Send Vulkan GPU calibration messages.

TracyD3D12.hpp

@@ -150,7 +150,7 @@ namespace tracy
 			memset(&item->gpuNewContext.thread, 0, sizeof(item->gpuNewContext.thread));
 			MemWrite(&item->gpuNewContext.period, 1E+09f / static_cast<float>(timestampFrequency));
 			MemWrite(&item->gpuNewContext.context, m_context);
-			MemWrite(&item->gpuNewContext.accuracyBits, uint8_t{ 0 });
+			MemWrite(&item->gpuNewContext.flags, uint8_t{ 0 });
 			MemWrite(&item->gpuNewContext.type, GpuContextType::Direct3D12);
 
 #ifdef TRACY_ON_DEMAND

TracyOpenCL.hpp

@@ -72,7 +72,7 @@ namespace tracy {
             MemWrite(&item->gpuNewContext.period, 1.0f);
             MemWrite(&item->gpuNewContext.type, GpuContextType::OpenCL);
             MemWrite(&item->gpuNewContext.context, (uint8_t) m_contextId);
-            MemWrite(&item->gpuNewContext.accuracyBits, (uint8_t)0);
+            MemWrite(&item->gpuNewContext.flags, (uint8_t)0);
 #ifdef TRACY_ON_DEMAND
             GetProfiler().DeferItem(*item);
 #endif

TracyOpenGL.hpp

@@ -110,7 +110,7 @@ public:
         MemWrite( &item->gpuNewContext.thread, thread );
         MemWrite( &item->gpuNewContext.period, period );
         MemWrite( &item->gpuNewContext.context, m_context );
-        MemWrite( &item->gpuNewContext.accuracyBits, (uint8_t)bits );
+        MemWrite( &item->gpuNewContext.flags, uint8_t( 0 ) );
         MemWrite( &item->gpuNewContext.type, GpuContextType::OpenGl );
 
 #ifdef TRACY_ON_DEMAND

TracyVulkan.hpp

@@ -102,21 +102,56 @@ public:
         vkQueueSubmit( queue, 1, &submitInfo, VK_NULL_HANDLE );
         vkQueueWaitIdle( queue );
 
-        vkBeginCommandBuffer( cmdbuf, &beginInfo );
+        int64_t tcpu, tgpu;
-        vkCmdWriteTimestamp( cmdbuf, VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT, m_query, 0 );
+        if( m_timeDomain == VK_TIME_DOMAIN_DEVICE_EXT )
-        vkEndCommandBuffer( cmdbuf );
+        {
-        vkQueueSubmit( queue, 1, &submitInfo, VK_NULL_HANDLE );
+            vkBeginCommandBuffer( cmdbuf, &beginInfo );
-        vkQueueWaitIdle( queue );
+            vkCmdWriteTimestamp( cmdbuf, VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT, m_query, 0 );
+            vkEndCommandBuffer( cmdbuf );
+            vkQueueSubmit( queue, 1, &submitInfo, VK_NULL_HANDLE );
+            vkQueueWaitIdle( queue );
 
-        int64_t tcpu = Profiler::GetTime();
+            tcpu = Profiler::GetTime();
-        int64_t tgpu;
+            vkGetQueryPoolResults( device, m_query, 0, 1, sizeof( tgpu ), &tgpu, sizeof( tgpu ), VK_QUERY_RESULT_64_BIT | VK_QUERY_RESULT_WAIT_BIT );
-        vkGetQueryPoolResults( device, m_query, 0, 1, sizeof( tgpu ), &tgpu, sizeof( tgpu ), VK_QUERY_RESULT_64_BIT | VK_QUERY_RESULT_WAIT_BIT );
 
-        vkBeginCommandBuffer( cmdbuf, &beginInfo );
+            vkBeginCommandBuffer( cmdbuf, &beginInfo );
-        vkCmdResetQueryPool( cmdbuf, m_query, 0, 1 );
+            vkCmdResetQueryPool( cmdbuf, m_query, 0, 1 );
-        vkEndCommandBuffer( cmdbuf );
+            vkEndCommandBuffer( cmdbuf );
-        vkQueueSubmit( queue, 1, &submitInfo, VK_NULL_HANDLE );
+            vkQueueSubmit( queue, 1, &submitInfo, VK_NULL_HANDLE );
-        vkQueueWaitIdle( queue );
+            vkQueueWaitIdle( queue );
+        }
+        else
+        {
+            enum { NumProbes = 32 };
+
+            VkCalibratedTimestampInfoEXT spec[2] = {
+                { VK_STRUCTURE_TYPE_CALIBRATED_TIMESTAMP_INFO_EXT, nullptr, VK_TIME_DOMAIN_DEVICE_EXT },
+                { VK_STRUCTURE_TYPE_CALIBRATED_TIMESTAMP_INFO_EXT, nullptr, m_timeDomain },
+            };
+            uint64_t ts[2];
+            uint64_t deviation[NumProbes];
+            for( int i=0; i<NumProbes; i++ )
+            {
+                _vkGetCalibratedTimestampsEXT( device, 2, spec, ts, deviation+i );
+            }
+            uint64_t minDeviation = deviation[0];
+            for( int i=1; i<NumProbes; i++ )
+            {
+                if( minDeviation > deviation[i] )
+                {
+                    minDeviation = deviation[i];
+                }
+            }
+            m_deviation = minDeviation * 3 / 2;
+
+            m_qpcToNs = int64_t( 1000000000. / GetFrequencyQpc() );
+
+            Calibrate( device, m_prevCalibration, tgpu );
+            tcpu = Profiler::GetTime();
+        }
+
+        uint8_t flags = 0;
+        if( m_timeDomain != VK_TIME_DOMAIN_DEVICE_EXT ) flags |= GpuContextCalibration;
 
         auto item = Profiler::QueueSerial();
         MemWrite( &item->hdr.type, QueueType::GpuNewContext );
@@ -125,7 +160,7 @@ public:
         memset( &item->gpuNewContext.thread, 0, sizeof( item->gpuNewContext.thread ) );
         MemWrite( &item->gpuNewContext.period, period );
         MemWrite( &item->gpuNewContext.context, m_context );
-        MemWrite( &item->gpuNewContext.accuracyBits, uint8_t( 0 ) );
+        MemWrite( &item->gpuNewContext.flags, flags );
         MemWrite( &item->gpuNewContext.type, GpuContextType::Vulkan );
 
 #ifdef TRACY_ON_DEMAND
@@ -153,6 +188,8 @@ public:
         {
             vkCmdResetQueryPool( cmdbuf, m_query, 0, m_queryCount );
             m_head = m_tail = 0;
+            int64_t tgpu;
+            if( m_timeDomain != VK_TIME_DOMAIN_DEVICE_EXT ) Calibrate( m_device, m_prevCalibration, tgpu );
             return;
         }
 #endif
@@ -184,6 +221,25 @@ public:
             Profiler::QueueSerialFinish();
         }
 
+        if( m_timeDomain != VK_TIME_DOMAIN_DEVICE_EXT )
+        {
+            int64_t tgpu, tcpu;
+            Calibrate( m_device, tcpu, tgpu );
+            const auto refCpu = Profiler::GetTime();
+            const auto delta = tcpu - m_prevCalibration;
+            if( delta > 0 )
+            {
+                m_prevCalibration = tcpu;
+                auto item = Profiler::QueueSerial();
+                MemWrite( &item->hdr.type, QueueType::GpuCalibration );
+                MemWrite( &item->gpuCalibration.gpuTime, tgpu );
+                MemWrite( &item->gpuCalibration.cpuTime, refCpu );
+                MemWrite( &item->gpuCalibration.cpuDelta, delta );
+                MemWrite( &item->gpuCalibration.context, m_context );
+                Profiler::QueueSerialFinish();
+            }
+        }
+
         vkCmdResetQueryPool( cmdbuf, m_query, m_tail, cnt );
 
         m_tail += cnt;

common/TracyProtocol.hpp

@@ -9,7 +9,7 @@ namespace tracy
 
 constexpr unsigned Lz4CompressBound( unsigned isize ) { return isize + ( isize / 255 ) + 16; }
 
-enum : uint32_t { ProtocolVersion = 36 };
+enum : uint32_t { ProtocolVersion = 37 };
 enum : uint32_t { BroadcastVersion = 1 };
 
 using lz4sz_t = uint32_t;

common/TracyQueue.hpp

@@ -56,6 +56,7 @@ enum class QueueType : uint8_t
     Terminate,
     KeepAlive,
     ThreadContext,
+    GpuCalibration,
     Crash,
     CrashReport,
     ZoneValidation,
@@ -268,6 +269,11 @@ enum class GpuContextType : uint8_t
     Direct3D12
 };
 
+enum GpuContextFlags : uint8_t
+{
+    GpuContextCalibration   = 1 << 0
+};
+
 struct QueueGpuNewContext
 {
     int64_t cpuTime;
@@ -275,7 +281,7 @@ struct QueueGpuNewContext
     uint64_t thread;
     float period;
     uint8_t context;
-    uint8_t accuracyBits;
+    GpuContextFlags flags;
     GpuContextType type;
 };
 
@@ -303,6 +309,14 @@ struct QueueGpuTime
     uint8_t context;
 };
 
+struct QueueGpuCalibration
+{
+    int64_t gpuTime;
+    int64_t cpuTime;
+    int64_t cpuDelta;
+    uint8_t context;
+};
+
 struct QueueMemAlloc
 {
     int64_t time;
@@ -477,6 +491,7 @@ struct QueueItem
         QueueGpuZoneBegin gpuZoneBegin;
         QueueGpuZoneEnd gpuZoneEnd;
         QueueGpuTime gpuTime;
+        QueueGpuCalibration gpuCalibration;
         QueueMemAlloc memAlloc;
         QueueMemFree memFree;
         QueueCallstackMemory callstackMemory;
@@ -553,6 +568,7 @@ static constexpr size_t QueueDataSize[] = {
     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 ),

server/TracyWorker.cpp

@@ -3957,6 +3957,9 @@ bool Worker::Process( const QueueItem& ev )
     case QueueType::GpuTime:
         ProcessGpuTime( ev.gpuTime );
         break;
+    case QueueType::GpuCalibration:
+        ProcessGpuCalibration( ev.gpuCalibration );
+        break;
     case QueueType::MemAlloc:
         ProcessMemAlloc( ev.memAlloc );
         break;
@@ -5061,6 +5064,11 @@ void Worker::ProcessGpuTime( const QueueGpuTime& ev )
     }
 }
 
+void Worker::ProcessGpuCalibration( const QueueGpuCalibration& ev )
+{
+
+}
+
 void Worker::ProcessMemAlloc( const QueueMemAlloc& ev )
 {
     const auto refTime = m_refTimeSerial + ev.time;

server/TracyWorker.hpp

@@ -643,6 +643,7 @@ private:
     tracy_force_inline void ProcessGpuZoneBeginCallstack( const QueueGpuZoneBegin& ev, bool serial );
     tracy_force_inline void ProcessGpuZoneEnd( const QueueGpuZoneEnd& ev, bool serial );
     tracy_force_inline void ProcessGpuTime( const QueueGpuTime& ev );
+    tracy_force_inline void ProcessGpuCalibration( const QueueGpuCalibration& ev );
     tracy_force_inline void ProcessMemAlloc( const QueueMemAlloc& ev );
     tracy_force_inline bool ProcessMemFree( const QueueMemFree& ev );
     tracy_force_inline void ProcessMemAllocCallstack( const QueueMemAlloc& ev );