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 "common/include/ProfilingException.hpp"
   9
  10 #include <armnn/Version.hpp>
  11
  12 #include <WallClockTimer.hpp>
  13
  14 #include <armnn/utility/Assert.hpp>
  15
  16 #include <fstream>
  17 #include <iostream>
  18 #include <limits>
  19
  20 namespace armnn
  21 {
  22
  23 namespace profiling
  24 {
  25
  26 namespace
  27 {
  28
  29 void ThrowIfCantGenerateNextUid(uint16_t uid, uint16_t cores = 0)
  30 {
  31     // Check that it is possible to generate the next UID without causing an overflow
  32     switch (cores)
  33     {
  34     case 0:
  35     case 1:
  36         // Number of cores not specified or set to 1 (a value of zero indicates the device is not capable of
  37         // running multiple parallel workloads and will not provide multiple streams of data for each event)
  38         if (uid == std::numeric_limits<uint16_t>::max())
  39         {
  40             throw RuntimeException("Generating the next UID for profiling would result in an overflow");
  41         }
  42         break;
  43     default: // cores > 1
  44         // Multiple cores available, as max_counter_uid has to be set to: counter_uid + cores - 1, the maximum
  45         // allowed value for a counter UID is consequently: uint16_t_max - cores + 1
  46         if (uid >= std::numeric_limits<uint16_t>::max() - cores + 1)
  47         {
  48             throw RuntimeException("Generating the next UID for profiling would result in an overflow");
  49         }
  50         break;
  51     }
  52 }
  53
  54 } // Anonymous namespace
  55
  56 uint16_t GetNextUid(bool peekOnly)
  57 {
  58     // The UID used for profiling objects and events. The first valid UID is 1, as 0 is a reserved value
  59     static uint16_t uid = 1;
  60
  61     // Check that it is possible to generate the next UID without causing an overflow (throws in case of error)
  62     ThrowIfCantGenerateNextUid(uid);
  63
  64     if (peekOnly)
  65     {
  66         // Peek only
  67         return uid;
  68     }
  69     else
  70     {
  71         // Get the next UID
  72         return uid++;
  73     }
  74 }
  75
  76 std::vector<uint16_t> GetNextCounterUids(uint16_t firstUid, uint16_t cores)
  77 {
  78     // Check that it is possible to generate the next counter UID without causing an overflow (throws in case of error)
  79     ThrowIfCantGenerateNextUid(firstUid, cores);
  80
  81     // Get the next counter UIDs
  82     size_t counterUidsSize = cores == 0 ? 1 : cores;
  83     std::vector<uint16_t> counterUids(counterUidsSize, 0);
  84     for (size_t i = 0; i < counterUidsSize; i++)
  85     {
  86         counterUids[i] = firstUid++;
  87     }
  88     return counterUids;
  89 }
  90
  91 void WriteBytes(const IPacketBufferPtr& packetBuffer, unsigned int offset,  const void* value, unsigned int valueSize)
  92 {
  93     ARMNN_ASSERT(packetBuffer);
  94
  95     WriteBytes(packetBuffer->GetWritableData(), offset, value, valueSize);
  96 }
  97
  98 uint32_t ConstructHeader(uint32_t packetFamily,
  99                          uint32_t packetId)
 100 {
 101     return (( packetFamily & 0x0000003F ) << 26 )|
 102            (( packetId & 0x000003FF )     << 16 );
 103 }
 104
 105 void WriteUint64(const std::unique_ptr<IPacketBuffer>& packetBuffer, unsigned int offset, uint64_t value)
 106 {
 107     ARMNN_ASSERT(packetBuffer);
 108
 109     WriteUint64(packetBuffer->GetWritableData(), offset, value);
 110 }
 111
 112 void WriteUint32(const IPacketBufferPtr& packetBuffer, unsigned int offset, uint32_t value)
 113 {
 114     ARMNN_ASSERT(packetBuffer);
 115
 116     WriteUint32(packetBuffer->GetWritableData(), offset, value);
 117 }
 118
 119 void WriteUint16(const IPacketBufferPtr& packetBuffer, unsigned int offset, uint16_t value)
 120 {
 121     ARMNN_ASSERT(packetBuffer);
 122
 123     WriteUint16(packetBuffer->GetWritableData(), offset, value);
 124 }
 125
 126 void WriteUint8(const IPacketBufferPtr& packetBuffer, unsigned int offset, uint8_t value)
 127 {
 128     ARMNN_ASSERT(packetBuffer);
 129
 130     WriteUint8(packetBuffer->GetWritableData(), offset, value);
 131 }
 132
 133 void WriteBytes(unsigned char* buffer, unsigned int offset, const void* value, unsigned int valueSize)
 134 {
 135     ARMNN_ASSERT(buffer);
 136     ARMNN_ASSERT(value);
 137
 138     for (unsigned int i = 0; i < valueSize; i++, offset++)
 139     {
 140         buffer[offset] = *(reinterpret_cast<const unsigned char*>(value) + i);
 141     }
 142 }
 143
 144 void WriteUint64(unsigned char* buffer, unsigned int offset, uint64_t value)
 145 {
 146     ARMNN_ASSERT(buffer);
 147
 148     buffer[offset]     = static_cast<unsigned char>(value & 0xFF);
 149     buffer[offset + 1] = static_cast<unsigned char>((value >> 8) & 0xFF);
 150     buffer[offset + 2] = static_cast<unsigned char>((value >> 16) & 0xFF);
 151     buffer[offset + 3] = static_cast<unsigned char>((value >> 24) & 0xFF);
 152     buffer[offset + 4] = static_cast<unsigned char>((value >> 32) & 0xFF);
 153     buffer[offset + 5] = static_cast<unsigned char>((value >> 40) & 0xFF);
 154     buffer[offset + 6] = static_cast<unsigned char>((value >> 48) & 0xFF);
 155     buffer[offset + 7] = static_cast<unsigned char>((value >> 56) & 0xFF);
 156 }
 157
 158 void WriteUint32(unsigned char* buffer, unsigned int offset, uint32_t value)
 159 {
 160     ARMNN_ASSERT(buffer);
 161
 162     buffer[offset]     = static_cast<unsigned char>(value & 0xFF);
 163     buffer[offset + 1] = static_cast<unsigned char>((value >> 8) & 0xFF);
 164     buffer[offset + 2] = static_cast<unsigned char>((value >> 16) & 0xFF);
 165     buffer[offset + 3] = static_cast<unsigned char>((value >> 24) & 0xFF);
 166 }
 167
 168 void WriteUint16(unsigned char* buffer, unsigned int offset, uint16_t value)
 169 {
 170     ARMNN_ASSERT(buffer);
 171
 172     buffer[offset]     = static_cast<unsigned char>(value & 0xFF);
 173     buffer[offset + 1] = static_cast<unsigned char>((value >> 8) & 0xFF);
 174 }
 175
 176 void WriteUint8(unsigned char* buffer, unsigned int offset, uint8_t value)
 177 {
 178     ARMNN_ASSERT(buffer);
 179
 180     buffer[offset] = static_cast<unsigned char>(value);
 181 }
 182
 183 void ReadBytes(const IPacketBufferPtr& packetBuffer, unsigned int offset, unsigned int valueSize, uint8_t outValue[])
 184 {
 185     ARMNN_ASSERT(packetBuffer);
 186
 187     ReadBytes(packetBuffer->GetReadableData(), offset, valueSize, outValue);
 188 }
 189
 190 uint64_t ReadUint64(const IPacketBufferPtr& packetBuffer, unsigned int offset)
 191 {
 192     ARMNN_ASSERT(packetBuffer);
 193
 194     return ReadUint64(packetBuffer->GetReadableData(), offset);
 195 }
 196
 197 uint32_t ReadUint32(const IPacketBufferPtr& packetBuffer, unsigned int offset)
 198 {
 199     ARMNN_ASSERT(packetBuffer);
 200
 201     return ReadUint32(packetBuffer->GetReadableData(), offset);
 202 }
 203
 204 uint16_t ReadUint16(const IPacketBufferPtr& packetBuffer, unsigned int offset)
 205 {
 206     ARMNN_ASSERT(packetBuffer);
 207
 208     return ReadUint16(packetBuffer->GetReadableData(), offset);
 209 }
 210
 211 uint8_t ReadUint8(const IPacketBufferPtr& packetBuffer, unsigned int offset)
 212 {
 213     ARMNN_ASSERT(packetBuffer);
 214
 215     return ReadUint8(packetBuffer->GetReadableData(), offset);
 216 }
 217
 218 void ReadBytes(const unsigned char* buffer, unsigned int offset, unsigned int valueSize, uint8_t outValue[])
 219 {
 220     ARMNN_ASSERT(buffer);
 221     ARMNN_ASSERT(outValue);
 222
 223     for (unsigned int i = 0; i < valueSize; i++, offset++)
 224     {
 225         outValue[i] = static_cast<uint8_t>(buffer[offset]);
 226     }
 227 }
 228
 229 uint64_t ReadUint64(const unsigned char* buffer, unsigned int offset)
 230 {
 231     ARMNN_ASSERT(buffer);
 232
 233     uint64_t value = 0;
 234     value  = static_cast<uint64_t>(buffer[offset]);
 235     value |= static_cast<uint64_t>(buffer[offset + 1]) << 8;
 236     value |= static_cast<uint64_t>(buffer[offset + 2]) << 16;
 237     value |= static_cast<uint64_t>(buffer[offset + 3]) << 24;
 238     value |= static_cast<uint64_t>(buffer[offset + 4]) << 32;
 239     value |= static_cast<uint64_t>(buffer[offset + 5]) << 40;
 240     value |= static_cast<uint64_t>(buffer[offset + 6]) << 48;
 241     value |= static_cast<uint64_t>(buffer[offset + 7]) << 56;
 242
 243     return value;
 244 }
 245
 246 uint32_t ReadUint32(const unsigned char* buffer, unsigned int offset)
 247 {
 248     ARMNN_ASSERT(buffer);
 249
 250     uint32_t value = 0;
 251     value  = static_cast<uint32_t>(buffer[offset]);
 252     value |= static_cast<uint32_t>(buffer[offset + 1]) << 8;
 253     value |= static_cast<uint32_t>(buffer[offset + 2]) << 16;
 254     value |= static_cast<uint32_t>(buffer[offset + 3]) << 24;
 255     return value;
 256 }
 257
 258 uint16_t ReadUint16(const unsigned char* buffer, unsigned int offset)
 259 {
 260     ARMNN_ASSERT(buffer);
 261
 262     uint32_t value = 0;
 263     value  = static_cast<uint32_t>(buffer[offset]);
 264     value |= static_cast<uint32_t>(buffer[offset + 1]) << 8;
 265     return static_cast<uint16_t>(value);
 266 }
 267
 268 uint8_t ReadUint8(const unsigned char* buffer, unsigned int offset)
 269 {
 270     ARMNN_ASSERT(buffer);
 271
 272     return buffer[offset];
 273 }
 274
 275 std::string GetSoftwareInfo()
 276 {
 277     return std::string("ArmNN");
 278 }
 279
 280 std::string GetHardwareVersion()
 281 {
 282     return std::string();
 283 }
 284
 285 std::string GetSoftwareVersion()
 286 {
 287     std::string armnnVersion(ARMNN_VERSION);
 288     std::string result = "Armnn " + armnnVersion.substr(2,2) + "." + armnnVersion.substr(4,2);
 289     return result;
 290 }
 291
 292 std::string GetProcessName()
 293 {
 294     std::ifstream comm("/proc/self/comm");
 295     std::string name;
 296     getline(comm, name);
 297     return name;
 298 }
 299
 300 // Calculate the actual length an SwString will be including the terminating null character
 301 // padding to bring it to the next uint32_t boundary but minus the leading uint32_t encoding
 302 // the size to allow the offset to be correctly updated when decoding a binary packet.
 303 uint32_t CalculateSizeOfPaddedSwString(const std::string& str)
 304 {
 305     std::vector<uint32_t> swTraceString;
 306     StringToSwTraceString<SwTraceCharPolicy>(str, swTraceString);
 307     unsigned int uint32_t_size = sizeof(uint32_t);
 308     uint32_t size = (boost::numeric_cast<uint32_t>(swTraceString.size()) - 1) * uint32_t_size;
 309     return size;
 310 }
 311
 312 // Read TimelineMessageDirectoryPacket from given IPacketBuffer and offset
 313 SwTraceMessage ReadSwTraceMessage(const unsigned char* packetBuffer, unsigned int& offset)
 314 {
 315     ARMNN_ASSERT(packetBuffer);
 316
 317     unsigned int uint32_t_size = sizeof(uint32_t);
 318
 319     SwTraceMessage swTraceMessage;
 320
 321     // Read the decl_id
 322     uint32_t readDeclId = ReadUint32(packetBuffer, offset);
 323     swTraceMessage.m_Id = readDeclId;
 324
 325     // SWTrace "namestring" format
 326     // length of the string (first 4 bytes) + string + null terminator
 327
 328     // Check the decl_name
 329     offset += uint32_t_size;
 330     uint32_t swTraceDeclNameLength = ReadUint32(packetBuffer, offset);
 331
 332     offset += uint32_t_size;
 333     std::vector<unsigned char> swTraceStringBuffer(swTraceDeclNameLength - 1);
 334     std::memcpy(swTraceStringBuffer.data(),
 335                 packetBuffer + offset, swTraceStringBuffer.size());
 336
 337     swTraceMessage.m_Name.assign(swTraceStringBuffer.begin(), swTraceStringBuffer.end()); // name
 338
 339     // Check the ui_name
 340     offset += CalculateSizeOfPaddedSwString(swTraceMessage.m_Name);
 341     uint32_t swTraceUINameLength = ReadUint32(packetBuffer, offset);
 342
 343     offset += uint32_t_size;
 344     swTraceStringBuffer.resize(swTraceUINameLength - 1);
 345     std::memcpy(swTraceStringBuffer.data(),
 346                 packetBuffer  + offset, swTraceStringBuffer.size());
 347
 348     swTraceMessage.m_UiName.assign(swTraceStringBuffer.begin(), swTraceStringBuffer.end()); // ui_name
 349
 350     // Check arg_types
 351     offset += CalculateSizeOfPaddedSwString(swTraceMessage.m_UiName);
 352     uint32_t swTraceArgTypesLength = ReadUint32(packetBuffer, offset);
 353
 354     offset += uint32_t_size;
 355     swTraceStringBuffer.resize(swTraceArgTypesLength - 1);
 356     std::memcpy(swTraceStringBuffer.data(),
 357                 packetBuffer  + offset, swTraceStringBuffer.size());
 358
 359     swTraceMessage.m_ArgTypes.assign(swTraceStringBuffer.begin(), swTraceStringBuffer.end()); // arg_types
 360
 361     std::string swTraceString(swTraceStringBuffer.begin(), swTraceStringBuffer.end());
 362
 363     // Check arg_names
 364     offset += CalculateSizeOfPaddedSwString(swTraceString);
 365     uint32_t swTraceArgNamesLength = ReadUint32(packetBuffer, offset);
 366
 367     offset += uint32_t_size;
 368     swTraceStringBuffer.resize(swTraceArgNamesLength - 1);
 369     std::memcpy(swTraceStringBuffer.data(),
 370                 packetBuffer  + offset, swTraceStringBuffer.size());
 371
 372     swTraceString.assign(swTraceStringBuffer.begin(), swTraceStringBuffer.end());
 373     std::stringstream stringStream(swTraceString);
 374     std::string argName;
 375     while (std::getline(stringStream, argName, ','))
 376     {
 377         swTraceMessage.m_ArgNames.push_back(argName);
 378     }
 379
 380     offset += CalculateSizeOfPaddedSwString(swTraceString);
 381
 382     return swTraceMessage;
 383 }
 384
 385 /// Creates a timeline packet header
 386 ///
 387 /// \params
 388 ///   packetFamiliy     Timeline Packet Family
 389 ///   packetClass       Timeline Packet Class
 390 ///   packetType        Timeline Packet Type
 391 ///   streamId          Stream identifier
 392 ///   seqeunceNumbered  When non-zero the 4 bytes following the header is a u32 sequence number
 393 ///   dataLength        Unsigned 24-bit integer. Length of data, in bytes. Zero is permitted
 394 ///
 395 /// \returns
 396 ///   Pair of uint32_t containing word0 and word1 of the header
 397 std::pair<uint32_t, uint32_t> CreateTimelinePacketHeader(uint32_t packetFamily,
 398                                                          uint32_t packetClass,
 399                                                          uint32_t packetType,
 400                                                          uint32_t streamId,
 401                                                          uint32_t sequenceNumbered,
 402                                                          uint32_t dataLength)
 403 {
 404     // Packet header word 0:
 405     // 26:31 [6] packet_family: timeline Packet Family, value 0b000001
 406     // 19:25 [7] packet_class: packet class
 407     // 16:18 [3] packet_type: packet type
 408     // 8:15  [8] reserved: all zeros
 409     // 0:7   [8] stream_id: stream identifier
 410     uint32_t packetHeaderWord0 = ((packetFamily & 0x0000003F) << 26) |
 411                                  ((packetClass  & 0x0000007F) << 19) |
 412                                  ((packetType   & 0x00000007) << 16) |
 413                                  ((streamId     & 0x00000007) <<  0);
 414
 415     // Packet header word 1:
 416     // 25:31 [7]  reserved: all zeros
 417     // 24    [1]  sequence_numbered: when non-zero the 4 bytes following the header is a u32 sequence number
 418     // 0:23  [24] data_length: unsigned 24-bit integer. Length of data, in bytes. Zero is permitted
 419     uint32_t packetHeaderWord1 = ((sequenceNumbered & 0x00000001) << 24) |
 420                                  ((dataLength       & 0x00FFFFFF) <<  0);
 421
 422     return std::make_pair(packetHeaderWord0, packetHeaderWord1);
 423 }
 424
 425 /// Creates a packet header for the timeline messages:
 426 /// * declareLabel
 427 /// * declareEntity
 428 /// * declareEventClass
 429 /// * declareRelationship
 430 /// * declareEvent
 431 ///
 432 /// \param
 433 ///   dataLength The length of the message body in bytes
 434 ///
 435 /// \returns
 436 ///   Pair of uint32_t containing word0 and word1 of the header
 437 std::pair<uint32_t, uint32_t> CreateTimelineMessagePacketHeader(unsigned int dataLength)
 438 {
 439     return CreateTimelinePacketHeader(1,           // Packet family
 440                                       0,           // Packet class
 441                                       1,           // Packet type
 442                                       0,           // Stream id
 443                                       0,           // Sequence number
 444                                       dataLength); // Data length
 445 }
 446
 447 TimelinePacketStatus WriteTimelineLabelBinaryPacket(uint64_t profilingGuid,
 448                                                     const std::string& label,
 449                                                     unsigned char* buffer,
 450                                                     unsigned int remainingBufferSize,
 451                                                     unsigned int& numberOfBytesWritten)
 452 {
 453     // Initialize the output value
 454     numberOfBytesWritten = 0;
 455
 456     // Check that the given buffer is valid
 457     if (buffer == nullptr || remainingBufferSize == 0)
 458     {
 459         return TimelinePacketStatus::BufferExhaustion;
 460     }
 461
 462     // Utils
 463     unsigned int uint32_t_size = sizeof(uint32_t);
 464     unsigned int uint64_t_size = sizeof(uint64_t);
 465
 466     // Convert the label into a SWTrace string
 467     std::vector<uint32_t> swTraceLabel;
 468     bool result = StringToSwTraceString<SwTraceCharPolicy>(label, swTraceLabel);
 469     if (!result)
 470     {
 471         return TimelinePacketStatus::Error;
 472     }
 473
 474     // Calculate the size of the SWTrace string label (in bytes)
 475     unsigned int swTraceLabelSize = boost::numeric_cast<unsigned int>(swTraceLabel.size()) * uint32_t_size;
 476
 477     // Calculate the length of the data (in bytes)
 478     unsigned int timelineLabelPacketDataLength = uint32_t_size +   // decl_Id
 479                                                  uint64_t_size +   // Profiling GUID
 480                                                  swTraceLabelSize; // Label
 481
 482     // Check whether the timeline binary packet fits in the given buffer
 483     if (timelineLabelPacketDataLength > remainingBufferSize)
 484     {
 485         return TimelinePacketStatus::BufferExhaustion;
 486     }
 487
 488     // Initialize the offset for writing in the buffer
 489     unsigned int offset = 0;
 490
 491     // Write decl_Id to the buffer
 492     WriteUint32(buffer, offset, 0u);
 493     offset += uint32_t_size;
 494
 495     // Write the timeline binary packet payload to the buffer
 496     WriteUint64(buffer, offset, profilingGuid); // Profiling GUID
 497     offset += uint64_t_size;
 498     for (uint32_t swTraceLabelWord : swTraceLabel)
 499     {
 500         WriteUint32(buffer, offset, swTraceLabelWord); // Label
 501         offset += uint32_t_size;
 502     }
 503
 504     // Update the number of bytes written
 505     numberOfBytesWritten = timelineLabelPacketDataLength;
 506
 507     return TimelinePacketStatus::Ok;
 508 }
 509
 510 TimelinePacketStatus WriteTimelineEntityBinary(uint64_t profilingGuid,
 511                                                unsigned char* buffer,
 512                                                unsigned int remainingBufferSize,
 513                                                unsigned int& numberOfBytesWritten)
 514 {
 515     // Initialize the output value
 516     numberOfBytesWritten = 0;
 517
 518     // Check that the given buffer is valid
 519     if (buffer == nullptr || remainingBufferSize == 0)
 520     {
 521         return TimelinePacketStatus::BufferExhaustion;
 522     }
 523
 524     // Utils
 525     unsigned int uint32_t_size = sizeof(uint32_t);
 526     unsigned int uint64_t_size = sizeof(uint64_t);
 527
 528     // Calculate the length of the data (in bytes)
 529     unsigned int timelineEntityDataLength = uint32_t_size + uint64_t_size;  // decl_id + Profiling GUID
 530
 531     // Check whether the timeline binary packet fits in the given buffer
 532     if (timelineEntityDataLength > remainingBufferSize)
 533     {
 534         return TimelinePacketStatus::BufferExhaustion;
 535     }
 536
 537     // Initialize the offset for writing in the buffer
 538     unsigned int offset = 0;
 539
 540     // Write the decl_Id to the buffer
 541     WriteUint32(buffer, offset, 1u);
 542     offset += uint32_t_size;
 543
 544     // Write the timeline binary packet payload to the buffer
 545     WriteUint64(buffer, offset, profilingGuid); // Profiling GUID
 546
 547     // Update the number of bytes written
 548     numberOfBytesWritten = timelineEntityDataLength;
 549
 550     return TimelinePacketStatus::Ok;
 551 }
 552
 553 TimelinePacketStatus WriteTimelineRelationshipBinary(ProfilingRelationshipType relationshipType,
 554                                                      uint64_t relationshipGuid,
 555                                                      uint64_t headGuid,
 556                                                      uint64_t tailGuid,
 557                                                      unsigned char* buffer,
 558                                                      unsigned int remainingBufferSize,
 559                                                      unsigned int& numberOfBytesWritten)
 560 {
 561     // Initialize the output value
 562     numberOfBytesWritten = 0;
 563
 564     // Check that the given buffer is valid
 565     if (buffer == nullptr || remainingBufferSize == 0)
 566     {
 567         return TimelinePacketStatus::BufferExhaustion;
 568     }
 569
 570     // Utils
 571     unsigned int uint32_t_size = sizeof(uint32_t);
 572     unsigned int uint64_t_size = sizeof(uint64_t);
 573
 574     // Calculate the length of the data (in bytes)
 575     unsigned int timelineRelationshipDataLength = uint32_t_size * 2 + // decl_id + Relationship Type
 576                                                   uint64_t_size * 3;  // Relationship GUID + Head GUID + tail GUID
 577
 578     // Check whether the timeline binary fits in the given buffer
 579     if (timelineRelationshipDataLength > remainingBufferSize)
 580     {
 581         return TimelinePacketStatus::BufferExhaustion;
 582     }
 583
 584     // Initialize the offset for writing in the buffer
 585     unsigned int offset = 0;
 586
 587     uint32_t relationshipTypeUint = 0;
 588
 589     switch (relationshipType)
 590     {
 591         case ProfilingRelationshipType::RetentionLink:
 592             relationshipTypeUint = 0;
 593             break;
 594         case ProfilingRelationshipType::ExecutionLink:
 595             relationshipTypeUint = 1;
 596             break;
 597         case ProfilingRelationshipType::DataLink:
 598             relationshipTypeUint = 2;
 599             break;
 600         case ProfilingRelationshipType::LabelLink:
 601             relationshipTypeUint = 3;
 602             break;
 603         default:
 604             throw InvalidArgumentException("Unknown relationship type given.");
 605     }
 606
 607     // Write the timeline binary payload to the buffer
 608     // decl_id of the timeline message
 609     uint32_t declId = 3;
 610     WriteUint32(buffer, offset, declId); // decl_id
 611     offset += uint32_t_size;
 612     WriteUint32(buffer, offset, relationshipTypeUint); // Relationship Type
 613     offset += uint32_t_size;
 614     WriteUint64(buffer, offset, relationshipGuid); // GUID of this relationship
 615     offset += uint64_t_size;
 616     WriteUint64(buffer, offset, headGuid); // head of relationship GUID
 617     offset += uint64_t_size;
 618     WriteUint64(buffer, offset, tailGuid); // tail of relationship GUID
 619
 620     // Update the number of bytes written
 621     numberOfBytesWritten = timelineRelationshipDataLength;
 622
 623     return TimelinePacketStatus::Ok;
 624 }
 625
 626 TimelinePacketStatus WriteTimelineMessageDirectoryPackage(unsigned char* buffer,
 627                                                           unsigned int remainingBufferSize,
 628                                                           unsigned int& numberOfBytesWritten)
 629 {
 630     // Initialize the output value
 631     numberOfBytesWritten = 0;
 632
 633     // Check that the given buffer is valid
 634     if (buffer == nullptr || remainingBufferSize == 0)
 635     {
 636         return TimelinePacketStatus::BufferExhaustion;
 637     }
 638
 639     // Utils
 640     unsigned int uint8_t_size  = sizeof(uint8_t);
 641     unsigned int uint32_t_size = sizeof(uint32_t);
 642     unsigned int uint64_t_size = sizeof(uint64_t);
 643     unsigned int threadId_size = sizeof(std::thread::id);
 644
 645     // The payload/data of the packet consists of swtrace event definitions encoded according
 646     // to the swtrace directory specification. The messages being the five defined below:
 647     //
 648     // |  decl_id  |     decl_name       |      ui_name          |  arg_types  |            arg_names                |
 649     // |-----------|---------------------|-----------------------|-------------|-------------------------------------|
 650     // |    0      |   declareLabel      |   declare label       |    ps       |  guid,value                         |
 651     // |    1      |   declareEntity     |   declare entity      |    p        |  guid                               |
 652     // |    2      | declareEventClass   |  declare event class  |    p        |  guid                               |
 653     // |    3      | declareRelationship | declare relationship  |    Ippp     |  relationshipType,relationshipGuid, |
 654     // |           |                     |                       |             |  headGuid,tailGuid                  |
 655     // |    4      |   declareEvent      |   declare event       |    @tp      |  timestamp,threadId,eventGuid       |
 656     std::vector<std::vector<std::string>> timelineDirectoryMessages
 657     {
 658         { "0", "declareLabel", "declare label", "ps", "guid,value" },
 659         { "1", "declareEntity", "declare entity", "p", "guid" },
 660         { "2", "declareEventClass", "declare event class", "p", "guid" },
 661         { "3", "declareRelationship", "declare relationship", "Ippp",
 662           "relationshipType,relationshipGuid,headGuid,tailGuid" },
 663         { "4", "declareEvent", "declare event", "@tp", "timestamp,threadId,eventGuid" }
 664     };
 665
 666     // Build the message declarations
 667     std::vector<uint32_t> swTraceBuffer;
 668     for (const auto& directoryComponent : timelineDirectoryMessages)
 669     {
 670         // decl_id
 671         uint32_t declId = 0;
 672         try
 673         {
 674             declId = boost::numeric_cast<uint32_t>(std::stoul(directoryComponent[0]));
 675         }
 676         catch (const std::exception&)
 677         {
 678             return TimelinePacketStatus::Error;
 679         }
 680         swTraceBuffer.push_back(declId);
 681
 682         bool result = true;
 683         result &= ConvertDirectoryComponent<SwTraceNameCharPolicy>(directoryComponent[1], swTraceBuffer); // decl_name
 684         result &= ConvertDirectoryComponent<SwTraceCharPolicy>    (directoryComponent[2], swTraceBuffer); // ui_name
 685         result &= ConvertDirectoryComponent<SwTraceTypeCharPolicy>(directoryComponent[3], swTraceBuffer); // arg_types
 686         result &= ConvertDirectoryComponent<SwTraceCharPolicy>    (directoryComponent[4], swTraceBuffer); // arg_names
 687         if (!result)
 688         {
 689             return TimelinePacketStatus::Error;
 690         }
 691     }
 692
 693     unsigned int dataLength = 3 * uint8_t_size +  // Stream header (3 bytes)
 694                               boost::numeric_cast<unsigned int>(swTraceBuffer.size()) *
 695                                   uint32_t_size; // Trace directory (5 messages)
 696
 697     // Calculate the timeline directory binary packet size (in bytes)
 698     unsigned int timelineDirectoryPacketSize = 2 * uint32_t_size + // Header (2 words)
 699                                                dataLength;         // Payload
 700
 701     // Check whether the timeline directory binary packet fits in the given buffer
 702     if (timelineDirectoryPacketSize > remainingBufferSize)
 703     {
 704         return TimelinePacketStatus::BufferExhaustion;
 705     }
 706
 707     // Create packet header
 708     auto packetHeader = CreateTimelinePacketHeader(1, 0, 0, 0, 0, boost::numeric_cast<uint32_t>(dataLength));
 709
 710     // Initialize the offset for writing in the buffer
 711     unsigned int offset = 0;
 712
 713     // Write the timeline binary packet header to the buffer
 714     WriteUint32(buffer, offset, packetHeader.first);
 715     offset += uint32_t_size;
 716     WriteUint32(buffer, offset, packetHeader.second);
 717     offset += uint32_t_size;
 718
 719     // Write the stream header
 720     uint8_t streamVersion = 4;
 721     uint8_t pointerBytes  = boost::numeric_cast<uint8_t>(uint64_t_size); // All GUIDs are uint64_t
 722     uint8_t threadIdBytes = boost::numeric_cast<uint8_t>(threadId_size);
 723     switch (threadIdBytes)
 724     {
 725     case 4: // Typically Windows and Android
 726     case 8: // Typically Linux
 727         break; // Valid values
 728     default:
 729         return TimelinePacketStatus::Error; // Invalid value
 730     }
 731     WriteUint8(buffer, offset, streamVersion);
 732     offset += uint8_t_size;
 733     WriteUint8(buffer, offset, pointerBytes);
 734     offset += uint8_t_size;
 735     WriteUint8(buffer, offset, threadIdBytes);
 736     offset += uint8_t_size;
 737
 738     // Write the SWTrace directory
 739     uint32_t numberOfDeclarations = boost::numeric_cast<uint32_t>(timelineDirectoryMessages.size());
 740     WriteUint32(buffer, offset, numberOfDeclarations); // Number of declarations
 741     offset += uint32_t_size;
 742     for (uint32_t i : swTraceBuffer)
 743     {
 744         WriteUint32(buffer, offset, i); // Message declarations
 745         offset += uint32_t_size;
 746     }
 747
 748     // Update the number of bytes written
 749     numberOfBytesWritten = timelineDirectoryPacketSize;
 750
 751     return TimelinePacketStatus::Ok;
 752 }
 753
 754 TimelinePacketStatus WriteTimelineEventClassBinary(uint64_t profilingGuid,
 755                                                    unsigned char* buffer,
 756                                                    unsigned int remainingBufferSize,
 757                                                    unsigned int& numberOfBytesWritten)
 758 {
 759     // Initialize the output value
 760     numberOfBytesWritten = 0;
 761
 762     // Check that the given buffer is valid
 763     if (buffer == nullptr || remainingBufferSize == 0)
 764     {
 765         return TimelinePacketStatus::BufferExhaustion;
 766     }
 767
 768     // Utils
 769     unsigned int uint32_t_size = sizeof(uint32_t);
 770     unsigned int uint64_t_size = sizeof(uint64_t);
 771
 772     // decl_id of the timeline message
 773     uint32_t declId = 2;
 774
 775     // Calculate the length of the data (in bytes)
 776     unsigned int dataSize = uint32_t_size + uint64_t_size; // decl_id + Profiling GUID
 777
 778     // Check whether the timeline binary fits in the given buffer
 779     if (dataSize > remainingBufferSize)
 780     {
 781         return TimelinePacketStatus::BufferExhaustion;
 782     }
 783
 784     // Initialize the offset for writing in the buffer
 785     unsigned int offset = 0;
 786
 787     // Write the timeline binary payload to the buffer
 788     WriteUint32(buffer, offset, declId);        // decl_id
 789     offset += uint32_t_size;
 790     WriteUint64(buffer, offset, profilingGuid); // Profiling GUID
 791
 792     // Update the number of bytes written
 793     numberOfBytesWritten = dataSize;
 794
 795     return TimelinePacketStatus::Ok;
 796 }
 797
 798 TimelinePacketStatus WriteTimelineEventBinary(uint64_t timestamp,
 799                                               std::thread::id threadId,
 800                                               uint64_t profilingGuid,
 801                                               unsigned char* buffer,
 802                                               unsigned int remainingBufferSize,
 803                                               unsigned int& numberOfBytesWritten)
 804 {
 805     // Initialize the output value
 806     numberOfBytesWritten = 0;
 807     // Check that the given buffer is valid
 808     if (buffer == nullptr || remainingBufferSize == 0)
 809     {
 810         return TimelinePacketStatus::BufferExhaustion;
 811     }
 812
 813     // Utils
 814     unsigned int uint32_t_size = sizeof(uint32_t);
 815     unsigned int uint64_t_size = sizeof(uint64_t);
 816     unsigned int threadId_size = sizeof(std::thread::id);
 817
 818     // decl_id of the timeline message
 819     uint32_t declId = 4;
 820
 821     // Calculate the length of the data (in bytes)
 822     unsigned int timelineEventDataLength = uint32_t_size + // decl_id
 823                                            uint64_t_size + // Timestamp
 824                                            threadId_size + // Thread id
 825                                            uint64_t_size;  // Profiling GUID
 826
 827     // Check whether the timeline binary packet fits in the given buffer
 828     if (timelineEventDataLength > remainingBufferSize)
 829     {
 830         return TimelinePacketStatus::BufferExhaustion;
 831     }
 832
 833     // Initialize the offset for writing in the buffer
 834     unsigned int offset = 0;
 835
 836     // Write the timeline binary payload to the buffer
 837     WriteUint32(buffer, offset, declId); // decl_id
 838     offset += uint32_t_size;
 839     WriteUint64(buffer, offset, timestamp); // Timestamp
 840     offset += uint64_t_size;
 841     WriteBytes(buffer, offset, &threadId, threadId_size); // Thread id
 842     offset += threadId_size;
 843     WriteUint64(buffer, offset, profilingGuid); // Profiling GUID
 844     offset += uint64_t_size;
 845     // Update the number of bytes written
 846     numberOfBytesWritten = timelineEventDataLength;
 847
 848     return TimelinePacketStatus::Ok;
 849 }
 850
 851 std::string CentreAlignFormatting(const std::string& stringToPass, const int spacingWidth)
 852 {
 853     std::stringstream outputStream, centrePadding;
 854     int padding = spacingWidth - static_cast<int>(stringToPass.size());
 855
 856     for (int i = 0; i < padding / 2; ++i)
 857     {
 858         centrePadding << " ";
 859     }
 860
 861     outputStream << centrePadding.str() << stringToPass << centrePadding.str();
 862
 863     if (padding > 0 && padding %2 != 0)
 864     {
 865         outputStream << " ";
 866     }
 867
 868     return outputStream.str();
 869 }
 870
 871 void PrintDeviceDetails(const std::pair<const unsigned short, std::unique_ptr<Device>>& devicePair)
 872 {
 873     std::string body;
 874
 875     body.append(CentreAlignFormatting(devicePair.second->m_Name, 20));
 876     body.append(" | ");
 877     body.append(CentreAlignFormatting(std::to_string(devicePair.first), 13));
 878     body.append(" | ");
 879     body.append(CentreAlignFormatting(std::to_string(devicePair.second->m_Cores), 10));
 880     body.append("\n");
 881
 882     std::cout << std::string(body.size(), '-') << "\n";
 883     std::cout<< body;
 884 }
 885
 886 void PrintCounterSetDetails(const std::pair<const unsigned short, std::unique_ptr<CounterSet>>& counterSetPair)
 887 {
 888     std::string body;
 889
 890     body.append(CentreAlignFormatting(counterSetPair.second->m_Name, 20));
 891     body.append(" | ");
 892     body.append(CentreAlignFormatting(std::to_string(counterSetPair.first), 13));
 893     body.append(" | ");
 894     body.append(CentreAlignFormatting(std::to_string(counterSetPair.second->m_Count), 10));
 895     body.append("\n");
 896
 897     std::cout << std::string(body.size(), '-') << "\n";
 898
 899     std::cout<< body;
 900 }
 901
 902 void PrintCounterDetails(std::shared_ptr<Counter>& counter)
 903 {
 904     std::string body;
 905
 906     body.append(CentreAlignFormatting(counter->m_Name, 20));
 907     body.append(" | ");
 908     body.append(CentreAlignFormatting(counter->m_Description, 50));
 909     body.append(" | ");
 910     body.append(CentreAlignFormatting(counter->m_Units, 14));
 911     body.append(" | ");
 912     body.append(CentreAlignFormatting(std::to_string(counter->m_Uid), 6));
 913     body.append(" | ");
 914     body.append(CentreAlignFormatting(std::to_string(counter->m_MaxCounterUid), 10));
 915     body.append(" | ");
 916     body.append(CentreAlignFormatting(std::to_string(counter->m_Class), 8));
 917     body.append(" | ");
 918     body.append(CentreAlignFormatting(std::to_string(counter->m_Interpolation), 14));
 919     body.append(" | ");
 920     body.append(CentreAlignFormatting(std::to_string(counter->m_Multiplier), 20));
 921     body.append(" | ");
 922     body.append(CentreAlignFormatting(std::to_string(counter->m_CounterSetUid), 16));
 923     body.append(" | ");
 924     body.append(CentreAlignFormatting(std::to_string(counter->m_DeviceUid), 14));
 925
 926     body.append("\n");
 927
 928     std::cout << std::string(body.size(), '-') << "\n";
 929
 930     std::cout << body;
 931 }
 932
 933 void PrintCategoryDetails(const std::unique_ptr<Category>& category,
 934                           std::unordered_map<unsigned short, std::shared_ptr<Counter>> counterMap)
 935 {
 936     std::string categoryBody;
 937     std::string categoryHeader;
 938
 939     categoryHeader.append(CentreAlignFormatting("Name", 20));
 940     categoryHeader.append(" | ");
 941     categoryHeader.append(CentreAlignFormatting("Event Count", 14));
 942     categoryHeader.append("\n");
 943
 944     categoryBody.append(CentreAlignFormatting(category->m_Name, 20));
 945     categoryBody.append(" | ");
 946     categoryBody.append(CentreAlignFormatting(std::to_string(category->m_Counters.size()), 14));
 947
 948     std::cout << "\n" << "\n";
 949     std::cout << CentreAlignFormatting("CATEGORY", static_cast<int>(categoryHeader.size()));
 950     std::cout << "\n";
 951     std::cout << std::string(categoryHeader.size(), '=') << "\n";
 952
 953     std::cout << categoryHeader;
 954
 955     std::cout << std::string(categoryBody.size(), '-') << "\n";
 956
 957     std::cout << categoryBody;
 958
 959     std::string counterHeader;
 960
 961     counterHeader.append(CentreAlignFormatting("Counter Name", 20));
 962     counterHeader.append(" | ");
 963     counterHeader.append(CentreAlignFormatting("Description", 50));
 964     counterHeader.append(" | ");
 965     counterHeader.append(CentreAlignFormatting("Units", 14));
 966     counterHeader.append(" | ");
 967     counterHeader.append(CentreAlignFormatting("UID", 6));
 968     counterHeader.append(" | ");
 969     counterHeader.append(CentreAlignFormatting("Max UID", 10));
 970     counterHeader.append(" | ");
 971     counterHeader.append(CentreAlignFormatting("Class", 8));
 972     counterHeader.append(" | ");
 973     counterHeader.append(CentreAlignFormatting("Interpolation", 14));
 974     counterHeader.append(" | ");
 975     counterHeader.append(CentreAlignFormatting("Multiplier", 20));
 976     counterHeader.append(" | ");
 977     counterHeader.append(CentreAlignFormatting("Counter set UID", 16));
 978     counterHeader.append(" | ");
 979     counterHeader.append(CentreAlignFormatting("Device UID", 14));
 980     counterHeader.append("\n");
 981
 982     std::cout << "\n" << "\n";
 983     std::cout << CentreAlignFormatting("EVENTS IN CATEGORY: " + category->m_Name,
 984                                        static_cast<int>(counterHeader.size()));
 985     std::cout << "\n";
 986     std::cout << std::string(counterHeader.size(), '=') << "\n";
 987     std::cout << counterHeader;
 988     for (auto& it: category->m_Counters) {
 989         auto search = counterMap.find(it);
 990         if(search != counterMap.end()) {
 991             PrintCounterDetails(search->second);
 992         }
 993     }
 994 }
 995
 996 void PrintCounterDirectory(ICounterDirectory& counterDirectory)
 997 {
 998     std::string devicesHeader;
 999
1000     devicesHeader.append(CentreAlignFormatting("Device name", 20));
1001     devicesHeader.append(" | ");
1002     devicesHeader.append(CentreAlignFormatting("UID", 13));
1003     devicesHeader.append(" | ");
1004     devicesHeader.append(CentreAlignFormatting("Cores", 10));
1005     devicesHeader.append("\n");
1006
1007     std::cout << "\n" << "\n";
1008     std::cout << CentreAlignFormatting("DEVICES", static_cast<int>(devicesHeader.size()));
1009     std::cout << "\n";
1010     std::cout << std::string(devicesHeader.size(), '=') << "\n";
1011     std::cout << devicesHeader;
1012     for (auto& it: counterDirectory.GetDevices()) {
1013         PrintDeviceDetails(it);
1014     }
1015
1016     std::string counterSetHeader;
1017
1018     counterSetHeader.append(CentreAlignFormatting("Counter set name", 20));
1019     counterSetHeader.append(" | ");
1020     counterSetHeader.append(CentreAlignFormatting("UID", 13));
1021     counterSetHeader.append(" | ");
1022     counterSetHeader.append(CentreAlignFormatting("Count", 10));
1023     counterSetHeader.append("\n");
1024
1025     std::cout << "\n" << "\n";
1026     std::cout << CentreAlignFormatting("COUNTER SETS", static_cast<int>(counterSetHeader.size()));
1027     std::cout << "\n";
1028     std::cout << std::string(counterSetHeader.size(), '=') << "\n";
1029
1030     std::cout << counterSetHeader;
1031
1032     for (auto& it: counterDirectory.GetCounterSets()) {
1033         PrintCounterSetDetails(it);
1034     }
1035
1036     auto counters = counterDirectory.GetCounters();
1037     for (auto& it: counterDirectory.GetCategories()) {
1038         PrintCategoryDetails(it, counters);
1039     }
1040     std::cout << "\n";
1041 }
1042
1043 uint64_t GetTimestamp()
1044 {
1045 #if USE_CLOCK_MONOTONIC_RAW
1046     using clock = MonotonicClockRaw;
1047 #else
1048     using clock = std::chrono::steady_clock;
1049 #endif
1050
1051     // Take a timestamp
1052     auto timestamp = std::chrono::duration_cast<std::chrono::nanoseconds>(clock::now().time_since_epoch());
1053
1054     return static_cast<uint64_t>(timestamp.count());
1055 }
1056
1057 Packet ReceivePacket(const unsigned char* buffer, uint32_t length)
1058 {
1059     if (buffer == nullptr)
1060     {
1061         throw armnnProfiling::ProfilingException("data buffer is nullptr");
1062     }
1063     if (length < 8)
1064     {
1065         throw armnnProfiling::ProfilingException("length of data buffer is less than 8");
1066     }
1067
1068     uint32_t metadataIdentifier = 0;
1069     std::memcpy(&metadataIdentifier, buffer, sizeof(metadataIdentifier));
1070
1071     uint32_t dataLength = 0;
1072     std::memcpy(&dataLength, buffer + 4u, sizeof(dataLength));
1073
1074     std::unique_ptr<unsigned char[]> packetData;
1075     if (dataLength > 0)
1076     {
1077         packetData = std::make_unique<unsigned char[]>(dataLength);
1078         std::memcpy(packetData.get(), buffer + 8u, dataLength);
1079     }
1080
1081     return Packet(metadataIdentifier, dataLength, packetData);
1082 }
1083
1084 } // namespace profiling
1085
1086 } // namespace armnn
1087
1088 namespace std
1089 {
1090
1091 bool operator==(const std::vector<uint8_t>& left, std::thread::id right)
1092 {
1093     return std::memcmp(left.data(), &right, left.size()) == 0;
1094 }
1095
1096 } // namespace std