Support hw sample times on server.

server/TracyEvent.hpp

@@ -158,6 +158,8 @@ private:
     uint8_t m_val[6];
 };
 
+struct Int48Sort { bool operator()( const Int48& lhs, const Int48& rhs ) { return lhs.Val() < rhs.Val(); }; };
+
 
 struct SourceLocationBase
 {
@@ -253,12 +255,33 @@ enum { SampleDataRangeSize = sizeof( SampleDataRange ) };
 
 struct HwSampleData
 {
-    uint32_t cycles;
+    SortedVector<Int48, Int48Sort> cycles;
-    uint32_t retired;
+    SortedVector<Int48, Int48Sort> retired;
-    uint32_t cacheRef;
+    SortedVector<Int48, Int48Sort> cacheRef;
-    uint32_t cacheMiss;
+    SortedVector<Int48, Int48Sort> cacheMiss;
-    uint32_t branchRetired;
+    SortedVector<Int48, Int48Sort> branchRetired;
-    uint32_t branchMiss;
+    SortedVector<Int48, Int48Sort> branchMiss;
+
+    bool is_sorted() const
+    {
+        return
+            cycles.is_sorted() &&
+            retired.is_sorted() &&
+            cacheRef.is_sorted() &&
+            cacheMiss.is_sorted() &&
+            branchRetired.is_sorted() &&
+            branchMiss.is_sorted();
+    }
+
+    void sort()
+    {
+        if( !cycles.is_sorted() ) cycles.sort();
+        if( !retired.is_sorted() ) retired.sort();
+        if( !cacheRef.is_sorted() ) cacheRef.sort();
+        if( !cacheMiss.is_sorted() ) cacheMiss.sort();
+        if( !branchRetired.is_sorted() ) branchRetired.sort();
+        if( !branchMiss.is_sorted() ) branchMiss.sort();
+    }
 };
 
 enum { HwSampleDataSize = sizeof( HwSampleData ) };

server/TracySourceView.cpp

@@ -2633,9 +2633,9 @@ void SourceView::RenderLine( const Tokenizer::Line& line, int lineNum, const Add
     DrawLine( draw, dpos + ImVec2( 0, ty+2 ), dpos + ImVec2( w, ty+2 ), 0x08FFFFFF );
 }
 
-static void PrintHwSampleTooltip( const HwSampleData& hw, bool hideFirstSeparator )
+static void PrintHwSampleTooltip( size_t cycles, size_t retired, size_t cacheRef, size_t cacheMiss, size_t branchRetired, size_t branchMiss, bool hideFirstSeparator )
 {
-    if( hw.cycles || hw.retired )
+    if( cycles || retired )
     {
         if( hideFirstSeparator )
         {
@@ -2645,17 +2645,17 @@ static void PrintHwSampleTooltip( const HwSampleData& hw, bool hideFirstSeparato
         {
             ImGui::Separator();
         }
-        if( hw.cycles && hw.retired )
+        if( cycles && retired )
         {
             char buf[32];
-            auto end = PrintFloat( buf, buf+32, float( hw.retired ) / hw.cycles, 2 );
+            auto end = PrintFloat( buf, buf+32, float( retired ) / cycles, 2 );
             *end = '\0';
             TextFocused( "IPC:", buf );
         }
-        if( hw.cycles ) TextFocused( "Cycles:", RealToString( hw.cycles ) );
+        if( cycles ) TextFocused( "Cycles:", RealToString( cycles ) );
-        if( hw.retired ) TextFocused( "Retirements:", RealToString( hw.retired ) );
+        if( retired ) TextFocused( "Retirements:", RealToString( retired ) );
     }
-    if( hw.cacheRef || hw.cacheMiss )
+    if( cacheRef || cacheMiss )
     {
         if( hideFirstSeparator )
         {
@@ -2665,17 +2665,17 @@ static void PrintHwSampleTooltip( const HwSampleData& hw, bool hideFirstSeparato
         {
             ImGui::Separator();
         }
-        if( hw.cacheRef )
+        if( cacheRef )
         {
             char buf[32];
-            auto end = PrintFloat( buf, buf+32, float( 100 * hw.cacheMiss ) / hw.cacheRef, 2 );
+            auto end = PrintFloat( buf, buf+32, float( 100 * cacheMiss ) / cacheRef, 2 );
             memcpy( end, "%", 2 );
             TextFocused( "Cache miss rate:", buf );
-            TextFocused( "Cache references:", RealToString( hw.cacheRef ) );
+            TextFocused( "Cache references:", RealToString( cacheRef ) );
         }
-        if( hw.cacheMiss ) TextFocused( "Cache misses:", RealToString( hw.cacheMiss ) );
+        if( cacheMiss ) TextFocused( "Cache misses:", RealToString( cacheMiss ) );
     }
-    if( hw.branchRetired || hw.branchMiss )
+    if( branchRetired || branchMiss )
     {
         if( hideFirstSeparator )
         {
@@ -2685,15 +2685,15 @@ static void PrintHwSampleTooltip( const HwSampleData& hw, bool hideFirstSeparato
         {
             ImGui::Separator();
         }
-        if( hw.branchRetired )
+        if( branchRetired )
         {
             char buf[32];
-            auto end = PrintFloat( buf, buf+32, float( 100 * hw.branchMiss ) / hw.branchRetired, 2 );
+            auto end = PrintFloat( buf, buf+32, float( 100 * branchMiss ) / branchRetired, 2 );
             memcpy( end, "%", 2 );
             TextFocused( "Branch mispredictions rate:", buf );
-            TextFocused( "Retired branches:", RealToString( hw.branchRetired ) );
+            TextFocused( "Retired branches:", RealToString( branchRetired ) );
         }
-        if( hw.branchMiss ) TextFocused( "Branch mispredictions:", RealToString( hw.branchMiss ) );
+        if( branchMiss ) TextFocused( "Branch mispredictions:", RealToString( branchMiss ) );
     }
 }
 
@@ -2719,6 +2719,18 @@ void SourceView::RenderAsmLine( AsmLine& line, const AddrStat& ipcnt, const Addr
 
     const auto asmIdx = &line - m_asm.data();
 
+    const auto hw = worker.GetHwSampleData( line.addr );
+    size_t cycles = 0, retired = 0, cacheRef = 0, cacheMiss = 0, branchRetired = 0, branchMiss = 0;
+    if( hw )
+    {
+        cycles = hw->cycles.size();
+        retired = hw->retired.size();
+        cacheRef = hw->cacheRef.size();
+        cacheMiss = hw->cacheMiss.size();
+        branchRetired = hw->branchRetired.size();
+        branchMiss = hw->branchMiss.size();
+    }
+
     const auto ts = ImGui::CalcTextSize( " " );
     if( iptotal.local + iptotal.ext != 0 )
     {
@@ -2732,7 +2744,7 @@ void SourceView::RenderAsmLine( AsmLine& line, const AddrStat& ipcnt, const Addr
                 {
                     if( m_font ) ImGui::PopFont();
                     ImGui::BeginTooltip();
-                    PrintHwSampleTooltip( *hw, true );
+                    PrintHwSampleTooltip( cycles, retired, cacheRef, cacheMiss, branchRetired, branchMiss, true );
                     ImGui::EndTooltip();
                     if( m_font ) ImGui::PushFont( m_font );
                 }
@@ -2788,7 +2800,7 @@ void SourceView::RenderAsmLine( AsmLine& line, const AddrStat& ipcnt, const Addr
                 }
 
                 const auto hw = worker.GetHwSampleData( line.addr );
-                if( hw ) PrintHwSampleTooltip( *hw, false );
+                if( hw ) PrintHwSampleTooltip( cycles, retired, cacheRef, cacheMiss, branchRetired, branchMiss, false );
 
                 const auto& stats = *worker.GetSymbolStats( symAddrParents );
                 if( !stats.parents.empty() )
@@ -2884,13 +2896,12 @@ void SourceView::RenderAsmLine( AsmLine& line, const AddrStat& ipcnt, const Addr
     const bool showHwSamples = m_hwSamples && worker.GetHwSampleCountAddress() != 0;
     if( showHwSamples )
     {
-        auto hw = worker.GetHwSampleData( line.addr );
         if( hw )
         {
-            if( hw->cycles != 0 )
+            if( cycles )
             {
-                const bool unreliable = hw->cycles < 10 || hw->retired < 10;
+                const bool unreliable = cycles < 10 || retired < 10;
-                const float ipc = float( hw->retired ) / hw->cycles;
+                const float ipc = float( retired ) / cycles;
                 uint32_t col = unreliable ? 0x44FFFFFF : GetGoodnessColor( ipc * 0.25f );
                 if( ipc >= 10 )
                 {
@@ -2913,8 +2924,8 @@ void SourceView::RenderAsmLine( AsmLine& line, const AddrStat& ipcnt, const Addr
                     ImGui::SameLine();
                     TextDisabledUnformatted( "Higher is better" );
                     ImGui::Separator();
-                    TextFocused( "Cycles:", RealToString( hw->cycles ) );
+                    TextFocused( "Cycles:", RealToString( cycles ) );
-                    TextFocused( "Retirements:", RealToString( hw->retired ) );
+                    TextFocused( "Retirements:", RealToString( retired ) );
                     if( unreliable ) TextColoredUnformatted( 0xFF4444FF, "Not enough samples for reliable data!" );
                     ImGui::EndTooltip();
                     if( m_font ) ImGui::PushFont( m_font );
@@ -2925,12 +2936,12 @@ void SourceView::RenderAsmLine( AsmLine& line, const AddrStat& ipcnt, const Addr
                 ImGui::ItemSize( ImVec2( 7 * ts.x, ts.y ) );
             }
             ImGui::SameLine( 0, 0 );
-            if( hw->branchRetired != 0 )
+            if( branchRetired )
             {
-                const bool unreliable = hw->branchRetired < 10;
+                const bool unreliable = branchRetired < 10;
-                const float rate = float( hw->branchMiss ) / hw->branchRetired;
+                const float rate = float( branchMiss ) / branchRetired;
                 uint32_t col = unreliable ? 0x44FFFFFF : GetGoodnessColor( 1.f - rate * 3.f );
-                if( hw->branchMiss == 0 )
+                if( branchMiss == 0 )
                 {
                     TextColoredUnformatted( col, "   0%  " );
                 }
@@ -2963,8 +2974,8 @@ void SourceView::RenderAsmLine( AsmLine& line, const AddrStat& ipcnt, const Addr
                     ImGui::SameLine();
                     TextDisabledUnformatted( "Lower is better" );
                     ImGui::Separator();
-                    TextFocused( "Retired branches:", RealToString( hw->branchRetired ) );
+                    TextFocused( "Retired branches:", RealToString( branchRetired ) );
-                    TextFocused( "Branch mispredictions:", RealToString( hw->branchMiss ) );
+                    TextFocused( "Branch mispredictions:", RealToString( branchMiss ) );
                     if( unreliable ) TextColoredUnformatted( 0xFF4444FF, "Not enough samples for reliable data!" );
                     ImGui::EndTooltip();
                     if( m_font ) ImGui::PushFont( m_font );
@@ -2975,12 +2986,12 @@ void SourceView::RenderAsmLine( AsmLine& line, const AddrStat& ipcnt, const Addr
                 ImGui::ItemSize( ImVec2( 7 * ts.x, ts.y ) );
             }
             ImGui::SameLine( 0, 0 );
-            if( hw->cacheRef != 0 )
+            if( cacheRef )
             {
-                const bool unreliable = hw->cacheRef < 10;
+                const bool unreliable = cacheRef < 10;
-                const float rate = float( hw->cacheMiss ) / hw->cacheRef;
+                const float rate = float( cacheMiss ) / cacheRef;
                 uint32_t col = unreliable ? 0x44FFFFFF : GetGoodnessColor( 1.f - rate * 3.f );
-                if( hw->cacheMiss == 0 )
+                if( cacheMiss == 0 )
                 {
                     TextColoredUnformatted( col, "   0%" );
                 }
@@ -3013,8 +3024,8 @@ void SourceView::RenderAsmLine( AsmLine& line, const AddrStat& ipcnt, const Addr
                     ImGui::SameLine();
                     TextDisabledUnformatted( "Lower is better" );
                     ImGui::Separator();
-                    TextFocused( "Cache references:", RealToString( hw->cacheRef ) );
+                    TextFocused( "Cache references:", RealToString( cacheRef ) );
-                    TextFocused( "Cache misses:", RealToString( hw->cacheMiss ) );
+                    TextFocused( "Cache misses:", RealToString( cacheMiss ) );
                     if( unreliable ) TextColoredUnformatted( 0xFF4444FF, "Not enough samples for reliable data!" );
                     ImGui::EndTooltip();
                     if( m_font ) ImGui::PushFont( m_font );

server/TracyWorker.cpp

@@ -237,6 +237,23 @@ static tracy_force_inline void UpdateLockRange( LockMap& lockmap, const LockEven
     if( range.end < lt ) range.end = lt;
 }
 
+template<int U>
+static void ReadHwSampleVec( FileRead& f, SortedVector<Int48, Int48Sort>& vec, Slab<U>& slab )
+{
+    uint64_t sz;
+    f.Read( sz );
+    if( sz != 0 )
+    {
+        int64_t refTime = 0;
+        vec.reserve_exact( sz, slab );
+        for( uint64_t i=0; i<sz; i++ )
+        {
+            vec[i] = ReadTimeOffset( f, refTime );
+        }
+    }
+}
+
+
 LoadProgress Worker::s_loadProgress;
 
 Worker::Worker( const char* addr, uint16_t port )
@@ -1707,9 +1724,14 @@ Worker::Worker( FileRead& f, EventType::Type eventMask, bool bgTasks )
         for( uint64_t i=0; i<sz; i++ )
         {
             uint64_t addr;
-            HwSampleData data;
+            f.Read( addr );
-            f.Read2( addr, data );
+            auto& data = m_data.hwSamples.emplace( addr, HwSampleData {} ).first->second;
-            m_data.hwSamples.emplace( addr, data );
+            ReadHwSampleVec( f, data.cycles, m_slab );
+            ReadHwSampleVec( f, data.retired, m_slab );
+            ReadHwSampleVec( f, data.cacheRef, m_slab );
+            ReadHwSampleVec( f, data.cacheMiss, m_slab );
+            ReadHwSampleVec( f, data.branchRetired, m_slab );
+            ReadHwSampleVec( f, data.branchMiss, m_slab );
         }
     }
 
@@ -6323,44 +6345,50 @@ void Worker::ProcessTidToPid( const QueueTidToPid& ev )
 
 void Worker::ProcessHwSampleCpuCycle( const QueueHwSample& ev )
 {
+    const auto time = TscTime( ev.time - m_data.baseTime );
     auto it = m_data.hwSamples.find( ev.ip );
     if( it == m_data.hwSamples.end() ) it = m_data.hwSamples.emplace( ev.ip, HwSampleData {} ).first;
-    it->second.cycles++;
+    it->second.cycles.push_back( time );
 }
 
 void Worker::ProcessHwSampleInstructionRetired( const QueueHwSample& ev )
 {
+    const auto time = TscTime( ev.time - m_data.baseTime );
     auto it = m_data.hwSamples.find( ev.ip );
     if( it == m_data.hwSamples.end() ) it = m_data.hwSamples.emplace( ev.ip, HwSampleData {} ).first;
-    it->second.retired++;
+    it->second.retired.push_back( time );
 }
 
 void Worker::ProcessHwSampleCacheReference( const QueueHwSample& ev )
 {
+    const auto time = TscTime( ev.time - m_data.baseTime );
     auto it = m_data.hwSamples.find( ev.ip );
     if( it == m_data.hwSamples.end() ) it = m_data.hwSamples.emplace( ev.ip, HwSampleData {} ).first;
-    it->second.cacheRef++;
+    it->second.cacheRef.push_back( time );
 }
 
 void Worker::ProcessHwSampleCacheMiss( const QueueHwSample& ev )
 {
+    const auto time = TscTime( ev.time - m_data.baseTime );
     auto it = m_data.hwSamples.find( ev.ip );
     if( it == m_data.hwSamples.end() ) it = m_data.hwSamples.emplace( ev.ip, HwSampleData {} ).first;
-    it->second.cacheMiss++;
+    it->second.cacheMiss.push_back( time );
 }
 
 void Worker::ProcessHwSampleBranchRetired( const QueueHwSample& ev )
 {
+    const auto time = TscTime( ev.time - m_data.baseTime );
     auto it = m_data.hwSamples.find( ev.ip );
     if( it == m_data.hwSamples.end() ) it = m_data.hwSamples.emplace( ev.ip, HwSampleData {} ).first;
-    it->second.branchRetired++;
+    it->second.branchRetired.push_back( time );
 }
 
 void Worker::ProcessHwSampleBranchMiss( const QueueHwSample& ev )
 {
+    const auto time = TscTime( ev.time - m_data.baseTime );
     auto it = m_data.hwSamples.find( ev.ip );
     if( it == m_data.hwSamples.end() ) it = m_data.hwSamples.emplace( ev.ip, HwSampleData {} ).first;
-    it->second.branchMiss++;
+    it->second.branchMiss.push_back( time );
 }
 
 void Worker::ProcessParamSetup( const QueueParamSetup& ev )
@@ -7023,6 +7051,21 @@ void Worker::Disconnect()
     m_disconnect = true;
 }
 
+static void WriteHwSampleVec( FileWrite& f, SortedVector<Int48, Int48Sort>& vec )
+{
+    uint64_t sz = vec.size();
+    f.Write( &sz, sizeof( sz ) );
+    if( sz != 0 )
+    {
+        if( !vec.is_sorted() ) vec.sort();
+        int64_t refTime = 0;
+        for( auto& v : vec )
+        {
+            WriteTimeOffset( f, refTime, v.Val() );
+        }
+    }
+}
+
 void Worker::Write( FileWrite& f, bool fiDict )
 {
     DoPostponedWork();
@@ -7563,7 +7606,12 @@ void Worker::Write( FileWrite& f, bool fiDict )
     for( auto& v : m_data.hwSamples )
     {
         f.Write( &v.first, sizeof( v.first ) );
-        f.Write( &v.second, sizeof( v.second ) );
+        WriteHwSampleVec( f, v.second.cycles );
+        WriteHwSampleVec( f, v.second.retired );
+        WriteHwSampleVec( f, v.second.cacheRef );
+        WriteHwSampleVec( f, v.second.cacheMiss );
+        WriteHwSampleVec( f, v.second.branchRetired );
+        WriteHwSampleVec( f, v.second.branchMiss );
     }
 
     sz = m_data.sourceFileCache.size();
@@ -7775,8 +7823,12 @@ uint64_t Worker::GetHwSampleCount() const
     uint64_t cnt = 0;
     for( auto& v : m_data.hwSamples )
     {
-        cnt += v.second.cycles;
+        cnt += v.second.cycles.size();
-        cnt += v.second.retired;
+        cnt += v.second.retired.size();
+        cnt += v.second.cacheRef.size();
+        cnt += v.second.cacheMiss.size();
+        cnt += v.second.branchRetired.size();
+        cnt += v.second.branchMiss.size();
     }
     return cnt;
 }