Add red color to Tracy profiler zones that cause CPU waiting to improve profiling

Source/Engine/Content/Asset.cpp

@@ -494,6 +494,9 @@ bool Asset::WaitForLoaded(double timeoutInMilliseconds) const
     }
 
     PROFILE_CPU();
+    ZoneColor(TracyWaitZoneColor);
+    const StringView path(GetPath());
+    ZoneText(*path, path.Length());
 
     Content::WaitForTask(loadingTask, timeoutInMilliseconds);
 

Source/Engine/Content/Content.cpp

@@ -1128,6 +1128,8 @@ void Content::WaitForTask(ContentLoadTask* loadingTask, double timeoutInMillisec
                             localQueue.Clear();
                         }
 
+                        PROFILE_CPU_NAMED("Inline");
+                        ZoneColor(0xffaaaaaa);
                         thread->Run(tmp);
                     }
                     else

Source/Engine/GraphicsDevice/DirectX/DX11/GPUSwapChainDX11.cpp

@@ -3,11 +3,12 @@
 #if GRAPHICS_API_DIRECTX11
 
 #include "GPUSwapChainDX11.h"
+#include "GPUContextDX11.h"
 #include "Engine/Platform/Window.h"
 #include "Engine/Graphics/RenderTools.h"
 #include "Engine/GraphicsDevice/DirectX/RenderToolsDX.h"
+#include "Engine/Profiler/ProfilerCPU.h"
 #include "Engine/Profiler/ProfilerMemory.h"
-#include "GPUContextDX11.h"
 
 GPUSwapChainDX11::GPUSwapChainDX11(GPUDeviceDX11* device, Window* window)
     : GPUResourceDX11(device, StringView::Empty)
@@ -140,6 +141,9 @@ GPUTextureView* GPUSwapChainDX11::GetBackBufferView()
 
 void GPUSwapChainDX11::Present(bool vsync)
 {
+    PROFILE_CPU();
+    ZoneColor(TracyWaitZoneColor);
+
     // Present frame
     ASSERT(_swapChain);
     UINT presentFlags = 0;

Source/Engine/GraphicsDevice/DirectX/DX12/CommandQueueDX12.cpp

@@ -6,6 +6,7 @@
 #include "GPUDeviceDX12.h"
 #include "Engine/Threading/Threading.h"
 #include "Engine/GraphicsDevice/DirectX/RenderToolsDX.h"
+#include "Engine/Profiler/ProfilerCPU.h"
 
 FenceDX12::FenceDX12(GPUDeviceDX12* device)
     : _currentValue(1)
@@ -64,12 +65,12 @@ void FenceDX12::WaitCPU(uint64 value)
 {
     if (IsFenceComplete(value))
         return;
-
+    PROFILE_CPU();
+    ZoneColor(TracyWaitZoneColor);
     ScopeLock lock(_locker);
 
     _fence->SetEventOnCompletion(value, _event);
     WaitForSingleObject(_event, INFINITE);
-
     _lastCompletedValue = _fence->GetCompletedValue();
 }
 

Source/Engine/GraphicsDevice/DirectX/DX12/GPUSwapChainDX12.cpp

@@ -6,6 +6,7 @@
 #include "GPUContextDX12.h"
 #include "../IncludeDirectXHeaders.h"
 #include "Engine/GraphicsDevice/DirectX/RenderToolsDX.h"
+#include "Engine/Profiler/ProfilerCPU.h"
 #include "Engine/Profiler/ProfilerMemory.h"
 
 void BackBufferDX12::Setup(GPUSwapChainDX12* window, ID3D12Resource* backbuffer)
@@ -364,6 +365,8 @@ void GPUSwapChainDX12::End(RenderTask* task)
 
 void GPUSwapChainDX12::Present(bool vsync)
 {
+    PROFILE_CPU();
+    ZoneColor(TracyWaitZoneColor);
 #if PLATFORM_XBOX_SCARLETT || PLATFORM_XBOX_ONE
     ID3D12Resource* backBuffer = _backBuffers[_currentFrameIndex].GetResource();
     D3D12XBOX_PRESENT_PLANE_PARAMETERS planeParameters = {};

Source/Engine/GraphicsDevice/Vulkan/GPUDeviceVulkan.cpp

@@ -2094,6 +2094,7 @@ void GPUDeviceVulkan::WaitForGPU()
     if (Device != VK_NULL_HANDLE)
     {
         PROFILE_CPU();
+        ZoneColor(TracyWaitZoneColor);
         VALIDATE_VULKAN_RESULT(vkDeviceWaitIdle(Device));
     }
 }

Source/Engine/GraphicsDevice/Vulkan/GPUSwapChainVulkan.cpp

@@ -553,6 +553,7 @@ void GPUSwapChainVulkan::Present(bool vsync)
     if (_acquiredImageIndex == -1)
         return;
     PROFILE_CPU();
+    ZoneColor(TracyWaitZoneColor);
 
     // Ensure that backbuffer has been acquired before presenting it to the window
     const auto backBuffer = (GPUTextureViewVulkan*)GetBackBufferView();

Source/Engine/Profiler/ProfilerCPU.h

@@ -412,3 +412,6 @@ struct TIsPODType<ProfilerCPU::Event>
 #define PROFILE_CPU_ACTOR(actor)
 
 #endif
+
+// CPU-wait zones can be marked with red color for better readability
+#define TracyWaitZoneColor 0xba1904

Source/Engine/Threading/JobSystem.cpp

@@ -385,6 +385,7 @@ void JobSystem::Wait(int64 label)
 {
 #if JOB_SYSTEM_ENABLED
     PROFILE_CPU();
+    ZoneColor(TracyWaitZoneColor);
 
     while (Platform::AtomicRead(&ExitFlag) == 0)
     {

Source/Engine/Threading/Task.cpp

@@ -40,6 +40,7 @@ void Task::Cancel()
 bool Task::Wait(double timeoutMilliseconds) const
 {
     PROFILE_CPU();
+    ZoneColor(TracyWaitZoneColor);
     const double startTime = Platform::GetTimeSeconds();
 
     // TODO: no active waiting! use a semaphore!
@@ -76,6 +77,7 @@ bool Task::Wait(double timeoutMilliseconds) const
 bool Task::WaitAll(const Span<Task*>& tasks, double timeoutMilliseconds)
 {
     PROFILE_CPU();
+    ZoneColor(TracyWaitZoneColor);
     for (int32 i = 0; i < tasks.Length(); i++)
     {
         if (tasks[i]->Wait())