |
| 125 | |
| 126 | cxxopts::Options optionsDesc(argv[0], "Arm NN Support Library for the Android Neural Networks API." |
| 127 | "The support library will convert Android NNAPI requests " |
| 128 | "and delegate them to available ArmNN backends."); |
| 129 | try |
| 130 | { |
| 131 | optionsDesc.add_options() |
| 132 | |
| 133 | ("a,enable-fast-math", |
| 134 | "Enables fast_math options in backends that support it. Using the fast_math flag can " |
| 135 | "lead to performance improvements but may result in reduced or different precision.", |
| 136 | cxxopts::value<bool>(m_FastMathEnabled)->default_value("false")) |
| 137 | |
| 138 | ("c,compute","Comma separated list of backends to run layers on. " |
| 139 | "Examples of possible values are: CpuRef, CpuAcc, GpuAcc", |
| 140 | cxxopts::value<std::vector<std::string>>(backends)) |
| 141 | |
| 142 | ("d,request-inputs-and-outputs-dump-dir", |
| 143 | "If non-empty, the directory where request inputs and outputs should be dumped", |
| 144 | cxxopts::value<std::string>(m_RequestInputsAndOutputsDumpDir)->default_value("")) |
| 145 | |
| 146 | ("f,fp16-enabled", "Enables support for relaxed computation from Float32 to Float16", |
| 147 | cxxopts::value<bool>(m_fp16Enabled)->default_value("false")) |
| 148 | |
| 149 | ("h,help", "Show this help", |
| 150 | cxxopts::value<bool>(showHelp)->default_value("false")->implicit_value("true")) |
| 151 | |
| 152 | ("m,cl-tuned-parameters-mode", |
| 153 | "If 'UseTunedParameters' (the default), will read CL tuned parameters from the file specified by " |
| 154 | "--cl-tuned-parameters-file. " |
| 155 | "If 'UpdateTunedParameters', will also find the optimum parameters when preparing new networks and update " |
| 156 | "the file accordingly.", |
| 157 | cxxopts::value<std::string>(clTunedParametersModeAsString)->default_value("UseTunedParameters")) |
| 158 | |
| 159 | ("g,mlgo-cl-tuned-parameters-file", |
| 160 | "If non-empty, the given file will be used to load/save MLGO CL tuned parameters. ", |
| 161 | cxxopts::value<std::string>(m_ClMLGOTunedParametersFile)->default_value("")) |
| 162 | |
| 163 | ("o,cl-tuning-level", |
| 164 | "exhaustive: all lws values are tested " |
| 165 | "normal: reduced number of lws values but enough to still have the performance really close to the " |
| 166 | "exhaustive approach " |
| 167 | "rapid: only 3 lws values should be tested for each kernel ", |
| 168 | cxxopts::value<std::string>(clTuningLevelAsString)->default_value("rapid")) |
| 169 | |
| 170 | ("p,gpu-profiling", "Turns GPU profiling on", |
| 171 | cxxopts::value<bool>(m_EnableGpuProfiling)->default_value("false")) |
| 172 | |
| 173 | ("q,cached-network-file", "If non-empty, the given file will be used to load/save cached network. " |
| 174 | "If save-cached-network option is given will save the cached network to given file." |
| 175 | "If save-cached-network option is not given will load the cached network from given " |
| 176 | "file.", |
| 177 | cxxopts::value<std::string>(m_CachedNetworkFilePath)->default_value("")) |
| 178 | |
| 179 | ("s,save-cached-network", |
| 180 | "Enables saving the cached network to the file given with cached-network-file option." |
| 181 | " See also --cached-network-file", |
| 182 | cxxopts::value<bool>(m_SaveCachedNetwork)->default_value("false")) |
| 183 | |
| 184 | ("number-of-threads", |
| 185 | "Assign the number of threads used by the CpuAcc backend. " |
| 186 | "Input value must be between 1 and 64. " |
| 187 | "Default is set to 0 (Backend will decide number of threads to use).", |
| 188 | cxxopts::value<unsigned int>(m_NumberOfThreads)->default_value("0")) |
| 189 | |
| 190 | ("t,cl-tuned-parameters-file", |
| 191 | "If non-empty, the given file will be used to load/save CL tuned parameters. " |
| 192 | "See also --cl-tuned-parameters-mode", |
| 193 | cxxopts::value<std::string>(m_ClTunedParametersFile)->default_value("")) |
| 194 | |
| 195 | ("u,unsupported-operations", |
| 196 | "If non-empty, a comma-separated list of operation indices which the driver will forcibly " |
| 197 | "consider unsupported", |
| 198 | cxxopts::value<std::string>(unsupportedOperationsAsString)->default_value("")) |
| 199 | |
| 200 | ("v,verbose-logging", "Turns verbose logging on", |
| 201 | cxxopts::value<bool>(m_VerboseLogging)->default_value("false")->implicit_value("true")) |
| 202 | |
| 203 | ("V,version", "Show version information", |
| 204 | cxxopts::value<bool>(showVersion)->default_value("false")->implicit_value("true")) |
| 205 | ; |
| 206 | } |
| 207 | catch (const std::exception& e) |
| 208 | { |
| 209 | VLOG(DRIVER) << "An error occurred attempting to construct options: " << e.what(); |
| 210 | std::cout << "An error occurred attempting to construct options: %s" << std::endl; |
| 211 | m_ExitCode = EXIT_FAILURE; |
| 212 | return; |
| 213 | } |
| 214 | |
| 215 | try |
| 216 | { |
| 217 | cxxopts::ParseResult result = optionsDesc.parse(argc, argv); |
| 218 | } |
| 219 | catch (const cxxopts::OptionException& e) |
| 220 | { |
| 221 | VLOG(DRIVER) << "An exception occurred attempting to parse program options: " << e.what(); |
| 222 | std::cout << optionsDesc.help() << std::endl |
| 223 | << "An exception occurred while parsing program options: " << std::endl |
| 224 | << e.what() << std::endl; |
| 225 | m_ShouldExit = true; |
| 226 | m_ExitCode = EXIT_FAILURE; |
| 227 | return; |
| 228 | } |
| 229 | if (showHelp) |
| 230 | { |
| 231 | VLOG(DRIVER) << "Showing help and exiting"; |
| 232 | std::cout << optionsDesc.help() << std::endl; |
| 233 | m_ShouldExit = true; |
| 234 | m_ExitCode = EXIT_SUCCESS; |
| 235 | return; |
| 236 | } |
| 237 | if (showVersion) |
| 238 | { |
| 239 | VLOG(DRIVER) << "Showing version and exiting"; |
| 240 | std::cout << "ArmNN Android NN driver for the Android Neural Networks API.\n" |
| 241 | "ArmNN v" << ARMNN_VERSION << std::endl; |
| 242 | m_ShouldExit = true; |
| 243 | m_ExitCode = EXIT_SUCCESS; |
| 244 | return; |
| 245 | } |
| 246 | |
| 247 | // Convert the string backend names into backendId's. |
| 248 | m_Backends.reserve(backends.size()); |
| 249 | for (auto&& backend : backends) |
| 250 | { |
| 251 | m_Backends.emplace_back(backend); |
| 252 | } |
| 253 | |
| 254 | // If no backends have been specified then the default value is GpuAcc. |
| 255 | if (backends.empty()) |
| 256 | { |
| 257 | VLOG(DRIVER) << "No backends have been specified:"; |
| 258 | std::cout << optionsDesc.help() << std::endl |
| 259 | << "Unable to start:" << std::endl |
| 260 | << "No backends have been specified" << std::endl; |
| 261 | m_ShouldExit = true; |
| 262 | m_ExitCode = EXIT_FAILURE; |
| 263 | return; |
| 264 | } |
| 265 | |
| 266 | if (!unsupportedOperationsAsString.empty()) |
| 267 | { |
| 268 | std::istringstream argStream(unsupportedOperationsAsString); |
| 269 | |
| 270 | std::string s; |
| 271 | while (!argStream.eof()) |
| 272 | { |
| 273 | std::getline(argStream, s, ','); |
| 274 | try |
| 275 | { |
| 276 | unsigned int operationIdx = std::stoi(s); |
| 277 | m_ForcedUnsupportedOperations.insert(operationIdx); |
| 278 | } |
| 279 | catch (const std::invalid_argument&) |
| 280 | { |
| 281 | VLOG(DRIVER) << "Ignoring invalid integer argument in -u/--unsupported-operations value: " << s.c_str(); |
| 282 | } |
| 283 | } |
| 284 | } |
| 285 | |
| 286 | if (!m_ClTunedParametersFile.empty()) |
| 287 | { |
| 288 | // The mode is only relevant if the file path has been provided |
| 289 | if (clTunedParametersModeAsString == "UseTunedParameters") |
| 290 | { |
| 291 | m_ClTunedParametersMode = armnn::IGpuAccTunedParameters::Mode::UseTunedParameters; |
| 292 | } |
| 293 | else if (clTunedParametersModeAsString == "UpdateTunedParameters") |
| 294 | { |
| 295 | m_ClTunedParametersMode = armnn::IGpuAccTunedParameters::Mode::UpdateTunedParameters; |
| 296 | } |
| 297 | else |
| 298 | { |
| 299 | VLOG(DRIVER) << "Requested unknown cl-tuned-parameters-mode " |
| 300 | << clTunedParametersModeAsString.c_str() << ". Defaulting to UseTunedParameters"; |
| 301 | } |
| 302 | |
| 303 | if (clTuningLevelAsString == "exhaustive") |
| 304 | { |
| 305 | m_ClTuningLevel = armnn::IGpuAccTunedParameters::TuningLevel::Exhaustive; |
| 306 | } |
| 307 | else if (clTuningLevelAsString == "normal") |
| 308 | { |
| 309 | m_ClTuningLevel = armnn::IGpuAccTunedParameters::TuningLevel::Normal; |
| 310 | } |
| 311 | else if (clTuningLevelAsString == "rapid") |
| 312 | { |
| 313 | m_ClTuningLevel = armnn::IGpuAccTunedParameters::TuningLevel::Rapid; |
| 314 | } |
| 315 | else |
| 316 | { |
| 317 | VLOG(DRIVER) << "Requested unknown cl-tuner-mode '%s'. " |
| 318 | "Defaulting to rapid" << clTuningLevelAsString.c_str(); |
| 319 | } |
| 320 | } |
| 321 | } |
| 322 | |
| 323 | } // namespace armnn_driver |