src/profiling/ProfilingUtils.cpp

   1 //
   2 // Copyright © 2017 Arm Ltd. All rights reserved.
   3 // SPDX-License-Identifier: MIT
   4 //
   5
   6 #include "ProfilingUtils.hpp"
   7
   8 #include <armnn/Version.hpp>
   9 #include <armnn/Conversion.hpp>
  10
  11 #include <WallClockTimer.hpp>
  12
  13 #include <boost/assert.hpp>
  14
  15 #include <fstream>
  16 #include <iostream>
  17 #include <limits>
  18
  19 namespace armnn
  20 {
  21
  22 namespace profiling
  23 {
  24
  25 namespace
  26 {
  27
  28 void ThrowIfCantGenerateNextUid(uint16_t uid, uint16_t cores = 0)
  29 {
  30     // Check that it is possible to generate the next UID without causing an overflow
  31     switch (cores)
  32     {
  33     case 0:
  34     case 1:
  35         // Number of cores not specified or set to 1 (a value of zero indicates the device is not capable of
  36         // running multiple parallel workloads and will not provide multiple streams of data for each event)
  37         if (uid == std::numeric_limits<uint16_t>::max())
  38         {
  39             throw RuntimeException("Generating the next UID for profiling would result in an overflow");
  40         }
  41         break;
  42     default: // cores > 1
  43         // Multiple cores available, as max_counter_uid has to be set to: counter_uid + cores - 1, the maximum
  44         // allowed value for a counter UID is consequently: uint16_t_max - cores + 1
  45         if (uid >= std::numeric_limits<uint16_t>::max() - cores + 1)
  46         {
  47             throw RuntimeException("Generating the next UID for profiling would result in an overflow");
  48         }
  49         break;
  50     }
  51 }
  52
  53 } // Anonymous namespace
  54
  55 uint16_t GetNextUid(bool peekOnly)
  56 {
  57     // The UID used for profiling objects and events. The first valid UID is 1, as 0 is a reserved value
  58     static uint16_t uid = 1;
  59
  60     // Check that it is possible to generate the next UID without causing an overflow (throws in case of error)
  61     ThrowIfCantGenerateNextUid(uid);
  62
  63     if (peekOnly)
  64     {
  65         // Peek only
  66         return uid;
  67     }
  68     else
  69     {
  70         // Get the next UID
  71         return uid++;
  72     }
  73 }
  74
  75 std::vector<uint16_t> GetNextCounterUids(uint16_t cores)
  76 {
  77     // The UID used for counters only. The first valid UID is 0
  78     static uint16_t counterUid = 0;
  79
  80     // Check that it is possible to generate the next counter UID without causing an overflow (throws in case of error)
  81     ThrowIfCantGenerateNextUid(counterUid, cores);
  82
  83     // Get the next counter UIDs
  84     size_t counterUidsSize = cores == 0 ? 1 : cores;
  85     std::vector<uint16_t> counterUids(counterUidsSize, 0);
  86     for (size_t i = 0; i < counterUidsSize; i++)
  87     {
  88         counterUids[i] = counterUid++;
  89     }
  90     return counterUids;
  91 }
  92
  93 void WriteBytes(const IPacketBufferPtr& packetBuffer, unsigned int offset,  const void* value, unsigned int valueSize)
  94 {
  95     BOOST_ASSERT(packetBuffer);
  96
  97     WriteBytes(packetBuffer->GetWritableData(), offset, value, valueSize);
  98 }
  99
 100 uint32_t ConstructHeader(uint32_t packetFamily,
 101                          uint32_t packetId)
 102 {
 103     return ((packetFamily & 0x3F) << 26)|
 104            ((packetId & 0x3FF) << 16);
 105 }
 106
 107 uint32_t ConstructHeader(uint32_t packetFamily,
 108                          uint32_t packetClass,
 109                          uint32_t packetType)
 110 {
 111     return ((packetFamily & 0x3F) << 26)|
 112            ((packetClass & 0x3FF) << 19)|
 113            ((packetType & 0x3FFF) << 16);
 114 }
 115
 116 void WriteUint64(const std::unique_ptr<IPacketBuffer>& packetBuffer, unsigned int offset, uint64_t value)
 117 {
 118     BOOST_ASSERT(packetBuffer);
 119
 120     WriteUint64(packetBuffer->GetWritableData(), offset, value);
 121 }
 122
 123 void WriteUint32(const IPacketBufferPtr& packetBuffer, unsigned int offset, uint32_t value)
 124 {
 125     BOOST_ASSERT(packetBuffer);
 126
 127     WriteUint32(packetBuffer->GetWritableData(), offset, value);
 128 }
 129
 130 void WriteUint16(const IPacketBufferPtr& packetBuffer, unsigned int offset, uint16_t value)
 131 {
 132     BOOST_ASSERT(packetBuffer);
 133
 134     WriteUint16(packetBuffer->GetWritableData(), offset, value);
 135 }
 136
 137 void WriteBytes(unsigned char* buffer, unsigned int offset, const void* value, unsigned int valueSize)
 138 {
 139     BOOST_ASSERT(buffer);
 140     BOOST_ASSERT(value);
 141
 142     for (unsigned int i = 0; i < valueSize; i++, offset++)
 143     {
 144         buffer[offset] = *(reinterpret_cast<const unsigned char*>(value) + i);
 145     }
 146 }
 147
 148 void WriteUint64(unsigned char* buffer, unsigned int offset, uint64_t value)
 149 {
 150     BOOST_ASSERT(buffer);
 151
 152     buffer[offset]     = static_cast<unsigned char>(value & 0xFF);
 153     buffer[offset + 1] = static_cast<unsigned char>((value >> 8) & 0xFF);
 154     buffer[offset + 2] = static_cast<unsigned char>((value >> 16) & 0xFF);
 155     buffer[offset + 3] = static_cast<unsigned char>((value >> 24) & 0xFF);
 156     buffer[offset + 4] = static_cast<unsigned char>((value >> 32) & 0xFF);
 157     buffer[offset + 5] = static_cast<unsigned char>((value >> 40) & 0xFF);
 158     buffer[offset + 6] = static_cast<unsigned char>((value >> 48) & 0xFF);
 159     buffer[offset + 7] = static_cast<unsigned char>((value >> 56) & 0xFF);
 160 }
 161
 162 void WriteUint32(unsigned char* buffer, unsigned int offset, uint32_t value)
 163 {
 164     BOOST_ASSERT(buffer);
 165
 166     buffer[offset]     = static_cast<unsigned char>(value & 0xFF);
 167     buffer[offset + 1] = static_cast<unsigned char>((value >> 8) & 0xFF);
 168     buffer[offset + 2] = static_cast<unsigned char>((value >> 16) & 0xFF);
 169     buffer[offset + 3] = static_cast<unsigned char>((value >> 24) & 0xFF);
 170 }
 171
 172 void WriteUint16(unsigned char* buffer, unsigned int offset, uint16_t value)
 173 {
 174     BOOST_ASSERT(buffer);
 175
 176     buffer[offset]     = static_cast<unsigned char>(value & 0xFF);
 177     buffer[offset + 1] = static_cast<unsigned char>((value >> 8) & 0xFF);
 178 }
 179
 180 void ReadBytes(const IPacketBufferPtr& packetBuffer, unsigned int offset, unsigned int valueSize, uint8_t outValue[])
 181 {
 182     BOOST_ASSERT(packetBuffer);
 183
 184     ReadBytes(packetBuffer->GetReadableData(), offset, valueSize, outValue);
 185 }
 186
 187 uint64_t ReadUint64(const IPacketBufferPtr& packetBuffer, unsigned int offset)
 188 {
 189     BOOST_ASSERT(packetBuffer);
 190
 191     return ReadUint64(packetBuffer->GetReadableData(), offset);
 192 }
 193
 194 uint32_t ReadUint32(const IPacketBufferPtr& packetBuffer, unsigned int offset)
 195 {
 196     BOOST_ASSERT(packetBuffer);
 197
 198     return ReadUint32(packetBuffer->GetReadableData(), offset);
 199 }
 200
 201 uint16_t ReadUint16(const IPacketBufferPtr& packetBuffer, unsigned int offset)
 202 {
 203     BOOST_ASSERT(packetBuffer);
 204
 205     return ReadUint16(packetBuffer->GetReadableData(), offset);
 206 }
 207
 208 uint8_t ReadUint8(const IPacketBufferPtr& packetBuffer, unsigned int offset)
 209 {
 210     BOOST_ASSERT(packetBuffer);
 211
 212     return ReadUint8(packetBuffer->GetReadableData(), offset);
 213 }
 214
 215 void ReadBytes(const unsigned char* buffer, unsigned int offset, unsigned int valueSize, uint8_t outValue[])
 216 {
 217     BOOST_ASSERT(buffer);
 218     BOOST_ASSERT(outValue);
 219
 220     for (unsigned int i = 0; i < valueSize; i++, offset++)
 221     {
 222         outValue[i] = static_cast<uint8_t>(buffer[offset]);
 223     }
 224 }
 225
 226 uint64_t ReadUint64(const unsigned char* buffer, unsigned int offset)
 227 {
 228     BOOST_ASSERT(buffer);
 229
 230     uint64_t value = 0;
 231     value  = static_cast<uint64_t>(buffer[offset]);
 232     value |= static_cast<uint64_t>(buffer[offset + 1]) << 8;
 233     value |= static_cast<uint64_t>(buffer[offset + 2]) << 16;
 234     value |= static_cast<uint64_t>(buffer[offset + 3]) << 24;
 235     value |= static_cast<uint64_t>(buffer[offset + 4]) << 32;
 236     value |= static_cast<uint64_t>(buffer[offset + 5]) << 40;
 237     value |= static_cast<uint64_t>(buffer[offset + 6]) << 48;
 238     value |= static_cast<uint64_t>(buffer[offset + 7]) << 56;
 239
 240     return value;
 241 }
 242
 243 uint32_t ReadUint32(const unsigned char* buffer, unsigned int offset)
 244 {
 245     BOOST_ASSERT(buffer);
 246
 247     uint32_t value = 0;
 248     value  = static_cast<uint32_t>(buffer[offset]);
 249     value |= static_cast<uint32_t>(buffer[offset + 1]) << 8;
 250     value |= static_cast<uint32_t>(buffer[offset + 2]) << 16;
 251     value |= static_cast<uint32_t>(buffer[offset + 3]) << 24;
 252     return value;
 253 }
 254
 255 uint16_t ReadUint16(const unsigned char* buffer, unsigned int offset)
 256 {
 257     BOOST_ASSERT(buffer);
 258
 259     uint32_t value = 0;
 260     value  = static_cast<uint32_t>(buffer[offset]);
 261     value |= static_cast<uint32_t>(buffer[offset + 1]) << 8;
 262     return static_cast<uint16_t>(value);
 263 }
 264
 265 uint8_t ReadUint8(const unsigned char* buffer, unsigned int offset)
 266 {
 267     BOOST_ASSERT(buffer);
 268
 269     return buffer[offset];
 270 }
 271
 272 std::string GetSoftwareInfo()
 273 {
 274     return std::string("ArmNN");
 275 }
 276
 277 std::string GetHardwareVersion()
 278 {
 279     return std::string();
 280 }
 281
 282 std::string GetSoftwareVersion()
 283 {
 284     std::string armnnVersion(ARMNN_VERSION);
 285     std::string result = "Armnn " + armnnVersion.substr(2,2) + "." + armnnVersion.substr(4,2);
 286     return result;
 287 }
 288
 289 std::string GetProcessName()
 290 {
 291     std::ifstream comm("/proc/self/comm");
 292     std::string name;
 293     getline(comm, name);
 294     return name;
 295 }
 296
 297 /// Creates a timeline packet header
 298 ///
 299 /// \params
 300 ///   packetFamiliy     Timeline Packet Family
 301 ///   packetClass       Timeline Packet Class
 302 ///   packetType        Timeline Packet Type
 303 ///   streamId          Stream identifier
 304 ///   seqeunceNumbered  When non-zero the 4 bytes following the header is a u32 sequence number
 305 ///   dataLength        Unsigned 24-bit integer. Length of data, in bytes. Zero is permitted
 306 ///
 307 /// \returns
 308 ///   Pair of uint32_t containing word0 and word1 of the header
 309 std::pair<uint32_t, uint32_t> CreateTimelinePacketHeader(uint32_t packetFamily,
 310                                                          uint32_t packetClass,
 311                                                          uint32_t packetType,
 312                                                          uint32_t streamId,
 313                                                          uint32_t sequenceNumbered,
 314                                                          uint32_t dataLength)
 315 {
 316     // Packet header word 0:
 317     // 26:31 [6] packet_family: timeline Packet Family, value 0b000001
 318     // 19:25 [7] packet_class: packet class
 319     // 16:18 [3] packet_type: packet type
 320     // 8:15  [8] reserved: all zeros
 321     // 0:7   [8] stream_id: stream identifier
 322     uint32_t packetHeaderWord0 = ((packetFamily & 0x0000003F) << 26) |
 323                                  ((packetClass  & 0x0000007F) << 19) |
 324                                  ((packetType   & 0x00000007) << 16) |
 325                                  ((streamId     & 0x00000007) <<  0);
 326
 327     // Packet header word 1:
 328     // 25:31 [7]  reserved: all zeros
 329     // 24    [1]  sequence_numbered: when non-zero the 4 bytes following the header is a u32 sequence number
 330     // 0:23  [24] data_length: unsigned 24-bit integer. Length of data, in bytes. Zero is permitted
 331     uint32_t packetHeaderWord1 = ((sequenceNumbered & 0x00000001) << 24) |
 332                                  ((dataLength       & 0x00FFFFFF) <<  0);
 333
 334     return std::make_pair(packetHeaderWord0, packetHeaderWord1);
 335 }
 336
 337 // Calculate the actual length an SwString will be including the terminating null character
 338 // padding to bring it to the next uint32_t boundary but minus the leading uint32_t encoding
 339 // the size to allow the offset to be correctly updated when decoding a binary packet.
 340 uint32_t CalculateSizeOfPaddedSwString(const std::string& str)
 341 {
 342     std::vector<uint32_t> swTraceString;
 343     StringToSwTraceString<SwTraceCharPolicy>(str, swTraceString);
 344     unsigned int uint32_t_size = sizeof(uint32_t);
 345     uint32_t size = (boost::numeric_cast<uint32_t>(swTraceString.size()) - 1) * uint32_t_size;
 346     return size;
 347 }
 348
 349 // Read TimelineMessageDirectoryPacket from given IPacketBuffer and offset
 350 SwTraceMessage ReadSwTraceMessage(const IPacketBufferPtr& packetBuffer, unsigned int& offset)
 351 {
 352     BOOST_ASSERT(packetBuffer);
 353
 354     unsigned int uint32_t_size = sizeof(uint32_t);
 355
 356     SwTraceMessage swTraceMessage;
 357
 358     // Read the decl_id
 359     uint32_t readDeclId = ReadUint32(packetBuffer, offset);
 360     swTraceMessage.id = readDeclId;
 361
 362     // SWTrace "namestring" format
 363     // length of the string (first 4 bytes) + string + null terminator
 364
 365     // Check the decl_name
 366     offset += uint32_t_size;
 367     uint32_t swTraceDeclNameLength = ReadUint32(packetBuffer, offset);
 368
 369     offset += uint32_t_size;
 370     std::vector<unsigned char> swTraceStringBuffer(swTraceDeclNameLength - 1);
 371     std::memcpy(swTraceStringBuffer.data(),
 372                 packetBuffer->GetReadableData()  + offset, swTraceStringBuffer.size());
 373
 374     swTraceMessage.name.assign(swTraceStringBuffer.begin(), swTraceStringBuffer.end()); // name
 375
 376     // Check the ui_name
 377     offset += CalculateSizeOfPaddedSwString(swTraceMessage.name);
 378     uint32_t swTraceUINameLength = ReadUint32(packetBuffer, offset);
 379
 380     offset += uint32_t_size;
 381     swTraceStringBuffer.resize(swTraceUINameLength - 1);
 382     std::memcpy(swTraceStringBuffer.data(),
 383                 packetBuffer->GetReadableData()  + offset, swTraceStringBuffer.size());
 384
 385     swTraceMessage.uiName.assign(swTraceStringBuffer.begin(), swTraceStringBuffer.end()); // ui_name
 386
 387     // Check arg_types
 388     offset += CalculateSizeOfPaddedSwString(swTraceMessage.uiName);
 389     uint32_t swTraceArgTypesLength = ReadUint32(packetBuffer, offset);
 390
 391     offset += uint32_t_size;
 392     swTraceStringBuffer.resize(swTraceArgTypesLength - 1);
 393     std::memcpy(swTraceStringBuffer.data(),
 394                 packetBuffer->GetReadableData()  + offset, swTraceStringBuffer.size());
 395
 396     swTraceMessage.argTypes.assign(swTraceStringBuffer.begin(), swTraceStringBuffer.end()); // arg_types
 397
 398     std::string swTraceString(swTraceStringBuffer.begin(), swTraceStringBuffer.end());
 399
 400     // Check arg_names
 401     offset += CalculateSizeOfPaddedSwString(swTraceString);
 402     uint32_t swTraceArgNamesLength = ReadUint32(packetBuffer, offset);
 403
 404     offset += uint32_t_size;
 405     swTraceStringBuffer.resize(swTraceArgNamesLength - 1);
 406     std::memcpy(swTraceStringBuffer.data(),
 407                 packetBuffer->GetReadableData()  + offset, swTraceStringBuffer.size());
 408
 409     swTraceString.assign(swTraceStringBuffer.begin(), swTraceStringBuffer.end());
 410     std::stringstream stringStream(swTraceString);
 411     std::string argName;
 412     while (std::getline(stringStream, argName, ','))
 413     {
 414         swTraceMessage.argNames.push_back(argName);
 415     }
 416
 417     offset += CalculateSizeOfPaddedSwString(swTraceString);
 418
 419     return swTraceMessage;
 420 }
 421
 422 /// Creates a packet header for the timeline messages:
 423 /// * declareLabel
 424 /// * declareEntity
 425 /// * declareEventClass
 426 /// * declareRelationship
 427 /// * declareEvent
 428 ///
 429 /// \param
 430 ///   dataLength The length of the message body in bytes
 431 ///
 432 /// \returns
 433 ///   Pair of uint32_t containing word0 and word1 of the header
 434 std::pair<uint32_t, uint32_t> CreateTimelineMessagePacketHeader(unsigned int dataLength)
 435 {
 436     return CreateTimelinePacketHeader(1,           // Packet family
 437                                       0,           // Packet class
 438                                       1,           // Packet type
 439                                       0,           // Stream id
 440                                       0,           // Sequence number
 441                                       dataLength); // Data length
 442 }
 443
 444 TimelinePacketStatus WriteTimelineLabelBinaryPacket(uint64_t profilingGuid,
 445                                                     const std::string& label,
 446                                                     unsigned char* buffer,
 447                                                     unsigned int bufferSize,
 448                                                     unsigned int& numberOfBytesWritten)
 449 {
 450     // Initialize the output value
 451     numberOfBytesWritten = 0;
 452
 453     // Check that the given buffer is valid
 454     if (buffer == nullptr || bufferSize == 0)
 455     {
 456         return TimelinePacketStatus::BufferExhaustion;
 457     }
 458
 459     // Utils
 460     unsigned int uint32_t_size = sizeof(uint32_t);
 461     unsigned int uint64_t_size = sizeof(uint64_t);
 462
 463     // Convert the label into a SWTrace string
 464     std::vector<uint32_t> swTraceLabel;
 465     bool result = StringToSwTraceString<SwTraceCharPolicy>(label, swTraceLabel);
 466     if (!result)
 467     {
 468         return TimelinePacketStatus::Error;
 469     }
 470
 471     // Calculate the size of the SWTrace string label (in bytes)
 472     unsigned int swTraceLabelSize = boost::numeric_cast<unsigned int>(swTraceLabel.size()) * uint32_t_size;
 473
 474     // Calculate the length of the data (in bytes)
 475     unsigned int timelineLabelPacketDataLength = uint32_t_size +   // decl_Id
 476                                                  uint64_t_size +   // Profiling GUID
 477                                                  swTraceLabelSize; // Label
 478
 479     // Calculate the timeline binary packet size (in bytes)
 480     unsigned int timelineLabelPacketSize = 2 * uint32_t_size +            // Header (2 words)
 481                                            timelineLabelPacketDataLength; // decl_Id + Profiling GUID + label
 482
 483     // Check whether the timeline binary packet fits in the given buffer
 484     if (timelineLabelPacketSize > bufferSize)
 485     {
 486         return TimelinePacketStatus::BufferExhaustion;
 487     }
 488
 489     // Create packet header
 490     std::pair<uint32_t, uint32_t> packetHeader = CreateTimelineMessagePacketHeader(timelineLabelPacketDataLength);
 491
 492     // Initialize the offset for writing in the buffer
 493     unsigned int offset = 0;
 494
 495     // Write the timeline binary packet header to the buffer
 496     WriteUint32(buffer, offset, packetHeader.first);
 497     offset += uint32_t_size;
 498     WriteUint32(buffer, offset, packetHeader.second);
 499     offset += uint32_t_size;
 500
 501     // Write decl_Id to the buffer
 502     WriteUint32(buffer, offset, 0u);
 503     offset += uint32_t_size;
 504
 505     // Write the timeline binary packet payload to the buffer
 506     WriteUint64(buffer, offset, profilingGuid); // Profiling GUID
 507     offset += uint64_t_size;
 508     for (uint32_t swTraceLabelWord : swTraceLabel)
 509     {
 510         WriteUint32(buffer, offset, swTraceLabelWord); // Label
 511         offset += uint32_t_size;
 512     }
 513
 514     // Update the number of bytes written
 515     numberOfBytesWritten = timelineLabelPacketSize;
 516
 517     return TimelinePacketStatus::Ok;
 518 }
 519
 520 TimelinePacketStatus WriteTimelineEntityBinaryPacket(uint64_t profilingGuid,
 521                                                      unsigned char* buffer,
 522                                                      unsigned int bufferSize,
 523                                                      unsigned int& numberOfBytesWritten)
 524 {
 525     // Initialize the output value
 526     numberOfBytesWritten = 0;
 527
 528     // Check that the given buffer is valid
 529     if (buffer == nullptr || bufferSize == 0)
 530     {
 531         return TimelinePacketStatus::BufferExhaustion;
 532     }
 533
 534     // Utils
 535     unsigned int uint32_t_size = sizeof(uint32_t);
 536     unsigned int uint64_t_size = sizeof(uint64_t);
 537
 538     // Calculate the length of the data (in bytes)
 539     unsigned int timelineEntityPacketDataLength = uint64_t_size;   // Profiling GUID
 540
 541
 542     // Calculate the timeline binary packet size (in bytes)
 543     unsigned int timelineEntityPacketSize = 2 * uint32_t_size +             // Header (2 words)
 544                                             uint32_t_size +                 // decl_Id
 545                                             timelineEntityPacketDataLength; // Profiling GUID
 546
 547     // Check whether the timeline binary packet fits in the given buffer
 548     if (timelineEntityPacketSize > bufferSize)
 549     {
 550         return TimelinePacketStatus::BufferExhaustion;
 551     }
 552
 553     // Create packet header
 554     std::pair<uint32_t, uint32_t> packetHeader = CreateTimelineMessagePacketHeader(timelineEntityPacketDataLength);
 555
 556     // Initialize the offset for writing in the buffer
 557     unsigned int offset = 0;
 558
 559     // Write the timeline binary packet header to the buffer
 560     WriteUint32(buffer, offset, packetHeader.first);
 561     offset += uint32_t_size;
 562     WriteUint32(buffer, offset, packetHeader.second);
 563     offset += uint32_t_size;
 564
 565     // Write the decl_Id to the buffer
 566     WriteUint32(buffer, offset, 1u);
 567     offset += uint32_t_size;
 568
 569     // Write the timeline binary packet payload to the buffer
 570     WriteUint64(buffer, offset, profilingGuid); // Profiling GUID
 571
 572     // Update the number of bytes written
 573     numberOfBytesWritten = timelineEntityPacketSize;
 574
 575     return TimelinePacketStatus::Ok;
 576 }
 577
 578 TimelinePacketStatus WriteTimelineRelationshipBinaryPacket(ProfilingRelationshipType relationshipType,
 579                                                            uint64_t relationshipGuid,
 580                                                            uint64_t headGuid,
 581                                                            uint64_t tailGuid,
 582                                                            unsigned char* buffer,
 583                                                            unsigned int bufferSize,
 584                                                            unsigned int& numberOfBytesWritten)
 585 {
 586     // Initialize the output value
 587     numberOfBytesWritten = 0;
 588
 589     // Check that the given buffer is valid
 590     if (buffer == nullptr || bufferSize == 0)
 591     {
 592         return TimelinePacketStatus::BufferExhaustion;
 593     }
 594
 595     // Utils
 596     unsigned int uint32_t_size = sizeof(uint32_t);
 597     unsigned int uint64_t_size = sizeof(uint64_t);
 598
 599     // Calculate the length of the data (in bytes)
 600     unsigned int timelineRelationshipPacketDataLength = uint32_t_size * 2 + // decl_id + Relationship Type
 601                                                         uint64_t_size * 3; // Relationship GUID + Head GUID + tail GUID
 602
 603     // Calculate the timeline binary packet size (in bytes)
 604     unsigned int timelineRelationshipPacketSize = 2 * uint32_t_size + // Header (2 words)
 605                                                   timelineRelationshipPacketDataLength;
 606
 607     // Check whether the timeline binary packet fits in the given buffer
 608     if (timelineRelationshipPacketSize > bufferSize)
 609     {
 610         return TimelinePacketStatus::BufferExhaustion;
 611     }
 612
 613     // Create packet header
 614     uint32_t dataLength = boost::numeric_cast<uint32_t>(timelineRelationshipPacketDataLength);
 615     std::pair<uint32_t, uint32_t> packetHeader = CreateTimelineMessagePacketHeader(dataLength);
 616
 617     // Initialize the offset for writing in the buffer
 618     unsigned int offset = 0;
 619
 620     // Write the timeline binary packet header to the buffer
 621     WriteUint32(buffer, offset, packetHeader.first);
 622     offset += uint32_t_size;
 623     WriteUint32(buffer, offset, packetHeader.second);
 624     offset += uint32_t_size;
 625
 626     uint32_t relationshipTypeUint = 0;
 627
 628     switch (relationshipType)
 629     {
 630         case ProfilingRelationshipType::RetentionLink:
 631             relationshipTypeUint = 0;
 632             break;
 633         case ProfilingRelationshipType::ExecutionLink:
 634             relationshipTypeUint = 1;
 635             break;
 636         case ProfilingRelationshipType::DataLink:
 637             relationshipTypeUint = 2;
 638             break;
 639         case ProfilingRelationshipType::LabelLink:
 640             relationshipTypeUint = 3;
 641             break;
 642         default:
 643             throw InvalidArgumentException("Unknown relationship type given.");
 644     }
 645
 646     // Write the timeline binary packet payload to the buffer
 647     // decl_id of the timeline message
 648     uint32_t declId = 3;
 649     WriteUint32(buffer, offset, declId); // decl_id
 650     offset += uint32_t_size;
 651     WriteUint32(buffer, offset, relationshipTypeUint); // Relationship Type
 652     offset += uint32_t_size;
 653     WriteUint64(buffer, offset, relationshipGuid); // GUID of this relationship
 654     offset += uint64_t_size;
 655     WriteUint64(buffer, offset, headGuid); // head of relationship GUID
 656     offset += uint64_t_size;
 657     WriteUint64(buffer, offset, tailGuid); // tail of relationship GUID
 658
 659     // Update the number of bytes written
 660     numberOfBytesWritten = timelineRelationshipPacketSize;
 661
 662     return TimelinePacketStatus::Ok;
 663 }
 664
 665 TimelinePacketStatus WriteTimelineMessageDirectoryPackage(unsigned char* buffer,
 666                                                           unsigned int bufferSize,
 667                                                           unsigned int& numberOfBytesWritten)
 668 {
 669     // Initialize the output value
 670     numberOfBytesWritten = 0;
 671
 672     // Check that the given buffer is valid
 673     if (buffer == nullptr || bufferSize == 0)
 674     {
 675         return TimelinePacketStatus::BufferExhaustion;
 676     }
 677
 678     // Utils
 679     unsigned int uint32_t_size = sizeof(uint32_t);
 680
 681     // The payload/data of the packet consists of swtrace event definitions encoded according
 682     // to the swtrace directory specification. The messages being the five defined below:
 683     // |  decl_id  |  decl_name          |    ui_name            |  arg_types  |  arg_names                          |
 684     // |-----------|---------------------|-----------------------|-------------|-------------------------------------|
 685     // |    0      |   declareLabel      |   declare label       |    ps       |  guid,value                         |
 686     // |    1      |   declareEntity     |   declare entity      |    p        |  guid                               |
 687     // |    2      | declareEventClass   |  declare event class  |    p        |  guid                               |
 688     // |    3      | declareRelationship | declare relationship  |    Ippp     |  relationshipType,relationshipGuid, |
 689     // |           |                     |                       |             |  headGuid,tailGuid                  |
 690     // |    4      |   declareEvent      |   declare event       |    @tp      |  timestamp,threadId,eventGuid       |
 691
 692     std::vector<std::vector<std::string>> timelineDirectoryMessages
 693     {
 694         {"declareLabel", "declare label", "ps", "guid,value"},
 695         {"declareEntity", "declare entity", "p", "guid"},
 696         {"declareEventClass", "declare event class", "p", "guid"},
 697         {"declareRelationship", "declare relationship", "Ippp", "relationshipType,relationshipGuid,headGuid,tailGuid"},
 698         {"declareEvent", "declare event", "@tp", "timestamp,threadId,eventGuid"}
 699     };
 700
 701     unsigned int messagesDataLength = 0u;
 702     std::vector<std::vector<std::vector<uint32_t>>> swTraceTimelineDirectoryMessages;
 703
 704     for (const auto& timelineDirectoryMessage : timelineDirectoryMessages)
 705     {
 706         messagesDataLength += uint32_t_size; // decl_id
 707
 708         std::vector<std::vector<uint32_t>> swTraceStringsVector;
 709         for (const auto& label : timelineDirectoryMessage)
 710         {
 711             std::vector<uint32_t> swTraceString;
 712             bool result = StringToSwTraceString<SwTraceCharPolicy>(label, swTraceString);
 713             if (!result)
 714             {
 715                 return TimelinePacketStatus::Error;
 716             }
 717
 718             messagesDataLength += boost::numeric_cast<unsigned int>(swTraceString.size()) * uint32_t_size;
 719             swTraceStringsVector.push_back(swTraceString);
 720         }
 721         swTraceTimelineDirectoryMessages.push_back(swTraceStringsVector);
 722     }
 723
 724     // Calculate the timeline directory binary packet size (in bytes)
 725     unsigned int timelineDirectoryPacketSize = 2 * uint32_t_size + // Header (2 words)
 726                                                messagesDataLength; // 5 messages length
 727
 728     // Check whether the timeline directory binary packet fits in the given buffer
 729     if (timelineDirectoryPacketSize > bufferSize)
 730     {
 731         return TimelinePacketStatus::BufferExhaustion;
 732     }
 733
 734     // Create packet header
 735     uint32_t dataLength = boost::numeric_cast<uint32_t>(messagesDataLength);
 736     std::pair<uint32_t, uint32_t> packetHeader = CreateTimelinePacketHeader(1, 0, 0, 0, 0, dataLength);
 737
 738     // Initialize the offset for writing in the buffer
 739     unsigned int offset = 0;
 740
 741     // Write the timeline binary packet header to the buffer
 742     WriteUint32(buffer, offset, packetHeader.first);
 743     offset += uint32_t_size;
 744     WriteUint32(buffer, offset, packetHeader.second);
 745     offset += uint32_t_size;
 746
 747     for (unsigned int i = 0u; i < swTraceTimelineDirectoryMessages.size(); ++i)
 748     {
 749         // Write the timeline binary packet payload to the buffer
 750         WriteUint32(buffer, offset, i); // decl_id
 751         offset += uint32_t_size;
 752
 753         for (std::vector<uint32_t> swTraceString : swTraceTimelineDirectoryMessages[i])
 754         {
 755             for (uint32_t swTraceDeclStringWord : swTraceString)
 756             {
 757                 WriteUint32(buffer, offset, swTraceDeclStringWord);
 758                 offset += uint32_t_size;
 759             }
 760         }
 761     }
 762
 763     // Update the number of bytes written
 764     numberOfBytesWritten = timelineDirectoryPacketSize;
 765
 766     return TimelinePacketStatus::Ok;
 767 }
 768
 769 TimelinePacketStatus WriteTimelineEventClassBinaryPacket(uint64_t profilingGuid,
 770                                                          unsigned char* buffer,
 771                                                          unsigned int bufferSize,
 772                                                          unsigned int& numberOfBytesWritten)
 773 {
 774     // Initialize the output value
 775     numberOfBytesWritten = 0;
 776
 777     // Check that the given buffer is valid
 778     if (buffer == nullptr || bufferSize == 0)
 779     {
 780         return TimelinePacketStatus::BufferExhaustion;
 781     }
 782
 783     // Utils
 784     unsigned int uint32_t_size = sizeof(uint32_t);
 785     unsigned int uint64_t_size = sizeof(uint64_t);
 786
 787     // decl_id of the timeline message
 788     uint32_t declId = 2;
 789
 790     // Calculate the length of the data (in bytes)
 791     unsigned int packetBodySize = uint32_t_size + uint64_t_size; // decl_id + Profiling GUID
 792
 793     // Calculate the timeline binary packet size (in bytes)
 794     unsigned int packetSize = 2 * uint32_t_size + // Header (2 words)
 795                               packetBodySize;     // Body
 796
 797     // Check whether the timeline binary packet fits in the given buffer
 798     if (packetSize > bufferSize)
 799     {
 800         return TimelinePacketStatus::BufferExhaustion;
 801     }
 802
 803     // Create packet header
 804     std::pair<uint32_t, uint32_t> packetHeader = CreateTimelineMessagePacketHeader(packetBodySize);
 805
 806     // Initialize the offset for writing in the buffer
 807     unsigned int offset = 0;
 808
 809     // Write the timeline binary packet header to the buffer
 810     WriteUint32(buffer, offset, packetHeader.first);
 811     offset += uint32_t_size;
 812     WriteUint32(buffer, offset, packetHeader.second);
 813     offset += uint32_t_size;
 814
 815     // Write the timeline binary packet payload to the buffer
 816     WriteUint32(buffer, offset, declId);        // decl_id
 817     offset += uint32_t_size;
 818     WriteUint64(buffer, offset, profilingGuid); // Profiling GUID
 819
 820     // Update the number of bytes written
 821     numberOfBytesWritten = packetSize;
 822
 823     return TimelinePacketStatus::Ok;
 824 }
 825
 826 TimelinePacketStatus WriteTimelineEventBinaryPacket(uint64_t timestamp,
 827                                                     std::thread::id threadId,
 828                                                     uint64_t profilingGuid,
 829                                                     unsigned char* buffer,
 830                                                     unsigned int bufferSize,
 831                                                     unsigned int& numberOfBytesWritten)
 832 {
 833     // Initialize the output value
 834     numberOfBytesWritten = 0;
 835
 836     // Check that the given buffer is valid
 837     if (buffer == nullptr || bufferSize == 0)
 838     {
 839         return TimelinePacketStatus::BufferExhaustion;
 840     }
 841
 842     // Utils
 843     unsigned int uint32_t_size = sizeof(uint32_t);
 844     unsigned int uint64_t_size = sizeof(uint64_t);
 845     unsigned int threadId_size = sizeof(std::thread::id);
 846
 847     // decl_id of the timeline message
 848     uint32_t declId = 4;
 849
 850     // Calculate the length of the data (in bytes)
 851     unsigned int timelineEventPacketDataLength = uint32_t_size + // decl_id
 852                                                  uint64_t_size + // Timestamp
 853                                                  threadId_size + // Thread id
 854                                                  uint64_t_size;  // Profiling GUID
 855
 856     // Calculate the timeline binary packet size (in bytes)
 857     unsigned int timelineEventPacketSize = 2 * uint32_t_size +            // Header (2 words)
 858                                            timelineEventPacketDataLength; // Timestamp + thread id + profiling GUID
 859
 860     // Check whether the timeline binary packet fits in the given buffer
 861     if (timelineEventPacketSize > bufferSize)
 862     {
 863         return TimelinePacketStatus::BufferExhaustion;
 864     }
 865
 866     // Create packet header
 867     std::pair<uint32_t, uint32_t> packetHeader = CreateTimelineMessagePacketHeader(timelineEventPacketDataLength);
 868
 869     // Initialize the offset for writing in the buffer
 870     unsigned int offset = 0;
 871
 872     // Write the timeline binary packet header to the buffer
 873     WriteUint32(buffer, offset, packetHeader.first);
 874     offset += uint32_t_size;
 875     WriteUint32(buffer, offset, packetHeader.second);
 876     offset += uint32_t_size;
 877
 878     // Write the timeline binary packet payload to the buffer
 879     WriteUint32(buffer, offset, declId); // decl_id
 880     offset += uint32_t_size;
 881     WriteUint64(buffer, offset, timestamp); // Timestamp
 882     offset += uint64_t_size;
 883     WriteBytes(buffer, offset, &threadId, threadId_size); // Thread id
 884     offset += threadId_size;
 885     WriteUint64(buffer, offset, profilingGuid); // Profiling GUID
 886     offset += uint64_t_size;
 887
 888     // Update the number of bytes written
 889     numberOfBytesWritten = timelineEventPacketSize;
 890
 891     return TimelinePacketStatus::Ok;
 892 }
 893
 894 std::string CentreAlignFormatting(const std::string& stringToPass, const int spacingWidth)
 895 {
 896     std::stringstream outputStream, centrePadding;
 897     int padding = spacingWidth - static_cast<int>(stringToPass.size());
 898
 899     for (int i = 0; i < padding / 2; ++i)
 900     {
 901         centrePadding << " ";
 902     }
 903
 904     outputStream << centrePadding.str() << stringToPass << centrePadding.str();
 905
 906     if (padding > 0 && padding %2 != 0)
 907     {
 908         outputStream << " ";
 909     }
 910
 911     return outputStream.str();
 912 }
 913
 914 void PrintDeviceDetails(const std::pair<const unsigned short, std::unique_ptr<Device>>& devicePair)
 915 {
 916     std::string body;
 917
 918     body.append(CentreAlignFormatting(devicePair.second->m_Name, 20));
 919     body.append(" | ");
 920     body.append(CentreAlignFormatting(std::to_string(devicePair.first), 13));
 921     body.append(" | ");
 922     body.append(CentreAlignFormatting(std::to_string(devicePair.second->m_Cores), 10));
 923     body.append("\n");
 924
 925     std::cout << std::string(body.size(), '-') << "\n";
 926     std::cout<< body;
 927 }
 928
 929 void PrintCounterSetDetails(const std::pair<const unsigned short, std::unique_ptr<CounterSet>>& counterSetPair)
 930 {
 931     std::string body;
 932
 933     body.append(CentreAlignFormatting(counterSetPair.second->m_Name, 20));
 934     body.append(" | ");
 935     body.append(CentreAlignFormatting(std::to_string(counterSetPair.first), 13));
 936     body.append(" | ");
 937     body.append(CentreAlignFormatting(std::to_string(counterSetPair.second->m_Count), 10));
 938     body.append("\n");
 939
 940     std::cout << std::string(body.size(), '-') << "\n";
 941
 942     std::cout<< body;
 943 }
 944
 945 void PrintCounterDetails(std::shared_ptr<Counter>& counter)
 946 {
 947     std::string body;
 948
 949     body.append(CentreAlignFormatting(counter->m_Name, 20));
 950     body.append(" | ");
 951     body.append(CentreAlignFormatting(counter->m_Description, 50));
 952     body.append(" | ");
 953     body.append(CentreAlignFormatting(counter->m_Units, 14));
 954     body.append(" | ");
 955     body.append(CentreAlignFormatting(std::to_string(counter->m_Uid), 6));
 956     body.append(" | ");
 957     body.append(CentreAlignFormatting(std::to_string(counter->m_MaxCounterUid), 10));
 958     body.append(" | ");
 959     body.append(CentreAlignFormatting(std::to_string(counter->m_Class), 8));
 960     body.append(" | ");
 961     body.append(CentreAlignFormatting(std::to_string(counter->m_Interpolation), 14));
 962     body.append(" | ");
 963     body.append(CentreAlignFormatting(std::to_string(counter->m_Multiplier), 20));
 964     body.append(" | ");
 965     body.append(CentreAlignFormatting(std::to_string(counter->m_CounterSetUid), 16));
 966     body.append(" | ");
 967     body.append(CentreAlignFormatting(std::to_string(counter->m_DeviceUid), 14));
 968
 969     body.append("\n");
 970
 971     std::cout << std::string(body.size(), '-') << "\n";
 972
 973     std::cout << body;
 974 }
 975
 976 void PrintCategoryDetails(const std::unique_ptr<Category>& category,
 977                           std::unordered_map<unsigned short, std::shared_ptr<Counter>> counterMap)
 978 {
 979     std::string categoryBody;
 980     std::string categoryHeader;
 981
 982     categoryHeader.append(CentreAlignFormatting("Name", 20));
 983     categoryHeader.append(" | ");
 984     categoryHeader.append(CentreAlignFormatting("Device", 12));
 985     categoryHeader.append(" | ");
 986     categoryHeader.append(CentreAlignFormatting("Counter set UID:", 16));
 987     categoryHeader.append(" | ");
 988     categoryHeader.append(CentreAlignFormatting("Event Count", 14));
 989     categoryHeader.append("\n");
 990
 991     categoryBody.append(CentreAlignFormatting(category->m_Name, 20));
 992     categoryBody.append(" | ");
 993     categoryBody.append(CentreAlignFormatting(std::to_string(category->m_DeviceUid), 12));
 994     categoryBody.append(" | ");
 995     categoryBody.append(CentreAlignFormatting(std::to_string(category->m_CounterSetUid), 16));
 996     categoryBody.append(" | ");
 997     categoryBody.append(CentreAlignFormatting(std::to_string(category->m_Counters.size()), 14));
 998
 999     std::cout << "\n" << "\n";
1000     std::cout << CentreAlignFormatting("CATEGORY", static_cast<int>(categoryHeader.size()));
1001     std::cout << "\n";
1002     std::cout << std::string(categoryHeader.size(), '=') << "\n";
1003
1004     std::cout << categoryHeader;
1005
1006     std::cout << std::string(categoryBody.size(), '-') << "\n";
1007
1008     std::cout << categoryBody;
1009
1010     std::string counterHeader;
1011
1012     counterHeader.append(CentreAlignFormatting("Counter Name", 20));
1013     counterHeader.append(" | ");
1014     counterHeader.append(CentreAlignFormatting("Description", 50));
1015     counterHeader.append(" | ");
1016     counterHeader.append(CentreAlignFormatting("Units", 14));
1017     counterHeader.append(" | ");
1018     counterHeader.append(CentreAlignFormatting("UID", 6));
1019     counterHeader.append(" | ");
1020     counterHeader.append(CentreAlignFormatting("Max UID", 10));
1021     counterHeader.append(" | ");
1022     counterHeader.append(CentreAlignFormatting("Class", 8));
1023     counterHeader.append(" | ");
1024     counterHeader.append(CentreAlignFormatting("Interpolation", 14));
1025     counterHeader.append(" | ");
1026     counterHeader.append(CentreAlignFormatting("Multiplier", 20));
1027     counterHeader.append(" | ");
1028     counterHeader.append(CentreAlignFormatting("Counter set UID", 16));
1029     counterHeader.append(" | ");
1030     counterHeader.append(CentreAlignFormatting("Device UID", 14));
1031     counterHeader.append("\n");
1032
1033     std::cout << "\n" << "\n";
1034     std::cout << CentreAlignFormatting("EVENTS IN CATEGORY: " + category->m_Name,
1035                                        static_cast<int>(counterHeader.size()));
1036     std::cout << "\n";
1037     std::cout << std::string(counterHeader.size(), '=') << "\n";
1038     std::cout << counterHeader;
1039     for (auto& it: category->m_Counters) {
1040         auto search = counterMap.find(it);
1041         if(search != counterMap.end()) {
1042             PrintCounterDetails(search->second);
1043         }
1044     }
1045 }
1046
1047 void PrintCounterDirectory(ICounterDirectory& counterDirectory)
1048 {
1049     std::string devicesHeader;
1050
1051     devicesHeader.append(CentreAlignFormatting("Device name", 20));
1052     devicesHeader.append(" | ");
1053     devicesHeader.append(CentreAlignFormatting("UID", 13));
1054     devicesHeader.append(" | ");
1055     devicesHeader.append(CentreAlignFormatting("Cores", 10));
1056     devicesHeader.append("\n");
1057
1058     std::cout << "\n" << "\n";
1059     std::cout << CentreAlignFormatting("DEVICES", static_cast<int>(devicesHeader.size()));
1060     std::cout << "\n";
1061     std::cout << std::string(devicesHeader.size(), '=') << "\n";
1062     std::cout << devicesHeader;
1063     for (auto& it: counterDirectory.GetDevices()) {
1064         PrintDeviceDetails(it);
1065     }
1066
1067     std::string counterSetHeader;
1068
1069     counterSetHeader.append(CentreAlignFormatting("Counter set name", 20));
1070     counterSetHeader.append(" | ");
1071     counterSetHeader.append(CentreAlignFormatting("UID", 13));
1072     counterSetHeader.append(" | ");
1073     counterSetHeader.append(CentreAlignFormatting("Count", 10));
1074     counterSetHeader.append("\n");
1075
1076     std::cout << "\n" << "\n";
1077     std::cout << CentreAlignFormatting("COUNTER SETS", static_cast<int>(counterSetHeader.size()));
1078     std::cout << "\n";
1079     std::cout << std::string(counterSetHeader.size(), '=') << "\n";
1080
1081     std::cout << counterSetHeader;
1082
1083     for (auto& it: counterDirectory.GetCounterSets()) {
1084         PrintCounterSetDetails(it);
1085     }
1086
1087     auto counters = counterDirectory.GetCounters();
1088     for (auto& it: counterDirectory.GetCategories()) {
1089         PrintCategoryDetails(it, counters);
1090     }
1091     std::cout << "\n";
1092 }
1093
1094 uint64_t GetTimestamp()
1095 {
1096 #if USE_CLOCK_MONOTONIC_RAW
1097     using clock = MonotonicClockRaw;
1098 #else
1099     using clock = std::chrono::steady_clock;
1100 #endif
1101
1102     // Take a timestamp
1103     auto timestamp = clock::now();
1104
1105     return static_cast<uint64_t>(timestamp.time_since_epoch().count());
1106 }
1107
1108 } // namespace profiling
1109
1110 } // namespace armnn
1111
1112 namespace std
1113 {
1114
1115 bool operator==(const std::vector<uint8_t>& left, std::thread::id right)
1116 {
1117     return std::memcmp(left.data(), &right, left.size()) == 0;
1118 }
1119
1120 } // namespace std