2 // Copyright © 2017 Arm Ltd. All rights reserved.
3 // SPDX-License-Identifier: MIT
6 #include "ProfilingUtils.hpp"
8 #include "common/include/ProfilingException.hpp"
10 #include <armnn/Version.hpp>
12 #include <WallClockTimer.hpp>
14 #include <armnn/utility/Assert.hpp>
29 void ThrowIfCantGenerateNextUid(uint16_t uid, uint16_t cores = 0)
31 // Check that it is possible to generate the next UID without causing an overflow
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())
40 throw RuntimeException("Generating the next UID for profiling would result in an overflow");
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)
48 throw RuntimeException("Generating the next UID for profiling would result in an overflow");
54 } // Anonymous namespace
56 uint16_t GetNextUid(bool peekOnly)
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;
61 // Check that it is possible to generate the next UID without causing an overflow (throws in case of error)
62 ThrowIfCantGenerateNextUid(uid);
76 std::vector<uint16_t> GetNextCounterUids(uint16_t firstUid, uint16_t cores)
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);
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++)
86 counterUids[i] = firstUid++;
91 void WriteBytes(const IPacketBufferPtr& packetBuffer, unsigned int offset, const void* value, unsigned int valueSize)
93 ARMNN_ASSERT(packetBuffer);
95 WriteBytes(packetBuffer->GetWritableData(), offset, value, valueSize);
98 uint32_t ConstructHeader(uint32_t packetFamily,
101 return (( packetFamily & 0x0000003F ) << 26 )|
102 (( packetId & 0x000003FF ) << 16 );
105 void WriteUint64(const std::unique_ptr<IPacketBuffer>& packetBuffer, unsigned int offset, uint64_t value)
107 ARMNN_ASSERT(packetBuffer);
109 WriteUint64(packetBuffer->GetWritableData(), offset, value);
112 void WriteUint32(const IPacketBufferPtr& packetBuffer, unsigned int offset, uint32_t value)
114 ARMNN_ASSERT(packetBuffer);
116 WriteUint32(packetBuffer->GetWritableData(), offset, value);
119 void WriteUint16(const IPacketBufferPtr& packetBuffer, unsigned int offset, uint16_t value)
121 ARMNN_ASSERT(packetBuffer);
123 WriteUint16(packetBuffer->GetWritableData(), offset, value);
126 void WriteUint8(const IPacketBufferPtr& packetBuffer, unsigned int offset, uint8_t value)
128 ARMNN_ASSERT(packetBuffer);
130 WriteUint8(packetBuffer->GetWritableData(), offset, value);
133 void WriteBytes(unsigned char* buffer, unsigned int offset, const void* value, unsigned int valueSize)
135 ARMNN_ASSERT(buffer);
138 for (unsigned int i = 0; i < valueSize; i++, offset++)
140 buffer[offset] = *(reinterpret_cast<const unsigned char*>(value) + i);
144 void WriteUint64(unsigned char* buffer, unsigned int offset, uint64_t value)
146 ARMNN_ASSERT(buffer);
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);
158 void WriteUint32(unsigned char* buffer, unsigned int offset, uint32_t value)
160 ARMNN_ASSERT(buffer);
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);
168 void WriteUint16(unsigned char* buffer, unsigned int offset, uint16_t value)
170 ARMNN_ASSERT(buffer);
172 buffer[offset] = static_cast<unsigned char>(value & 0xFF);
173 buffer[offset + 1] = static_cast<unsigned char>((value >> 8) & 0xFF);
176 void WriteUint8(unsigned char* buffer, unsigned int offset, uint8_t value)
178 ARMNN_ASSERT(buffer);
180 buffer[offset] = static_cast<unsigned char>(value);
183 void ReadBytes(const IPacketBufferPtr& packetBuffer, unsigned int offset, unsigned int valueSize, uint8_t outValue[])
185 ARMNN_ASSERT(packetBuffer);
187 ReadBytes(packetBuffer->GetReadableData(), offset, valueSize, outValue);
190 uint64_t ReadUint64(const IPacketBufferPtr& packetBuffer, unsigned int offset)
192 ARMNN_ASSERT(packetBuffer);
194 return ReadUint64(packetBuffer->GetReadableData(), offset);
197 uint32_t ReadUint32(const IPacketBufferPtr& packetBuffer, unsigned int offset)
199 ARMNN_ASSERT(packetBuffer);
201 return ReadUint32(packetBuffer->GetReadableData(), offset);
204 uint16_t ReadUint16(const IPacketBufferPtr& packetBuffer, unsigned int offset)
206 ARMNN_ASSERT(packetBuffer);
208 return ReadUint16(packetBuffer->GetReadableData(), offset);
211 uint8_t ReadUint8(const IPacketBufferPtr& packetBuffer, unsigned int offset)
213 ARMNN_ASSERT(packetBuffer);
215 return ReadUint8(packetBuffer->GetReadableData(), offset);
218 void ReadBytes(const unsigned char* buffer, unsigned int offset, unsigned int valueSize, uint8_t outValue[])
220 ARMNN_ASSERT(buffer);
221 ARMNN_ASSERT(outValue);
223 for (unsigned int i = 0; i < valueSize; i++, offset++)
225 outValue[i] = static_cast<uint8_t>(buffer[offset]);
229 uint64_t ReadUint64(const unsigned char* buffer, unsigned int offset)
231 ARMNN_ASSERT(buffer);
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;
246 uint32_t ReadUint32(const unsigned char* buffer, unsigned int offset)
248 ARMNN_ASSERT(buffer);
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;
258 uint16_t ReadUint16(const unsigned char* buffer, unsigned int offset)
260 ARMNN_ASSERT(buffer);
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);
268 uint8_t ReadUint8(const unsigned char* buffer, unsigned int offset)
270 ARMNN_ASSERT(buffer);
272 return buffer[offset];
275 std::string GetSoftwareInfo()
277 return std::string("ArmNN");
280 std::string GetHardwareVersion()
282 return std::string();
285 std::string GetSoftwareVersion()
287 std::string armnnVersion(ARMNN_VERSION);
288 std::string result = "Armnn " + armnnVersion.substr(2,2) + "." + armnnVersion.substr(4,2);
292 std::string GetProcessName()
294 std::ifstream comm("/proc/self/comm");
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)
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;
312 // Read TimelineMessageDirectoryPacket from given IPacketBuffer and offset
313 SwTraceMessage ReadSwTraceMessage(const unsigned char* packetBuffer, unsigned int& offset)
315 ARMNN_ASSERT(packetBuffer);
317 unsigned int uint32_t_size = sizeof(uint32_t);
319 SwTraceMessage swTraceMessage;
322 uint32_t readDeclId = ReadUint32(packetBuffer, offset);
323 swTraceMessage.m_Id = readDeclId;
325 // SWTrace "namestring" format
326 // length of the string (first 4 bytes) + string + null terminator
328 // Check the decl_name
329 offset += uint32_t_size;
330 uint32_t swTraceDeclNameLength = ReadUint32(packetBuffer, offset);
332 offset += uint32_t_size;
333 std::vector<unsigned char> swTraceStringBuffer(swTraceDeclNameLength - 1);
334 std::memcpy(swTraceStringBuffer.data(),
335 packetBuffer + offset, swTraceStringBuffer.size());
337 swTraceMessage.m_Name.assign(swTraceStringBuffer.begin(), swTraceStringBuffer.end()); // name
340 offset += CalculateSizeOfPaddedSwString(swTraceMessage.m_Name);
341 uint32_t swTraceUINameLength = ReadUint32(packetBuffer, offset);
343 offset += uint32_t_size;
344 swTraceStringBuffer.resize(swTraceUINameLength - 1);
345 std::memcpy(swTraceStringBuffer.data(),
346 packetBuffer + offset, swTraceStringBuffer.size());
348 swTraceMessage.m_UiName.assign(swTraceStringBuffer.begin(), swTraceStringBuffer.end()); // ui_name
351 offset += CalculateSizeOfPaddedSwString(swTraceMessage.m_UiName);
352 uint32_t swTraceArgTypesLength = ReadUint32(packetBuffer, offset);
354 offset += uint32_t_size;
355 swTraceStringBuffer.resize(swTraceArgTypesLength - 1);
356 std::memcpy(swTraceStringBuffer.data(),
357 packetBuffer + offset, swTraceStringBuffer.size());
359 swTraceMessage.m_ArgTypes.assign(swTraceStringBuffer.begin(), swTraceStringBuffer.end()); // arg_types
361 std::string swTraceString(swTraceStringBuffer.begin(), swTraceStringBuffer.end());
364 offset += CalculateSizeOfPaddedSwString(swTraceString);
365 uint32_t swTraceArgNamesLength = ReadUint32(packetBuffer, offset);
367 offset += uint32_t_size;
368 swTraceStringBuffer.resize(swTraceArgNamesLength - 1);
369 std::memcpy(swTraceStringBuffer.data(),
370 packetBuffer + offset, swTraceStringBuffer.size());
372 swTraceString.assign(swTraceStringBuffer.begin(), swTraceStringBuffer.end());
373 std::stringstream stringStream(swTraceString);
375 while (std::getline(stringStream, argName, ','))
377 swTraceMessage.m_ArgNames.push_back(argName);
380 offset += CalculateSizeOfPaddedSwString(swTraceString);
382 return swTraceMessage;
385 /// Creates a timeline packet header
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
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,
401 uint32_t sequenceNumbered,
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);
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);
422 return std::make_pair(packetHeaderWord0, packetHeaderWord1);
425 /// Creates a packet header for the timeline messages:
428 /// * declareEventClass
429 /// * declareRelationship
433 /// dataLength The length of the message body in bytes
436 /// Pair of uint32_t containing word0 and word1 of the header
437 std::pair<uint32_t, uint32_t> CreateTimelineMessagePacketHeader(unsigned int dataLength)
439 return CreateTimelinePacketHeader(1, // Packet family
443 0, // Sequence number
444 dataLength); // Data length
447 TimelinePacketStatus WriteTimelineLabelBinaryPacket(uint64_t profilingGuid,
448 const std::string& label,
449 unsigned char* buffer,
450 unsigned int remainingBufferSize,
451 unsigned int& numberOfBytesWritten)
453 // Initialize the output value
454 numberOfBytesWritten = 0;
456 // Check that the given buffer is valid
457 if (buffer == nullptr || remainingBufferSize == 0)
459 return TimelinePacketStatus::BufferExhaustion;
463 unsigned int uint32_t_size = sizeof(uint32_t);
464 unsigned int uint64_t_size = sizeof(uint64_t);
466 // Convert the label into a SWTrace string
467 std::vector<uint32_t> swTraceLabel;
468 bool result = StringToSwTraceString<SwTraceCharPolicy>(label, swTraceLabel);
471 return TimelinePacketStatus::Error;
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;
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
482 // Check whether the timeline binary packet fits in the given buffer
483 if (timelineLabelPacketDataLength > remainingBufferSize)
485 return TimelinePacketStatus::BufferExhaustion;
488 // Initialize the offset for writing in the buffer
489 unsigned int offset = 0;
491 // Write decl_Id to the buffer
492 WriteUint32(buffer, offset, 0u);
493 offset += uint32_t_size;
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)
500 WriteUint32(buffer, offset, swTraceLabelWord); // Label
501 offset += uint32_t_size;
504 // Update the number of bytes written
505 numberOfBytesWritten = timelineLabelPacketDataLength;
507 return TimelinePacketStatus::Ok;
510 TimelinePacketStatus WriteTimelineEntityBinary(uint64_t profilingGuid,
511 unsigned char* buffer,
512 unsigned int remainingBufferSize,
513 unsigned int& numberOfBytesWritten)
515 // Initialize the output value
516 numberOfBytesWritten = 0;
518 // Check that the given buffer is valid
519 if (buffer == nullptr || remainingBufferSize == 0)
521 return TimelinePacketStatus::BufferExhaustion;
525 unsigned int uint32_t_size = sizeof(uint32_t);
526 unsigned int uint64_t_size = sizeof(uint64_t);
528 // Calculate the length of the data (in bytes)
529 unsigned int timelineEntityDataLength = uint32_t_size + uint64_t_size; // decl_id + Profiling GUID
531 // Check whether the timeline binary packet fits in the given buffer
532 if (timelineEntityDataLength > remainingBufferSize)
534 return TimelinePacketStatus::BufferExhaustion;
537 // Initialize the offset for writing in the buffer
538 unsigned int offset = 0;
540 // Write the decl_Id to the buffer
541 WriteUint32(buffer, offset, 1u);
542 offset += uint32_t_size;
544 // Write the timeline binary packet payload to the buffer
545 WriteUint64(buffer, offset, profilingGuid); // Profiling GUID
547 // Update the number of bytes written
548 numberOfBytesWritten = timelineEntityDataLength;
550 return TimelinePacketStatus::Ok;
553 TimelinePacketStatus WriteTimelineRelationshipBinary(ProfilingRelationshipType relationshipType,
554 uint64_t relationshipGuid,
557 unsigned char* buffer,
558 unsigned int remainingBufferSize,
559 unsigned int& numberOfBytesWritten)
561 // Initialize the output value
562 numberOfBytesWritten = 0;
564 // Check that the given buffer is valid
565 if (buffer == nullptr || remainingBufferSize == 0)
567 return TimelinePacketStatus::BufferExhaustion;
571 unsigned int uint32_t_size = sizeof(uint32_t);
572 unsigned int uint64_t_size = sizeof(uint64_t);
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
578 // Check whether the timeline binary fits in the given buffer
579 if (timelineRelationshipDataLength > remainingBufferSize)
581 return TimelinePacketStatus::BufferExhaustion;
584 // Initialize the offset for writing in the buffer
585 unsigned int offset = 0;
587 uint32_t relationshipTypeUint = 0;
589 switch (relationshipType)
591 case ProfilingRelationshipType::RetentionLink:
592 relationshipTypeUint = 0;
594 case ProfilingRelationshipType::ExecutionLink:
595 relationshipTypeUint = 1;
597 case ProfilingRelationshipType::DataLink:
598 relationshipTypeUint = 2;
600 case ProfilingRelationshipType::LabelLink:
601 relationshipTypeUint = 3;
604 throw InvalidArgumentException("Unknown relationship type given.");
607 // Write the timeline binary payload to the buffer
608 // decl_id of the timeline message
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
620 // Update the number of bytes written
621 numberOfBytesWritten = timelineRelationshipDataLength;
623 return TimelinePacketStatus::Ok;
626 TimelinePacketStatus WriteTimelineMessageDirectoryPackage(unsigned char* buffer,
627 unsigned int remainingBufferSize,
628 unsigned int& numberOfBytesWritten)
630 // Initialize the output value
631 numberOfBytesWritten = 0;
633 // Check that the given buffer is valid
634 if (buffer == nullptr || remainingBufferSize == 0)
636 return TimelinePacketStatus::BufferExhaustion;
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);
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:
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
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" }
666 // Build the message declarations
667 std::vector<uint32_t> swTraceBuffer;
668 for (const auto& directoryComponent : timelineDirectoryMessages)
674 declId = boost::numeric_cast<uint32_t>(std::stoul(directoryComponent[0]));
676 catch (const std::exception&)
678 return TimelinePacketStatus::Error;
680 swTraceBuffer.push_back(declId);
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
689 return TimelinePacketStatus::Error;
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)
697 // Calculate the timeline directory binary packet size (in bytes)
698 unsigned int timelineDirectoryPacketSize = 2 * uint32_t_size + // Header (2 words)
699 dataLength; // Payload
701 // Check whether the timeline directory binary packet fits in the given buffer
702 if (timelineDirectoryPacketSize > remainingBufferSize)
704 return TimelinePacketStatus::BufferExhaustion;
707 // Create packet header
708 auto packetHeader = CreateTimelinePacketHeader(1, 0, 0, 0, 0, boost::numeric_cast<uint32_t>(dataLength));
710 // Initialize the offset for writing in the buffer
711 unsigned int offset = 0;
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;
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)
725 case 4: // Typically Windows and Android
726 case 8: // Typically Linux
727 break; // Valid values
729 return TimelinePacketStatus::Error; // Invalid value
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;
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)
744 WriteUint32(buffer, offset, i); // Message declarations
745 offset += uint32_t_size;
748 // Update the number of bytes written
749 numberOfBytesWritten = timelineDirectoryPacketSize;
751 return TimelinePacketStatus::Ok;
754 TimelinePacketStatus WriteTimelineEventClassBinary(uint64_t profilingGuid,
755 unsigned char* buffer,
756 unsigned int remainingBufferSize,
757 unsigned int& numberOfBytesWritten)
759 // Initialize the output value
760 numberOfBytesWritten = 0;
762 // Check that the given buffer is valid
763 if (buffer == nullptr || remainingBufferSize == 0)
765 return TimelinePacketStatus::BufferExhaustion;
769 unsigned int uint32_t_size = sizeof(uint32_t);
770 unsigned int uint64_t_size = sizeof(uint64_t);
772 // decl_id of the timeline message
775 // Calculate the length of the data (in bytes)
776 unsigned int dataSize = uint32_t_size + uint64_t_size; // decl_id + Profiling GUID
778 // Check whether the timeline binary fits in the given buffer
779 if (dataSize > remainingBufferSize)
781 return TimelinePacketStatus::BufferExhaustion;
784 // Initialize the offset for writing in the buffer
785 unsigned int offset = 0;
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
792 // Update the number of bytes written
793 numberOfBytesWritten = dataSize;
795 return TimelinePacketStatus::Ok;
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)
805 // Initialize the output value
806 numberOfBytesWritten = 0;
807 // Check that the given buffer is valid
808 if (buffer == nullptr || remainingBufferSize == 0)
810 return TimelinePacketStatus::BufferExhaustion;
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);
818 // decl_id of the timeline message
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
827 // Check whether the timeline binary packet fits in the given buffer
828 if (timelineEventDataLength > remainingBufferSize)
830 return TimelinePacketStatus::BufferExhaustion;
833 // Initialize the offset for writing in the buffer
834 unsigned int offset = 0;
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;
848 return TimelinePacketStatus::Ok;
851 std::string CentreAlignFormatting(const std::string& stringToPass, const int spacingWidth)
853 std::stringstream outputStream, centrePadding;
854 int padding = spacingWidth - static_cast<int>(stringToPass.size());
856 for (int i = 0; i < padding / 2; ++i)
858 centrePadding << " ";
861 outputStream << centrePadding.str() << stringToPass << centrePadding.str();
863 if (padding > 0 && padding %2 != 0)
868 return outputStream.str();
871 void PrintDeviceDetails(const std::pair<const unsigned short, std::unique_ptr<Device>>& devicePair)
875 body.append(CentreAlignFormatting(devicePair.second->m_Name, 20));
877 body.append(CentreAlignFormatting(std::to_string(devicePair.first), 13));
879 body.append(CentreAlignFormatting(std::to_string(devicePair.second->m_Cores), 10));
882 std::cout << std::string(body.size(), '-') << "\n";
886 void PrintCounterSetDetails(const std::pair<const unsigned short, std::unique_ptr<CounterSet>>& counterSetPair)
890 body.append(CentreAlignFormatting(counterSetPair.second->m_Name, 20));
892 body.append(CentreAlignFormatting(std::to_string(counterSetPair.first), 13));
894 body.append(CentreAlignFormatting(std::to_string(counterSetPair.second->m_Count), 10));
897 std::cout << std::string(body.size(), '-') << "\n";
902 void PrintCounterDetails(std::shared_ptr<Counter>& counter)
906 body.append(CentreAlignFormatting(counter->m_Name, 20));
908 body.append(CentreAlignFormatting(counter->m_Description, 50));
910 body.append(CentreAlignFormatting(counter->m_Units, 14));
912 body.append(CentreAlignFormatting(std::to_string(counter->m_Uid), 6));
914 body.append(CentreAlignFormatting(std::to_string(counter->m_MaxCounterUid), 10));
916 body.append(CentreAlignFormatting(std::to_string(counter->m_Class), 8));
918 body.append(CentreAlignFormatting(std::to_string(counter->m_Interpolation), 14));
920 body.append(CentreAlignFormatting(std::to_string(counter->m_Multiplier), 20));
922 body.append(CentreAlignFormatting(std::to_string(counter->m_CounterSetUid), 16));
924 body.append(CentreAlignFormatting(std::to_string(counter->m_DeviceUid), 14));
928 std::cout << std::string(body.size(), '-') << "\n";
933 void PrintCategoryDetails(const std::unique_ptr<Category>& category,
934 std::unordered_map<unsigned short, std::shared_ptr<Counter>> counterMap)
936 std::string categoryBody;
937 std::string categoryHeader;
939 categoryHeader.append(CentreAlignFormatting("Name", 20));
940 categoryHeader.append(" | ");
941 categoryHeader.append(CentreAlignFormatting("Event Count", 14));
942 categoryHeader.append("\n");
944 categoryBody.append(CentreAlignFormatting(category->m_Name, 20));
945 categoryBody.append(" | ");
946 categoryBody.append(CentreAlignFormatting(std::to_string(category->m_Counters.size()), 14));
948 std::cout << "\n" << "\n";
949 std::cout << CentreAlignFormatting("CATEGORY", static_cast<int>(categoryHeader.size()));
951 std::cout << std::string(categoryHeader.size(), '=') << "\n";
953 std::cout << categoryHeader;
955 std::cout << std::string(categoryBody.size(), '-') << "\n";
957 std::cout << categoryBody;
959 std::string counterHeader;
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");
982 std::cout << "\n" << "\n";
983 std::cout << CentreAlignFormatting("EVENTS IN CATEGORY: " + category->m_Name,
984 static_cast<int>(counterHeader.size()));
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);
996 void PrintCounterDirectory(ICounterDirectory& counterDirectory)
998 std::string devicesHeader;
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");
1007 std::cout << "\n" << "\n";
1008 std::cout << CentreAlignFormatting("DEVICES", static_cast<int>(devicesHeader.size()));
1010 std::cout << std::string(devicesHeader.size(), '=') << "\n";
1011 std::cout << devicesHeader;
1012 for (auto& it: counterDirectory.GetDevices()) {
1013 PrintDeviceDetails(it);
1016 std::string counterSetHeader;
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");
1025 std::cout << "\n" << "\n";
1026 std::cout << CentreAlignFormatting("COUNTER SETS", static_cast<int>(counterSetHeader.size()));
1028 std::cout << std::string(counterSetHeader.size(), '=') << "\n";
1030 std::cout << counterSetHeader;
1032 for (auto& it: counterDirectory.GetCounterSets()) {
1033 PrintCounterSetDetails(it);
1036 auto counters = counterDirectory.GetCounters();
1037 for (auto& it: counterDirectory.GetCategories()) {
1038 PrintCategoryDetails(it, counters);
1043 uint64_t GetTimestamp()
1045 #if USE_CLOCK_MONOTONIC_RAW
1046 using clock = MonotonicClockRaw;
1048 using clock = std::chrono::steady_clock;
1052 auto timestamp = std::chrono::duration_cast<std::chrono::nanoseconds>(clock::now().time_since_epoch());
1054 return static_cast<uint64_t>(timestamp.count());
1057 Packet ReceivePacket(const unsigned char* buffer, uint32_t length)
1059 if (buffer == nullptr)
1061 throw armnnProfiling::ProfilingException("data buffer is nullptr");
1065 throw armnnProfiling::ProfilingException("length of data buffer is less than 8");
1068 uint32_t metadataIdentifier = 0;
1069 std::memcpy(&metadataIdentifier, buffer, sizeof(metadataIdentifier));
1071 uint32_t dataLength = 0;
1072 std::memcpy(&dataLength, buffer + 4u, sizeof(dataLength));
1074 std::unique_ptr<unsigned char[]> packetData;
1077 packetData = std::make_unique<unsigned char[]>(dataLength);
1078 std::memcpy(packetData.get(), buffer + 8u, dataLength);
1081 return Packet(metadataIdentifier, dataLength, packetData);
1084 } // namespace profiling
1086 } // namespace armnn
1091 bool operator==(const std::vector<uint8_t>& left, std::thread::id right)
1093 return std::memcmp(left.data(), &right, left.size()) == 0;