Aron Virginas-Tar | 69362cc | 2018-11-22 15:04:42 +0000 | [diff] [blame] | 1 | // |
| 2 | // Copyright © 2017 Arm Ltd. All rights reserved. |
| 3 | // SPDX-License-Identifier: MIT |
| 4 | // |
| 5 | |
| 6 | #include "JsonPrinterTestImpl.hpp" |
| 7 | |
| 8 | #include <Profiling.hpp> |
| 9 | |
| 10 | #include <armnn/Descriptors.hpp> |
| 11 | #include <armnn/IRuntime.hpp> |
| 12 | #include <armnn/INetwork.hpp> |
| 13 | |
| 14 | #include <boost/algorithm/string.hpp> |
| 15 | #include <boost/lexical_cast.hpp> |
| 16 | #include <boost/test/unit_test.hpp> |
| 17 | |
| 18 | #include <sstream> |
| 19 | #include <stack> |
| 20 | #include <string> |
| 21 | |
| 22 | inline bool AreMatchingPair(const char opening, const char closing) |
| 23 | { |
| 24 | return (opening == '{' && closing == '}') || (opening == '[' && closing == ']'); |
| 25 | } |
| 26 | |
| 27 | bool AreParenthesesMatching(const std::string& exp) |
| 28 | { |
| 29 | std::stack<char> expStack; |
| 30 | for (size_t i = 0; i < exp.length(); ++i) |
| 31 | { |
| 32 | if (exp[i] == '{' || exp[i] == '[') |
| 33 | { |
| 34 | expStack.push(exp[i]); |
| 35 | } |
| 36 | else if (exp[i] == '}' || exp[i] == ']') |
| 37 | { |
| 38 | if (expStack.empty() || !AreMatchingPair(expStack.top(), exp[i])) |
| 39 | { |
| 40 | return false; |
| 41 | } |
| 42 | else |
| 43 | { |
| 44 | expStack.pop(); |
| 45 | } |
| 46 | } |
| 47 | } |
| 48 | return expStack.empty(); |
| 49 | } |
| 50 | |
| 51 | std::vector<double> ExtractMeasurements(const std::string& exp) |
| 52 | { |
| 53 | std::vector<double> numbers; |
| 54 | bool inArray = false; |
| 55 | std::string numberString; |
| 56 | for (size_t i = 0; i < exp.size(); ++i) |
| 57 | { |
| 58 | if (exp[i] == '[') |
| 59 | { |
| 60 | inArray = true; |
| 61 | } |
| 62 | else if (exp[i] == ']' && inArray) |
| 63 | { |
| 64 | try |
| 65 | { |
| 66 | boost::trim_if(numberString, boost::is_any_of("\t,\n")); |
| 67 | numbers.push_back(std::stod(numberString)); |
| 68 | } |
| 69 | catch (std::invalid_argument const& e) |
| 70 | { |
| 71 | BOOST_FAIL("Could not convert measurements to double: " + numberString); |
| 72 | } |
| 73 | |
| 74 | numberString.clear(); |
| 75 | inArray = false; |
| 76 | } |
| 77 | else if (exp[i] == ',' && inArray) |
| 78 | { |
| 79 | try |
| 80 | { |
| 81 | boost::trim_if(numberString, boost::is_any_of("\t,\n")); |
| 82 | numbers.push_back(std::stod(numberString)); |
| 83 | } |
| 84 | catch (std::invalid_argument const& e) |
| 85 | { |
| 86 | BOOST_FAIL("Could not convert measurements to double: " + numberString); |
| 87 | } |
| 88 | numberString.clear(); |
| 89 | } |
| 90 | else if (exp[i] != '[' && inArray && exp[i] != ',' && exp[i] != ' ') |
| 91 | { |
| 92 | numberString += exp[i]; |
| 93 | } |
| 94 | } |
| 95 | return numbers; |
| 96 | } |
| 97 | |
| 98 | std::vector<std::string> ExtractSections(const std::string& exp) |
| 99 | { |
| 100 | std::vector<std::string> sections; |
| 101 | |
| 102 | std::stack<size_t> s; |
| 103 | for (size_t i = 0; i < exp.size(); i++) |
| 104 | { |
| 105 | if (exp.at(i) == '{') |
| 106 | { |
| 107 | s.push(i); |
| 108 | } |
| 109 | else if (exp.at(i) == '}') |
| 110 | { |
| 111 | size_t from = s.top(); |
| 112 | s.pop(); |
| 113 | sections.push_back(exp.substr(from, i - from + 1)); |
| 114 | } |
| 115 | } |
| 116 | |
| 117 | return sections; |
| 118 | } |
| 119 | |
| 120 | std::string GetSoftmaxProfilerJson(const std::vector<armnn::BackendId>& backends) |
| 121 | { |
| 122 | using namespace armnn; |
| 123 | |
| 124 | BOOST_CHECK(!backends.empty()); |
| 125 | |
| 126 | ProfilerManager& profilerManager = armnn::ProfilerManager::GetInstance(); |
| 127 | |
| 128 | // Create runtime in which test will run |
| 129 | IRuntime::CreationOptions options; |
| 130 | options.m_EnableGpuProfiling = backends.front() == armnn::Compute::GpuAcc; |
| 131 | IRuntimePtr runtime(IRuntime::Create(options)); |
| 132 | |
| 133 | // build up the structure of the network |
| 134 | INetworkPtr net(INetwork::Create()); |
| 135 | |
| 136 | IConnectableLayer* input = net->AddInputLayer(0, "input"); |
| 137 | IConnectableLayer* softmax = net->AddSoftmaxLayer(SoftmaxDescriptor(), "softmax"); |
| 138 | IConnectableLayer* output = net->AddOutputLayer(0, "output"); |
| 139 | |
| 140 | input->GetOutputSlot(0).Connect(softmax->GetInputSlot(0)); |
| 141 | softmax->GetOutputSlot(0).Connect(output->GetInputSlot(0)); |
| 142 | |
| 143 | // set the tensors in the network |
| 144 | TensorInfo inputTensorInfo(TensorShape({1, 5}), DataType::QuantisedAsymm8); |
| 145 | inputTensorInfo.SetQuantizationOffset(100); |
| 146 | inputTensorInfo.SetQuantizationScale(10000.0f); |
| 147 | input->GetOutputSlot(0).SetTensorInfo(inputTensorInfo); |
| 148 | |
| 149 | TensorInfo outputTensorInfo(TensorShape({1, 5}), DataType::QuantisedAsymm8); |
| 150 | outputTensorInfo.SetQuantizationOffset(0); |
| 151 | outputTensorInfo.SetQuantizationScale(1.0f / 256.0f); |
| 152 | softmax->GetOutputSlot(0).SetTensorInfo(outputTensorInfo); |
| 153 | |
| 154 | // optimize the network |
| 155 | IOptimizedNetworkPtr optNet = Optimize(*net, backends, runtime->GetDeviceSpec()); |
| 156 | if(!optNet) |
| 157 | { |
| 158 | BOOST_FAIL("Error occurred during Optimization, Optimize() returned nullptr."); |
| 159 | } |
| 160 | // load it into the runtime |
| 161 | NetworkId netId; |
| 162 | auto error = runtime->LoadNetwork(netId, std::move(optNet)); |
| 163 | BOOST_TEST(error == Status::Success); |
| 164 | |
| 165 | // create structures for input & output |
| 166 | std::vector<uint8_t> inputData |
| 167 | { |
| 168 | 1, 10, 3, 200, 5 |
| 169 | // one of inputs is sufficiently larger than the others to saturate softmax |
| 170 | }; |
| 171 | std::vector<uint8_t> outputData(5); |
| 172 | |
| 173 | armnn::InputTensors inputTensors |
| 174 | { |
| 175 | {0, armnn::ConstTensor(runtime->GetInputTensorInfo(netId, 0), inputData.data())} |
| 176 | }; |
| 177 | armnn::OutputTensors outputTensors |
| 178 | { |
| 179 | {0, armnn::Tensor(runtime->GetOutputTensorInfo(netId, 0), outputData.data())} |
| 180 | }; |
| 181 | |
| 182 | runtime->GetProfiler(netId)->EnableProfiling(true); |
| 183 | |
| 184 | // do the inferences |
| 185 | runtime->EnqueueWorkload(netId, inputTensors, outputTensors); |
| 186 | runtime->EnqueueWorkload(netId, inputTensors, outputTensors); |
| 187 | runtime->EnqueueWorkload(netId, inputTensors, outputTensors); |
| 188 | |
| 189 | // retrieve the Profiler.Print() output |
| 190 | std::stringstream ss; |
| 191 | profilerManager.GetProfiler()->Print(ss); |
| 192 | |
| 193 | return ss.str(); |
| 194 | } |
| 195 | |
| 196 | inline void ValidateProfilerJson(std::string& result, const std::string& testData) |
| 197 | { |
| 198 | // ensure all measurements are greater than zero |
| 199 | std::vector<double> measurementsVector = ExtractMeasurements(result); |
| 200 | BOOST_CHECK(!measurementsVector.empty()); |
| 201 | |
| 202 | // check sections contain raw and unit tags |
| 203 | // first ensure Parenthesis are balanced |
| 204 | if (AreParenthesesMatching(result)) |
| 205 | { |
| 206 | // remove parent sections that will not have raw or unit tag |
| 207 | std::vector<std::string> sectionVector = ExtractSections(result); |
| 208 | for (size_t i = 0; i < sectionVector.size(); ++i) |
| 209 | { |
| 210 | if (boost::contains(sectionVector[i], "\"ArmNN\":") |
| 211 | || boost::contains(sectionVector[i], "\"inference_measurements\":")) |
| 212 | { |
| 213 | sectionVector.erase(sectionVector.begin() + static_cast<int>(i)); |
| 214 | } |
| 215 | } |
| 216 | BOOST_CHECK(!sectionVector.empty()); |
| 217 | |
| 218 | BOOST_CHECK(std::all_of(sectionVector.begin(), sectionVector.end(), |
| 219 | [](std::string i) { return boost::contains(i, "\"raw\":"); })); |
| 220 | |
| 221 | BOOST_CHECK(std::all_of(sectionVector.begin(), sectionVector.end(), |
| 222 | [](std::string i) { return boost::contains(i, "\"unit\":"); })); |
| 223 | } |
| 224 | |
| 225 | // remove the time measurements as they vary from test to test |
| 226 | result.erase(std::remove_if (result.begin(),result.end(), |
| 227 | [](char c) { return c == '.'; }), result.end()); |
| 228 | result.erase(std::remove_if (result.begin(), result.end(), &isdigit), result.end()); |
| 229 | result.erase(std::remove_if (result.begin(),result.end(), |
| 230 | [](char c) { return c == '\t'; }), result.end()); |
| 231 | |
| 232 | BOOST_CHECK(boost::contains(result, "ArmNN")); |
| 233 | BOOST_CHECK(boost::contains(result, "inference_measurements")); |
| 234 | BOOST_CHECK(boost::contains(result, "layer_measurements")); |
| 235 | BOOST_CHECK_EQUAL(result, testData); |
| 236 | |
| 237 | // ensure no spare parenthesis present in print output |
| 238 | BOOST_CHECK(AreParenthesesMatching(result)); |
| 239 | } |
| 240 | |
| 241 | void RunSoftmaxProfilerJsonPrinterTest(const std::vector<armnn::BackendId>& backends) |
| 242 | { |
| 243 | // setup the test fixture and obtain JSON Printer result |
| 244 | std::string result = GetSoftmaxProfilerJson(backends); |
| 245 | |
| 246 | std::string backend = "Ref"; |
nikraj01 | a121de3 | 2019-05-29 10:51:05 +0100 | [diff] [blame] | 247 | std::string testName = "SoftmaxWorkload_Execute"; |
Aron Virginas-Tar | 69362cc | 2018-11-22 15:04:42 +0000 | [diff] [blame] | 248 | std::string changeLine31 = "\n},\n\"CopyMemGeneric_Execute\": {"; |
| 249 | std::string changeLine39 = "us\""; |
| 250 | std::string changeLine40; |
| 251 | std::string changeLine45; |
| 252 | |
| 253 | const armnn::BackendId& firstBackend = backends.at(0); |
| 254 | if (firstBackend == armnn::Compute::GpuAcc) |
| 255 | { |
| 256 | backend = "Cl"; |
nikraj01 | a121de3 | 2019-05-29 10:51:05 +0100 | [diff] [blame] | 257 | testName = "SoftmaxUintWorkload_Execute"; |
Aron Virginas-Tar | 69362cc | 2018-11-22 15:04:42 +0000 | [diff] [blame] | 258 | changeLine31 = ",\n\"OpenClKernelTimer/: softmax_layer_max_shift_exp_sum_quantized_serial GWS[,,]\": {"; |
| 259 | changeLine39 = R"(us" |
| 260 | }, |
| 261 | "OpenClKernelTimer/: softmax_layer_norm_quantized GWS[,,]": { |
| 262 | "raw": [ |
| 263 | , |
| 264 | , |
| 265 | |
| 266 | ], |
| 267 | "unit": "us")"; |
| 268 | |
| 269 | changeLine40 = R"( |
| 270 | }, |
| 271 | "CopyMemGeneric_Execute": { |
| 272 | "raw": [ |
| 273 | , |
| 274 | , |
| 275 | |
| 276 | ], |
| 277 | "unit": "us")"; |
| 278 | changeLine45 = "}\n"; |
| 279 | } |
| 280 | else if (firstBackend == armnn::Compute::CpuAcc) |
| 281 | { |
| 282 | backend = "Neon"; |
nikraj01 | a121de3 | 2019-05-29 10:51:05 +0100 | [diff] [blame] | 283 | testName = "SoftmaxUintWorkload_Execute"; |
Aron Virginas-Tar | 69362cc | 2018-11-22 15:04:42 +0000 | [diff] [blame] | 284 | changeLine31 = ",\n\"NeonKernelTimer/: NEFillBorderKernel\": {"; |
| 285 | changeLine39 = R"(us" |
| 286 | }, |
| 287 | "NeonKernelTimer/: NELogitsDMaxKernel": { |
| 288 | "raw": [ |
| 289 | , |
| 290 | , |
| 291 | |
| 292 | ], |
| 293 | "unit": "us" |
| 294 | }, |
| 295 | "NeonKernelTimer/: NELogitsDSoftmaxKernel": { |
| 296 | "raw": [ |
| 297 | , |
| 298 | , |
| 299 | |
| 300 | ], |
| 301 | "unit": "us")"; |
| 302 | changeLine40 = R"( |
| 303 | }, |
| 304 | "CopyMemGeneric_Execute": { |
| 305 | "raw": [ |
| 306 | , |
| 307 | , |
| 308 | |
| 309 | ], |
| 310 | "unit": "us")"; |
| 311 | changeLine45 = "}\n"; |
| 312 | } |
| 313 | |
| 314 | std::string testData = R"({ |
| 315 | "ArmNN": { |
| 316 | "inference_measurements": { |
| 317 | "raw": [ |
| 318 | , |
| 319 | , |
| 320 | |
| 321 | ], |
| 322 | "unit": "us", |
| 323 | "layer_measurements": { |
| 324 | "raw": [ |
| 325 | , |
| 326 | , |
| 327 | |
| 328 | ], |
| 329 | "unit": "us", |
| 330 | "CopyMemGeneric_Execute": { |
| 331 | "raw": [ |
| 332 | , |
| 333 | , |
| 334 | |
| 335 | ], |
| 336 | "unit": "us" |
| 337 | }, |
nikraj01 | a121de3 | 2019-05-29 10:51:05 +0100 | [diff] [blame] | 338 | ")" + backend + testName + R"(": { |
Aron Virginas-Tar | 69362cc | 2018-11-22 15:04:42 +0000 | [diff] [blame] | 339 | "raw": [ |
| 340 | , |
| 341 | , |
| 342 | |
| 343 | ], |
| 344 | "unit": "us")" + changeLine31 + R"( |
| 345 | "raw": [ |
| 346 | , |
| 347 | , |
| 348 | |
| 349 | ], |
| 350 | "unit": ")" + changeLine39 + R"( |
| 351 | })" + changeLine40 + R"( |
| 352 | } |
| 353 | } |
| 354 | } |
| 355 | } |
| 356 | )" + changeLine45 + R"()"; |
| 357 | |
| 358 | // validate the JSON Printer result |
| 359 | ValidateProfilerJson(result, testData); |
| 360 | } |