Process allocated source location GPU zones.

2 files changed, 47 insertions, 3 deletions

diff --git a/server/TracyWorker.cpp b/server/TracyWorker.cpp
index abf7a192..1abb975c 100644
--- a/server/TracyWorker.cpp
+++ b/server/TracyWorker.cpp
@@ -4206,6 +4206,12 @@ bool Worker::Process( const QueueItem& ev )
     case QueueType::GpuZoneBeginCallstack:
         ProcessGpuZoneBeginCallstack( ev.gpuZoneBegin, false );
         break;
+    case QueueType::GpuZoneBeginAllocSrcLoc:
+        ProcessGpuZoneBeginAllocSrcLoc( ev.gpuZoneBeginLean, false );
+        break;
+    case QueueType::GpuZoneBeginAllocSrcLocCallstack:
+        ProcessGpuZoneBeginAllocSrcLocCallstack( ev.gpuZoneBeginLean, false );
+        break;
     case QueueType::GpuZoneEnd:
         ProcessGpuZoneEnd( ev.gpuZoneEnd, false );
         break;
@@ -4215,6 +4221,12 @@ bool Worker::Process( const QueueItem& ev )
     case QueueType::GpuZoneBeginCallstackSerial:
         ProcessGpuZoneBeginCallstack( ev.gpuZoneBegin, true );
         break;
+    case QueueType::GpuZoneBeginAllocSrcLocSerial:
+        ProcessGpuZoneBeginAllocSrcLoc( ev.gpuZoneBeginLean, true );
+        break;
+    case QueueType::GpuZoneBeginAllocSrcLocCallstackSerial:
+        ProcessGpuZoneBeginAllocSrcLocCallstack( ev.gpuZoneBeginLean, true );
+        break;
     case QueueType::GpuZoneEndSerial:
         ProcessGpuZoneEnd( ev.gpuZoneEnd, true );
         break;
@@ -5173,13 +5185,26 @@ void Worker::ProcessGpuNewContext( const QueueGpuNewContext& ev )
 
 void Worker::ProcessGpuZoneBeginImpl( GpuEvent* zone, const QueueGpuZoneBegin& ev, bool serial )
 {
+    CheckSourceLocation( ev.srcloc );
+    zone->SetSrcLoc( ShrinkSourceLocation( ev.srcloc ) );
+    ProcessGpuZoneBeginImplCommon( zone, ev, serial );
+}
+
+void Worker::ProcessGpuZoneBeginAllocSrcLocImpl( GpuEvent* zone, const QueueGpuZoneBeginLean& ev, bool serial )
+{
+    assert( m_pendingSourceLocationPayload != 0 );
+    zone->SetSrcLoc( m_pendingSourceLocationPayload );
+    ProcessGpuZoneBeginImplCommon( zone, ev, time );
+    m_pendingSourceLocationPayload = 0;
+}
+
+void Worker::ProcessGpuZoneBeginImplCommon( GpuEvent* zone, const QueueGpuZoneBeginLean& ev, bool serial )
+{
     m_data.gpuCnt++;
 
     auto ctx = m_gpuCtxMap[ev.context].get();
     assert( ctx );
 
-    CheckSourceLocation( ev.srcloc );
-
     int64_t cpuTime;
     if( serial )
     {
@@ -5196,7 +5221,6 @@ void Worker::ProcessGpuZoneBeginImpl( GpuEvent* zone, const QueueGpuZoneBegin& e
     zone->SetCpuEnd( -1 );
     zone->SetGpuStart( -1 );
     zone->SetGpuEnd( -1 );
-    zone->SetSrcLoc( ShrinkSourceLocation( ev.srcloc ) );
     zone->callstack.SetVal( 0 );
     zone->SetChild( -1 );
 
@@ -5258,6 +5282,22 @@ void Worker::ProcessGpuZoneBeginCallstack( const QueueGpuZoneBegin& ev, bool ser
     it->second = 0;
 }
 
+void Worker::ProcessGpuZoneBeginAllocSrcLoc( const QueueGpuZoneBeginLean& ev, bool serial )
+{
+    auto zone = m_slab.Alloc<GpuEvent>();
+    ProcessGpuZoneBeginAllocSrcLocImpl( zone, ev, serial );
+}
+
+void Worker::ProcessGpuZoneBeginAllocSrcLocCallstack( const QueueGpuZoneBeginLean& ev, bool serial )
+{
+    auto zone = m_slab.Alloc<GpuEvent>();
+    ProcessGpuZoneBeginAllocSrcLocImpl( zone, ev, serial );
+    auto it = m_nextCallstack.find( m_threadCtx );
+    assert( it != m_nextCallstack.end() );
+    zone->callstack.SetVal( it->second );
+    it->second = 0;
+}
+
 void Worker::ProcessGpuZoneEnd( const QueueGpuZoneEnd& ev, bool serial )
 {
     auto ctx = m_gpuCtxMap[ev.context];
diff --git a/server/TracyWorker.hpp b/server/TracyWorker.hpp
index 2fe6ba00..6593c4ee 100644
--- a/server/TracyWorker.hpp
+++ b/server/TracyWorker.hpp
@@ -635,6 +635,8 @@ private:
     tracy_force_inline void ProcessGpuNewContext( const QueueGpuNewContext& ev );
     tracy_force_inline void ProcessGpuZoneBegin( const QueueGpuZoneBegin& ev, bool serial );
     tracy_force_inline void ProcessGpuZoneBeginCallstack( const QueueGpuZoneBegin& ev, bool serial );
+    tracy_force_inline void ProcessGpuZoneBeginAllocSrcLoc( const QueueGpuZoneBeginLean& ev, bool serial );
+    tracy_force_inline void ProcessGpuZoneBeginAllocSrcLocCallstack( const QueueGpuZoneBeginLean& 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 );
@@ -666,6 +668,8 @@ private:
     tracy_force_inline void ProcessZoneBeginImpl( ZoneEvent* zone, const QueueZoneBegin& ev );
     tracy_force_inline void ProcessZoneBeginAllocSrcLocImpl( ZoneEvent* zone, const QueueZoneBeginLean& ev );
     tracy_force_inline void ProcessGpuZoneBeginImpl( GpuEvent* zone, const QueueGpuZoneBegin& ev, bool serial );
+    tracy_force_inline void ProcessGpuZoneBeginAllocSrcLocImpl( GpuEvent* zone, const QueueGpuZoneBeginLean& ev, bool serial );
+    tracy_force_inline void ProcessGpuZoneBeginImplCommon( GpuEvent* zone, const QueueGpuZoneBeginLean& ev, bool serial );
     tracy_force_inline MemEvent* ProcessMemAllocImpl( uint64_t memname, MemData& memdata, const QueueMemAlloc& ev );
     tracy_force_inline MemEvent* ProcessMemFreeImpl( uint64_t memname, MemData& memdata, const QueueMemFree& ev );