Add Tracy profiler support

Source/ThirdParty/tracy/common/TracyAlign.hpp

@@ -0,0 +1,25 @@
+#ifndef __TRACYALIGN_HPP__
+#define __TRACYALIGN_HPP__
+
+#include <string.h>
+
+namespace tracy
+{
+
+template<typename T>
+tracy_force_inline T MemRead( const void* ptr )
+{
+    T val;
+    memcpy( &val, ptr, sizeof( T ) );
+    return val;
+}
+
+template<typename T>
+tracy_force_inline void MemWrite( void* ptr, T val )
+{
+    memcpy( ptr, &val, sizeof( T ) );
+}
+
+}
+
+#endif

Source/ThirdParty/tracy/common/TracyAlloc.hpp

@@ -0,0 +1,42 @@
+#ifndef __TRACYALLOC_HPP__
+#define __TRACYALLOC_HPP__
+
+#include <stdlib.h>
+
+#ifdef TRACY_ENABLE
+#  include "../client/tracy_rpmalloc.hpp"
+#endif
+
+namespace tracy
+{
+
+static inline void* tracy_malloc( size_t size )
+{
+#ifdef TRACY_ENABLE
+    return rpmalloc( size );
+#else
+    return malloc( size );
+#endif
+}
+
+static inline void tracy_free( void* ptr )
+{
+#ifdef TRACY_ENABLE
+    rpfree( ptr );
+#else
+    free( ptr );
+#endif
+}
+
+static inline void* tracy_realloc( void* ptr, size_t size )
+{
+#ifdef TRACY_ENABLE
+    return rprealloc( ptr, size );
+#else
+    return realloc( ptr, size );
+#endif
+}
+
+}
+
+#endif

Source/ThirdParty/tracy/common/TracyMutex.hpp

@@ -0,0 +1,24 @@
+#ifndef __TRACYMUTEX_HPP__
+#define __TRACYMUTEX_HPP__
+
+#if defined _MSC_VER
+
+#  include <shared_mutex>
+
+namespace tracy
+{
+using TracyMutex = std::shared_mutex;
+}
+
+#else
+
+#include <mutex>
+
+namespace tracy
+{
+using TracyMutex = std::mutex;
+}
+
+#endif
+
+#endif

Source/ThirdParty/tracy/common/TracyProtocol.hpp

@@ -0,0 +1,128 @@
+#ifndef __TRACYPROTOCOL_HPP__
+#define __TRACYPROTOCOL_HPP__
+
+#include <limits>
+#include <stdint.h>
+
+namespace tracy
+{
+
+constexpr unsigned Lz4CompressBound( unsigned isize ) { return isize + ( isize / 255 ) + 16; }
+
+enum : uint32_t { ProtocolVersion = 46 };
+enum : uint16_t { BroadcastVersion = 2 };
+
+using lz4sz_t = uint32_t;
+
+enum { TargetFrameSize = 256 * 1024 };
+enum { LZ4Size = Lz4CompressBound( TargetFrameSize ) };
+static_assert( LZ4Size <= std::numeric_limits<lz4sz_t>::max(), "LZ4Size greater than lz4sz_t" );
+static_assert( TargetFrameSize * 2 >= 64 * 1024, "Not enough space for LZ4 stream buffer" );
+
+enum { HandshakeShibbolethSize = 8 };
+static const char HandshakeShibboleth[HandshakeShibbolethSize] = { 'T', 'r', 'a', 'c', 'y', 'P', 'r', 'f' };
+
+enum HandshakeStatus : uint8_t
+{
+    HandshakePending,
+    HandshakeWelcome,
+    HandshakeProtocolMismatch,
+    HandshakeNotAvailable,
+    HandshakeDropped
+};
+
+enum { WelcomeMessageProgramNameSize = 64 };
+enum { WelcomeMessageHostInfoSize = 1024 };
+
+#pragma pack( 1 )
+
+// Must increase left query space after handling!
+enum ServerQuery : uint8_t
+{
+    ServerQueryTerminate,
+    ServerQueryString,
+    ServerQueryThreadString,
+    ServerQuerySourceLocation,
+    ServerQueryPlotName,
+    ServerQueryCallstackFrame,
+    ServerQueryFrameName,
+    ServerQueryDisconnect,
+    ServerQueryExternalName,
+    ServerQueryParameter,
+    ServerQuerySymbol,
+    ServerQuerySymbolCode,
+    ServerQueryCodeLocation,
+    ServerQuerySourceCode,
+    ServerQueryDataTransfer,
+    ServerQueryDataTransferPart
+};
+
+struct ServerQueryPacket
+{
+    ServerQuery type;
+    uint64_t ptr;
+    uint32_t extra;
+};
+
+enum { ServerQueryPacketSize = sizeof( ServerQueryPacket ) };
+
+
+enum CpuArchitecture : uint8_t
+{
+    CpuArchUnknown,
+    CpuArchX86,
+    CpuArchX64,
+    CpuArchArm32,
+    CpuArchArm64
+};
+
+
+struct WelcomeMessage
+{
+    double timerMul;
+    int64_t initBegin;
+    int64_t initEnd;
+    uint64_t delay;
+    uint64_t resolution;
+    uint64_t epoch;
+    uint64_t exectime;
+    uint64_t pid;
+    int64_t samplingPeriod;
+    uint8_t onDemand;
+    uint8_t isApple;
+    uint8_t cpuArch;
+    uint8_t codeTransfer;
+    char cpuManufacturer[12];
+    uint32_t cpuId;
+    char programName[WelcomeMessageProgramNameSize];
+    char hostInfo[WelcomeMessageHostInfoSize];
+};
+
+enum { WelcomeMessageSize = sizeof( WelcomeMessage ) };
+
+
+struct OnDemandPayloadMessage
+{
+    uint64_t frames;
+    uint64_t currentTime;
+};
+
+enum { OnDemandPayloadMessageSize = sizeof( OnDemandPayloadMessage ) };
+
+
+struct BroadcastMessage
+{
+    uint16_t broadcastVersion;
+    uint16_t listenPort;
+    uint32_t protocolVersion;
+    int32_t activeTime;        // in seconds
+    char programName[WelcomeMessageProgramNameSize];
+};
+
+enum { BroadcastMessageSize = sizeof( BroadcastMessage ) };
+
+#pragma pack()
+
+}
+
+#endif

Source/ThirdParty/tracy/common/TracyQueue.hpp

@@ -0,0 +1,678 @@
+#ifndef __TRACYQUEUE_HPP__
+#define __TRACYQUEUE_HPP__
+
+#include <stdint.h>
+
+namespace tracy
+{
+
+enum class QueueType : uint8_t
+{
+    ZoneText,
+    ZoneName,
+    Message,
+    MessageColor,
+    MessageCallstack,
+    MessageColorCallstack,
+    MessageAppInfo,
+    ZoneBeginAllocSrcLoc,
+    ZoneBeginAllocSrcLocCallstack,
+    CallstackSerial,
+    Callstack,
+    CallstackAlloc,
+    CallstackSample,
+    FrameImage,
+    ZoneBegin,
+    ZoneBeginCallstack,
+    ZoneEnd,
+    LockWait,
+    LockObtain,
+    LockRelease,
+    LockSharedWait,
+    LockSharedObtain,
+    LockSharedRelease,
+    LockName,
+    MemAlloc,
+    MemAllocNamed,
+    MemFree,
+    MemFreeNamed,
+    MemAllocCallstack,
+    MemAllocCallstackNamed,
+    MemFreeCallstack,
+    MemFreeCallstackNamed,
+    GpuZoneBegin,
+    GpuZoneBeginCallstack,
+    GpuZoneBeginAllocSrcLoc,
+    GpuZoneBeginAllocSrcLocCallstack,
+    GpuZoneEnd,
+    GpuZoneBeginSerial,
+    GpuZoneBeginCallstackSerial,
+    GpuZoneBeginAllocSrcLocSerial,
+    GpuZoneBeginAllocSrcLocCallstackSerial,
+    GpuZoneEndSerial,
+    PlotData,
+    ContextSwitch,
+    ThreadWakeup,
+    GpuTime,
+    GpuContextName,
+    Terminate,
+    KeepAlive,
+    ThreadContext,
+    GpuCalibration,
+    Crash,
+    CrashReport,
+    ZoneValidation,
+    ZoneColor,
+    ZoneValue,
+    FrameMarkMsg,
+    FrameMarkMsgStart,
+    FrameMarkMsgEnd,
+    SourceLocation,
+    LockAnnounce,
+    LockTerminate,
+    LockMark,
+    MessageLiteral,
+    MessageLiteralColor,
+    MessageLiteralCallstack,
+    MessageLiteralColorCallstack,
+    GpuNewContext,
+    CallstackFrameSize,
+    CallstackFrame,
+    SymbolInformation,
+    CodeInformation,
+    SysTimeReport,
+    TidToPid,
+    PlotConfig,
+    ParamSetup,
+    AckServerQueryNoop,
+    AckSourceCodeNotAvailable,
+    CpuTopology,
+    SingleStringData,
+    SecondStringData,
+    MemNamePayload,
+    StringData,
+    ThreadName,
+    PlotName,
+    SourceLocationPayload,
+    CallstackPayload,
+    CallstackAllocPayload,
+    FrameName,
+    FrameImageData,
+    ExternalName,
+    ExternalThreadName,
+    SymbolCode,
+    SourceCode,
+    NUM_TYPES
+};
+
+#pragma pack( 1 )
+
+struct QueueThreadContext
+{
+    uint64_t thread;
+};
+
+struct QueueZoneBeginLean
+{
+    int64_t time;
+};
+
+struct QueueZoneBegin : public QueueZoneBeginLean
+{
+    uint64_t srcloc;    // ptr
+};
+
+struct QueueZoneEnd
+{
+    int64_t time;
+};
+
+struct QueueZoneValidation
+{
+    uint32_t id;
+};
+
+struct QueueZoneColor
+{
+    uint8_t r;
+    uint8_t g;
+    uint8_t b;
+};
+
+struct QueueZoneValue
+{
+    uint64_t value;
+};
+
+struct QueueStringTransfer
+{
+    uint64_t ptr;
+};
+
+struct QueueFrameMark
+{
+    int64_t time;
+    uint64_t name;      // ptr
+};
+
+struct QueueFrameImage
+{
+    uint32_t frame;
+    uint16_t w;
+    uint16_t h;
+    uint8_t flip;
+};
+
+struct QueueFrameImageFat : public QueueFrameImage
+{
+    uint64_t image;     // ptr
+};
+
+struct QueueSourceLocation
+{
+    uint64_t name;
+    uint64_t function;  // ptr
+    uint64_t file;      // ptr
+    uint32_t line;
+    uint8_t r;
+    uint8_t g;
+    uint8_t b;
+};
+
+struct QueueZoneTextFat
+{
+    uint64_t text;      // ptr
+    uint16_t size;
+};
+
+enum class LockType : uint8_t
+{
+    Lockable,
+    SharedLockable
+};
+
+struct QueueLockAnnounce
+{
+    uint32_t id;
+    int64_t time;
+    uint64_t lckloc;    // ptr
+    LockType type;
+};
+
+struct QueueLockTerminate
+{
+    uint32_t id;
+    int64_t time;
+};
+
+struct QueueLockWait
+{
+    uint64_t thread;
+    uint32_t id;
+    int64_t time;
+};
+
+struct QueueLockObtain
+{
+    uint64_t thread;
+    uint32_t id;
+    int64_t time;
+};
+
+struct QueueLockRelease
+{
+    uint64_t thread;
+    uint32_t id;
+    int64_t time;
+};
+
+struct QueueLockMark
+{
+    uint64_t thread;
+    uint32_t id;
+    uint64_t srcloc;    // ptr
+};
+
+struct QueueLockName
+{
+    uint32_t id;
+};
+
+struct QueueLockNameFat : public QueueLockName
+{
+    uint64_t name;      // ptr
+    uint16_t size;
+};
+
+enum class PlotDataType : uint8_t
+{
+    Float,
+    Double,
+    Int
+};
+
+struct QueuePlotData
+{
+    uint64_t name;      // ptr
+    int64_t time;
+    PlotDataType type;
+    union
+    {
+        double d;
+        float f;
+        int64_t i;
+    } data;
+};
+
+struct QueueMessage
+{
+    int64_t time;
+};
+
+struct QueueMessageColor : public QueueMessage
+{
+    uint8_t r;
+    uint8_t g;
+    uint8_t b;
+};
+
+struct QueueMessageLiteral : public QueueMessage
+{
+    uint64_t text;      // ptr
+};
+
+struct QueueMessageColorLiteral : public QueueMessageColor
+{
+    uint64_t text;      // ptr
+};
+
+struct QueueMessageFat : public QueueMessage
+{
+    uint64_t text;      // ptr
+    uint16_t size;
+};
+
+struct QueueMessageColorFat : public QueueMessageColor
+{
+    uint64_t text;      // ptr
+    uint16_t size;
+};
+
+// Don't change order, only add new entries at the end, this is also used on trace dumps!
+enum class GpuContextType : uint8_t
+{
+    Invalid,
+    OpenGl,
+    Vulkan,
+    OpenCL,
+    Direct3D12
+};
+
+enum GpuContextFlags : uint8_t
+{
+    GpuContextCalibration   = 1 << 0
+};
+
+struct QueueGpuNewContext
+{
+    int64_t cpuTime;
+    int64_t gpuTime;
+    uint64_t thread;
+    float period;
+    uint8_t context;
+    GpuContextFlags flags;
+    GpuContextType type;
+};
+
+struct QueueGpuZoneBeginLean
+{
+    int64_t cpuTime;
+    uint64_t thread;
+    uint16_t queryId;
+    uint8_t context;
+};
+
+struct QueueGpuZoneBegin : public QueueGpuZoneBeginLean
+{
+    uint64_t srcloc;
+};
+
+struct QueueGpuZoneEnd
+{
+    int64_t cpuTime;
+    uint64_t thread;
+    uint16_t queryId;
+    uint8_t context;
+};
+
+struct QueueGpuTime
+{
+    int64_t gpuTime;
+    uint16_t queryId;
+    uint8_t context;
+};
+
+struct QueueGpuCalibration
+{
+    int64_t gpuTime;
+    int64_t cpuTime;
+    int64_t cpuDelta;
+    uint8_t context;
+};
+
+struct QueueGpuContextName
+{
+    uint8_t context;
+};
+
+struct QueueGpuContextNameFat : public QueueGpuContextName
+{
+    uint64_t ptr;
+    uint16_t size;
+};
+
+struct QueueMemNamePayload
+{
+    uint64_t name;
+};
+
+struct QueueMemAlloc
+{
+    int64_t time;
+    uint64_t thread;
+    uint64_t ptr;
+    char size[6];
+};
+
+struct QueueMemFree
+{
+    int64_t time;
+    uint64_t thread;
+    uint64_t ptr;
+};
+
+struct QueueCallstackFat
+{
+    uint64_t ptr;
+};
+
+struct QueueCallstackAllocFat
+{
+    uint64_t ptr;
+    uint64_t nativePtr;
+};
+
+struct QueueCallstackSample
+{
+    int64_t time;
+    uint64_t thread;
+};
+
+struct QueueCallstackSampleFat : public QueueCallstackSample
+{
+    uint64_t ptr;
+};
+
+struct QueueCallstackFrameSize
+{
+    uint64_t ptr;
+    uint8_t size;
+};
+
+struct QueueCallstackFrame
+{
+    uint32_t line;
+    uint64_t symAddr;
+    uint32_t symLen;
+};
+
+struct QueueSymbolInformation
+{
+    uint32_t line;
+    uint64_t symAddr;
+};
+
+struct QueueCodeInformation
+{
+    uint64_t ptr;
+    uint32_t line;
+};
+
+struct QueueCrashReport
+{
+    int64_t time;
+    uint64_t text;      // ptr
+};
+
+struct QueueSysTime
+{
+    int64_t time;
+    float sysTime;
+};
+
+struct QueueContextSwitch
+{
+    int64_t time;
+    uint64_t oldThread;
+    uint64_t newThread;
+    uint8_t cpu;
+    uint8_t reason;
+    uint8_t state;
+};
+
+struct QueueThreadWakeup
+{
+    int64_t time;
+    uint64_t thread;
+};
+
+struct QueueTidToPid
+{
+    uint64_t tid;
+    uint64_t pid;
+};
+
+struct QueuePlotConfig
+{
+    uint64_t name;      // ptr
+    uint8_t type;
+};
+
+struct QueueParamSetup
+{
+    uint32_t idx;
+    uint64_t name;      // ptr
+    uint8_t isBool;
+    int32_t val;
+};
+
+struct QueueCpuTopology
+{
+    uint32_t package;
+    uint32_t core;
+    uint32_t thread;
+};
+
+struct QueueHeader
+{
+    union
+    {
+        QueueType type;
+        uint8_t idx;
+    };
+};
+
+struct QueueItem
+{
+    QueueHeader hdr;
+    union
+    {
+        QueueThreadContext threadCtx;
+        QueueZoneBegin zoneBegin;
+        QueueZoneBeginLean zoneBeginLean;
+        QueueZoneEnd zoneEnd;
+        QueueZoneValidation zoneValidation;
+        QueueZoneColor zoneColor;
+        QueueZoneValue zoneValue;
+        QueueStringTransfer stringTransfer;
+        QueueFrameMark frameMark;
+        QueueFrameImage frameImage;
+        QueueFrameImageFat frameImageFat;
+        QueueSourceLocation srcloc;
+        QueueZoneTextFat zoneTextFat;
+        QueueLockAnnounce lockAnnounce;
+        QueueLockTerminate lockTerminate;
+        QueueLockWait lockWait;
+        QueueLockObtain lockObtain;
+        QueueLockRelease lockRelease;
+        QueueLockMark lockMark;
+        QueueLockName lockName;
+        QueueLockNameFat lockNameFat;
+        QueuePlotData plotData;
+        QueueMessage message;
+        QueueMessageColor messageColor;
+        QueueMessageLiteral messageLiteral;
+        QueueMessageColorLiteral messageColorLiteral;
+        QueueMessageFat messageFat;
+        QueueMessageColorFat messageColorFat;
+        QueueGpuNewContext gpuNewContext;
+        QueueGpuZoneBegin gpuZoneBegin;
+        QueueGpuZoneBeginLean gpuZoneBeginLean;
+        QueueGpuZoneEnd gpuZoneEnd;
+        QueueGpuTime gpuTime;
+        QueueGpuCalibration gpuCalibration;
+        QueueGpuContextName gpuContextName;
+        QueueGpuContextNameFat gpuContextNameFat;
+        QueueMemAlloc memAlloc;
+        QueueMemFree memFree;
+        QueueMemNamePayload memName;
+        QueueCallstackFat callstackFat;
+        QueueCallstackAllocFat callstackAllocFat;
+        QueueCallstackSample callstackSample;
+        QueueCallstackSampleFat callstackSampleFat;
+        QueueCallstackFrameSize callstackFrameSize;
+        QueueCallstackFrame callstackFrame;
+        QueueSymbolInformation symbolInformation;
+        QueueCodeInformation codeInformation;
+        QueueCrashReport crashReport;
+        QueueSysTime sysTime;
+        QueueContextSwitch contextSwitch;
+        QueueThreadWakeup threadWakeup;
+        QueueTidToPid tidToPid;
+        QueuePlotConfig plotConfig;
+        QueueParamSetup paramSetup;
+        QueueCpuTopology cpuTopology;
+    };
+};
+#pragma pack()
+
+
+enum { QueueItemSize = sizeof( QueueItem ) };
+
+static constexpr size_t QueueDataSize[] = {
+    sizeof( QueueHeader ),                                  // zone text
+    sizeof( QueueHeader ),                                  // zone name
+    sizeof( QueueHeader ) + sizeof( QueueMessage ),
+    sizeof( QueueHeader ) + sizeof( QueueMessageColor ),
+    sizeof( QueueHeader ) + sizeof( QueueMessage ),         // callstack
+    sizeof( QueueHeader ) + sizeof( QueueMessageColor ),    // callstack
+    sizeof( QueueHeader ) + sizeof( QueueMessage ),         // app info
+    sizeof( QueueHeader ) + sizeof( QueueZoneBeginLean ),   // allocated source location
+    sizeof( QueueHeader ) + sizeof( QueueZoneBeginLean ),   // allocated source location, callstack
+    sizeof( QueueHeader ),                                  // callstack memory
+    sizeof( QueueHeader ),                                  // callstack
+    sizeof( QueueHeader ),                                  // callstack alloc
+    sizeof( QueueHeader ) + sizeof( QueueCallstackSample ),
+    sizeof( QueueHeader ) + sizeof( QueueFrameImage ),
+    sizeof( QueueHeader ) + sizeof( QueueZoneBegin ),
+    sizeof( QueueHeader ) + sizeof( QueueZoneBegin ),       // callstack
+    sizeof( QueueHeader ) + sizeof( QueueZoneEnd ),
+    sizeof( QueueHeader ) + sizeof( QueueLockWait ),
+    sizeof( QueueHeader ) + sizeof( QueueLockObtain ),
+    sizeof( QueueHeader ) + sizeof( QueueLockRelease ),
+    sizeof( QueueHeader ) + sizeof( QueueLockWait ),        // shared
+    sizeof( QueueHeader ) + sizeof( QueueLockObtain ),      // shared
+    sizeof( QueueHeader ) + sizeof( QueueLockRelease ),     // shared
+    sizeof( QueueHeader ) + sizeof( QueueLockName ),
+    sizeof( QueueHeader ) + sizeof( QueueMemAlloc ),
+    sizeof( QueueHeader ) + sizeof( QueueMemAlloc ),        // named
+    sizeof( QueueHeader ) + sizeof( QueueMemFree ),
+    sizeof( QueueHeader ) + sizeof( QueueMemFree ),         // named
+    sizeof( QueueHeader ) + sizeof( QueueMemAlloc ),        // callstack
+    sizeof( QueueHeader ) + sizeof( QueueMemAlloc ),        // callstack, named
+    sizeof( QueueHeader ) + sizeof( QueueMemFree ),         // callstack
+    sizeof( QueueHeader ) + sizeof( QueueMemFree ),         // callstack, named
+    sizeof( QueueHeader ) + sizeof( QueueGpuZoneBegin ),
+    sizeof( QueueHeader ) + sizeof( QueueGpuZoneBegin ),    // callstack
+    sizeof( QueueHeader ) + sizeof( QueueGpuZoneBeginLean ),// allocated source location
+    sizeof( QueueHeader ) + sizeof( QueueGpuZoneBeginLean ),// allocated source location, callstack
+    sizeof( QueueHeader ) + sizeof( QueueGpuZoneEnd ),
+    sizeof( QueueHeader ) + sizeof( QueueGpuZoneBegin ),    // serial
+    sizeof( QueueHeader ) + sizeof( QueueGpuZoneBegin ),    // serial, callstack
+    sizeof( QueueHeader ) + sizeof( QueueGpuZoneBeginLean ),// serial, allocated source location
+    sizeof( QueueHeader ) + sizeof( QueueGpuZoneBeginLean ),// serial, allocated source location, callstack
+    sizeof( QueueHeader ) + sizeof( QueueGpuZoneEnd ),      // serial
+    sizeof( QueueHeader ) + sizeof( QueuePlotData ),
+    sizeof( QueueHeader ) + sizeof( QueueContextSwitch ),
+    sizeof( QueueHeader ) + sizeof( QueueThreadWakeup ),
+    sizeof( QueueHeader ) + sizeof( QueueGpuTime ),
+    sizeof( QueueHeader ) + sizeof( QueueGpuContextName ),
+    // above items must be first
+    sizeof( QueueHeader ),                                  // terminate
+    sizeof( QueueHeader ),                                  // keep alive
+    sizeof( QueueHeader ) + sizeof( QueueThreadContext ),
+    sizeof( QueueHeader ) + sizeof( QueueGpuCalibration ),
+    sizeof( QueueHeader ),                                  // crash
+    sizeof( QueueHeader ) + sizeof( QueueCrashReport ),
+    sizeof( QueueHeader ) + sizeof( QueueZoneValidation ),
+    sizeof( QueueHeader ) + sizeof( QueueZoneColor ),
+    sizeof( QueueHeader ) + sizeof( QueueZoneValue ),
+    sizeof( QueueHeader ) + sizeof( QueueFrameMark ),       // continuous frames
+    sizeof( QueueHeader ) + sizeof( QueueFrameMark ),       // start
+    sizeof( QueueHeader ) + sizeof( QueueFrameMark ),       // end
+    sizeof( QueueHeader ) + sizeof( QueueSourceLocation ),
+    sizeof( QueueHeader ) + sizeof( QueueLockAnnounce ),
+    sizeof( QueueHeader ) + sizeof( QueueLockTerminate ),
+    sizeof( QueueHeader ) + sizeof( QueueLockMark ),
+    sizeof( QueueHeader ) + sizeof( QueueMessageLiteral ),
+    sizeof( QueueHeader ) + sizeof( QueueMessageColorLiteral ),
+    sizeof( QueueHeader ) + sizeof( QueueMessageLiteral ),  // callstack
+    sizeof( QueueHeader ) + sizeof( QueueMessageColorLiteral ), // callstack
+    sizeof( QueueHeader ) + sizeof( QueueGpuNewContext ),
+    sizeof( QueueHeader ) + sizeof( QueueCallstackFrameSize ),
+    sizeof( QueueHeader ) + sizeof( QueueCallstackFrame ),
+    sizeof( QueueHeader ) + sizeof( QueueSymbolInformation ),
+    sizeof( QueueHeader ) + sizeof( QueueCodeInformation ),
+    sizeof( QueueHeader ) + sizeof( QueueSysTime ),
+    sizeof( QueueHeader ) + sizeof( QueueTidToPid ),
+    sizeof( QueueHeader ) + sizeof( QueuePlotConfig ),
+    sizeof( QueueHeader ) + sizeof( QueueParamSetup ),
+    sizeof( QueueHeader ),                                  // server query acknowledgement
+    sizeof( QueueHeader ),                                  // source code not available
+    sizeof( QueueHeader ) + sizeof( QueueCpuTopology ),
+    sizeof( QueueHeader ),                                  // single string data
+    sizeof( QueueHeader ),                                  // second string data
+    sizeof( QueueHeader ) + sizeof( QueueMemNamePayload ),
+    // keep all QueueStringTransfer below
+    sizeof( QueueHeader ) + sizeof( QueueStringTransfer ),  // string data
+    sizeof( QueueHeader ) + sizeof( QueueStringTransfer ),  // thread name
+    sizeof( QueueHeader ) + sizeof( QueueStringTransfer ),  // plot name
+    sizeof( QueueHeader ) + sizeof( QueueStringTransfer ),  // allocated source location payload
+    sizeof( QueueHeader ) + sizeof( QueueStringTransfer ),  // callstack payload
+    sizeof( QueueHeader ) + sizeof( QueueStringTransfer ),  // callstack alloc payload
+    sizeof( QueueHeader ) + sizeof( QueueStringTransfer ),  // frame name
+    sizeof( QueueHeader ) + sizeof( QueueStringTransfer ),  // frame image data
+    sizeof( QueueHeader ) + sizeof( QueueStringTransfer ),  // external name
+    sizeof( QueueHeader ) + sizeof( QueueStringTransfer ),  // external thread name
+    sizeof( QueueHeader ) + sizeof( QueueStringTransfer ),  // symbol code
+    sizeof( QueueHeader ) + sizeof( QueueStringTransfer ),  // source code
+};
+
+static_assert( QueueItemSize == 32, "Queue item size not 32 bytes" );
+static_assert( sizeof( QueueDataSize ) / sizeof( size_t ) == (uint8_t)QueueType::NUM_TYPES, "QueueDataSize mismatch" );
+static_assert( sizeof( void* ) <= sizeof( uint64_t ), "Pointer size > 8 bytes" );
+static_assert( sizeof( void* ) == sizeof( uintptr_t ), "Pointer size != uintptr_t" );
+
+}
+
+#endif

Source/ThirdParty/tracy/common/TracySocket.cpp

@@ -0,0 +1,748 @@
+#include <assert.h>
+#include <inttypes.h>
+#include <new>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include <sys/types.h>
+
+#include "TracyAlloc.hpp"
+#include "TracySocket.hpp"
+
+#ifdef _WIN32
+#  ifndef NOMINMAX
+#    define NOMINMAX
+#  endif
+#  include <winsock2.h>
+#  include <ws2tcpip.h>
+#  ifdef _MSC_VER
+#    pragma warning(disable:4244)
+#    pragma warning(disable:4267)
+#  endif
+#  define poll WSAPoll
+#else
+#  include <arpa/inet.h>
+#  include <sys/socket.h>
+#  include <sys/param.h>
+#  include <errno.h>
+#  include <fcntl.h>
+#  include <netinet/in.h>
+#  include <netdb.h>
+#  include <unistd.h>
+#  include <poll.h>
+#endif
+
+#ifndef MSG_NOSIGNAL
+#  define MSG_NOSIGNAL 0
+#endif
+
+namespace tracy
+{
+
+#ifdef _WIN32
+typedef SOCKET socket_t;
+#else
+typedef int socket_t;
+#endif
+
+#ifdef _WIN32
+struct __wsinit
+{
+    __wsinit()
+    {
+        WSADATA wsaData;
+        if( WSAStartup( MAKEWORD( 2, 2 ), &wsaData ) != 0 )
+        {
+            fprintf( stderr, "Cannot init winsock.\n" );
+            exit( 1 );
+        }
+    }
+};
+
+void InitWinSock()
+{
+    static __wsinit init;
+}
+#endif
+
+
+enum { BufSize = 128 * 1024 };
+
+Socket::Socket()
+    : m_buf( (char*)tracy_malloc( BufSize ) )
+    , m_bufPtr( nullptr )
+    , m_sock( -1 )
+    , m_bufLeft( 0 )
+    , m_ptr( nullptr )
+{
+#ifdef _WIN32
+    InitWinSock();
+#endif
+}
+
+Socket::Socket( int sock )
+    : m_buf( (char*)tracy_malloc( BufSize ) )
+    , m_bufPtr( nullptr )
+    , m_sock( sock )
+    , m_bufLeft( 0 )
+    , m_ptr( nullptr )
+{
+}
+
+Socket::~Socket()
+{
+    tracy_free( m_buf );
+    if( m_sock.load( std::memory_order_relaxed ) != -1 )
+    {
+        Close();
+    }
+    if( m_ptr )
+    {
+        freeaddrinfo( m_res );
+#ifdef _WIN32
+        closesocket( m_connSock );
+#else
+        close( m_connSock );
+#endif
+    }
+}
+
+bool Socket::Connect( const char* addr, uint16_t port )
+{
+    assert( !IsValid() );
+
+    if( m_ptr )
+    {
+        const auto c = connect( m_connSock, m_ptr->ai_addr, m_ptr->ai_addrlen );
+        if( c == -1 )
+        {
+#if defined _WIN32
+            const auto err = WSAGetLastError();
+            if( err == WSAEALREADY || err == WSAEINPROGRESS ) return false;
+            if( err != WSAEISCONN )
+            {
+                freeaddrinfo( m_res );
+                closesocket( m_connSock );
+                m_ptr = nullptr;
+                return false;
+            }
+#else
+            const auto err = errno;
+            if( err == EALREADY || err == EINPROGRESS ) return false;
+            if( err != EISCONN )
+            {
+                freeaddrinfo( m_res );
+                close( m_connSock );
+                m_ptr = nullptr;
+                return false;
+            }
+#endif
+        }
+
+#if defined _WIN32
+        u_long nonblocking = 0;
+        ioctlsocket( m_connSock, FIONBIO, &nonblocking );
+#else
+        int flags = fcntl( m_connSock, F_GETFL, 0 );
+        fcntl( m_connSock, F_SETFL, flags & ~O_NONBLOCK );
+#endif
+        m_sock.store( m_connSock, std::memory_order_relaxed );
+        freeaddrinfo( m_res );
+        m_ptr = nullptr;
+        return true;
+    }
+
+    struct addrinfo hints;
+    struct addrinfo *res, *ptr;
+
+    memset( &hints, 0, sizeof( hints ) );
+    hints.ai_family = AF_UNSPEC;
+    hints.ai_socktype = SOCK_STREAM;
+
+    char portbuf[32];
+    sprintf( portbuf, "%" PRIu16, port );
+
+    if( getaddrinfo( addr, portbuf, &hints, &res ) != 0 ) return false;
+    int sock = 0;
+    for( ptr = res; ptr; ptr = ptr->ai_next )
+    {
+        if( ( sock = socket( ptr->ai_family, ptr->ai_socktype, ptr->ai_protocol ) ) == -1 ) continue;
+#if defined __APPLE__
+        int val = 1;
+        setsockopt( sock, SOL_SOCKET, SO_NOSIGPIPE, &val, sizeof( val ) );
+#endif
+#if defined _WIN32
+        u_long nonblocking = 1;
+        ioctlsocket( sock, FIONBIO, &nonblocking );
+#else
+        int flags = fcntl( sock, F_GETFL, 0 );
+        fcntl( sock, F_SETFL, flags | O_NONBLOCK );
+#endif
+        if( connect( sock, ptr->ai_addr, ptr->ai_addrlen ) == 0 )
+        {
+            break;
+        }
+        else
+        {
+#if defined _WIN32
+            const auto err = WSAGetLastError();
+            if( err != WSAEWOULDBLOCK )
+            {
+                closesocket( sock );
+                continue;
+            }
+#else
+            if( errno != EINPROGRESS )
+            {
+                close( sock );
+                continue;
+            }
+#endif
+        }
+        m_res = res;
+        m_ptr = ptr;
+        m_connSock = sock;
+        return false;
+    }
+    freeaddrinfo( res );
+    if( !ptr ) return false;
+
+#if defined _WIN32
+    u_long nonblocking = 0;
+    ioctlsocket( sock, FIONBIO, &nonblocking );
+#else
+    int flags = fcntl( sock, F_GETFL, 0 );
+    fcntl( sock, F_SETFL, flags & ~O_NONBLOCK );
+#endif
+
+    m_sock.store( sock, std::memory_order_relaxed );
+    return true;
+}
+
+bool Socket::ConnectBlocking( const char* addr, uint16_t port )
+{
+    assert( !IsValid() );
+    assert( !m_ptr );
+
+    struct addrinfo hints;
+    struct addrinfo *res, *ptr;
+
+    memset( &hints, 0, sizeof( hints ) );
+    hints.ai_family = AF_UNSPEC;
+    hints.ai_socktype = SOCK_STREAM;
+
+    char portbuf[32];
+    sprintf( portbuf, "%" PRIu16, port );
+
+    if( getaddrinfo( addr, portbuf, &hints, &res ) != 0 ) return false;
+    int sock = 0;
+    for( ptr = res; ptr; ptr = ptr->ai_next )
+    {
+        if( ( sock = socket( ptr->ai_family, ptr->ai_socktype, ptr->ai_protocol ) ) == -1 ) continue;
+#if defined __APPLE__
+        int val = 1;
+        setsockopt( sock, SOL_SOCKET, SO_NOSIGPIPE, &val, sizeof( val ) );
+#endif
+        if( connect( sock, ptr->ai_addr, ptr->ai_addrlen ) == -1 )
+        {
+#ifdef _WIN32
+            closesocket( sock );
+#else
+            close( sock );
+#endif
+            continue;
+        }
+        break;
+    }
+    freeaddrinfo( res );
+    if( !ptr ) return false;
+
+    m_sock.store( sock, std::memory_order_relaxed );
+    return true;
+}
+
+void Socket::Close()
+{
+    const auto sock = m_sock.load( std::memory_order_relaxed );
+    assert( sock != -1 );
+#ifdef _WIN32
+    closesocket( sock );
+#else
+    close( sock );
+#endif
+    m_sock.store( -1, std::memory_order_relaxed );
+}
+
+int Socket::Send( const void* _buf, int len )
+{
+    const auto sock = m_sock.load( std::memory_order_relaxed );
+    auto buf = (const char*)_buf;
+    assert( sock != -1 );
+    auto start = buf;
+    while( len > 0 )
+    {
+        auto ret = send( sock, buf, len, MSG_NOSIGNAL );
+        if( ret == -1 ) return -1;
+        len -= ret;
+        buf += ret;
+    }
+    return int( buf - start );
+}
+
+int Socket::GetSendBufSize()
+{
+    const auto sock = m_sock.load( std::memory_order_relaxed );
+    int bufSize;
+#if defined _WIN32
+    int sz = sizeof( bufSize );
+    getsockopt( sock, SOL_SOCKET, SO_SNDBUF, (char*)&bufSize, &sz );
+#else
+    socklen_t sz = sizeof( bufSize );
+    getsockopt( sock, SOL_SOCKET, SO_SNDBUF, &bufSize, &sz );
+#endif
+    return bufSize;
+}
+
+int Socket::RecvBuffered( void* buf, int len, int timeout )
+{
+    if( len <= m_bufLeft )
+    {
+        memcpy( buf, m_bufPtr, len );
+        m_bufPtr += len;
+        m_bufLeft -= len;
+        return len;
+    }
+
+    if( m_bufLeft > 0 )
+    {
+        memcpy( buf, m_bufPtr, m_bufLeft );
+        const auto ret = m_bufLeft;
+        m_bufLeft = 0;
+        return ret;
+    }
+
+    if( len >= BufSize ) return Recv( buf, len, timeout );
+
+    m_bufLeft = Recv( m_buf, BufSize, timeout );
+    if( m_bufLeft <= 0 ) return m_bufLeft;
+
+    const auto sz = len < m_bufLeft ? len : m_bufLeft;
+    memcpy( buf, m_buf, sz );
+    m_bufPtr = m_buf + sz;
+    m_bufLeft -= sz;
+    return sz;
+}
+
+int Socket::Recv( void* _buf, int len, int timeout )
+{
+    const auto sock = m_sock.load( std::memory_order_relaxed );
+    auto buf = (char*)_buf;
+
+    struct pollfd fd;
+    fd.fd = (socket_t)sock;
+    fd.events = POLLIN;
+
+    if( poll( &fd, 1, timeout ) > 0 )
+    {
+        return recv( sock, buf, len, 0 );
+    }
+    else
+    {
+        return -1;
+    }
+}
+
+int Socket::ReadUpTo( void* _buf, int len, int timeout )
+{
+    const auto sock = m_sock.load( std::memory_order_relaxed );
+    auto buf = (char*)_buf;
+
+    int rd = 0;
+    while( len > 0 )
+    {
+        const auto res = recv( sock, buf, len, 0 );
+        if( res == 0 ) break;
+        if( res == -1 ) return -1;
+        len -= res;
+        rd += res;
+        buf += res;
+    }
+    return rd;
+}
+
+bool Socket::Read( void* buf, int len, int timeout )
+{
+    auto cbuf = (char*)buf;
+    while( len > 0 )
+    {
+        if( !ReadImpl( cbuf, len, timeout ) ) return false;
+    }
+    return true;
+}
+
+bool Socket::ReadImpl( char*& buf, int& len, int timeout )
+{
+    const auto sz = RecvBuffered( buf, len, timeout );
+    switch( sz )
+    {
+    case 0:
+        return false;
+    case -1:
+#ifdef _WIN32
+    {
+        auto err = WSAGetLastError();
+        if( err == WSAECONNABORTED || err == WSAECONNRESET ) return false;
+    }
+#endif
+    break;
+    default:
+        len -= sz;
+        buf += sz;
+        break;
+    }
+    return true;
+}
+
+bool Socket::ReadRaw( void* _buf, int len, int timeout )
+{
+    auto buf = (char*)_buf;
+    while( len > 0 )
+    {
+        const auto sz = Recv( buf, len, timeout );
+        if( sz <= 0 ) return false;
+        len -= sz;
+        buf += sz;
+    }
+    return true;
+}
+
+bool Socket::HasData()
+{
+    const auto sock = m_sock.load( std::memory_order_relaxed );
+    if( m_bufLeft > 0 ) return true;
+
+    struct pollfd fd;
+    fd.fd = (socket_t)sock;
+    fd.events = POLLIN;
+
+    return poll( &fd, 1, 0 ) > 0;
+}
+
+bool Socket::IsValid() const
+{
+    return m_sock.load( std::memory_order_relaxed ) >= 0;
+}
+
+
+ListenSocket::ListenSocket()
+    : m_sock( -1 )
+{
+#ifdef _WIN32
+    InitWinSock();
+#endif
+}
+
+ListenSocket::~ListenSocket()
+{
+    if( m_sock != -1 ) Close();
+}
+
+static int addrinfo_and_socket_for_family( uint16_t port, int ai_family, struct addrinfo** res )
+{
+    struct addrinfo hints;
+    memset( &hints, 0, sizeof( hints ) );
+    hints.ai_family = ai_family;
+    hints.ai_socktype = SOCK_STREAM;
+#ifndef TRACY_ONLY_LOCALHOST
+    const char* onlyLocalhost = getenv( "TRACY_ONLY_LOCALHOST" );
+    if( !onlyLocalhost || onlyLocalhost[0] != '1' )
+    {
+        hints.ai_flags = AI_PASSIVE;
+    }
+#endif
+    char portbuf[32];
+    sprintf( portbuf, "%" PRIu16, port );
+    if( getaddrinfo( nullptr, portbuf, &hints, res ) != 0 ) return -1;
+    int sock = socket( (*res)->ai_family, (*res)->ai_socktype, (*res)->ai_protocol );
+    if (sock == -1) freeaddrinfo( *res );
+    return sock;
+}
+
+bool ListenSocket::Listen( uint16_t port, int backlog )
+{
+    assert( m_sock == -1 );
+
+    struct addrinfo* res = nullptr;
+
+#if !defined TRACY_ONLY_IPV4 && !defined TRACY_ONLY_LOCALHOST
+    const char* onlyIPv4 = getenv( "TRACY_ONLY_IPV4" );
+    if( !onlyIPv4 || onlyIPv4[0] != '1' )
+    {
+        m_sock = addrinfo_and_socket_for_family( port, AF_INET6, &res );
+    }
+#endif
+    if (m_sock == -1)
+    {
+        // IPV6 protocol may not be available/is disabled. Try to create a socket
+        // with the IPV4 protocol
+        m_sock = addrinfo_and_socket_for_family( port, AF_INET, &res );
+        if( m_sock == -1 ) return false;
+    }
+#if defined _WIN32 || defined __CYGWIN__
+    unsigned long val = 0;
+    setsockopt( m_sock, IPPROTO_IPV6, IPV6_V6ONLY, (const char*)&val, sizeof( val ) );
+#elif defined BSD
+    int val = 0;
+    setsockopt( m_sock, IPPROTO_IPV6, IPV6_V6ONLY, (const char*)&val, sizeof( val ) );
+    val = 1;
+    setsockopt( m_sock, SOL_SOCKET, SO_REUSEADDR, &val, sizeof( val ) );
+#else
+    int val = 1;
+    setsockopt( m_sock, SOL_SOCKET, SO_REUSEADDR, &val, sizeof( val ) );
+#endif
+    if( bind( m_sock, res->ai_addr, res->ai_addrlen ) == -1 ) { freeaddrinfo( res ); Close(); return false; }
+    if( listen( m_sock, backlog ) == -1 ) { freeaddrinfo( res ); Close(); return false; }
+    freeaddrinfo( res );
+    return true;
+}
+
+Socket* ListenSocket::Accept()
+{
+    struct sockaddr_storage remote;
+    socklen_t sz = sizeof( remote );
+
+    struct pollfd fd;
+    fd.fd = (socket_t)m_sock;
+    fd.events = POLLIN;
+
+    if( poll( &fd, 1, 10 ) > 0 )
+    {
+        int sock = accept( m_sock, (sockaddr*)&remote, &sz);
+        if( sock == -1 ) return nullptr;
+
+#if defined __APPLE__
+        int val = 1;
+        setsockopt( sock, SOL_SOCKET, SO_NOSIGPIPE, &val, sizeof( val ) );
+#endif
+
+        auto ptr = (Socket*)tracy_malloc( sizeof( Socket ) );
+        new(ptr) Socket( sock );
+        return ptr;
+    }
+    else
+    {
+        return nullptr;
+    }
+}
+
+void ListenSocket::Close()
+{
+    assert( m_sock != -1 );
+#ifdef _WIN32
+    closesocket( m_sock );
+#else
+    close( m_sock );
+#endif
+    m_sock = -1;
+}
+
+UdpBroadcast::UdpBroadcast()
+    : m_sock( -1 )
+{
+#ifdef _WIN32
+    InitWinSock();
+#endif
+}
+
+UdpBroadcast::~UdpBroadcast()
+{
+    if( m_sock != -1 ) Close();
+}
+
+bool UdpBroadcast::Open( const char* addr, uint16_t port )
+{
+    assert( m_sock == -1 );
+
+    struct addrinfo hints;
+    struct addrinfo *res, *ptr;
+
+    memset( &hints, 0, sizeof( hints ) );
+    hints.ai_family = AF_INET;
+    hints.ai_socktype = SOCK_DGRAM;
+
+    char portbuf[32];
+    sprintf( portbuf, "%" PRIu16, port );
+
+    if( getaddrinfo( addr, portbuf, &hints, &res ) != 0 ) return false;
+    int sock = 0;
+    for( ptr = res; ptr; ptr = ptr->ai_next )
+    {
+        if( ( sock = socket( ptr->ai_family, ptr->ai_socktype, ptr->ai_protocol ) ) == -1 ) continue;
+#if defined __APPLE__
+        int val = 1;
+        setsockopt( sock, SOL_SOCKET, SO_NOSIGPIPE, &val, sizeof( val ) );
+#endif
+#if defined _WIN32
+        unsigned long broadcast = 1;
+        if( setsockopt( sock, SOL_SOCKET, SO_BROADCAST, (const char*)&broadcast, sizeof( broadcast ) ) == -1 )
+#else
+        int broadcast = 1;
+        if( setsockopt( sock, SOL_SOCKET, SO_BROADCAST, &broadcast, sizeof( broadcast ) ) == -1 )
+#endif
+        {
+#ifdef _WIN32
+            closesocket( sock );
+#else
+            close( sock );
+#endif
+            continue;
+        }
+        break;
+    }
+    freeaddrinfo( res );
+    if( !ptr ) return false;
+
+    m_sock = sock;
+    inet_pton( AF_INET, addr, &m_addr );
+    return true;
+}
+
+void UdpBroadcast::Close()
+{
+    assert( m_sock != -1 );
+#ifdef _WIN32
+    closesocket( m_sock );
+#else
+    close( m_sock );
+#endif
+    m_sock = -1;
+}
+
+int UdpBroadcast::Send( uint16_t port, const void* data, int len )
+{
+    assert( m_sock != -1 );
+    struct sockaddr_in addr;
+    addr.sin_family = AF_INET;
+    addr.sin_port = htons( port );
+    addr.sin_addr.s_addr = m_addr;
+    return sendto( m_sock, (const char*)data, len, MSG_NOSIGNAL, (sockaddr*)&addr, sizeof( addr ) );
+}
+
+IpAddress::IpAddress()
+    : m_number( 0 )
+{
+    *m_text = '\0';
+}
+
+IpAddress::~IpAddress()
+{
+}
+
+void IpAddress::Set( const struct sockaddr& addr )
+{
+#if defined _WIN32 && ( !defined NTDDI_WIN10 || NTDDI_VERSION < NTDDI_WIN10 )
+    struct sockaddr_in tmp;
+    memcpy( &tmp, &addr, sizeof( tmp ) );
+    auto ai = &tmp;
+#else
+    auto ai = (const struct sockaddr_in*)&addr;
+#endif
+    inet_ntop( AF_INET, &ai->sin_addr, m_text, 17 );
+    m_number = ai->sin_addr.s_addr;
+}
+
+UdpListen::UdpListen()
+    : m_sock( -1 )
+{
+#ifdef _WIN32
+    InitWinSock();
+#endif
+}
+
+UdpListen::~UdpListen()
+{
+    if( m_sock != -1 ) Close();
+}
+
+bool UdpListen::Listen( uint16_t port )
+{
+    assert( m_sock == -1 );
+
+    int sock;
+    if( ( sock = socket( AF_INET, SOCK_DGRAM, 0 ) ) == -1 ) return false;
+
+#if defined __APPLE__
+    int val = 1;
+    setsockopt( sock, SOL_SOCKET, SO_NOSIGPIPE, &val, sizeof( val ) );
+#endif
+#if defined _WIN32
+    unsigned long reuse = 1;
+    setsockopt( m_sock, SOL_SOCKET, SO_REUSEADDR, (const char*)&reuse, sizeof( reuse ) );
+#else
+    int reuse = 1;
+    setsockopt( m_sock, SOL_SOCKET, SO_REUSEADDR, &reuse, sizeof( reuse ) );
+#endif
+#if defined _WIN32
+    unsigned long broadcast = 1;
+    if( setsockopt( sock, SOL_SOCKET, SO_BROADCAST, (const char*)&broadcast, sizeof( broadcast ) ) == -1 )
+#else
+    int broadcast = 1;
+    if( setsockopt( sock, SOL_SOCKET, SO_BROADCAST, &broadcast, sizeof( broadcast ) ) == -1 )
+#endif
+    {
+#ifdef _WIN32
+        closesocket( sock );
+#else
+        close( sock );
+#endif
+        return false;
+    }
+
+    struct sockaddr_in addr;
+    addr.sin_family = AF_INET;
+    addr.sin_port = htons( port );
+    addr.sin_addr.s_addr = INADDR_ANY;
+
+    if( bind( sock, (sockaddr*)&addr, sizeof( addr ) ) == -1 )
+    {
+#ifdef _WIN32
+        closesocket( sock );
+#else
+        close( sock );
+#endif
+        return false;
+    }
+
+    m_sock = sock;
+    return true;
+}
+
+void UdpListen::Close()
+{
+    assert( m_sock != -1 );
+#ifdef _WIN32
+    closesocket( m_sock );
+#else
+    close( m_sock );
+#endif
+    m_sock = -1;
+}
+
+const char* UdpListen::Read( size_t& len, IpAddress& addr, int timeout )
+{
+    static char buf[2048];
+
+    struct pollfd fd;
+    fd.fd = (socket_t)m_sock;
+    fd.events = POLLIN;
+    if( poll( &fd, 1, timeout ) <= 0 ) return nullptr;
+
+    sockaddr sa;
+    socklen_t salen = sizeof( struct sockaddr );
+    len = (size_t)recvfrom( m_sock, buf, 2048, 0, &sa, &salen );
+    addr.Set( sa );
+
+    return buf;
+}
+
+}

Source/ThirdParty/tracy/common/TracySocket.hpp

@@ -0,0 +1,154 @@
+#ifndef __TRACYSOCKET_HPP__
+#define __TRACYSOCKET_HPP__
+
+#include <atomic>
+#include <stdint.h>
+
+struct addrinfo;
+struct sockaddr;
+
+namespace tracy
+{
+
+#ifdef _WIN32
+void InitWinSock();
+#endif
+
+class Socket
+{
+public:
+    Socket();
+    Socket( int sock );
+    ~Socket();
+
+    bool Connect( const char* addr, uint16_t port );
+    bool ConnectBlocking( const char* addr, uint16_t port );
+    void Close();
+
+    int Send( const void* buf, int len );
+    int GetSendBufSize();
+
+    int ReadUpTo( void* buf, int len, int timeout );
+    bool Read( void* buf, int len, int timeout );
+
+    template<typename ShouldExit>
+    bool Read( void* buf, int len, int timeout, ShouldExit exitCb )
+    {
+        auto cbuf = (char*)buf;
+        while( len > 0 )
+        {
+            if( exitCb() ) return false;
+            if( !ReadImpl( cbuf, len, timeout ) ) return false;
+        }
+        return true;
+    }
+
+    bool ReadRaw( void* buf, int len, int timeout );
+    bool HasData();
+    bool IsValid() const;
+
+    Socket( const Socket& ) = delete;
+    Socket( Socket&& ) = delete;
+    Socket& operator=( const Socket& ) = delete;
+    Socket& operator=( Socket&& ) = delete;
+
+private:
+    int RecvBuffered( void* buf, int len, int timeout );
+    int Recv( void* buf, int len, int timeout );
+
+    bool ReadImpl( char*& buf, int& len, int timeout );
+
+    char* m_buf;
+    char* m_bufPtr;
+    std::atomic<int> m_sock;
+    int m_bufLeft;
+
+    struct addrinfo *m_res;
+    struct addrinfo *m_ptr;
+    int m_connSock;
+};
+
+class ListenSocket
+{
+public:
+    ListenSocket();
+    ~ListenSocket();
+
+    bool Listen( uint16_t port, int backlog );
+    Socket* Accept();
+    void Close();
+
+    ListenSocket( const ListenSocket& ) = delete;
+    ListenSocket( ListenSocket&& ) = delete;
+    ListenSocket& operator=( const ListenSocket& ) = delete;
+    ListenSocket& operator=( ListenSocket&& ) = delete;
+
+private:
+    int m_sock;
+};
+
+class UdpBroadcast
+{
+public:
+    UdpBroadcast();
+    ~UdpBroadcast();
+
+    bool Open( const char* addr, uint16_t port );
+    void Close();
+
+    int Send( uint16_t port, const void* data, int len );
+
+    UdpBroadcast( const UdpBroadcast& ) = delete;
+    UdpBroadcast( UdpBroadcast&& ) = delete;
+    UdpBroadcast& operator=( const UdpBroadcast& ) = delete;
+    UdpBroadcast& operator=( UdpBroadcast&& ) = delete;
+
+private:
+    int m_sock;
+    uint32_t m_addr;
+};
+
+class IpAddress
+{
+public:
+    IpAddress();
+    ~IpAddress();
+
+    void Set( const struct sockaddr& addr );
+
+    uint32_t GetNumber() const { return m_number; }
+    const char* GetText() const { return m_text; }
+
+    IpAddress( const IpAddress& ) = delete;
+    IpAddress( IpAddress&& ) = delete;
+    IpAddress& operator=( const IpAddress& ) = delete;
+    IpAddress& operator=( IpAddress&& ) = delete;
+
+private:
+    uint32_t m_number;
+    char m_text[17];
+};
+
+class UdpListen
+{
+public:
+    UdpListen();
+    ~UdpListen();
+
+    bool Listen( uint16_t port );
+    void Close();
+
+    const char* Read( size_t& len, IpAddress& addr, int timeout );
+
+    UdpListen( const UdpListen& ) = delete;
+    UdpListen( UdpListen&& ) = delete;
+    UdpListen& operator=( const UdpListen& ) = delete;
+    UdpListen& operator=( UdpListen&& ) = delete;
+
+private:
+    int m_sock;
+};
+
+}
+
+#endif

Source/ThirdParty/tracy/common/TracySystem.cpp

@@ -0,0 +1,239 @@
+#if defined _MSC_VER || defined __CYGWIN__ || defined _WIN32
+# ifndef WIN32_LEAN_AND_MEAN
+#  define WIN32_LEAN_AND_MEAN
+# endif
+# ifndef NOMINMAX
+#  define NOMINMAX
+# endif
+#endif
+#ifdef _MSC_VER
+#  pragma warning(disable:4996)
+#endif
+#if defined _WIN32 || defined __CYGWIN__
+#  include <windows.h>
+#else
+#  include <pthread.h>
+#  include <string.h>
+#  include <unistd.h>
+#endif
+
+#ifdef __linux__
+#  ifdef __ANDROID__
+#    include <sys/types.h>
+#  else
+#    include <sys/syscall.h>
+#  endif
+#  include <fcntl.h>
+#elif defined __FreeBSD__
+#  include <sys/thr.h>
+#elif defined __NetBSD__ || defined __DragonFly__
+#  include <sys/lwp.h>
+#endif
+
+#ifdef __MINGW32__
+#  define __STDC_FORMAT_MACROS
+#endif
+#include <inttypes.h>
+#include <stdio.h>
+#include <stdlib.h>
+
+#include "TracySystem.hpp"
+
+#if defined _WIN32 || defined __CYGWIN__
+extern "C" typedef HRESULT (WINAPI *t_SetThreadDescription)( HANDLE, PCWSTR );
+extern "C" typedef HRESULT (WINAPI *t_GetThreadDescription)( HANDLE, PWSTR* );
+#endif
+
+#ifdef TRACY_ENABLE
+#  include <atomic>
+#  include "TracyAlloc.hpp"
+#endif
+
+namespace tracy
+{
+
+namespace detail
+{
+
+TRACY_API uint64_t GetThreadHandleImpl()
+{
+#if defined _WIN32 || defined __CYGWIN__
+    static_assert( sizeof( decltype( GetCurrentThreadId() ) ) <= sizeof( uint64_t ), "Thread handle too big to fit in protocol" );
+    return uint64_t( GetCurrentThreadId() );
+#elif defined __APPLE__
+    uint64_t id;
+    pthread_threadid_np( pthread_self(), &id );
+    return id;
+#elif defined __ANDROID__
+    return (uint64_t)gettid();
+#elif defined __linux__
+    return (uint64_t)syscall( SYS_gettid );
+#elif defined __FreeBSD__
+    long id;
+    thr_self( &id );
+    return id;
+#elif defined __NetBSD__
+    return _lwp_self();
+#elif defined __DragonFly__
+    return lwp_gettid();
+#elif defined __OpenBSD__
+    return getthrid();
+#else
+    static_assert( sizeof( decltype( pthread_self() ) ) <= sizeof( uint64_t ), "Thread handle too big to fit in protocol" );
+    return uint64_t( pthread_self() );
+#endif
+
+}
+
+}
+
+#ifdef TRACY_ENABLE
+struct ThreadNameData
+{
+    uint64_t id;
+    const char* name;
+    ThreadNameData* next;
+};
+std::atomic<ThreadNameData*>& GetThreadNameData();
+TRACY_API void InitRPMallocThread();
+#endif
+
+TRACY_API void SetThreadName( const char* name )
+{
+#if defined _WIN32 || defined __CYGWIN__
+    static auto _SetThreadDescription = (t_SetThreadDescription)GetProcAddress( GetModuleHandleA( "kernel32.dll" ), "SetThreadDescription" );
+    if( _SetThreadDescription )
+    {
+        wchar_t buf[256];
+        mbstowcs( buf, name, 256 );
+        _SetThreadDescription( GetCurrentThread(), buf );
+    }
+    else
+    {
+#  if defined _MSC_VER
+        const DWORD MS_VC_EXCEPTION=0x406D1388;
+#    pragma pack( push, 8 )
+        struct THREADNAME_INFO
+        {
+            DWORD dwType;
+            LPCSTR szName;
+            DWORD dwThreadID;
+            DWORD dwFlags;
+        };
+#    pragma pack(pop)
+
+        DWORD ThreadId = GetCurrentThreadId();
+        THREADNAME_INFO info;
+        info.dwType = 0x1000;
+        info.szName = name;
+        info.dwThreadID = ThreadId;
+        info.dwFlags = 0;
+
+        __try
+        {
+            RaiseException( MS_VC_EXCEPTION, 0, sizeof(info)/sizeof(ULONG_PTR), (ULONG_PTR*)&info );
+        }
+        __except(EXCEPTION_EXECUTE_HANDLER)
+        {
+        }
+#  endif
+    }
+#elif defined _GNU_SOURCE && !defined __EMSCRIPTEN__ && !defined __CYGWIN__
+    {
+        const auto sz = strlen( name );
+        if( sz <= 15 )
+        {
+            pthread_setname_np( pthread_self(), name );
+        }
+        else
+        {
+            char buf[16];
+            memcpy( buf, name, 15 );
+            buf[15] = '\0';
+            pthread_setname_np( pthread_self(), buf );
+        }
+    }
+#endif
+#ifdef TRACY_ENABLE
+    {
+        InitRPMallocThread();
+        const auto sz = strlen( name );
+        char* buf = (char*)tracy_malloc( sz+1 );
+        memcpy( buf, name, sz );
+        buf[sz] = '\0';
+        auto data = (ThreadNameData*)tracy_malloc( sizeof( ThreadNameData ) );
+        data->id = detail::GetThreadHandleImpl();
+        data->name = buf;
+        data->next = GetThreadNameData().load( std::memory_order_relaxed );
+        while( !GetThreadNameData().compare_exchange_weak( data->next, data, std::memory_order_release, std::memory_order_relaxed ) ) {}
+    }
+#endif
+}
+
+TRACY_API const char* GetThreadName( uint64_t id )
+{
+    static char buf[256];
+#ifdef TRACY_ENABLE
+    auto ptr = GetThreadNameData().load( std::memory_order_relaxed );
+    while( ptr )
+    {
+        if( ptr->id == id )
+        {
+            return ptr->name;
+        }
+        ptr = ptr->next;
+    }
+#else
+#  if defined _WIN32 || defined __CYGWIN__
+    static auto _GetThreadDescription = (t_GetThreadDescription)GetProcAddress( GetModuleHandleA( "kernel32.dll" ), "GetThreadDescription" );
+    if( _GetThreadDescription )
+    {
+        auto hnd = OpenThread( THREAD_QUERY_LIMITED_INFORMATION, FALSE, (DWORD)id );
+        if( hnd != 0 )
+        {
+            PWSTR tmp;
+            _GetThreadDescription( hnd, &tmp );
+            auto ret = wcstombs( buf, tmp, 256 );
+            CloseHandle( hnd );
+            if( ret != 0 )
+            {
+                return buf;
+            }
+        }
+    }
+#  elif defined __linux__
+    int cs, fd;
+    char path[32];
+#   ifdef __ANDROID__
+    int tid = gettid();
+#   else
+    int tid = (int) syscall( SYS_gettid );
+#   endif
+    snprintf( path, sizeof( path ), "/proc/self/task/%d/comm", tid );
+    sprintf( buf, "%" PRIu64, id );
+#   ifndef __ANDROID__
+    pthread_setcancelstate( PTHREAD_CANCEL_DISABLE, &cs );
+#   endif
+    if ( ( fd = open( path, O_RDONLY ) ) > 0) {
+        int len = read( fd, buf, 255 );
+        if( len > 0 )
+        {
+            buf[len] = 0;
+            if( len > 1 && buf[len-1] == '\n' )
+            {
+                buf[len-1] = 0;
+            }
+        }
+        close( fd );
+    }
+#   ifndef __ANDROID__
+    pthread_setcancelstate( cs, 0 );
+#   endif
+    return buf;
+#  endif
+#endif
+    sprintf( buf, "%" PRIu64, id );
+    return buf;
+}
+
+}

Source/ThirdParty/tracy/common/TracySystem.hpp

@@ -0,0 +1,95 @@
+#ifndef __TRACYSYSTEM_HPP__
+#define __TRACYSYSTEM_HPP__
+
+#include <stdint.h>
+
+// Tracy -> Flax integration:
+// - use LZ4 from Flax
+// - use engine symbols export
+// - use engine types and macros
+// - remove AddVectoredExceptionHandler from win32 to prevent messing with Flax crashes reporting
+// - hide implementation from includers to reduce compilation overhead
+// - optimize includes (faster compilation)
+// - remove some features (colors, frame image, dx1 compression)
+#include "Engine/Core/Types/BaseTypes.h"
+#define TRACY_API FLAXENGINE_API
+#define tracy_force_inline FORCE_INLINE
+#define tracy_no_inline FORCE_NOINLINE
+
+#ifndef TracyConcat
+#  define TracyConcat(x,y) TracyConcatIndirect(x,y)
+#endif
+#ifndef TracyConcatIndirect
+#  define TracyConcatIndirect(x,y) x##y
+#endif
+
+namespace tracy
+{
+enum class PlotFormatType : uint8_t
+{
+    Number,
+    Memory,
+    Percentage
+};
+
+typedef void(*ParameterCallback)( uint32_t idx, int32_t val );
+
+struct TRACY_API SourceLocationData
+{
+    const char* name;
+    const char* function;
+    const char* file;
+    uint32_t line;
+    uint32_t color;
+};
+
+class TRACY_API ScopedZone
+{
+public:
+    ScopedZone( const ScopedZone& ) = delete;
+    ScopedZone( ScopedZone&& ) = delete;
+    ScopedZone& operator=( const ScopedZone& ) = delete;
+    ScopedZone& operator=( ScopedZone&& ) = delete;
+
+    ScopedZone( const SourceLocationData* srcloc, bool is_active = true );
+    ScopedZone( const SourceLocationData* srcloc, int depth, bool is_active = true );
+    ScopedZone( uint32_t line, const char* source, size_t sourceSz, const char* function, size_t functionSz, const char* name, size_t nameSz, bool is_active = true );
+    ScopedZone( uint32_t line, const char* source, size_t sourceSz, const char* function, size_t functionSz, const char* name, size_t nameSz, int depth, bool is_active = true );
+
+    ~ScopedZone();
+
+    void Text( const char* txt, size_t size );
+    void Name( const char* txt, size_t size );
+    void Name( const Char* txt, size_t size );
+    void Color( uint32_t color );
+    void Value( uint64_t value );
+    bool IsActive() const;
+
+private:
+    const bool m_active;
+
+#ifdef TRACY_ON_DEMAND
+    uint64_t m_connectionId;
+#endif
+};
+
+namespace detail
+{
+TRACY_API uint64_t GetThreadHandleImpl();
+}
+
+#ifdef TRACY_ENABLE
+TRACY_API uint64_t GetThreadHandle();
+#else
+static inline uint64_t GetThreadHandle()
+{
+    return detail::GetThreadHandleImpl();
+}
+#endif
+
+TRACY_API void SetThreadName( const char* name );
+TRACY_API const char* GetThreadName( uint64_t id );
+
+}
+
+#endif