Add **GPU profiling support to Tracy integration**

2025-07-30 19:08:45 +02:00
parent 8fcbef863e
commit 5e4d564338
26 changed files with 2716 additions and 46 deletions
--- a/Source/Engine/Engine/Engine.cpp
+++ b/Source/Engine/Engine/Engine.cpp
@@ -247,7 +247,6 @@ int32 Engine::Main(const Char* cmdLine)
        {
            OnDraw();
            Time::OnEndDraw();
-            FrameMark;
        }
    }

@@ -397,6 +396,11 @@ void Engine::OnLateUpdate()

 void Engine::OnDraw()
 {
+#if COMPILE_WITH_PROFILER
+    // Auto-enable GPU events when Tracy got connected
+    if (!ProfilerGPU::EventsEnabled && TracyIsConnected)
+        ProfilerGPU::EventsEnabled = true;
+#endif
    PROFILE_CPU_NAMED("Draw");

    // Begin frame rendering
@@ -411,6 +415,7 @@ void Engine::OnDraw()
    device->Draw();

    // End frame rendering
+    FrameMark;
 #if COMPILE_WITH_PROFILER
    ProfilerGPU::EndFrame();
 #endif
--- a/Source/Engine/Graphics/GPUContext.cpp
+++ b/Source/Engine/Graphics/GPUContext.cpp
@@ -69,6 +69,10 @@ void GPUContext::FrameEnd()
    FlushState();
 }

+void GPUContext::OnPresent()
+{
+}
+
 void GPUContext::BindSR(int32 slot, GPUTexture* t)
 {
    ASSERT_LOW_LAYER(t == nullptr || t->ResidentMipLevels() == 0 || t->IsShaderResource());
--- a/Source/Engine/Graphics/GPUContext.h
+++ b/Source/Engine/Graphics/GPUContext.h
@@ -148,6 +148,11 @@ public:
    /// </summary>
    virtual void FrameEnd();

+    /// <summary>
+    /// Called after performing final swapchain presentation and submitting all GPU commands.
+    /// </summary>
+    virtual void OnPresent();
+
 public:
 #if GPU_ALLOW_PROFILE_EVENTS
    /// <summary>
--- a/Source/Engine/Graphics/GPUDevice.cpp
+++ b/Source/Engine/Graphics/GPUDevice.cpp
@@ -646,6 +646,7 @@ void GPUDevice::DrawEnd()
    const double presentEnd = Platform::GetTimeSeconds();
    ProfilerGPU::OnPresentTime((float)((presentEnd - presentStart) * 1000.0));
 #endif
+    GetMainContext()->OnPresent();

    _wasVSyncUsed = anyVSync;
    _isRendering = false;
--- a/Source/Engine/Graphics/Graphics.Build.cs
+++ b/Source/Engine/Graphics/Graphics.Build.cs
@@ -20,6 +20,12 @@ public abstract class GraphicsDeviceBaseModule : EngineModule
            // Enables GPU diagnostic tools (debug layer etc.)
            options.PublicDefinitions.Add("GPU_ENABLE_DIAGNOSTICS");
        }
+
+        if (Profiler.Use(options) && tracy.GPU && true)
+        {
+            // Enables GPU profiling with Tracy
+            options.PrivateDefinitions.Add("GPU_ENABLE_TRACY");
+        }
    }

    /// <inheritdoc />
--- a/Source/Engine/GraphicsDevice/DirectX/DX11/GPUContextDX11.cpp
+++ b/Source/Engine/GraphicsDevice/DirectX/DX11/GPUContextDX11.cpp
@@ -65,10 +65,17 @@ GPUContextDX11::GPUContextDX11(GPUDeviceDX11* device, ID3D11DeviceContext* conte
    _maxUASlots = GPU_MAX_UA_BINDED;
    if (_device->GetRendererType() != RendererType::DirectX11)
        _maxUASlots = 1;
+
+#if GPU_ENABLE_TRACY
+    _tracyContext = tracy::CreateD3D11Context(device->GetDevice(), context);
+#endif
 }

 GPUContextDX11::~GPUContextDX11()
 {
+#if GPU_ENABLE_TRACY
+    tracy::DestroyD3D11Context(_tracyContext);
+#endif
 #if GPU_ALLOW_PROFILE_EVENTS
    SAFE_RELEASE(_userDefinedAnnotations);
 #endif
@@ -139,16 +146,35 @@ void GPUContextDX11::FrameBegin()
    _context->CSSetSamplers(0, ARRAY_COUNT(samplers), samplers);
 }

+void GPUContextDX11::OnPresent()
+{
+    GPUContext::OnPresent();
+
+#if GPU_ENABLE_TRACY
+    tracy::CollectD3D11Context(_tracyContext);
+#endif
+}
+
 #if GPU_ALLOW_PROFILE_EVENTS

 void GPUContextDX11::EventBegin(const Char* name)
 {
    if (_userDefinedAnnotations)
        _userDefinedAnnotations->BeginEvent(name);
+
+#if GPU_ENABLE_TRACY
+    char buffer[60];
+    int32 bufferSize = StringUtils::Copy(buffer, name, sizeof(buffer));
+    tracy::BeginD3D11ZoneScope(_tracyZone, _tracyContext, buffer, bufferSize);
+#endif
 }

 void GPUContextDX11::EventEnd()
 {
+#if GPU_ENABLE_TRACY
+    tracy::EndD3D11ZoneScope(_tracyZone);
+#endif
+
    if (_userDefinedAnnotations)
        _userDefinedAnnotations->EndEvent();
 }
--- a/Source/Engine/GraphicsDevice/DirectX/DX11/GPUContextDX11.h
+++ b/Source/Engine/GraphicsDevice/DirectX/DX11/GPUContextDX11.h
@@ -6,6 +6,7 @@
 #include "GPUDeviceDX11.h"
 #include "GPUPipelineStateDX11.h"
 #include "../IncludeDirectXHeaders.h"
+#include <ThirdParty/tracy/tracy/TracyD3D11.hpp>

 #if GRAPHICS_API_DIRECTX11

@@ -23,6 +24,10 @@ private:
    ID3D11DeviceContext* _context;
 #if GPU_ALLOW_PROFILE_EVENTS
    ID3DUserDefinedAnnotation* _userDefinedAnnotations;
+#endif
+#if COMPILE_WITH_PROFILER
+    void* _tracyContext;
+    byte _tracyZone[TracyD3D11ZoneSize];
 #endif
    int32 _maxUASlots;

@@ -110,6 +115,7 @@ public:

    // [GPUContext]
    void FrameBegin() override;
+    void OnPresent() override;
 #if GPU_ALLOW_PROFILE_EVENTS
    void EventBegin(const Char* name) override;
    void EventEnd() override;
--- a/Source/Engine/GraphicsDevice/DirectX/DX12/GPUContextDX12.cpp
+++ b/Source/Engine/GraphicsDevice/DirectX/DX12/GPUContextDX12.cpp
@@ -99,10 +99,16 @@ GPUContextDX12::GPUContextDX12(GPUDeviceDX12* device, D3D12_COMMAND_LIST_TYPE ty
 #if GPU_ENABLE_RESOURCE_NAMING
    _commandList->SetName(TEXT("GPUContextDX12::CommandList"));
 #endif
+#if GPU_ENABLE_TRACY
+    _tracyContext = tracy::CreateD3D12Context(device->GetDevice(), _device->GetCommandQueue()->GetCommandQueue());
+#endif
 }

 GPUContextDX12::~GPUContextDX12()
 {
+#if GPU_ENABLE_TRACY
+    tracy::DestroyD3D12Context(_tracyContext);
+#endif
    DX_SAFE_RELEASE_CHECK(_commandList, 0);
 }

@@ -706,6 +712,15 @@ void GPUContextDX12::FrameEnd()
    FrameFenceValues[0] = Execute(false);
 }

+void GPUContextDX12::OnPresent()
+{
+    GPUContext::OnPresent();
+
+#if GPU_ENABLE_TRACY
+    tracy::CollectD3D12Context(_tracyContext);
+#endif
+}
+
 #if GPU_ALLOW_PROFILE_EVENTS

 void GPUContextDX12::EventBegin(const Char* name)
@@ -713,10 +728,22 @@ void GPUContextDX12::EventBegin(const Char* name)
 #if USE_PIX
    PIXBeginEvent(_commandList, 0, name);
 #endif
+
+#if GPU_ENABLE_TRACY
+    char buffer[60];
+    int32 bufferSize = StringUtils::Copy(buffer, name, sizeof(buffer));
+    auto& zone = _tracyZones.AddOne();
+    tracy::BeginD3D12ZoneScope(zone.Data, _tracyContext, _commandList, buffer, bufferSize);
+#endif
 }

 void GPUContextDX12::EventEnd()
 {
+#if GPU_ENABLE_TRACY
+    tracy::EndD3D12ZoneScope(_tracyZones.Last().Data);
+    _tracyZones.RemoveLast();
+#endif
+
 #if USE_PIX
    PIXEndEvent(_commandList);
 #endif
--- a/Source/Engine/GraphicsDevice/DirectX/DX12/GPUContextDX12.h
+++ b/Source/Engine/GraphicsDevice/DirectX/DX12/GPUContextDX12.h
@@ -6,6 +6,7 @@
 #include "IShaderResourceDX12.h"
 #include "DescriptorHeapDX12.h"
 #include "../IncludeDirectXHeaders.h"
+#include <ThirdParty/tracy/tracy/TracyD3D12.hpp>

 #if GRAPHICS_API_DIRECTX12

@@ -71,6 +72,12 @@ private:
    GPUConstantBufferDX12* _cbHandles[GPU_MAX_CB_BINDED];
    GPUSamplerDX12* _samplers[GPU_MAX_SAMPLER_BINDED - GPU_STATIC_SAMPLERS_COUNT];

+#if COMPILE_WITH_PROFILER
+    void* _tracyContext;
+    struct TracyZone { byte Data[TracyD3D12ZoneSize]; };
+    Array<TracyZone, InlinedAllocation<32>> _tracyZones;
+#endif
+
 public:

    GPUContextDX12(GPUDeviceDX12* device, D3D12_COMMAND_LIST_TYPE type);
@@ -154,6 +161,7 @@ public:
    // [GPUContext]
    void FrameBegin() override;
    void FrameEnd() override;
+    void OnPresent() override;
 #if GPU_ALLOW_PROFILE_EVENTS
    void EventBegin(const Char* name) override;
    void EventEnd() override;
--- a/Source/Engine/GraphicsDevice/Vulkan/CmdBufferVulkan.cpp
+++ b/Source/Engine/GraphicsDevice/Vulkan/CmdBufferVulkan.cpp
@@ -49,10 +49,19 @@ void CmdBufferVulkan::End()
    PROFILE_CPU();
    ASSERT(IsOutsideRenderPass());

-#if GPU_ALLOW_PROFILE_EVENTS && VK_EXT_debug_utils
+#if GPU_ALLOW_PROFILE_EVENTS
    // End remaining events
    while (_eventsBegin--)
-        vkCmdEndDebugUtilsLabelEXT(GetHandle());
+    {
+#if VK_EXT_debug_utils
+        if (vkCmdEndDebugUtilsLabelEXT)
+            vkCmdEndDebugUtilsLabelEXT(GetHandle());
+#endif
+#if GPU_ENABLE_TRACY
+        tracy::EndVkZoneScope(_tracyZones.Last().Data);
+        _tracyZones.RemoveLast();
+#endif
+    }
 #endif

    VALIDATE_VULKAN_RESULT(vkEndCommandBuffer(GetHandle()));
@@ -85,39 +94,43 @@ void CmdBufferVulkan::EndRenderPass()

 #if GPU_ALLOW_PROFILE_EVENTS

-void CmdBufferVulkan::BeginEvent(const Char* name)
+void CmdBufferVulkan::BeginEvent(const Char* name, void* tracyContext)
 {
-#if VK_EXT_debug_utils
-    if (!vkCmdBeginDebugUtilsLabelEXT)
-        return;
-
    _eventsBegin++;

-    // Convert to ANSI
-    char buffer[101];
-    int32 i = 0;
-    while (i < 100 && name[i])
-    {
-        buffer[i] = (char)name[i];
-        i++;
-    }
-    buffer[i] = 0;
+    char buffer[60];
+    int32 bufferSize = StringUtils::Copy(buffer, name, sizeof(buffer));

-    VkDebugUtilsLabelEXT label;
-    RenderToolsVulkan::ZeroStruct(label, VK_STRUCTURE_TYPE_DEBUG_UTILS_LABEL_EXT);
-    label.pLabelName = buffer;
-    vkCmdBeginDebugUtilsLabelEXT(GetHandle(), &label);
+#if GPU_ENABLE_TRACY
+    auto& zone = _tracyZones.AddOne();
+    tracy::BeginVkZoneScope(zone.Data, tracyContext, GetHandle(), buffer, bufferSize);
+#endif
+
+#if VK_EXT_debug_utils
+    if (vkCmdBeginDebugUtilsLabelEXT)
+    {
+        VkDebugUtilsLabelEXT label;
+        RenderToolsVulkan::ZeroStruct(label, VK_STRUCTURE_TYPE_DEBUG_UTILS_LABEL_EXT);
+        label.pLabelName = buffer;
+        vkCmdBeginDebugUtilsLabelEXT(GetHandle(), &label);
+    }
 #endif
 }

 void CmdBufferVulkan::EndEvent()
 {
-#if VK_EXT_debug_utils
-    if (_eventsBegin == 0 || !vkCmdEndDebugUtilsLabelEXT)
+    if (_eventsBegin == 0)
        return;
    _eventsBegin--;

-    vkCmdEndDebugUtilsLabelEXT(GetHandle());
+#if VK_EXT_debug_utils
+    if (vkCmdEndDebugUtilsLabelEXT)
+        vkCmdEndDebugUtilsLabelEXT(GetHandle());
+#endif
+
+#if GPU_ENABLE_TRACY
+    tracy::EndVkZoneScope(_tracyZones.Last().Data);
+    _tracyZones.RemoveLast();
 #endif
 }

--- a/Source/Engine/GraphicsDevice/Vulkan/CmdBufferVulkan.h
+++ b/Source/Engine/GraphicsDevice/Vulkan/CmdBufferVulkan.h
@@ -5,6 +5,7 @@
 #include "GPUDeviceVulkan.h"
 #include "Engine/Core/Types/BaseTypes.h"
 #include "Engine/Core/Collections/Array.h"
+#include <ThirdParty/tracy/tracy/TracyVulkan.hpp>

 #if GRAPHICS_API_VULKAN

@@ -42,6 +43,8 @@ private:
    FenceVulkan* _fence;
 #if GPU_ALLOW_PROFILE_EVENTS
    int32 _eventsBegin = 0;
+    struct TracyZone { byte Data[TracyVulkanZoneSize]; };
+    Array<TracyZone, InlinedAllocation<32>> _tracyZones;
 #endif

    // The latest value when command buffer was submitted.
@@ -129,7 +132,7 @@ public:
    }

 #if GPU_ALLOW_PROFILE_EVENTS
-    void BeginEvent(const Char* name);
+    void BeginEvent(const Char* name, void* tracyContext);
    void EndEvent();
 #endif

--- a/Source/Engine/GraphicsDevice/Vulkan/GPUContextVulkan.cpp
+++ b/Source/Engine/GraphicsDevice/Vulkan/GPUContextVulkan.cpp
@@ -4,6 +4,7 @@

 #include "GPUContextVulkan.h"
 #include "CmdBufferVulkan.h"
+#include "GPUAdapterVulkan.h"
 #include "RenderToolsVulkan.h"
 #include "Engine/Core/Math/Color.h"
 #include "Engine/Core/Math/Rectangle.h"
@@ -15,6 +16,7 @@
 #include "Engine/Profiler/RenderStats.h"
 #include "GPUShaderProgramVulkan.h"
 #include "GPUTextureVulkan.h"
+#include "QueueVulkan.h"
 #include "Engine/Graphics/PixelFormatExtensions.h"
 #include "Engine/Debug/Exceptions/NotImplementedException.h"

@@ -107,10 +109,37 @@ GPUContextVulkan::GPUContextVulkan(GPUDeviceVulkan* device, QueueVulkan* queue)
    _handlesSizes[(int32)SpirvShaderResourceBindingType::SRV] = GPU_MAX_SR_BINDED;
    _handlesSizes[(int32)SpirvShaderResourceBindingType::UAV] = GPU_MAX_UA_BINDED;
 #endif
+
+#if GPU_ENABLE_TRACY
+#if VK_EXT_calibrated_timestamps && VK_EXT_host_query_reset
+    // Use calibrated timestamps extension
+    if (vkResetQueryPoolEXT && vkGetCalibratedTimestampsEXT)
+    {
+        _tracyContext = tracy::CreateVkContext(_device->Adapter->Gpu, _device->Device, vkResetQueryPoolEXT, vkGetPhysicalDeviceCalibrateableTimeDomainsEXT, vkGetCalibratedTimestampsEXT);
+    }
+    else
+#endif
+    {
+        // Use immediate command buffer for Tracy initialization
+        VkCommandBufferAllocateInfo cmdInfo;
+        RenderToolsVulkan::ZeroStruct(cmdInfo, VK_STRUCTURE_TYPE_COMMAND_BUFFER_ALLOCATE_INFO);
+        cmdInfo.level = VK_COMMAND_BUFFER_LEVEL_PRIMARY;
+        cmdInfo.commandPool = _cmdBufferManager->GetHandle();
+        cmdInfo.commandBufferCount = 1;
+        VkCommandBuffer tracyCmdBuffer;
+        vkAllocateCommandBuffers(_device->Device, &cmdInfo, &tracyCmdBuffer);
+        _tracyContext = tracy::CreateVkContext(_device->Adapter->Gpu, _device->Device, _queue->GetHandle(), tracyCmdBuffer, vkGetPhysicalDeviceCalibrateableTimeDomainsEXT, vkGetCalibratedTimestampsEXT);
+        vkQueueWaitIdle(_queue->GetHandle());
+        vkFreeCommandBuffers(_device->Device, _cmdBufferManager->GetHandle(), 1, &tracyCmdBuffer);
+    }
+#endif
 }

 GPUContextVulkan::~GPUContextVulkan()
 {
+#if GPU_ENABLE_TRACY
+    tracy::DestroyVkContext(_tracyContext);
+#endif
    for (int32 i = 0; i < _descriptorPools.Count(); i++)
    {
        _descriptorPools[i].ClearDelete();
@@ -679,15 +708,9 @@ void GPUContextVulkan::OnDrawCall()
    // Bind descriptors sets to the graphics pipeline
    if (pipelineState->HasDescriptorsPerStageMask)
    {
-        vkCmdBindDescriptorSets(
-            cmdBuffer->GetHandle(),
-            VK_PIPELINE_BIND_POINT_GRAPHICS,
-            pipelineState->GetLayout()->Handle,
-            0,
-            pipelineState->DescriptorSetHandles.Count(),
-            pipelineState->DescriptorSetHandles.Get(),
-            pipelineState->DynamicOffsets.Count(),
-            pipelineState->DynamicOffsets.Get());
+        auto& descriptorSets = pipelineState->DescriptorSetHandles;
+        auto& dynamicOffsets = pipelineState->DynamicOffsets;
+        vkCmdBindDescriptorSets(cmdBuffer->GetHandle(), VK_PIPELINE_BIND_POINT_GRAPHICS, pipelineState->GetLayout()->Handle, 0, descriptorSets.Count(), descriptorSets.Get(), dynamicOffsets.Count(), dynamicOffsets.Get());
    }

    _rtDirtyFlag = false;
@@ -748,6 +771,11 @@ void GPUContextVulkan::FrameEnd()
    // Execute any queued layout transitions that weren't already handled by the render pass
    FlushBarriers();

+#if GPU_ENABLE_TRACY
+    if (cmdBuffer)
+        tracy::CollectVkContext(_tracyContext, cmdBuffer->GetHandle());
+#endif
+
    // Base
    GPUContext::FrameEnd();
 }
@@ -757,7 +785,12 @@ void GPUContextVulkan::FrameEnd()
 void GPUContextVulkan::EventBegin(const Char* name)
 {
    const auto cmdBuffer = _cmdBufferManager->GetCmdBuffer();
-    cmdBuffer->BeginEvent(name);
+#if COMPILE_WITH_PROFILER
+    void* tracyContext = _tracyContext;
+#else
+    void* tracyContext = nullptr;
+#endif
+    cmdBuffer->BeginEvent(name, tracyContext);
 }

 void GPUContextVulkan::EventEnd()
--- a/Source/Engine/GraphicsDevice/Vulkan/GPUContextVulkan.h
+++ b/Source/Engine/GraphicsDevice/Vulkan/GPUContextVulkan.h
@@ -94,6 +94,9 @@ private:
 #if ENABLE_ASSERTION
    uint32 _handlesSizes[(int32)SpirvShaderResourceBindingType::MAX];
 #endif
+#if COMPILE_WITH_PROFILER
+    void* _tracyContext;
+#endif

    typedef Array<DescriptorPoolVulkan*> DescriptorPoolArray;
    Dictionary<uint32, DescriptorPoolArray> _descriptorPools;
--- a/Source/Engine/GraphicsDevice/Vulkan/GPUDeviceVulkan.Layers.cpp
+++ b/Source/Engine/GraphicsDevice/Vulkan/GPUDeviceVulkan.Layers.cpp
@@ -62,6 +62,10 @@ static const char* GDeviceExtensions[] =
 #endif
 #if VK_KHR_sampler_mirror_clamp_to_edge
    VK_KHR_SAMPLER_MIRROR_CLAMP_TO_EDGE_EXTENSION_NAME,
+#endif
+#if GPU_ENABLE_TRACY && VK_EXT_calibrated_timestamps && VK_EXT_host_query_reset
+    VK_EXT_CALIBRATED_TIMESTAMPS_EXTENSION_NAME,
+    VK_EXT_HOST_QUERY_RESET_EXTENSION_NAME,
 #endif
    nullptr
 };
--- a/Source/Engine/GraphicsDevice/Vulkan/GPUSwapChainVulkan.cpp
+++ b/Source/Engine/GraphicsDevice/Vulkan/GPUSwapChainVulkan.cpp
@@ -424,6 +424,7 @@ GPUSwapChainVulkan::Status GPUSwapChainVulkan::Present(QueueVulkan* presentQueue
 {
    if (_currentImageIndex == -1)
        return Status::Ok;
+    PROFILE_CPU_NAMED("vkQueuePresentKHR");

    VkPresentInfoKHR presentInfo;
    RenderToolsVulkan::ZeroStruct(presentInfo, VK_STRUCTURE_TYPE_PRESENT_INFO_KHR);
@@ -506,7 +507,7 @@ int32 GPUSwapChainVulkan::TryPresent(Function<int32(GPUSwapChainVulkan*, void*)>

 int32 GPUSwapChainVulkan::AcquireNextImage(SemaphoreVulkan*& outSemaphore)
 {
-    PROFILE_CPU();
+    PROFILE_CPU_NAMED("vkAcquireNextImageKHR");
    ASSERT(_swapChain && _backBuffers.HasItems());

    uint32 imageIndex = _currentImageIndex;
@@ -514,13 +515,7 @@ int32 GPUSwapChainVulkan::AcquireNextImage(SemaphoreVulkan*& outSemaphore)
    _semaphoreIndex = (_semaphoreIndex + 1) % _backBuffers.Count();
    const auto semaphore = _backBuffers[_semaphoreIndex].ImageAcquiredSemaphore;

-    const VkResult result = vkAcquireNextImageKHR(
-        _device->Device,
-        _swapChain,
-        UINT64_MAX,
-        semaphore->GetHandle(),
-        VK_NULL_HANDLE,
-        &imageIndex);
+    const VkResult result = vkAcquireNextImageKHR(_device->Device, _swapChain, UINT64_MAX, semaphore->GetHandle(), VK_NULL_HANDLE, &imageIndex);
    if (result == VK_ERROR_OUT_OF_DATE_KHR)
    {
        _semaphoreIndex = prevSemaphoreIndex;
--- a/Source/Engine/Platform/Base/StringUtilsBase.cpp
+++ b/Source/Engine/Platform/Base/StringUtilsBase.cpp
@@ -16,6 +16,18 @@ constexpr char DirectorySeparatorChar = '\\';
 constexpr char AltDirectorySeparatorChar = '/';
 constexpr char VolumeSeparatorChar = ':';

+int32 StringUtils::Copy(char* dst, const Char* src, int32 count)
+{
+    int32 i = 0;
+    while (i < count && src[i])
+    {
+        dst[i] = (char)src[i];
+        i++;
+    }
+    dst[i] = 0;
+    return i;
+}
+
 const Char* StringUtils::FindIgnoreCase(const Char* str, const Char* toFind)
 {
    if (toFind == nullptr || str == nullptr)
--- a/Source/Engine/Platform/StringUtils.h
+++ b/Source/Engine/Platform/StringUtils.h
@@ -125,6 +125,9 @@ public:
    // Copies the string (count is maximum amount of characters to copy).
    static Char* Copy(Char* dst, const Char* src, int32 count);

+    // Copies the string (count is maximum amount of characters to copy). Returns amount of copied elements (excluding null terminator character).
+    static int32 Copy(char* dst, const Char* src, int32 count);
+
    // Finds specific sub-string in the input string. Returns the first found position in the input string or nulll if failed.
    static const Char* Find(const Char* str, const Char* toFind);

--- a/Source/ThirdParty/tracy/client/TracyProfiler.cpp
+++ b/Source/ThirdParty/tracy/client/TracyProfiler.cpp
@@ -1400,6 +1400,7 @@ TRACY_API LuaZoneState& GetLuaZoneState() { return s_luaZoneState; }
 #  endif
 #endif

+TRACY_API bool IsConnected() { return GetProfiler().IsConnected(); }
 TRACY_API bool ProfilerAvailable() { return s_instance != nullptr; }
 TRACY_API bool ProfilerAllocatorAvailable() { return !RpThreadShutdown; }

--- a/Source/ThirdParty/tracy/common/TracyColor.hpp
+++ b/Source/ThirdParty/tracy/common/TracyColor.hpp
@@ -0,0 +1,690 @@
+#ifndef __TRACYCOLOR_HPP__
+#define __TRACYCOLOR_HPP__
+
+namespace tracy
+{
+struct Color
+{
+enum ColorType
+{
+    Snow = 0xfffafa,
+    GhostWhite = 0xf8f8ff,
+    WhiteSmoke = 0xf5f5f5,
+    Gainsboro = 0xdcdcdc,
+    FloralWhite = 0xfffaf0,
+    OldLace = 0xfdf5e6,
+    Linen = 0xfaf0e6,
+    AntiqueWhite = 0xfaebd7,
+    PapayaWhip = 0xffefd5,
+    BlanchedAlmond = 0xffebcd,
+    Bisque = 0xffe4c4,
+    PeachPuff = 0xffdab9,
+    NavajoWhite = 0xffdead,
+    Moccasin = 0xffe4b5,
+    Cornsilk = 0xfff8dc,
+    Ivory = 0xfffff0,
+    LemonChiffon = 0xfffacd,
+    Seashell = 0xfff5ee,
+    Honeydew = 0xf0fff0,
+    MintCream = 0xf5fffa,
+    Azure = 0xf0ffff,
+    AliceBlue = 0xf0f8ff,
+    Lavender = 0xe6e6fa,
+    LavenderBlush = 0xfff0f5,
+    MistyRose = 0xffe4e1,
+    White = 0xffffff,
+    Black = 0x000000,
+    DarkSlateGray = 0x2f4f4f,
+    DarkSlateGrey = 0x2f4f4f,
+    DimGray = 0x696969,
+    DimGrey = 0x696969,
+    SlateGray = 0x708090,
+    SlateGrey = 0x708090,
+    LightSlateGray = 0x778899,
+    LightSlateGrey = 0x778899,
+    Gray = 0xbebebe,
+    Grey = 0xbebebe,
+    X11Gray = 0xbebebe,
+    X11Grey = 0xbebebe,
+    WebGray = 0x808080,
+    WebGrey = 0x808080,
+    LightGrey = 0xd3d3d3,
+    LightGray = 0xd3d3d3,
+    MidnightBlue = 0x191970,
+    Navy = 0x000080,
+    NavyBlue = 0x000080,
+    CornflowerBlue = 0x6495ed,
+    DarkSlateBlue = 0x483d8b,
+    SlateBlue = 0x6a5acd,
+    MediumSlateBlue = 0x7b68ee,
+    LightSlateBlue = 0x8470ff,
+    MediumBlue = 0x0000cd,
+    RoyalBlue = 0x4169e1,
+    Blue = 0x0000ff,
+    DodgerBlue = 0x1e90ff,
+    DeepSkyBlue = 0x00bfff,
+    SkyBlue = 0x87ceeb,
+    LightSkyBlue = 0x87cefa,
+    SteelBlue = 0x4682b4,
+    LightSteelBlue = 0xb0c4de,
+    LightBlue = 0xadd8e6,
+    PowderBlue = 0xb0e0e6,
+    PaleTurquoise = 0xafeeee,
+    DarkTurquoise = 0x00ced1,
+    MediumTurquoise = 0x48d1cc,
+    Turquoise = 0x40e0d0,
+    Cyan = 0x00ffff,
+    Aqua = 0x00ffff,
+    LightCyan = 0xe0ffff,
+    CadetBlue = 0x5f9ea0,
+    MediumAquamarine = 0x66cdaa,
+    Aquamarine = 0x7fffd4,
+    DarkGreen = 0x006400,
+    DarkOliveGreen = 0x556b2f,
+    DarkSeaGreen = 0x8fbc8f,
+    SeaGreen = 0x2e8b57,
+    MediumSeaGreen = 0x3cb371,
+    LightSeaGreen = 0x20b2aa,
+    PaleGreen = 0x98fb98,
+    SpringGreen = 0x00ff7f,
+    LawnGreen = 0x7cfc00,
+    Green = 0x00ff00,
+    Lime = 0x00ff00,
+    X11Green = 0x00ff00,
+    WebGreen = 0x008000,
+    Chartreuse = 0x7fff00,
+    MediumSpringGreen = 0x00fa9a,
+    GreenYellow = 0xadff2f,
+    LimeGreen = 0x32cd32,
+    YellowGreen = 0x9acd32,
+    ForestGreen = 0x228b22,
+    OliveDrab = 0x6b8e23,
+    DarkKhaki = 0xbdb76b,
+    Khaki = 0xf0e68c,
+    PaleGoldenrod = 0xeee8aa,
+    LightGoldenrodYellow = 0xfafad2,
+    LightYellow = 0xffffe0,
+    Yellow = 0xffff00,
+    Gold = 0xffd700,
+    LightGoldenrod = 0xeedd82,
+    Goldenrod = 0xdaa520,
+    DarkGoldenrod = 0xb8860b,
+    RosyBrown = 0xbc8f8f,
+    IndianRed = 0xcd5c5c,
+    SaddleBrown = 0x8b4513,
+    Sienna = 0xa0522d,
+    Peru = 0xcd853f,
+    Burlywood = 0xdeb887,
+    Beige = 0xf5f5dc,
+    Wheat = 0xf5deb3,
+    SandyBrown = 0xf4a460,
+    Tan = 0xd2b48c,
+    Chocolate = 0xd2691e,
+    Firebrick = 0xb22222,
+    Brown = 0xa52a2a,
+    DarkSalmon = 0xe9967a,
+    Salmon = 0xfa8072,
+    LightSalmon = 0xffa07a,
+    Orange = 0xffa500,
+    DarkOrange = 0xff8c00,
+    Coral = 0xff7f50,
+    LightCoral = 0xf08080,
+    Tomato = 0xff6347,
+    OrangeRed = 0xff4500,
+    Red = 0xff0000,
+    HotPink = 0xff69b4,
+    DeepPink = 0xff1493,
+    Pink = 0xffc0cb,
+    LightPink = 0xffb6c1,
+    PaleVioletRed = 0xdb7093,
+    Maroon = 0xb03060,
+    X11Maroon = 0xb03060,
+    WebMaroon = 0x800000,
+    MediumVioletRed = 0xc71585,
+    VioletRed = 0xd02090,
+    Magenta = 0xff00ff,
+    Fuchsia = 0xff00ff,
+    Violet = 0xee82ee,
+    Plum = 0xdda0dd,
+    Orchid = 0xda70d6,
+    MediumOrchid = 0xba55d3,
+    DarkOrchid = 0x9932cc,
+    DarkViolet = 0x9400d3,
+    BlueViolet = 0x8a2be2,
+    Purple = 0xa020f0,
+    X11Purple = 0xa020f0,
+    WebPurple = 0x800080,
+    MediumPurple = 0x9370db,
+    Thistle = 0xd8bfd8,
+    Snow1 = 0xfffafa,
+    Snow2 = 0xeee9e9,
+    Snow3 = 0xcdc9c9,
+    Snow4 = 0x8b8989,
+    Seashell1 = 0xfff5ee,
+    Seashell2 = 0xeee5de,
+    Seashell3 = 0xcdc5bf,
+    Seashell4 = 0x8b8682,
+    AntiqueWhite1 = 0xffefdb,
+    AntiqueWhite2 = 0xeedfcc,
+    AntiqueWhite3 = 0xcdc0b0,
+    AntiqueWhite4 = 0x8b8378,
+    Bisque1 = 0xffe4c4,
+    Bisque2 = 0xeed5b7,
+    Bisque3 = 0xcdb79e,
+    Bisque4 = 0x8b7d6b,
+    PeachPuff1 = 0xffdab9,
+    PeachPuff2 = 0xeecbad,
+    PeachPuff3 = 0xcdaf95,
+    PeachPuff4 = 0x8b7765,
+    NavajoWhite1 = 0xffdead,
+    NavajoWhite2 = 0xeecfa1,
+    NavajoWhite3 = 0xcdb38b,
+    NavajoWhite4 = 0x8b795e,
+    LemonChiffon1 = 0xfffacd,
+    LemonChiffon2 = 0xeee9bf,
+    LemonChiffon3 = 0xcdc9a5,
+    LemonChiffon4 = 0x8b8970,
+    Cornsilk1 = 0xfff8dc,
+    Cornsilk2 = 0xeee8cd,
+    Cornsilk3 = 0xcdc8b1,
+    Cornsilk4 = 0x8b8878,
+    Ivory1 = 0xfffff0,
+    Ivory2 = 0xeeeee0,
+    Ivory3 = 0xcdcdc1,
+    Ivory4 = 0x8b8b83,
+    Honeydew1 = 0xf0fff0,
+    Honeydew2 = 0xe0eee0,
+    Honeydew3 = 0xc1cdc1,
+    Honeydew4 = 0x838b83,
+    LavenderBlush1 = 0xfff0f5,
+    LavenderBlush2 = 0xeee0e5,
+    LavenderBlush3 = 0xcdc1c5,
+    LavenderBlush4 = 0x8b8386,
+    MistyRose1 = 0xffe4e1,
+    MistyRose2 = 0xeed5d2,
+    MistyRose3 = 0xcdb7b5,
+    MistyRose4 = 0x8b7d7b,
+    Azure1 = 0xf0ffff,
+    Azure2 = 0xe0eeee,
+    Azure3 = 0xc1cdcd,
+    Azure4 = 0x838b8b,
+    SlateBlue1 = 0x836fff,
+    SlateBlue2 = 0x7a67ee,
+    SlateBlue3 = 0x6959cd,
+    SlateBlue4 = 0x473c8b,
+    RoyalBlue1 = 0x4876ff,
+    RoyalBlue2 = 0x436eee,
+    RoyalBlue3 = 0x3a5fcd,
+    RoyalBlue4 = 0x27408b,
+    Blue1 = 0x0000ff,
+    Blue2 = 0x0000ee,
+    Blue3 = 0x0000cd,
+    Blue4 = 0x00008b,
+    DodgerBlue1 = 0x1e90ff,
+    DodgerBlue2 = 0x1c86ee,
+    DodgerBlue3 = 0x1874cd,
+    DodgerBlue4 = 0x104e8b,
+    SteelBlue1 = 0x63b8ff,
+    SteelBlue2 = 0x5cacee,
+    SteelBlue3 = 0x4f94cd,
+    SteelBlue4 = 0x36648b,
+    DeepSkyBlue1 = 0x00bfff,
+    DeepSkyBlue2 = 0x00b2ee,
+    DeepSkyBlue3 = 0x009acd,
+    DeepSkyBlue4 = 0x00688b,
+    SkyBlue1 = 0x87ceff,
+    SkyBlue2 = 0x7ec0ee,
+    SkyBlue3 = 0x6ca6cd,
+    SkyBlue4 = 0x4a708b,
+    LightSkyBlue1 = 0xb0e2ff,
+    LightSkyBlue2 = 0xa4d3ee,
+    LightSkyBlue3 = 0x8db6cd,
+    LightSkyBlue4 = 0x607b8b,
+    SlateGray1 = 0xc6e2ff,
+    SlateGray2 = 0xb9d3ee,
+    SlateGray3 = 0x9fb6cd,
+    SlateGray4 = 0x6c7b8b,
+    LightSteelBlue1 = 0xcae1ff,
+    LightSteelBlue2 = 0xbcd2ee,
+    LightSteelBlue3 = 0xa2b5cd,
+    LightSteelBlue4 = 0x6e7b8b,
+    LightBlue1 = 0xbfefff,
+    LightBlue2 = 0xb2dfee,
+    LightBlue3 = 0x9ac0cd,
+    LightBlue4 = 0x68838b,
+    LightCyan1 = 0xe0ffff,
+    LightCyan2 = 0xd1eeee,
+    LightCyan3 = 0xb4cdcd,
+    LightCyan4 = 0x7a8b8b,
+    PaleTurquoise1 = 0xbbffff,
+    PaleTurquoise2 = 0xaeeeee,
+    PaleTurquoise3 = 0x96cdcd,
+    PaleTurquoise4 = 0x668b8b,
+    CadetBlue1 = 0x98f5ff,
+    CadetBlue2 = 0x8ee5ee,
+    CadetBlue3 = 0x7ac5cd,
+    CadetBlue4 = 0x53868b,
+    Turquoise1 = 0x00f5ff,
+    Turquoise2 = 0x00e5ee,
+    Turquoise3 = 0x00c5cd,
+    Turquoise4 = 0x00868b,
+    Cyan1 = 0x00ffff,
+    Cyan2 = 0x00eeee,
+    Cyan3 = 0x00cdcd,
+    Cyan4 = 0x008b8b,
+    DarkSlateGray1 = 0x97ffff,
+    DarkSlateGray2 = 0x8deeee,
+    DarkSlateGray3 = 0x79cdcd,
+    DarkSlateGray4 = 0x528b8b,
+    Aquamarine1 = 0x7fffd4,
+    Aquamarine2 = 0x76eec6,
+    Aquamarine3 = 0x66cdaa,
+    Aquamarine4 = 0x458b74,
+    DarkSeaGreen1 = 0xc1ffc1,
+    DarkSeaGreen2 = 0xb4eeb4,
+    DarkSeaGreen3 = 0x9bcd9b,
+    DarkSeaGreen4 = 0x698b69,
+    SeaGreen1 = 0x54ff9f,
+    SeaGreen2 = 0x4eee94,
+    SeaGreen3 = 0x43cd80,
+    SeaGreen4 = 0x2e8b57,
+    PaleGreen1 = 0x9aff9a,
+    PaleGreen2 = 0x90ee90,
+    PaleGreen3 = 0x7ccd7c,
+    PaleGreen4 = 0x548b54,
+    SpringGreen1 = 0x00ff7f,
+    SpringGreen2 = 0x00ee76,
+    SpringGreen3 = 0x00cd66,
+    SpringGreen4 = 0x008b45,
+    Green1 = 0x00ff00,
+    Green2 = 0x00ee00,
+    Green3 = 0x00cd00,
+    Green4 = 0x008b00,
+    Chartreuse1 = 0x7fff00,
+    Chartreuse2 = 0x76ee00,
+    Chartreuse3 = 0x66cd00,
+    Chartreuse4 = 0x458b00,
+    OliveDrab1 = 0xc0ff3e,
+    OliveDrab2 = 0xb3ee3a,
+    OliveDrab3 = 0x9acd32,
+    OliveDrab4 = 0x698b22,
+    DarkOliveGreen1 = 0xcaff70,
+    DarkOliveGreen2 = 0xbcee68,
+    DarkOliveGreen3 = 0xa2cd5a,
+    DarkOliveGreen4 = 0x6e8b3d,
+    Khaki1 = 0xfff68f,
+    Khaki2 = 0xeee685,
+    Khaki3 = 0xcdc673,
+    Khaki4 = 0x8b864e,
+    LightGoldenrod1 = 0xffec8b,
+    LightGoldenrod2 = 0xeedc82,
+    LightGoldenrod3 = 0xcdbe70,
+    LightGoldenrod4 = 0x8b814c,
+    LightYellow1 = 0xffffe0,
+    LightYellow2 = 0xeeeed1,
+    LightYellow3 = 0xcdcdb4,
+    LightYellow4 = 0x8b8b7a,
+    Yellow1 = 0xffff00,
+    Yellow2 = 0xeeee00,
+    Yellow3 = 0xcdcd00,
+    Yellow4 = 0x8b8b00,
+    Gold1 = 0xffd700,
+    Gold2 = 0xeec900,
+    Gold3 = 0xcdad00,
+    Gold4 = 0x8b7500,
+    Goldenrod1 = 0xffc125,
+    Goldenrod2 = 0xeeb422,
+    Goldenrod3 = 0xcd9b1d,
+    Goldenrod4 = 0x8b6914,
+    DarkGoldenrod1 = 0xffb90f,
+    DarkGoldenrod2 = 0xeead0e,
+    DarkGoldenrod3 = 0xcd950c,
+    DarkGoldenrod4 = 0x8b6508,
+    RosyBrown1 = 0xffc1c1,
+    RosyBrown2 = 0xeeb4b4,
+    RosyBrown3 = 0xcd9b9b,
+    RosyBrown4 = 0x8b6969,
+    IndianRed1 = 0xff6a6a,
+    IndianRed2 = 0xee6363,
+    IndianRed3 = 0xcd5555,
+    IndianRed4 = 0x8b3a3a,
+    Sienna1 = 0xff8247,
+    Sienna2 = 0xee7942,
+    Sienna3 = 0xcd6839,
+    Sienna4 = 0x8b4726,
+    Burlywood1 = 0xffd39b,
+    Burlywood2 = 0xeec591,
+    Burlywood3 = 0xcdaa7d,
+    Burlywood4 = 0x8b7355,
+    Wheat1 = 0xffe7ba,
+    Wheat2 = 0xeed8ae,
+    Wheat3 = 0xcdba96,
+    Wheat4 = 0x8b7e66,
+    Tan1 = 0xffa54f,
+    Tan2 = 0xee9a49,
+    Tan3 = 0xcd853f,
+    Tan4 = 0x8b5a2b,
+    Chocolate1 = 0xff7f24,
+    Chocolate2 = 0xee7621,
+    Chocolate3 = 0xcd661d,
+    Chocolate4 = 0x8b4513,
+    Firebrick1 = 0xff3030,
+    Firebrick2 = 0xee2c2c,
+    Firebrick3 = 0xcd2626,
+    Firebrick4 = 0x8b1a1a,
+    Brown1 = 0xff4040,
+    Brown2 = 0xee3b3b,
+    Brown3 = 0xcd3333,
+    Brown4 = 0x8b2323,
+    Salmon1 = 0xff8c69,
+    Salmon2 = 0xee8262,
+    Salmon3 = 0xcd7054,
+    Salmon4 = 0x8b4c39,
+    LightSalmon1 = 0xffa07a,
+    LightSalmon2 = 0xee9572,
+    LightSalmon3 = 0xcd8162,
+    LightSalmon4 = 0x8b5742,
+    Orange1 = 0xffa500,
+    Orange2 = 0xee9a00,
+    Orange3 = 0xcd8500,
+    Orange4 = 0x8b5a00,
+    DarkOrange1 = 0xff7f00,
+    DarkOrange2 = 0xee7600,
+    DarkOrange3 = 0xcd6600,
+    DarkOrange4 = 0x8b4500,
+    Coral1 = 0xff7256,
+    Coral2 = 0xee6a50,
+    Coral3 = 0xcd5b45,
+    Coral4 = 0x8b3e2f,
+    Tomato1 = 0xff6347,
+    Tomato2 = 0xee5c42,
+    Tomato3 = 0xcd4f39,
+    Tomato4 = 0x8b3626,
+    OrangeRed1 = 0xff4500,
+    OrangeRed2 = 0xee4000,
+    OrangeRed3 = 0xcd3700,
+    OrangeRed4 = 0x8b2500,
+    Red1 = 0xff0000,
+    Red2 = 0xee0000,
+    Red3 = 0xcd0000,
+    Red4 = 0x8b0000,
+    DeepPink1 = 0xff1493,
+    DeepPink2 = 0xee1289,
+    DeepPink3 = 0xcd1076,
+    DeepPink4 = 0x8b0a50,
+    HotPink1 = 0xff6eb4,
+    HotPink2 = 0xee6aa7,
+    HotPink3 = 0xcd6090,
+    HotPink4 = 0x8b3a62,
+    Pink1 = 0xffb5c5,
+    Pink2 = 0xeea9b8,
+    Pink3 = 0xcd919e,
+    Pink4 = 0x8b636c,
+    LightPink1 = 0xffaeb9,
+    LightPink2 = 0xeea2ad,
+    LightPink3 = 0xcd8c95,
+    LightPink4 = 0x8b5f65,
+    PaleVioletRed1 = 0xff82ab,
+    PaleVioletRed2 = 0xee799f,
+    PaleVioletRed3 = 0xcd6889,
+    PaleVioletRed4 = 0x8b475d,
+    Maroon1 = 0xff34b3,
+    Maroon2 = 0xee30a7,
+    Maroon3 = 0xcd2990,
+    Maroon4 = 0x8b1c62,
+    VioletRed1 = 0xff3e96,
+    VioletRed2 = 0xee3a8c,
+    VioletRed3 = 0xcd3278,
+    VioletRed4 = 0x8b2252,
+    Magenta1 = 0xff00ff,
+    Magenta2 = 0xee00ee,
+    Magenta3 = 0xcd00cd,
+    Magenta4 = 0x8b008b,
+    Orchid1 = 0xff83fa,
+    Orchid2 = 0xee7ae9,
+    Orchid3 = 0xcd69c9,
+    Orchid4 = 0x8b4789,
+    Plum1 = 0xffbbff,
+    Plum2 = 0xeeaeee,
+    Plum3 = 0xcd96cd,
+    Plum4 = 0x8b668b,
+    MediumOrchid1 = 0xe066ff,
+    MediumOrchid2 = 0xd15fee,
+    MediumOrchid3 = 0xb452cd,
+    MediumOrchid4 = 0x7a378b,
+    DarkOrchid1 = 0xbf3eff,
+    DarkOrchid2 = 0xb23aee,
+    DarkOrchid3 = 0x9a32cd,
+    DarkOrchid4 = 0x68228b,
+    Purple1 = 0x9b30ff,
+    Purple2 = 0x912cee,
+    Purple3 = 0x7d26cd,
+    Purple4 = 0x551a8b,
+    MediumPurple1 = 0xab82ff,
+    MediumPurple2 = 0x9f79ee,
+    MediumPurple3 = 0x8968cd,
+    MediumPurple4 = 0x5d478b,
+    Thistle1 = 0xffe1ff,
+    Thistle2 = 0xeed2ee,
+    Thistle3 = 0xcdb5cd,
+    Thistle4 = 0x8b7b8b,
+    Gray0 = 0x000000,
+    Grey0 = 0x000000,
+    Gray1 = 0x030303,
+    Grey1 = 0x030303,
+    Gray2 = 0x050505,
+    Grey2 = 0x050505,
+    Gray3 = 0x080808,
+    Grey3 = 0x080808,
+    Gray4 = 0x0a0a0a,
+    Grey4 = 0x0a0a0a,
+    Gray5 = 0x0d0d0d,
+    Grey5 = 0x0d0d0d,
+    Gray6 = 0x0f0f0f,
+    Grey6 = 0x0f0f0f,
+    Gray7 = 0x121212,
+    Grey7 = 0x121212,
+    Gray8 = 0x141414,
+    Grey8 = 0x141414,
+    Gray9 = 0x171717,
+    Grey9 = 0x171717,
+    Gray10 = 0x1a1a1a,
+    Grey10 = 0x1a1a1a,
+    Gray11 = 0x1c1c1c,
+    Grey11 = 0x1c1c1c,
+    Gray12 = 0x1f1f1f,
+    Grey12 = 0x1f1f1f,
+    Gray13 = 0x212121,
+    Grey13 = 0x212121,
+    Gray14 = 0x242424,
+    Grey14 = 0x242424,
+    Gray15 = 0x262626,
+    Grey15 = 0x262626,
+    Gray16 = 0x292929,
+    Grey16 = 0x292929,
+    Gray17 = 0x2b2b2b,
+    Grey17 = 0x2b2b2b,
+    Gray18 = 0x2e2e2e,
+    Grey18 = 0x2e2e2e,
+    Gray19 = 0x303030,
+    Grey19 = 0x303030,
+    Gray20 = 0x333333,
+    Grey20 = 0x333333,
+    Gray21 = 0x363636,
+    Grey21 = 0x363636,
+    Gray22 = 0x383838,
+    Grey22 = 0x383838,
+    Gray23 = 0x3b3b3b,
+    Grey23 = 0x3b3b3b,
+    Gray24 = 0x3d3d3d,
+    Grey24 = 0x3d3d3d,
+    Gray25 = 0x404040,
+    Grey25 = 0x404040,
+    Gray26 = 0x424242,
+    Grey26 = 0x424242,
+    Gray27 = 0x454545,
+    Grey27 = 0x454545,
+    Gray28 = 0x474747,
+    Grey28 = 0x474747,
+    Gray29 = 0x4a4a4a,
+    Grey29 = 0x4a4a4a,
+    Gray30 = 0x4d4d4d,
+    Grey30 = 0x4d4d4d,
+    Gray31 = 0x4f4f4f,
+    Grey31 = 0x4f4f4f,
+    Gray32 = 0x525252,
+    Grey32 = 0x525252,
+    Gray33 = 0x545454,
+    Grey33 = 0x545454,
+    Gray34 = 0x575757,
+    Grey34 = 0x575757,
+    Gray35 = 0x595959,
+    Grey35 = 0x595959,
+    Gray36 = 0x5c5c5c,
+    Grey36 = 0x5c5c5c,
+    Gray37 = 0x5e5e5e,
+    Grey37 = 0x5e5e5e,
+    Gray38 = 0x616161,
+    Grey38 = 0x616161,
+    Gray39 = 0x636363,
+    Grey39 = 0x636363,
+    Gray40 = 0x666666,
+    Grey40 = 0x666666,
+    Gray41 = 0x696969,
+    Grey41 = 0x696969,
+    Gray42 = 0x6b6b6b,
+    Grey42 = 0x6b6b6b,
+    Gray43 = 0x6e6e6e,
+    Grey43 = 0x6e6e6e,
+    Gray44 = 0x707070,
+    Grey44 = 0x707070,
+    Gray45 = 0x737373,
+    Grey45 = 0x737373,
+    Gray46 = 0x757575,
+    Grey46 = 0x757575,
+    Gray47 = 0x787878,
+    Grey47 = 0x787878,
+    Gray48 = 0x7a7a7a,
+    Grey48 = 0x7a7a7a,
+    Gray49 = 0x7d7d7d,
+    Grey49 = 0x7d7d7d,
+    Gray50 = 0x7f7f7f,
+    Grey50 = 0x7f7f7f,
+    Gray51 = 0x828282,
+    Grey51 = 0x828282,
+    Gray52 = 0x858585,
+    Grey52 = 0x858585,
+    Gray53 = 0x878787,
+    Grey53 = 0x878787,
+    Gray54 = 0x8a8a8a,
+    Grey54 = 0x8a8a8a,
+    Gray55 = 0x8c8c8c,
+    Grey55 = 0x8c8c8c,
+    Gray56 = 0x8f8f8f,
+    Grey56 = 0x8f8f8f,
+    Gray57 = 0x919191,
+    Grey57 = 0x919191,
+    Gray58 = 0x949494,
+    Grey58 = 0x949494,
+    Gray59 = 0x969696,
+    Grey59 = 0x969696,
+    Gray60 = 0x999999,
+    Grey60 = 0x999999,
+    Gray61 = 0x9c9c9c,
+    Grey61 = 0x9c9c9c,
+    Gray62 = 0x9e9e9e,
+    Grey62 = 0x9e9e9e,
+    Gray63 = 0xa1a1a1,
+    Grey63 = 0xa1a1a1,
+    Gray64 = 0xa3a3a3,
+    Grey64 = 0xa3a3a3,
+    Gray65 = 0xa6a6a6,
+    Grey65 = 0xa6a6a6,
+    Gray66 = 0xa8a8a8,
+    Grey66 = 0xa8a8a8,
+    Gray67 = 0xababab,
+    Grey67 = 0xababab,
+    Gray68 = 0xadadad,
+    Grey68 = 0xadadad,
+    Gray69 = 0xb0b0b0,
+    Grey69 = 0xb0b0b0,
+    Gray70 = 0xb3b3b3,
+    Grey70 = 0xb3b3b3,
+    Gray71 = 0xb5b5b5,
+    Grey71 = 0xb5b5b5,
+    Gray72 = 0xb8b8b8,
+    Grey72 = 0xb8b8b8,
+    Gray73 = 0xbababa,
+    Grey73 = 0xbababa,
+    Gray74 = 0xbdbdbd,
+    Grey74 = 0xbdbdbd,
+    Gray75 = 0xbfbfbf,
+    Grey75 = 0xbfbfbf,
+    Gray76 = 0xc2c2c2,
+    Grey76 = 0xc2c2c2,
+    Gray77 = 0xc4c4c4,
+    Grey77 = 0xc4c4c4,
+    Gray78 = 0xc7c7c7,
+    Grey78 = 0xc7c7c7,
+    Gray79 = 0xc9c9c9,
+    Grey79 = 0xc9c9c9,
+    Gray80 = 0xcccccc,
+    Grey80 = 0xcccccc,
+    Gray81 = 0xcfcfcf,
+    Grey81 = 0xcfcfcf,
+    Gray82 = 0xd1d1d1,
+    Grey82 = 0xd1d1d1,
+    Gray83 = 0xd4d4d4,
+    Grey83 = 0xd4d4d4,
+    Gray84 = 0xd6d6d6,
+    Grey84 = 0xd6d6d6,
+    Gray85 = 0xd9d9d9,
+    Grey85 = 0xd9d9d9,
+    Gray86 = 0xdbdbdb,
+    Grey86 = 0xdbdbdb,
+    Gray87 = 0xdedede,
+    Grey87 = 0xdedede,
+    Gray88 = 0xe0e0e0,
+    Grey88 = 0xe0e0e0,
+    Gray89 = 0xe3e3e3,
+    Grey89 = 0xe3e3e3,
+    Gray90 = 0xe5e5e5,
+    Grey90 = 0xe5e5e5,
+    Gray91 = 0xe8e8e8,
+    Grey91 = 0xe8e8e8,
+    Gray92 = 0xebebeb,
+    Grey92 = 0xebebeb,
+    Gray93 = 0xededed,
+    Grey93 = 0xededed,
+    Gray94 = 0xf0f0f0,
+    Grey94 = 0xf0f0f0,
+    Gray95 = 0xf2f2f2,
+    Grey95 = 0xf2f2f2,
+    Gray96 = 0xf5f5f5,
+    Grey96 = 0xf5f5f5,
+    Gray97 = 0xf7f7f7,
+    Grey97 = 0xf7f7f7,
+    Gray98 = 0xfafafa,
+    Grey98 = 0xfafafa,
+    Gray99 = 0xfcfcfc,
+    Grey99 = 0xfcfcfc,
+    Gray100 = 0xffffff,
+    Grey100 = 0xffffff,
+    DarkGrey = 0xa9a9a9,
+    DarkGray = 0xa9a9a9,
+    DarkBlue = 0x00008b,
+    DarkCyan = 0x008b8b,
+    DarkMagenta = 0x8b008b,
+    DarkRed = 0x8b0000,
+    LightGreen = 0x90ee90,
+    Crimson = 0xdc143c,
+    Indigo = 0x4b0082,
+    Olive = 0x808000,
+    RebeccaPurple = 0x663399,
+    Silver = 0xc0c0c0,
+    Teal = 0x008080,
+};
+};
+}
+
+#endif
--- a/Source/ThirdParty/tracy/common/TracySystem.cpp
+++ b/Source/ThirdParty/tracy/common/TracySystem.cpp
@@ -351,3 +351,20 @@ TRACY_API void ___tracy_set_thread_name( const char* name ) { tracy::SetThreadNa
 #ifdef __cplusplus
 }
 #endif
+
+// Inset graphics integration (within Tracy module)
+#define TRACY_GPU_IMPL 1
+#if TRACY_GPU_D3D11
+#include <ThirdParty/tracy/tracy/TracyD3D11.hpp>
+static_assert(sizeof(tracy::D3D11ZoneScope) <= TracyD3D11ZoneSize, "Invalid zone size");
+#endif
+#if TRACY_GPU_D3D12
+#include <ThirdParty/tracy/tracy/TracyD3D12.hpp>
+static_assert(sizeof(tracy::D3D12ZoneScope) <= TracyD3D12ZoneSize, "Invalid zone size");
+#endif
+#if TRACY_GPU_VULKAN
+#define GRAPHICS_API_VULKAN 1
+#include "Engine/GraphicsDevice/Vulkan/IncludeVulkanHeaders.h"
+#include <ThirdParty/tracy/tracy/TracyVulkan.hpp>
+static_assert(sizeof(tracy::VkCtxScope) <= TracyVulkanZoneSize, "Invalid zone size");
+#endif
--- a/Source/ThirdParty/tracy/tracy.Build.cs
+++ b/Source/ThirdParty/tracy/tracy.Build.cs
@@ -15,6 +15,11 @@ public class tracy : ThirdPartyModule
    /// </summary>
    public static bool OnDemand = true;

+    /// <summary>
+    /// Enables GPU profiling.
+    /// </summary>
+    public static bool GPU = true;
+
    /// <inheritdoc />
    public override void Init()
    {
@@ -56,8 +61,25 @@ public class tracy : ThirdPartyModule
            options.PrivateDefinitions.Add("TRACY_USE_MALLOC");
            options.PrivateDefinitions.Add("TRACY_ONLY_IPV4");
            options.PrivateDefinitions.Add("TRACY_NO_PIPE");
+            options.PrivateDefinitions.Add("TRACY_NO_CODE_TRANSFER");
            break;
        }
+
+        if (GPU)
+        {
+            // Ask Graphics module which graphics backends are active
+            var graphics = new Graphics();
+            graphics.FilePath = FilePath;
+            graphics.FolderPath = FolderPath;
+            var graphicsOptions = (BuildOptions)options.Clone();
+            graphics.Setup(graphicsOptions);
+            if (graphicsOptions.PrivateDependencies.Contains("GraphicsDeviceDX11"))
+                options.PrivateDefinitions.Add("TRACY_GPU_D3D11");
+            if (graphicsOptions.PrivateDependencies.Contains("GraphicsDeviceDX12"))
+                options.PrivateDefinitions.Add("TRACY_GPU_D3D12");
+            if (graphicsOptions.PrivateDependencies.Contains("GraphicsDeviceVulkan"))
+                options.PrivateDefinitions.Add("TRACY_GPU_VULKAN");
+        }
    }

    /// <inheritdoc />
--- a/Source/ThirdParty/tracy/tracy/Tracy.hpp
+++ b/Source/ThirdParty/tracy/tracy/Tracy.hpp
@@ -119,6 +119,8 @@

 namespace tracy
 {
+TRACY_API bool IsConnected();
+
 class TRACY_API Profiler
 {
 public:
@@ -143,7 +145,6 @@ public:
    static void MemAllocCallstackNamed( const void* ptr, size_t size, int depth, bool secure, const char* name );
    static void MemFreeCallstackNamed( const void* ptr, int depth, bool secure, const char* name );
    static void SendCallstack( int depth );
-    static void ParameterRegister( ParameterCallback cb );
    static void ParameterRegister( ParameterCallback cb, void* data );
    static void ParameterSetup( uint32_t idx, const char* name, bool isBool, int32_t val );
 };
@@ -255,7 +256,7 @@ public:
 #define TracySourceCallbackRegister( cb, data ) tracy::Profiler::SourceCallbackRegister( cb, data )
 #define TracyParameterRegister( cb, data ) tracy::Profiler::ParameterRegister( cb, data )
 #define TracyParameterSetup( idx, name, isBool, val ) tracy::Profiler::ParameterSetup( idx, name, isBool, val )
-#define TracyIsConnected tracy::GetProfiler().IsConnected()
+#define TracyIsConnected tracy::IsConnected()
 #define TracySetProgramName( name ) tracy::GetProfiler().SetProgramName( name );

 #ifdef TRACY_FIBERS
--- a/Source/ThirdParty/tracy/tracy/TracyD3D11.hpp
+++ b/Source/ThirdParty/tracy/tracy/TracyD3D11.hpp
@@ -0,0 +1,456 @@
+#ifndef __TRACYD3D11_HPP__
+#define __TRACYD3D11_HPP__
+
+#define TracyD3D11ZoneSize 16
+
+#ifndef TRACY_ENABLE
+
+#define TracyD3D11Context(device,queue) nullptr
+#define TracyD3D11Destroy(ctx)
+#define TracyD3D11ContextName(ctx, name, size)
+
+#define TracyD3D11NewFrame(ctx)
+
+#define TracyD3D11Zone(ctx, name)
+#define TracyD3D11ZoneC(ctx, name, color)
+#define TracyD3D11NamedZone(ctx, varname, name, active)
+#define TracyD3D11NamedZoneC(ctx, varname, name, color, active)
+#define TracyD3D11ZoneTransient(ctx, varname, name, active)
+
+#define TracyD3D11ZoneS(ctx, name, depth)
+#define TracyD3D11ZoneCS(ctx, name, color, depth)
+#define TracyD3D11NamedZoneS(ctx, varname, name, depth, active)
+#define TracyD3D11NamedZoneCS(ctx, varname, name, color, depth, active)
+#define TracyD3D11ZoneTransientS(ctx, varname, name, depth, active)
+
+#define TracyD3D11Collect(ctx)
+
+namespace tracy
+{
+class D3D11ZoneScope {};
+}
+
+using TracyD3D11Ctx = void*;
+
+#elif TRACY_GPU_IMPL
+
+#include <atomic>
+#include <assert.h>
+#include <stdlib.h>
+
+#include "../client/TracyProfiler.hpp"
+#include "../client/TracyCallstack.hpp"
+#include "../common/TracyYield.hpp"
+#include "../common/TracyColor.hpp"
+
+#include <d3d11.h>
+
+#define TRACY_CALLSTACK 0
+#define TracyMessageLC( txt, color ) tracy::Profiler::MessageColor( txt, color, TRACY_CALLSTACK )
+#define TracyD3D11Panic(msg, ...) do { assert(false && "TracyD3D11: " msg); TracyMessageLC("TracyD3D11: " msg, tracy::Color::Red4); __VA_ARGS__; } while(false);
+
+namespace tracy
+{
+
+class D3D11Ctx
+{
+    friend class D3D11ZoneScope;
+
+    static constexpr uint32_t MaxQueries = 64 * 1024;
+
+    enum CollectMode { POLL, BLOCK };
+
+public:
+    tracy_force_inline D3D11Ctx( ID3D11Device* device, ID3D11DeviceContext* devicectx )
+    {
+        // TODO: consider calling ID3D11Device::GetImmediateContext() instead of passing it as an argument
+        m_device = device;
+        device->AddRef();
+        m_immediateDevCtx = devicectx;
+        devicectx->AddRef();
+
+        {
+            D3D11_QUERY_DESC desc = { };
+            desc.Query = D3D11_QUERY_TIMESTAMP_DISJOINT;
+            if (FAILED(m_device->CreateQuery(&desc, &m_disjointQuery)))
+            {
+                TracyD3D11Panic("unable to create disjoint timestamp query.", return);
+            }
+        }
+
+        for (ID3D11Query*& query : m_queries)
+        {
+            D3D11_QUERY_DESC desc = { };
+            desc.Query = D3D11_QUERY_TIMESTAMP;
+            if (FAILED(m_device->CreateQuery(&desc, &query)))
+            {
+                TracyD3D11Panic("unable to create timestamp query.", return);
+            }
+        }
+
+        // Calibrate CPU and GPU timestamps
+        int64_t tcpu = 0;
+        int64_t tgpu = 0;
+        for (int attempts = 0; attempts < 50; attempts++)
+        {
+            m_immediateDevCtx->Begin(m_disjointQuery);
+            m_immediateDevCtx->End(m_queries[0]);
+            m_immediateDevCtx->End(m_disjointQuery);
+
+            int64_t tcpu0 = Profiler::GetTime();
+            WaitForQuery(m_disjointQuery);
+            // NOTE: one would expect that by waiting for the enclosing disjoint query to finish,
+            // all timestamp queries within would also be readily available, but that does not
+            // seem to be the case here... See https://github.com/wolfpld/tracy/issues/947
+            WaitForQuery(m_queries[0]);
+            int64_t tcpu1 = Profiler::GetTime();
+
+            D3D11_QUERY_DATA_TIMESTAMP_DISJOINT disjoint = { };
+            if (m_immediateDevCtx->GetData(m_disjointQuery, &disjoint, sizeof(disjoint), 0) != S_OK)
+            {
+                TracyMessageLC("TracyD3D11: unable to query GPU timestamp; retrying...", tracy::Color::Tomato);
+                continue;
+            }
+
+            if (disjoint.Disjoint)
+                continue;
+
+            UINT64 timestamp = 0;
+            if (m_immediateDevCtx->GetData(m_queries[0], &timestamp, sizeof(timestamp), 0) != S_OK)
+                continue;   // this should never happen (we waited for the query to finish above)
+
+            tcpu = tcpu0 + (tcpu1 - tcpu0) * 1 / 2;
+            tgpu = timestamp * (1000000000 / disjoint.Frequency);
+            break;
+        }
+
+        // ready to roll
+        m_contextId = GetGpuCtxCounter().fetch_add(1);
+        m_immediateDevCtx->Begin(m_disjointQuery);
+        m_previousCheckpoint = m_nextCheckpoint = 0;
+
+        auto* item = Profiler::QueueSerial();
+        MemWrite( &item->hdr.type, QueueType::GpuNewContext );
+        MemWrite( &item->gpuNewContext.cpuTime, tcpu );
+        MemWrite( &item->gpuNewContext.gpuTime, tgpu );
+        MemWrite( &item->gpuNewContext.thread, uint32_t(0) );   // #TODO: why not GetThreadHandle()?
+        MemWrite( &item->gpuNewContext.period, 1.0f );
+        MemWrite( &item->gpuNewContext.context, m_contextId);
+        MemWrite( &item->gpuNewContext.flags, uint8_t(0) );
+        MemWrite( &item->gpuNewContext.type, GpuContextType::Direct3D11 );
+
+#ifdef TRACY_ON_DEMAND
+        GetProfiler().DeferItem( *item );
+#endif
+
+        Profiler::QueueSerialFinish();
+    }
+
+    tracy_force_inline ~D3D11Ctx()
+    {
+        // collect all pending timestamps before destroying everything
+        do
+        {
+            Collect(BLOCK);
+        } while (m_previousCheckpoint != m_queryCounter);
+
+        for (ID3D11Query* query : m_queries)
+        {
+            query->Release();
+        }
+        m_immediateDevCtx->End(m_disjointQuery);
+        m_disjointQuery->Release();
+        m_immediateDevCtx->Release();
+        m_device->Release();
+    }
+
+    tracy_force_inline void Name( const char* name, uint16_t len )
+    {
+        auto ptr = (char*)tracy_malloc( len );
+        memcpy( ptr, name, len );
+
+        auto item = Profiler::QueueSerial();
+        MemWrite( &item->hdr.type, QueueType::GpuContextName );
+        MemWrite( &item->gpuContextNameFat.context, m_contextId );
+        MemWrite( &item->gpuContextNameFat.ptr, (uint64_t)ptr );
+        MemWrite( &item->gpuContextNameFat.size, len );
+#ifdef TRACY_ON_DEMAND
+        GetProfiler().DeferItem( *item );
+#endif
+        Profiler::QueueSerialFinish();
+    }
+
+    void Collect(CollectMode mode = POLL)
+    {
+#ifdef TRACY_ON_DEMAND
+        if( !GetProfiler().IsConnected() )
+        {
+            m_previousCheckpoint = m_nextCheckpoint = m_queryCounter;
+            return;
+        }
+#endif
+
+        if (m_previousCheckpoint == m_nextCheckpoint)
+        {
+            uintptr_t nextCheckpoint = m_queryCounter;
+            if (nextCheckpoint == m_nextCheckpoint)
+            {
+                return;
+            }
+            m_nextCheckpoint = nextCheckpoint;
+            m_immediateDevCtx->End(m_disjointQuery);
+        }
+
+        if (mode == CollectMode::BLOCK)
+        {
+            WaitForQuery(m_disjointQuery);
+        }
+
+        D3D11_QUERY_DATA_TIMESTAMP_DISJOINT disjoint = { };
+        if (m_immediateDevCtx->GetData(m_disjointQuery, &disjoint, sizeof(disjoint), D3D11_ASYNC_GETDATA_DONOTFLUSH) != S_OK)
+        {
+            return;
+        }
+
+        if (disjoint.Disjoint == TRUE)
+        {
+            m_previousCheckpoint = m_nextCheckpoint;
+            TracyD3D11Panic("disjoint timestamps detected; dropping.");
+            return;
+        }
+
+        auto begin = m_previousCheckpoint;
+        auto end = m_nextCheckpoint;
+        for (auto i = begin; i != end; ++i)
+        {
+            uint32_t k = RingIndex(i);
+            UINT64 timestamp = 0;
+            if (m_immediateDevCtx->GetData(m_queries[k], &timestamp, sizeof(timestamp), 0) != S_OK)
+            {
+                TracyD3D11Panic("timestamp expected to be ready, but it was not!");
+                break;
+            }
+            timestamp *= (1000000000ull / disjoint.Frequency);
+            auto* item = Profiler::QueueSerial();
+            MemWrite(&item->hdr.type, QueueType::GpuTime);
+            MemWrite(&item->gpuTime.gpuTime, static_cast<int64_t>(timestamp));
+            MemWrite(&item->gpuTime.queryId, static_cast<uint16_t>(k));
+            MemWrite(&item->gpuTime.context, m_contextId);
+            Profiler::QueueSerialFinish();
+        }
+
+        // disjoint timestamp queries should only be invoked once per frame or less
+        // https://learn.microsoft.com/en-us/windows/win32/api/d3d11/ne-d3d11-d3d11_query
+        m_immediateDevCtx->Begin(m_disjointQuery);
+        m_previousCheckpoint = m_nextCheckpoint;
+    }
+
+private:
+    tracy_force_inline uint32_t RingIndex(uintptr_t index)
+    {
+        index %= MaxQueries;
+        return static_cast<uint32_t>(index);
+    }
+
+    tracy_force_inline uint32_t RingCount(uintptr_t begin, uintptr_t end)
+    {
+        // wrap-around safe: all unsigned
+        uintptr_t count = end - begin;
+        return static_cast<uint32_t>(count);
+    }
+
+    tracy_force_inline uint32_t NextQueryId()
+    {
+        auto id = m_queryCounter++;
+        if (RingCount(m_previousCheckpoint, id) >= MaxQueries)
+        {
+            TracyD3D11Panic("too many pending timestamp queries.");
+            // #TODO: return some sentinel value; ideally a "hidden" query index
+        }
+        return RingIndex(id);
+    }
+
+    tracy_force_inline ID3D11Query* GetQueryObjectFromId(uint32_t id)
+    {
+        return m_queries[id];
+    }
+
+    tracy_force_inline void WaitForQuery(ID3D11Query* query)
+    {
+        m_immediateDevCtx->Flush();
+        while (m_immediateDevCtx->GetData(query, nullptr, 0, 0) != S_OK)
+            YieldThread();  // busy-wait :-( attempt to reduce power usage with _mm_pause() & friends...
+    }
+
+    tracy_force_inline uint8_t GetContextId() const
+    {
+        return m_contextId;
+    }
+
+    ID3D11Device* m_device = nullptr;
+    ID3D11DeviceContext* m_immediateDevCtx = nullptr;
+
+    ID3D11Query* m_queries[MaxQueries];
+    ID3D11Query* m_disjointQuery = nullptr;
+
+    uint8_t m_contextId = 255;  // NOTE: apparently, 255 means invalid id; is this documented anywhere?
+
+    uintptr_t m_queryCounter = 0;
+
+    uintptr_t m_previousCheckpoint = 0;
+    uintptr_t m_nextCheckpoint = 0;
+};
+
+class D3D11ZoneScope
+{
+public:
+    tracy_force_inline D3D11ZoneScope( D3D11Ctx* ctx, const SourceLocationData* srcloc, bool active )
+        : D3D11ZoneScope(ctx, active)
+    {
+        if( !m_active ) return;
+
+        auto* item = Profiler::QueueSerial();
+        WriteQueueItem(item, QueueType::GpuZoneBeginSerial, reinterpret_cast<uint64_t>(srcloc));
+    }
+
+    tracy_force_inline D3D11ZoneScope( D3D11Ctx* ctx, const SourceLocationData* srcloc, int32_t depth, bool active )
+        : D3D11ZoneScope(ctx, active)
+    {
+        if( !m_active ) return;
+
+        if( depth > 0 && has_callstack() )
+        {
+            auto* item = Profiler::QueueSerialCallstack(Callstack(depth));
+            WriteQueueItem(item, QueueType::GpuZoneBeginCallstackSerial, reinterpret_cast<uint64_t>(srcloc));
+        }
+        else
+        {
+            auto* item = Profiler::QueueSerial();
+            WriteQueueItem(item, QueueType::GpuZoneBeginSerial, reinterpret_cast<uint64_t>(srcloc));
+        }
+    }
+
+    tracy_force_inline D3D11ZoneScope(D3D11Ctx* ctx, uint32_t line, const char* source, size_t sourceSz, const char* function, size_t functionSz, const char* name, size_t nameSz, bool active)
+        : D3D11ZoneScope(ctx, active)
+    {
+        if( !m_active ) return;
+
+        const auto sourceLocation = Profiler::AllocSourceLocation(line, source, sourceSz, function, functionSz, name, nameSz);
+
+        auto* item = Profiler::QueueSerial();
+        WriteQueueItem(item, QueueType::GpuZoneBeginAllocSrcLocSerial, sourceLocation);
+    }
+
+    tracy_force_inline D3D11ZoneScope(D3D11Ctx* ctx, uint32_t line, const char* source, size_t sourceSz, const char* function, size_t functionSz, const char* name, size_t nameSz, int32_t depth, bool active)
+        : D3D11ZoneScope(ctx, active)
+    {
+        if( !m_active ) return;
+
+        const auto sourceLocation = Profiler::AllocSourceLocation(line, source, sourceSz, function, functionSz, name, nameSz);
+
+        if ( depth > 0 && has_callstack() )
+        {
+            auto* item = Profiler::QueueSerialCallstack(Callstack(depth));
+            WriteQueueItem(item, QueueType::GpuZoneBeginAllocSrcLocCallstackSerial, sourceLocation);
+        }
+        else
+        {
+            auto* item = Profiler::QueueSerial();
+            WriteQueueItem(item, QueueType::GpuZoneBeginAllocSrcLocSerial, sourceLocation);
+        }
+    }
+
+    tracy_force_inline ~D3D11ZoneScope()
+    {
+        if( !m_active ) return;
+
+        const auto queryId = m_ctx->NextQueryId();
+        m_ctx->m_immediateDevCtx->End(m_ctx->GetQueryObjectFromId(queryId));
+
+        auto* item = Profiler::QueueSerial();
+        MemWrite( &item->hdr.type, QueueType::GpuZoneEndSerial );
+        MemWrite( &item->gpuZoneEnd.cpuTime, Profiler::GetTime() );
+        MemWrite( &item->gpuZoneEnd.thread, GetThreadHandle() );
+        MemWrite( &item->gpuZoneEnd.queryId, uint16_t( queryId ) );
+        MemWrite( &item->gpuZoneEnd.context, m_ctx->GetContextId() );
+        Profiler::QueueSerialFinish();
+    }
+
+private:
+    tracy_force_inline D3D11ZoneScope( D3D11Ctx* ctx, bool active )
+#ifdef TRACY_ON_DEMAND
+        : m_active( active && GetProfiler().IsConnected() )
+#else
+        : m_active( active )
+#endif
+    {
+        if( !m_active ) return;
+        m_ctx = ctx;
+    }
+
+    void WriteQueueItem(tracy::QueueItem* item, tracy::QueueType queueItemType, uint64_t sourceLocation)
+    {
+        const auto queryId = m_ctx->NextQueryId();
+        m_ctx->m_immediateDevCtx->End(m_ctx->GetQueryObjectFromId(queryId));
+
+        MemWrite( &item->hdr.type, queueItemType);
+        MemWrite( &item->gpuZoneBegin.cpuTime, Profiler::GetTime() );
+        MemWrite( &item->gpuZoneBegin.srcloc, sourceLocation );
+        MemWrite( &item->gpuZoneBegin.thread, GetThreadHandle() );
+        MemWrite( &item->gpuZoneBegin.queryId, uint16_t( queryId ) );
+        MemWrite( &item->gpuZoneBegin.context, m_ctx->GetContextId() );
+        Profiler::QueueSerialFinish();
+    }
+
+    const bool m_active;
+
+    D3D11Ctx* m_ctx;
+};
+
+void* CreateD3D11Context( ID3D11Device* device, ID3D11DeviceContext* devicectx )
+{
+    auto ctx = (D3D11Ctx*)tracy_malloc( sizeof( D3D11Ctx ) );
+    new(ctx) D3D11Ctx( device, devicectx );
+    ctx->Name("D3D11", 5);
+    return ctx;
+}
+
+void CollectD3D11Context(void* ctx)
+{
+    ((D3D11Ctx*)ctx)->Collect();
+}
+
+void DestroyD3D11Context(void* ctx )
+{
+    ((D3D11Ctx*)ctx)->~D3D11Ctx();
+    tracy_free( ctx );
+}
+
+void BeginD3D11ZoneScope(void* zone, void* ctx, const char* name, size_t nameLen)
+{
+    new(zone) tracy::D3D11ZoneScope{ (tracy::D3D11Ctx*)ctx, 0, 0, 0, 0, 0, name, nameLen, true };
+}
+
+void EndD3D11ZoneScope(void* zone)
+{
+    ((tracy::D3D11ZoneScope*)zone)->~D3D11ZoneScope();
+}
+}
+
+#undef TracyD3D11Panic
+
+#else
+
+// Forward declarations to be used in engine
+namespace tracy
+{
+    extern void* CreateD3D11Context(ID3D11Device* device, ID3D11DeviceContext* devicectx);
+    extern void DestroyD3D11Context(void* ctx);
+    extern void CollectD3D11Context(void* ctx);
+    extern void BeginD3D11ZoneScope(void* zone, void* ctx, const char* name, size_t nameLen);
+    extern void EndD3D11ZoneScope(void* zone);
+}
+
+#endif
+
+#endif
--- a/Source/ThirdParty/tracy/tracy/TracyD3D12.hpp
+++ b/Source/ThirdParty/tracy/tracy/TracyD3D12.hpp
@@ -0,0 +1,529 @@
+#ifndef __TRACYD3D12_HPP__
+#define __TRACYD3D12_HPP__
+
+#define TracyD3D12ZoneSize 32
+
+#ifndef TRACY_ENABLE
+
+#define TracyD3D12Context(device, queue) nullptr
+#define TracyD3D12Destroy(ctx)
+#define TracyD3D12ContextName(ctx, name, size)
+
+#define TracyD3D12NewFrame(ctx)
+
+#define TracyD3D12Zone(ctx, cmdList, name)
+#define TracyD3D12ZoneC(ctx, cmdList, name, color)
+#define TracyD3D12NamedZone(ctx, varname, cmdList, name, active)
+#define TracyD3D12NamedZoneC(ctx, varname, cmdList, name, color, active)
+#define TracyD3D12ZoneTransient(ctx, varname, cmdList, name, active)
+
+#define TracyD3D12ZoneS(ctx, cmdList, name, depth)
+#define TracyD3D12ZoneCS(ctx, cmdList, name, color, depth)
+#define TracyD3D12NamedZoneS(ctx, varname, cmdList, name, depth, active)
+#define TracyD3D12NamedZoneCS(ctx, varname, cmdList, name, color, depth, active)
+#define TracyD3D12ZoneTransientS(ctx, varname, cmdList, name, depth, active)
+
+#define TracyD3D12Collect(ctx)
+
+namespace tracy
+{
+    class D3D12ZoneScope {};
+}
+
+using TracyD3D12Ctx = void*;
+
+#elif TRACY_GPU_IMPL
+
+#include "../client/TracyProfiler.hpp"
+#include "../client/TracyProfiler.hpp"
+#include "../client/TracyCallstack.hpp"
+#include "../common/TracyColor.hpp"
+
+#include <cstdlib>
+#include <cassert>
+#include <d3d12.h>
+#include <dxgi.h>
+#include <queue>
+
+#define TRACY_CALLSTACK 0
+#define TracyMessageLC( txt, color ) tracy::Profiler::MessageColor( txt, color, TRACY_CALLSTACK )
+#define TracyD3D12Panic(msg, ...) do { assert(false && "TracyD3D12: " msg); TracyMessageLC("TracyD3D12: " msg, tracy::Color::Red4); __VA_ARGS__; } while(false);
+
+namespace tracy
+{
+
+    struct D3D12QueryPayload
+    {
+        uint32_t m_queryIdStart = 0;
+        uint32_t m_queryCount = 0;
+    };
+
+    // Command queue context.
+    class D3D12QueueCtx
+    {
+        friend class D3D12ZoneScope;
+
+        ID3D12Device* m_device = nullptr;
+        ID3D12CommandQueue* m_queue = nullptr;
+        uint8_t m_contextId = 255;  // TODO: apparently, 255 means "invalid id"; is this documented somewhere?
+        ID3D12QueryHeap* m_queryHeap = nullptr;
+        ID3D12Resource* m_readbackBuffer = nullptr;
+
+        // In-progress payload.
+        uint32_t m_queryLimit = 0;
+        std::atomic<uint32_t> m_queryCounter = 0;
+        uint32_t m_previousQueryCounter = 0;
+
+        uint32_t m_activePayload = 0;
+        ID3D12Fence* m_payloadFence = nullptr;
+        std::queue<D3D12QueryPayload> m_payloadQueue;
+
+        UINT64 m_prevCalibrationTicksCPU = 0;
+
+        void RecalibrateClocks()
+        {
+            UINT64 cpuTimestamp;
+            UINT64 gpuTimestamp;
+            if (FAILED(m_queue->GetClockCalibration(&gpuTimestamp, &cpuTimestamp)))
+            {
+                TracyD3D12Panic("failed to obtain queue clock calibration counters.", return);
+            }
+
+            int64_t cpuDeltaTicks = cpuTimestamp - m_prevCalibrationTicksCPU;
+            if (cpuDeltaTicks > 0)
+            {
+                static const int64_t nanosecodsPerTick = int64_t(1000000000) / GetFrequencyQpc();
+                int64_t cpuDeltaNS = cpuDeltaTicks * nanosecodsPerTick;
+                // Save the device cpu timestamp, not the Tracy profiler timestamp:
+                m_prevCalibrationTicksCPU = cpuTimestamp;
+
+                cpuTimestamp = Profiler::GetTime();
+
+                auto* item = Profiler::QueueSerial();
+                MemWrite(&item->hdr.type, QueueType::GpuCalibration);
+                MemWrite(&item->gpuCalibration.gpuTime, gpuTimestamp);
+                MemWrite(&item->gpuCalibration.cpuTime, cpuTimestamp);
+                MemWrite(&item->gpuCalibration.cpuDelta, cpuDeltaNS);
+                MemWrite(&item->gpuCalibration.context, GetId());
+                SubmitQueueItem(item);
+            }
+        }
+
+        tracy_force_inline void SubmitQueueItem(tracy::QueueItem* item)
+        {
+#ifdef TRACY_ON_DEMAND
+            GetProfiler().DeferItem(*item);
+#endif
+            Profiler::QueueSerialFinish();
+        }
+
+    public:
+        D3D12QueueCtx(ID3D12Device* device, ID3D12CommandQueue* queue)
+            : m_device(device)
+            , m_queue(queue)
+        {
+            // Verify we support timestamp queries on this queue.
+
+            if (queue->GetDesc().Type == D3D12_COMMAND_LIST_TYPE_COPY)
+            {
+                D3D12_FEATURE_DATA_D3D12_OPTIONS3 featureData{};
+
+                HRESULT hr = device->CheckFeatureSupport(D3D12_FEATURE_D3D12_OPTIONS3, &featureData, sizeof(featureData));
+                if (FAILED(hr) || (featureData.CopyQueueTimestampQueriesSupported == FALSE))
+                {
+                    TracyD3D12Panic("Platform does not support profiling of copy queues.", return);
+                }
+            }
+
+            static constexpr uint32_t MaxQueries = 64 * 1024;  // Must be even, because queries are (begin, end) pairs
+            m_queryLimit = MaxQueries;
+
+            D3D12_QUERY_HEAP_DESC heapDesc{};
+            heapDesc.Type = queue->GetDesc().Type == D3D12_COMMAND_LIST_TYPE_COPY ? D3D12_QUERY_HEAP_TYPE_COPY_QUEUE_TIMESTAMP : D3D12_QUERY_HEAP_TYPE_TIMESTAMP;
+            heapDesc.Count = m_queryLimit;
+            heapDesc.NodeMask = 0;  // #TODO: Support multiple adapters.
+
+            while (FAILED(device->CreateQueryHeap(&heapDesc, IID_PPV_ARGS(&m_queryHeap))))
+            {
+                m_queryLimit /= 2;
+                heapDesc.Count = m_queryLimit;
+            }
+
+            // Create a readback buffer, which will be used as a destination for the query data.
+
+            D3D12_RESOURCE_DESC readbackBufferDesc{};
+            readbackBufferDesc.Alignment = 0;
+            readbackBufferDesc.Dimension = D3D12_RESOURCE_DIMENSION_BUFFER;
+            readbackBufferDesc.Width = m_queryLimit * sizeof(uint64_t);
+            readbackBufferDesc.Height = 1;
+            readbackBufferDesc.DepthOrArraySize = 1;
+            readbackBufferDesc.Format = DXGI_FORMAT_UNKNOWN;
+            readbackBufferDesc.Layout = D3D12_TEXTURE_LAYOUT_ROW_MAJOR;  // Buffers are always row major.
+            readbackBufferDesc.MipLevels = 1;
+            readbackBufferDesc.SampleDesc.Count = 1;
+            readbackBufferDesc.SampleDesc.Quality = 0;
+            readbackBufferDesc.Flags = D3D12_RESOURCE_FLAG_NONE;
+
+            D3D12_HEAP_PROPERTIES readbackHeapProps{};
+            readbackHeapProps.Type = D3D12_HEAP_TYPE_READBACK;
+            readbackHeapProps.CPUPageProperty = D3D12_CPU_PAGE_PROPERTY_UNKNOWN;
+            readbackHeapProps.MemoryPoolPreference = D3D12_MEMORY_POOL_UNKNOWN;
+            readbackHeapProps.CreationNodeMask = 0;
+            readbackHeapProps.VisibleNodeMask = 0;  // #TODO: Support multiple adapters.
+
+            if (FAILED(device->CreateCommittedResource(&readbackHeapProps, D3D12_HEAP_FLAG_NONE, &readbackBufferDesc, D3D12_RESOURCE_STATE_COPY_DEST, nullptr, IID_PPV_ARGS(&m_readbackBuffer))))
+            {
+                TracyD3D12Panic("Failed to create query readback buffer.", return);
+            }
+
+            if (FAILED(device->CreateFence(0, D3D12_FENCE_FLAG_NONE, IID_PPV_ARGS(&m_payloadFence))))
+            {
+                TracyD3D12Panic("Failed to create payload fence.", return);
+            }
+
+            float period = [queue]()
+            {
+                uint64_t timestampFrequency;
+                if (FAILED(queue->GetTimestampFrequency(&timestampFrequency)))
+                {
+                    return 0.0f;
+                }
+                return static_cast<float>( 1E+09 / static_cast<double>(timestampFrequency) );
+            }();
+
+            if (period == 0.0f)
+            {
+                TracyD3D12Panic("Failed to get timestamp frequency.", return);
+            }
+
+            uint64_t cpuTimestamp;
+            uint64_t gpuTimestamp;
+            if (FAILED(queue->GetClockCalibration(&gpuTimestamp, &cpuTimestamp)))
+            {
+                TracyD3D12Panic("Failed to get queue clock calibration.", return);
+            }
+
+            // Save the device cpu timestamp, not the profiler's timestamp.
+            m_prevCalibrationTicksCPU = cpuTimestamp;
+
+            cpuTimestamp = Profiler::GetTime();
+
+            // all checked: ready to roll
+            m_contextId = GetGpuCtxCounter().fetch_add(1);
+
+            auto* item = Profiler::QueueSerial();
+            MemWrite(&item->hdr.type, QueueType::GpuNewContext);
+            MemWrite(&item->gpuNewContext.cpuTime, cpuTimestamp);
+            MemWrite(&item->gpuNewContext.gpuTime, gpuTimestamp);
+            MemWrite(&item->gpuNewContext.thread, decltype(item->gpuNewContext.thread)(0)); // #TODO: why 0 instead of GetThreadHandle()?
+            MemWrite(&item->gpuNewContext.period, period);
+            MemWrite(&item->gpuNewContext.context, GetId());
+            MemWrite(&item->gpuNewContext.flags, GpuContextCalibration);
+            MemWrite(&item->gpuNewContext.type, GpuContextType::Direct3D12);
+            SubmitQueueItem(item);
+        }
+
+        ~D3D12QueueCtx()
+        {
+            // collect all pending timestamps
+            while (m_payloadFence->GetCompletedValue() != m_activePayload)
+                /* busy-wait ... */;
+            Collect();
+            m_payloadFence->Release();
+            m_readbackBuffer->Release();
+            m_queryHeap->Release();
+        }
+
+
+        void NewFrame()
+        {
+            uint32_t queryCounter = m_queryCounter.exchange(0);
+            m_payloadQueue.emplace(D3D12QueryPayload{ m_previousQueryCounter, queryCounter });
+            m_previousQueryCounter += queryCounter;
+
+            if (m_previousQueryCounter >= m_queryLimit)
+            {
+                m_previousQueryCounter -= m_queryLimit;
+            }
+
+            m_queue->Signal(m_payloadFence, ++m_activePayload);
+        }
+
+        void Name( const char* name, uint16_t len )
+        {
+            auto ptr = (char*)tracy_malloc( len );
+            memcpy( ptr, name, len );
+
+            auto item = Profiler::QueueSerial();
+            MemWrite( &item->hdr.type, QueueType::GpuContextName );
+            MemWrite( &item->gpuContextNameFat.context, GetId());
+            MemWrite( &item->gpuContextNameFat.ptr, (uint64_t)ptr );
+            MemWrite( &item->gpuContextNameFat.size, len );
+            SubmitQueueItem(item);
+        }
+
+        void Collect()
+        {
+#ifdef TRACY_ON_DEMAND
+            if (!GetProfiler().IsConnected())
+            {
+                m_queryCounter = 0;
+
+                return;
+            }
+#endif
+
+            // Find out what payloads are available.
+            const auto newestReadyPayload = m_payloadFence->GetCompletedValue();
+            const auto payloadCount = m_payloadQueue.size() - (m_activePayload - newestReadyPayload);
+
+            if (!payloadCount)
+            {
+                return;  // No payloads are available yet, exit out.
+            }
+
+            D3D12_RANGE mapRange{ 0, m_queryLimit * sizeof(uint64_t) };
+
+            // Map the readback buffer so we can fetch the query data from the GPU.
+            void* readbackBufferMapping = nullptr;
+
+            if (FAILED(m_readbackBuffer->Map(0, &mapRange, &readbackBufferMapping)))
+            {
+                TracyD3D12Panic("Failed to map readback buffer.", return);
+            }
+
+            auto* timestampData = static_cast<uint64_t*>(readbackBufferMapping);
+
+            for (uint32_t i = 0; i < payloadCount; ++i)
+            {
+                const auto& payload = m_payloadQueue.front();
+
+                for (uint32_t j = 0; j < payload.m_queryCount; ++j)
+                {
+                    const auto counter = (payload.m_queryIdStart + j) % m_queryLimit;
+                    const auto timestamp = timestampData[counter];
+                    const auto queryId = counter;
+
+                    auto* item = Profiler::QueueSerial();
+                    MemWrite(&item->hdr.type, QueueType::GpuTime);
+                    MemWrite(&item->gpuTime.gpuTime, timestamp);
+                    MemWrite(&item->gpuTime.queryId, static_cast<uint16_t>(queryId));
+                    MemWrite(&item->gpuTime.context, GetId());
+
+                    Profiler::QueueSerialFinish();
+                }
+
+                m_payloadQueue.pop();
+            }
+
+            m_readbackBuffer->Unmap(0, nullptr);
+
+            // Recalibrate to account for drift.
+            RecalibrateClocks();
+        }
+
+    private:
+        tracy_force_inline uint32_t NextQueryId()
+        {
+            uint32_t queryCounter = m_queryCounter.fetch_add(2);
+            if (queryCounter >= m_queryLimit)
+            {
+                TracyD3D12Panic("Submitted too many GPU queries! Consider increasing MaxQueries.");
+                // #TODO: consider returning an invalid id or sentinel value here
+            }
+
+            const uint32_t id = (m_previousQueryCounter + queryCounter) % m_queryLimit;
+
+            return id;
+        }
+
+        tracy_force_inline uint8_t GetId() const
+        {
+            return m_contextId;
+        }
+    };
+
+    class D3D12ZoneScope
+    {
+        const bool m_active;
+        D3D12QueueCtx* m_ctx = nullptr;
+        ID3D12GraphicsCommandList* m_cmdList = nullptr;
+        uint32_t m_queryId = 0;  // Used for tracking in nested zones.
+
+        tracy_force_inline void WriteQueueItem(QueueItem* item, QueueType type, uint64_t srcLocation)
+        {
+            MemWrite(&item->hdr.type, type);
+            MemWrite(&item->gpuZoneBegin.cpuTime, Profiler::GetTime());
+            MemWrite(&item->gpuZoneBegin.srcloc, srcLocation);
+            MemWrite(&item->gpuZoneBegin.thread, GetThreadHandle());
+            MemWrite(&item->gpuZoneBegin.queryId, static_cast<uint16_t>(m_queryId));
+            MemWrite(&item->gpuZoneBegin.context, m_ctx->GetId());
+            Profiler::QueueSerialFinish();
+        }
+
+        tracy_force_inline D3D12ZoneScope(D3D12QueueCtx* ctx, ID3D12GraphicsCommandList* cmdList, bool active)
+#ifdef TRACY_ON_DEMAND
+            : m_active(active&& GetProfiler().IsConnected())
+#else
+            : m_active(active)
+#endif
+        {
+            if (!m_active) return;
+
+            m_ctx = ctx;
+            m_cmdList = cmdList;
+
+            m_queryId = m_ctx->NextQueryId();
+            m_cmdList->EndQuery(m_ctx->m_queryHeap, D3D12_QUERY_TYPE_TIMESTAMP, m_queryId);
+        }
+
+    public:
+        tracy_force_inline D3D12ZoneScope(D3D12QueueCtx* ctx, ID3D12GraphicsCommandList* cmdList, const SourceLocationData* srcLocation, bool active)
+            : D3D12ZoneScope(ctx, cmdList, active)
+        {
+            if (!m_active) return;
+
+            auto* item = Profiler::QueueSerial();
+            WriteQueueItem(item, QueueType::GpuZoneBeginSerial, reinterpret_cast<uint64_t>(srcLocation));
+        }
+
+        tracy_force_inline D3D12ZoneScope(D3D12QueueCtx* ctx, ID3D12GraphicsCommandList* cmdList, const SourceLocationData* srcLocation, int32_t depth, bool active)
+            : D3D12ZoneScope(ctx, cmdList, active)
+        {
+            if (!m_active) return;
+
+            auto* item = Profiler::QueueSerialCallstack(Callstack(depth));
+            WriteQueueItem(item, QueueType::GpuZoneBeginCallstackSerial, reinterpret_cast<uint64_t>(srcLocation));
+        }
+
+        tracy_force_inline D3D12ZoneScope(D3D12QueueCtx* ctx, uint32_t line, const char* source, size_t sourceSz, const char* function, size_t functionSz, const char* name, size_t nameSz, ID3D12GraphicsCommandList* cmdList, bool active)
+            : D3D12ZoneScope(ctx, cmdList, active)
+        {
+            if (!m_active) return;
+
+            const auto sourceLocation = Profiler::AllocSourceLocation(line, source, sourceSz, function, functionSz, name, nameSz);
+
+            auto* item = Profiler::QueueSerial();
+            WriteQueueItem(item, QueueType::GpuZoneBeginAllocSrcLocSerial, sourceLocation);
+        }
+
+        tracy_force_inline D3D12ZoneScope(D3D12QueueCtx* ctx, uint32_t line, const char* source, size_t sourceSz, const char* function, size_t functionSz, const char* name, size_t nameSz, ID3D12GraphicsCommandList* cmdList, int32_t depth, bool active)
+            : D3D12ZoneScope(ctx, cmdList, active)
+        {
+            if (!m_active) return;
+
+            const auto sourceLocation = Profiler::AllocSourceLocation(line, source, sourceSz, function, functionSz, name, nameSz);
+
+            auto* item = Profiler::QueueSerialCallstack(Callstack(depth));
+            WriteQueueItem(item, QueueType::GpuZoneBeginAllocSrcLocCallstackSerial, sourceLocation);
+        }
+
+        tracy_force_inline ~D3D12ZoneScope()
+        {
+            if (!m_active) return;
+
+            const auto queryId = m_queryId + 1;  // Our end query slot is immediately after the begin slot.
+            m_cmdList->EndQuery(m_ctx->m_queryHeap, D3D12_QUERY_TYPE_TIMESTAMP, queryId);
+
+            auto* item = Profiler::QueueSerial();
+            MemWrite(&item->hdr.type, QueueType::GpuZoneEndSerial);
+            MemWrite(&item->gpuZoneEnd.cpuTime, Profiler::GetTime());
+            MemWrite(&item->gpuZoneEnd.thread, GetThreadHandle());
+            MemWrite(&item->gpuZoneEnd.queryId, static_cast<uint16_t>(queryId));
+            MemWrite(&item->gpuZoneEnd.context, m_ctx->GetId());
+            Profiler::QueueSerialFinish();
+
+            m_cmdList->ResolveQueryData(m_ctx->m_queryHeap, D3D12_QUERY_TYPE_TIMESTAMP, m_queryId, 2, m_ctx->m_readbackBuffer, m_queryId * sizeof(uint64_t));
+        }
+    };
+
+	void* CreateD3D12Context(ID3D12Device* device, ID3D12CommandQueue* queue)
+    {
+        auto* ctx = static_cast<D3D12QueueCtx*>(tracy_malloc(sizeof(D3D12QueueCtx)));
+        new (ctx) D3D12QueueCtx{ device, queue };
+        ctx->Name("D3D12", 5);
+        return ctx;
+    }
+
+    void CollectD3D12Context(void* ctx)
+    {
+        ((D3D12QueueCtx*)ctx)->Collect();
+        ((D3D12QueueCtx*)ctx)->NewFrame();
+    }
+
+	void DestroyD3D12Context(void* ctx)
+    {
+        ((D3D12QueueCtx*)ctx)->~D3D12QueueCtx();
+        tracy_free(ctx);
+    }
+
+	void BeginD3D12ZoneScope(void* zone, void* ctx, ID3D12GraphicsCommandList* cmdList, const char* name, size_t nameLen)
+	{
+        new(zone) tracy::D3D12ZoneScope{ (tracy::D3D12QueueCtx*)ctx, 0, 0, 0, 0, 0, name, nameLen, cmdList, true };
+	}
+
+	void EndD3D12ZoneScope(void* zone)
+	{
+        ((tracy::D3D12ZoneScope*)zone)->~D3D12ZoneScope();
+	}
+}
+
+#undef TracyD3D12Panic
+
+using TracyD3D12Ctx = tracy::D3D12QueueCtx*;
+
+#define TracyD3D12Context(device, queue) tracy::CreateD3D12Context(device, queue);
+#define TracyD3D12Destroy(ctx) tracy::DestroyD3D12Context(ctx);
+#define TracyD3D12ContextName(ctx, name, size) ctx->Name(name, size);
+
+#define TracyD3D12NewFrame(ctx) ctx->NewFrame();
+
+#define TracyD3D12UnnamedZone ___tracy_gpu_d3d12_zone
+#define TracyD3D12SrcLocSymbol TracyConcat(__tracy_d3d12_source_location,TracyLine)
+#define TracyD3D12SrcLocObject(name, color) static constexpr tracy::SourceLocationData TracyD3D12SrcLocSymbol { name, TracyFunction, TracyFile, (uint32_t)TracyLine, color };
+
+#if defined TRACY_HAS_CALLSTACK && defined TRACY_CALLSTACK
+#  define TracyD3D12Zone(ctx, cmdList, name) TracyD3D12NamedZoneS(ctx, TracyD3D12UnnamedZone, cmdList, name, TRACY_CALLSTACK, true)
+#  define TracyD3D12ZoneC(ctx, cmdList, name, color) TracyD3D12NamedZoneCS(ctx, TracyD3D12UnnamedZone, cmdList, name, color, TRACY_CALLSTACK, true)
+#  define TracyD3D12NamedZone(ctx, varname, cmdList, name, active) TracyD3D12SrcLocObject(name, 0); tracy::D3D12ZoneScope varname{ ctx, cmdList, &TracyD3D12SrcLocSymbol, TRACY_CALLSTACK, active };
+#  define TracyD3D12NamedZoneC(ctx, varname, cmdList, name, color, active) TracyD3D12SrcLocObject(name, color); tracy::D3D12ZoneScope varname{ ctx, cmdList, &TracyD3D12SrcLocSymbol, TRACY_CALLSTACK, active };
+#  define TracyD3D12ZoneTransient(ctx, varname, cmdList, name, active) TracyD3D12ZoneTransientS(ctx, varname, cmdList, name, TRACY_CALLSTACK, active)
+#else
+#  define TracyD3D12Zone(ctx, cmdList, name) TracyD3D12NamedZone(ctx, TracyD3D12UnnamedZone, cmdList, name, true)
+#  define TracyD3D12ZoneC(ctx, cmdList, name, color) TracyD3D12NamedZoneC(ctx, TracyD3D12UnnamedZone, cmdList, name, color, true)
+#  define TracyD3D12NamedZone(ctx, varname, cmdList, name, active) TracyD3D12SrcLocObject(name, 0); tracy::D3D12ZoneScope varname{ ctx, cmdList, &TracyD3D12SrcLocSymbol, active };
+#  define TracyD3D12NamedZoneC(ctx, varname, cmdList, name, color, active) TracyD3D12SrcLocObject(name, color); tracy::D3D12ZoneScope varname{ ctx, cmdList, &TracyD3D12SrcLocSymbol, active };
+#  define TracyD3D12ZoneTransient(ctx, varname, cmdList, name, active) tracy::D3D12ZoneScope varname{ ctx, TracyLine, TracyFile, strlen(TracyFile), TracyFunction, strlen(TracyFunction), name, strlen(name), cmdList, active };
+#endif
+
+#ifdef TRACY_HAS_CALLSTACK
+#  define TracyD3D12ZoneS(ctx, cmdList, name, depth) TracyD3D12NamedZoneS(ctx, TracyD3D12UnnamedZone, cmdList, name, depth, true)
+#  define TracyD3D12ZoneCS(ctx, cmdList, name, color, depth) TracyD3D12NamedZoneCS(ctx, TracyD3D12UnnamedZone, cmdList, name, color, depth, true)
+#  define TracyD3D12NamedZoneS(ctx, varname, cmdList, name, depth, active) TracyD3D12SrcLocObject(name, 0); tracy::D3D12ZoneScope varname{ ctx, cmdList, &TracyD3D12SrcLocSymbol, depth, active };
+#  define TracyD3D12NamedZoneCS(ctx, varname, cmdList, name, color, depth, active) TracyD3D12SrcLocObject(name, color); tracy::D3D12ZoneScope varname{ ctx, cmdList, &TracyD3D12SrcLocSymbol, depth, active };
+#  define TracyD3D12ZoneTransientS(ctx, varname, cmdList, name, depth, active) tracy::D3D12ZoneScope varname{ ctx, TracyLine, TracyFile, strlen(TracyFile), TracyFunction, strlen(TracyFunction), name, strlen(name), cmdList, depth, active };
+#else
+#  define TracyD3D12ZoneS(ctx, cmdList, name, depth) TracyD3D12Zone(ctx, cmdList, name)
+#  define TracyD3D12ZoneCS(ctx, cmdList, name, color, depth) TracyD3D12Zone(ctx, cmdList, name, color)
+#  define TracyD3D12NamedZoneS(ctx, varname, cmdList, name, depth, active) TracyD3D12NamedZone(ctx, varname, cmdList, name, active)
+#  define TracyD3D12NamedZoneCS(ctx, varname, cmdList, name, color, depth, active) TracyD3D12NamedZoneC(ctx, varname, cmdList, name, color, active)
+#  define TracyD3D12ZoneTransientS(ctx, varname, cmdList, name, depth, active) TracyD3D12ZoneTransient(ctx, varname, cmdList, name, active)
+#endif
+
+#define TracyD3D12Collect(ctx) ctx->Collect();
+
+#else
+
+// Forward declarations to be used in engine
+namespace tracy
+{
+    extern void* CreateD3D12Context(ID3D12Device* device, ID3D12CommandQueue* queue);
+    extern void DestroyD3D12Context(void* ctx);
+    extern void CollectD3D12Context(void* ctx);
+    extern void BeginD3D12ZoneScope(void* zone, void* ctx, ID3D12GraphicsCommandList* cmdList, const char* name, size_t nameLen);
+    extern void EndD3D12ZoneScope(void* zone);
+}
+
+#endif
+
+#endif
--- a/Source/ThirdParty/tracy/tracy/TracyVulkan.hpp
+++ b/Source/ThirdParty/tracy/tracy/TracyVulkan.hpp
@@ -0,0 +1,779 @@
+#ifndef __TRACYVULKAN_HPP__
+#define __TRACYVULKAN_HPP__
+
+#define TracyVulkanZoneSize 24
+
+#if !defined TRACY_ENABLE
+
+#define TracyVkContext(x,y,z,w) nullptr
+#define TracyVkContextCalibrated(x,y,z,w,a,b) nullptr
+#if defined VK_EXT_host_query_reset
+#define TracyVkContextHostCalibrated(x,y,z,w,a) nullptr
+#endif
+#define TracyVkDestroy(x)
+#define TracyVkContextName(c,x,y)
+#define TracyVkNamedZone(c,x,y,z,w)
+#define TracyVkNamedZoneC(c,x,y,z,w,a)
+#define TracyVkZone(c,x,y)
+#define TracyVkZoneC(c,x,y,z)
+#define TracyVkZoneTransient(c,x,y,z,w)
+#define TracyVkCollect(c,x)
+
+#define TracyVkNamedZoneS(c,x,y,z,w,a)
+#define TracyVkNamedZoneCS(c,x,y,z,w,v,a)
+#define TracyVkZoneS(c,x,y,z)
+#define TracyVkZoneCS(c,x,y,z,w)
+#define TracyVkZoneTransientS(c,x,y,z,w,a)
+
+namespace tracy
+{
+class VkCtxScope {};
+}
+
+using TracyVkCtx = void*;
+
+#elif TRACY_GPU_VULKAN
+
+#if !defined VK_NULL_HANDLE
+#  error "You must include Vulkan headers before including TracyVulkan.hpp"
+#endif
+
+#include <assert.h>
+#include <stdlib.h>
+#include "../client/TracyProfiler.hpp"
+#include "../client/TracyCallstack.hpp"
+
+#include <atomic>
+
+namespace tracy
+{
+
+#if defined TRACY_VK_USE_SYMBOL_TABLE
+#define LoadVkDeviceCoreSymbols(Operation) \
+    Operation(vkBeginCommandBuffer) \
+    Operation(vkCmdResetQueryPool) \
+    Operation(vkCmdWriteTimestamp) \
+    Operation(vkCreateQueryPool) \
+    Operation(vkDestroyQueryPool) \
+    Operation(vkEndCommandBuffer) \
+    Operation(vkGetQueryPoolResults) \
+    Operation(vkQueueSubmit) \
+    Operation(vkQueueWaitIdle) \
+    Operation(vkResetQueryPool)
+
+#define LoadVkDeviceExtensionSymbols(Operation) \
+    Operation(vkGetCalibratedTimestampsEXT)
+
+#define LoadVkInstanceExtensionSymbols(Operation) \
+    Operation(vkGetPhysicalDeviceCalibrateableTimeDomainsEXT)
+
+#define LoadVkInstanceCoreSymbols(Operation) \
+    Operation(vkGetPhysicalDeviceProperties)
+
+struct VkSymbolTable
+{
+#define MAKE_PFN(name) PFN_##name name;
+    LoadVkDeviceCoreSymbols(MAKE_PFN)
+    LoadVkDeviceExtensionSymbols(MAKE_PFN)
+    LoadVkInstanceExtensionSymbols(MAKE_PFN)
+    LoadVkInstanceCoreSymbols(MAKE_PFN)
+#undef MAKE_PFN
+};
+
+#define VK_FUNCTION_WRAPPER(callSignature) m_symbols.callSignature
+#define CONTEXT_VK_FUNCTION_WRAPPER(callSignature) m_ctx->m_symbols.callSignature
+#else
+#define VK_FUNCTION_WRAPPER(callSignature) callSignature
+#define CONTEXT_VK_FUNCTION_WRAPPER(callSignature) callSignature
+#endif
+
+class VkCtx
+{
+    friend class VkCtxScope;
+
+    enum { QueryCount = 64 * 1024 };
+
+public:
+#if defined TRACY_VK_USE_SYMBOL_TABLE
+    VkCtx( VkInstance instance, VkPhysicalDevice physdev, VkDevice device, VkQueue queue, VkCommandBuffer cmdbuf, PFN_vkGetInstanceProcAddr instanceProcAddr, PFN_vkGetDeviceProcAddr deviceProcAddr, bool calibrated )
+#else
+    VkCtx( VkPhysicalDevice physdev, VkDevice device, VkQueue queue, VkCommandBuffer cmdbuf, PFN_vkGetPhysicalDeviceCalibrateableTimeDomainsEXT vkGetPhysicalDeviceCalibrateableTimeDomainsEXT, PFN_vkGetCalibratedTimestampsEXT vkGetCalibratedTimestampsEXT)
+#endif
+        : m_device( device )
+        , m_timeDomain( VK_TIME_DOMAIN_DEVICE_EXT )
+        , m_context( GetGpuCtxCounter().fetch_add( 1, std::memory_order_relaxed ) )
+        , m_head( 0 )
+        , m_tail( 0 )
+        , m_oldCnt( 0 )
+        , m_queryCount( QueryCount )
+#if !defined TRACY_VK_USE_SYMBOL_TABLE
+        , m_vkGetCalibratedTimestampsEXT( vkGetCalibratedTimestampsEXT )
+#endif
+    {
+        assert( m_context != 255 );
+
+#if defined TRACY_VK_USE_SYMBOL_TABLE
+        PopulateSymbolTable(instance, instanceProcAddr, deviceProcAddr);
+        if ( calibrated )
+        {
+            m_vkGetCalibratedTimestampsEXT = m_symbols.vkGetCalibratedTimestampsEXT;
+        }
+
+#endif
+
+        if( VK_FUNCTION_WRAPPER( vkGetPhysicalDeviceCalibrateableTimeDomainsEXT ) && m_vkGetCalibratedTimestampsEXT )
+        {
+            FindAvailableTimeDomains( physdev, VK_FUNCTION_WRAPPER( vkGetPhysicalDeviceCalibrateableTimeDomainsEXT ) );
+        }
+
+        CreateQueryPool();
+
+        VkCommandBufferBeginInfo beginInfo = {};
+        beginInfo.sType = VK_STRUCTURE_TYPE_COMMAND_BUFFER_BEGIN_INFO;
+        beginInfo.flags = VK_COMMAND_BUFFER_USAGE_ONE_TIME_SUBMIT_BIT;
+
+        VkSubmitInfo submitInfo = {};
+        submitInfo.sType = VK_STRUCTURE_TYPE_SUBMIT_INFO;
+        submitInfo.commandBufferCount = 1;
+        submitInfo.pCommandBuffers = &cmdbuf;
+
+        VK_FUNCTION_WRAPPER( vkBeginCommandBuffer( cmdbuf, &beginInfo ) );
+        VK_FUNCTION_WRAPPER( vkCmdResetQueryPool( cmdbuf, m_query, 0, m_queryCount ) );
+        VK_FUNCTION_WRAPPER( vkEndCommandBuffer( cmdbuf ) );
+        VK_FUNCTION_WRAPPER( vkQueueSubmit( queue, 1, &submitInfo, VK_NULL_HANDLE ) );
+        VK_FUNCTION_WRAPPER( vkQueueWaitIdle( queue ) );
+
+        int64_t tcpu, tgpu;
+        if( m_timeDomain == VK_TIME_DOMAIN_DEVICE_EXT )
+        {
+            VK_FUNCTION_WRAPPER( vkBeginCommandBuffer( cmdbuf, &beginInfo ) );
+            VK_FUNCTION_WRAPPER( vkCmdWriteTimestamp( cmdbuf, VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT, m_query, 0 ) );
+            VK_FUNCTION_WRAPPER( vkEndCommandBuffer( cmdbuf ) );
+            VK_FUNCTION_WRAPPER( vkQueueSubmit( queue, 1, &submitInfo, VK_NULL_HANDLE ) );
+            VK_FUNCTION_WRAPPER( vkQueueWaitIdle( queue ) );
+
+            tcpu = Profiler::GetTime();
+            VK_FUNCTION_WRAPPER( vkGetQueryPoolResults( device, m_query, 0, 1, sizeof( tgpu ), &tgpu, sizeof( tgpu ), VK_QUERY_RESULT_64_BIT | VK_QUERY_RESULT_WAIT_BIT ) );
+
+            VK_FUNCTION_WRAPPER( vkBeginCommandBuffer( cmdbuf, &beginInfo ) );
+            VK_FUNCTION_WRAPPER( vkCmdResetQueryPool( cmdbuf, m_query, 0, 1 ) );
+            VK_FUNCTION_WRAPPER( vkEndCommandBuffer( cmdbuf ) );
+            VK_FUNCTION_WRAPPER( vkQueueSubmit( queue, 1, &submitInfo, VK_NULL_HANDLE ) );
+            VK_FUNCTION_WRAPPER( vkQueueWaitIdle( queue ) );
+        }
+        else
+        {
+            FindCalibratedTimestampDeviation();
+            Calibrate( device, m_prevCalibration, tgpu );
+            tcpu = Profiler::GetTime();
+        }
+
+        WriteInitialItem( physdev, tcpu, tgpu );
+
+        m_res = (int64_t*)tracy_malloc( sizeof( int64_t ) * m_queryCount );
+    }
+
+#if defined VK_EXT_host_query_reset
+    /**
+     * This alternative constructor does not use command buffers and instead uses functionality from
+     * VK_EXT_host_query_reset (core with 1.2 and non-optional) and VK_EXT_calibrated_timestamps. This requires
+     * the physical device to have another time domain apart from DEVICE to be calibrateable.
+     */
+#if defined TRACY_VK_USE_SYMBOL_TABLE
+    VkCtx( VkInstance instance, VkPhysicalDevice physdev, VkDevice device, PFN_vkGetInstanceProcAddr instanceProcAddr, PFN_vkGetDeviceProcAddr deviceProcAddr )
+#else
+    VkCtx( VkPhysicalDevice physdev, VkDevice device, PFN_vkResetQueryPoolEXT vkResetQueryPool, PFN_vkGetPhysicalDeviceCalibrateableTimeDomainsEXT vkGetPhysicalDeviceCalibrateableTimeDomainsEXT, PFN_vkGetCalibratedTimestampsEXT vkGetCalibratedTimestampsEXT )
+#endif
+        : m_device( device )
+        , m_timeDomain( VK_TIME_DOMAIN_DEVICE_EXT )
+        , m_context( GetGpuCtxCounter().fetch_add(1, std::memory_order_relaxed) )
+        , m_head( 0 )
+        , m_tail( 0 )
+        , m_oldCnt( 0 )
+        , m_queryCount( QueryCount )
+#if !defined TRACY_VK_USE_SYMBOL_TABLE
+        , m_vkGetCalibratedTimestampsEXT( vkGetCalibratedTimestampsEXT )
+#endif
+    {
+        assert( m_context != 255);
+
+#if defined TRACY_VK_USE_SYMBOL_TABLE
+        PopulateSymbolTable(instance, instanceProcAddr, deviceProcAddr);
+        m_vkGetCalibratedTimestampsEXT = m_symbols.vkGetCalibratedTimestampsEXT;
+#endif
+
+        assert( VK_FUNCTION_WRAPPER( vkResetQueryPool ) != nullptr );
+        assert( VK_FUNCTION_WRAPPER( vkGetPhysicalDeviceCalibrateableTimeDomainsEXT ) != nullptr );
+        assert( VK_FUNCTION_WRAPPER( vkGetCalibratedTimestampsEXT ) != nullptr );
+
+        FindAvailableTimeDomains( physdev, VK_FUNCTION_WRAPPER( vkGetPhysicalDeviceCalibrateableTimeDomainsEXT ) );
+
+        // We require a host time domain to be available to properly calibrate.
+        FindCalibratedTimestampDeviation();
+        int64_t tgpu;
+        Calibrate( device, m_prevCalibration, tgpu );
+        int64_t tcpu = Profiler::GetTime();
+
+        CreateQueryPool();
+        VK_FUNCTION_WRAPPER( vkResetQueryPool( device, m_query, 0, m_queryCount ) );
+
+        WriteInitialItem( physdev, tcpu, tgpu );
+
+        // We need the buffer to be twice as large for availability values
+        size_t resSize = sizeof( int64_t ) * m_queryCount * 2;
+        m_res = (int64_t*)tracy_malloc( resSize );
+    }
+#endif
+
+    ~VkCtx()
+    {
+        tracy_free( m_res );
+        VK_FUNCTION_WRAPPER( vkDestroyQueryPool( m_device, m_query, nullptr ) );
+    }
+
+    void Name( const char* name, uint16_t len )
+    {
+        auto ptr = (char*)tracy_malloc( len );
+        memcpy( ptr, name, len );
+
+        auto item = Profiler::QueueSerial();
+        MemWrite( &item->hdr.type, QueueType::GpuContextName );
+        MemWrite( &item->gpuContextNameFat.context, m_context );
+        MemWrite( &item->gpuContextNameFat.ptr, (uint64_t)ptr );
+        MemWrite( &item->gpuContextNameFat.size, len );
+#ifdef TRACY_ON_DEMAND
+        GetProfiler().DeferItem( *item );
+#endif
+        Profiler::QueueSerialFinish();
+    }
+
+    void Collect( VkCommandBuffer cmdbuf )
+    {
+        const uint64_t head = m_head.load(std::memory_order_relaxed);
+        if( m_tail == head ) return;
+
+#ifdef TRACY_ON_DEMAND
+        if( !GetProfiler().IsConnected() )
+        {
+            VK_FUNCTION_WRAPPER( vkCmdResetQueryPool( cmdbuf, m_query, 0, m_queryCount ) );
+            m_tail = head;
+            m_oldCnt = 0;
+            int64_t tgpu;
+            if( m_timeDomain != VK_TIME_DOMAIN_DEVICE_EXT ) Calibrate( m_device, m_prevCalibration, tgpu );
+            return;
+        }
+#endif
+        assert( head > m_tail );
+
+        const unsigned int wrappedTail = (unsigned int)( m_tail % m_queryCount );
+
+        unsigned int cnt;
+        if( m_oldCnt != 0 )
+        {
+            cnt = m_oldCnt;
+            m_oldCnt = 0;
+        }
+        else
+        {
+            cnt = (unsigned int)( head - m_tail );
+            assert( cnt <= m_queryCount );
+            if( wrappedTail + cnt > m_queryCount )
+            {
+                cnt = m_queryCount - wrappedTail;
+            }
+        }
+
+
+        VK_FUNCTION_WRAPPER( vkGetQueryPoolResults( m_device, m_query, wrappedTail, cnt, sizeof( int64_t ) * m_queryCount * 2, m_res, sizeof( int64_t ) * 2, VK_QUERY_RESULT_64_BIT | VK_QUERY_RESULT_WITH_AVAILABILITY_BIT ) );
+
+        for( unsigned int idx=0; idx<cnt; idx++ )
+        {
+            int64_t avail = m_res[idx * 2 + 1];
+            if( avail == 0 )
+            {
+                m_oldCnt = cnt - idx;
+                cnt = idx;
+
+                break;
+            }
+
+            auto item = Profiler::QueueSerial();
+            MemWrite( &item->hdr.type, QueueType::GpuTime );
+            MemWrite( &item->gpuTime.gpuTime, m_res[idx * 2] );
+            MemWrite( &item->gpuTime.queryId, uint16_t( wrappedTail + idx ) );
+            MemWrite( &item->gpuTime.context, m_context );
+            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();
+            }
+        }
+
+        VK_FUNCTION_WRAPPER( vkCmdResetQueryPool( cmdbuf, m_query, wrappedTail, cnt ) );
+
+        m_tail += cnt;
+    }
+
+    tracy_force_inline unsigned int NextQueryId()
+    {
+        const uint64_t id = m_head.fetch_add(1, std::memory_order_relaxed);
+        return id % m_queryCount;
+    }
+
+    tracy_force_inline uint8_t GetId() const
+    {
+        return m_context;
+    }
+
+    tracy_force_inline VkQueryPool GetQueryPool() const
+    {
+         return m_query;
+    }
+
+private:
+    tracy_force_inline void Calibrate( VkDevice device, int64_t& tCpu, int64_t& tGpu )
+    {
+        assert( m_timeDomain != VK_TIME_DOMAIN_DEVICE_EXT );
+        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;
+        do
+        {
+            m_vkGetCalibratedTimestampsEXT( device, 2, spec, ts, &deviation );
+        }
+        while( deviation > m_deviation );
+
+#if defined _WIN32
+        tGpu = ts[0];
+        tCpu = ts[1] * m_qpcToNs;
+#elif defined __linux__ && defined CLOCK_MONOTONIC_RAW
+        tGpu = ts[0];
+        tCpu = ts[1];
+#else
+        assert( false );
+#endif
+    }
+
+    tracy_force_inline void CreateQueryPool()
+    {
+        VkQueryPoolCreateInfo poolInfo = {};
+        poolInfo.sType = VK_STRUCTURE_TYPE_QUERY_POOL_CREATE_INFO;
+        poolInfo.queryCount = m_queryCount;
+        poolInfo.queryType = VK_QUERY_TYPE_TIMESTAMP;
+        while ( VK_FUNCTION_WRAPPER( vkCreateQueryPool( m_device, &poolInfo, nullptr, &m_query ) != VK_SUCCESS ) )
+        {
+            m_queryCount /= 2;
+            poolInfo.queryCount = m_queryCount;
+        }
+    }
+
+    tracy_force_inline void FindAvailableTimeDomains( VkPhysicalDevice physicalDevice, PFN_vkGetPhysicalDeviceCalibrateableTimeDomainsEXT _vkGetPhysicalDeviceCalibrateableTimeDomainsEXT )
+    {
+        uint32_t num;
+        _vkGetPhysicalDeviceCalibrateableTimeDomainsEXT( physicalDevice, &num, nullptr );
+        if(num > 4) num = 4;
+        VkTimeDomainEXT data[4];
+        _vkGetPhysicalDeviceCalibrateableTimeDomainsEXT( physicalDevice, &num, data );
+        VkTimeDomainEXT supportedDomain = (VkTimeDomainEXT)-1;
+#if defined _WIN32
+        supportedDomain = VK_TIME_DOMAIN_QUERY_PERFORMANCE_COUNTER_EXT;
+#elif defined __linux__ && defined CLOCK_MONOTONIC_RAW
+        supportedDomain = VK_TIME_DOMAIN_CLOCK_MONOTONIC_RAW_EXT;
+#endif
+        for( uint32_t i=0; i<num; i++ ) {
+            if(data[i] == supportedDomain) {
+                m_timeDomain = data[i];
+                break;
+            }
+        }
+    }
+
+    tracy_force_inline void FindCalibratedTimestampDeviation()
+    {
+        assert( m_timeDomain != VK_TIME_DOMAIN_DEVICE_EXT );
+        constexpr size_t 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( size_t i=0; i<NumProbes; i++ ) {
+            m_vkGetCalibratedTimestampsEXT( m_device, 2, spec, ts, deviation + i );
+        }
+        uint64_t minDeviation = deviation[0];
+        for( size_t i=1; i<NumProbes; i++ ) {
+            if ( minDeviation > deviation[i] ) {
+                minDeviation = deviation[i];
+            }
+        }
+        m_deviation = minDeviation * 3 / 2;
+
+#if defined _WIN32
+        m_qpcToNs = int64_t( 1000000000. / GetFrequencyQpc() );
+#endif
+    }
+
+    tracy_force_inline void WriteInitialItem( VkPhysicalDevice physdev, int64_t tcpu, int64_t tgpu )
+    {
+        uint8_t flags = 0;
+        if( m_timeDomain != VK_TIME_DOMAIN_DEVICE_EXT ) flags |= GpuContextCalibration;
+
+        VkPhysicalDeviceProperties prop;
+        VK_FUNCTION_WRAPPER( vkGetPhysicalDeviceProperties( physdev, &prop ) );
+        const float period = prop.limits.timestampPeriod;
+
+        auto item = Profiler::QueueSerial();
+        MemWrite( &item->hdr.type, QueueType::GpuNewContext );
+        MemWrite( &item->gpuNewContext.cpuTime, tcpu );
+        MemWrite( &item->gpuNewContext.gpuTime, tgpu );
+        memset( &item->gpuNewContext.thread, 0, sizeof( item->gpuNewContext.thread ) );
+        MemWrite( &item->gpuNewContext.period, period );
+        MemWrite( &item->gpuNewContext.context, m_context );
+        MemWrite( &item->gpuNewContext.flags, flags );
+        MemWrite( &item->gpuNewContext.type, GpuContextType::Vulkan );
+
+#ifdef TRACY_ON_DEMAND
+        GetProfiler().DeferItem( *item );
+#endif
+        Profiler::QueueSerialFinish();
+    }
+
+#if defined TRACY_VK_USE_SYMBOL_TABLE
+    void PopulateSymbolTable( VkInstance instance, PFN_vkGetInstanceProcAddr instanceProcAddr, PFN_vkGetDeviceProcAddr deviceProcAddr )
+    {
+#define VK_GET_DEVICE_SYMBOL( name ) \
+        (PFN_##name)deviceProcAddr( m_device, #name );
+#define VK_LOAD_DEVICE_SYMBOL( name ) \
+        m_symbols.name = VK_GET_DEVICE_SYMBOL( name );
+#define VK_GET_INSTANCE_SYMBOL( name ) \
+        (PFN_##name)instanceProcAddr( instance, #name );
+#define VK_LOAD_INSTANCE_SYMBOL( name ) \
+        m_symbols.name = VK_GET_INSTANCE_SYMBOL( name );
+
+        LoadVkDeviceCoreSymbols( VK_LOAD_DEVICE_SYMBOL )
+        LoadVkDeviceExtensionSymbols( VK_LOAD_DEVICE_SYMBOL )
+        LoadVkInstanceExtensionSymbols( VK_LOAD_INSTANCE_SYMBOL )
+        LoadVkInstanceCoreSymbols( VK_LOAD_INSTANCE_SYMBOL )
+#undef VK_GET_DEVICE_SYMBOL
+#undef VK_LOAD_DEVICE_SYMBOL
+#undef VK_GET_INSTANCE_SYMBOL
+#undef VK_LOAD_INSTANCE_SYMBOL
+    }
+#endif
+
+    VkDevice m_device;
+    VkQueryPool m_query;
+    VkTimeDomainEXT m_timeDomain;
+#if defined TRACY_VK_USE_SYMBOL_TABLE
+    VkSymbolTable m_symbols;
+#endif
+    uint64_t m_deviation;
+#ifdef _WIN32
+    int64_t m_qpcToNs;
+#endif
+    int64_t m_prevCalibration;
+    uint8_t m_context;
+
+    std::atomic<uint64_t> m_head;
+    uint64_t m_tail;
+    unsigned int m_oldCnt;
+    unsigned int m_queryCount;
+
+    int64_t* m_res;
+
+    PFN_vkGetCalibratedTimestampsEXT m_vkGetCalibratedTimestampsEXT;
+};
+
+class VkCtxScope
+{
+public:
+    tracy_force_inline VkCtxScope( VkCtx* ctx, const SourceLocationData* srcloc, VkCommandBuffer cmdbuf, bool is_active )
+#ifdef TRACY_ON_DEMAND
+        : m_active( is_active && GetProfiler().IsConnected() )
+#else
+        : m_active( is_active )
+#endif
+    {
+        if( !m_active ) return;
+        m_cmdbuf = cmdbuf;
+        m_ctx = ctx;
+
+        const auto queryId = ctx->NextQueryId();
+        CONTEXT_VK_FUNCTION_WRAPPER( vkCmdWriteTimestamp( cmdbuf, VK_PIPELINE_STAGE_BOTTOM_OF_PIPE_BIT, ctx->m_query, queryId ) );
+
+        auto item = Profiler::QueueSerial();
+        MemWrite( &item->hdr.type, QueueType::GpuZoneBeginSerial );
+        MemWrite( &item->gpuZoneBegin.cpuTime, Profiler::GetTime() );
+        MemWrite( &item->gpuZoneBegin.srcloc, (uint64_t)srcloc );
+        MemWrite( &item->gpuZoneBegin.thread, GetThreadHandle() );
+        MemWrite( &item->gpuZoneBegin.queryId, uint16_t( queryId ) );
+        MemWrite( &item->gpuZoneBegin.context, ctx->GetId() );
+        Profiler::QueueSerialFinish();
+    }
+
+    tracy_force_inline VkCtxScope( VkCtx* ctx, const SourceLocationData* srcloc, VkCommandBuffer cmdbuf, int32_t depth, bool is_active )
+#ifdef TRACY_ON_DEMAND
+        : m_active( is_active && GetProfiler().IsConnected() )
+#else
+        : m_active( is_active )
+#endif
+    {
+        if( !m_active ) return;
+        m_cmdbuf = cmdbuf;
+        m_ctx = ctx;
+
+        const auto queryId = ctx->NextQueryId();
+        CONTEXT_VK_FUNCTION_WRAPPER( vkCmdWriteTimestamp( cmdbuf, VK_PIPELINE_STAGE_BOTTOM_OF_PIPE_BIT, ctx->m_query, queryId ) );
+
+        QueueItem *item;
+        if( depth > 0 && has_callstack() )
+        {
+            item = Profiler::QueueSerialCallstack( Callstack( depth ) );
+            MemWrite( &item->hdr.type, QueueType::GpuZoneBeginCallstackSerial );
+        }
+        else
+        {
+            item = Profiler::QueueSerial();
+            MemWrite( &item->hdr.type, QueueType::GpuZoneBeginSerial );
+        }
+        MemWrite( &item->gpuZoneBegin.cpuTime, Profiler::GetTime() );
+        MemWrite( &item->gpuZoneBegin.srcloc, (uint64_t)srcloc );
+        MemWrite( &item->gpuZoneBegin.thread, GetThreadHandle() );
+        MemWrite( &item->gpuZoneBegin.queryId, uint16_t( queryId ) );
+        MemWrite( &item->gpuZoneBegin.context, ctx->GetId() );
+        Profiler::QueueSerialFinish();
+    }
+
+    tracy_force_inline VkCtxScope( VkCtx* ctx, uint32_t line, const char* source, size_t sourceSz, const char* function, size_t functionSz, const char* name, size_t nameSz, VkCommandBuffer cmdbuf, bool is_active )
+#ifdef TRACY_ON_DEMAND
+        : m_active( is_active && GetProfiler().IsConnected() )
+#else
+        : m_active( is_active )
+#endif
+    {
+        if( !m_active ) return;
+        m_cmdbuf = cmdbuf;
+        m_ctx = ctx;
+
+        const auto queryId = ctx->NextQueryId();
+        CONTEXT_VK_FUNCTION_WRAPPER( vkCmdWriteTimestamp( cmdbuf, VK_PIPELINE_STAGE_BOTTOM_OF_PIPE_BIT, ctx->m_query, queryId ) );
+
+        const auto srcloc = Profiler::AllocSourceLocation( line, source, sourceSz, function, functionSz, name, nameSz );
+        auto item = Profiler::QueueSerial();
+        MemWrite( &item->hdr.type, QueueType::GpuZoneBeginAllocSrcLocSerial );
+        MemWrite( &item->gpuZoneBegin.cpuTime, Profiler::GetTime() );
+        MemWrite( &item->gpuZoneBegin.srcloc, srcloc );
+        MemWrite( &item->gpuZoneBegin.thread, GetThreadHandle() );
+        MemWrite( &item->gpuZoneBegin.queryId, uint16_t( queryId ) );
+        MemWrite( &item->gpuZoneBegin.context, ctx->GetId() );
+        Profiler::QueueSerialFinish();
+    }
+
+    tracy_force_inline VkCtxScope( VkCtx* ctx, uint32_t line, const char* source, size_t sourceSz, const char* function, size_t functionSz, const char* name, size_t nameSz, VkCommandBuffer cmdbuf, int32_t depth, bool is_active )
+#ifdef TRACY_ON_DEMAND
+        : m_active( is_active && GetProfiler().IsConnected() )
+#else
+        : m_active( is_active )
+#endif
+    {
+        if( !m_active ) return;
+        m_cmdbuf = cmdbuf;
+        m_ctx = ctx;
+
+        const auto queryId = ctx->NextQueryId();
+        CONTEXT_VK_FUNCTION_WRAPPER( vkCmdWriteTimestamp( cmdbuf, VK_PIPELINE_STAGE_BOTTOM_OF_PIPE_BIT, ctx->m_query, queryId ) );
+
+        const auto srcloc = Profiler::AllocSourceLocation( line, source, sourceSz, function, functionSz, name, nameSz );
+        QueueItem *item;
+        if( depth > 0 && has_callstack() )
+        {
+            item = Profiler::QueueSerialCallstack( Callstack( depth ) );
+            MemWrite( &item->hdr.type, QueueType::GpuZoneBeginAllocSrcLocCallstackSerial );
+        }
+        else
+        {
+            item = Profiler::QueueSerial();
+            MemWrite( &item->hdr.type, QueueType::GpuZoneBeginAllocSrcLocSerial );
+        }
+        MemWrite( &item->gpuZoneBegin.cpuTime, Profiler::GetTime() );
+        MemWrite( &item->gpuZoneBegin.srcloc, srcloc );
+        MemWrite( &item->gpuZoneBegin.thread, GetThreadHandle() );
+        MemWrite( &item->gpuZoneBegin.queryId, uint16_t( queryId ) );
+        MemWrite( &item->gpuZoneBegin.context, ctx->GetId() );
+        Profiler::QueueSerialFinish();
+    }
+
+    tracy_force_inline ~VkCtxScope()
+    {
+        if( !m_active ) return;
+
+        const auto queryId = m_ctx->NextQueryId();
+        CONTEXT_VK_FUNCTION_WRAPPER( vkCmdWriteTimestamp( m_cmdbuf, VK_PIPELINE_STAGE_BOTTOM_OF_PIPE_BIT, m_ctx->m_query, queryId ) );
+
+        auto item = Profiler::QueueSerial();
+        MemWrite( &item->hdr.type, QueueType::GpuZoneEndSerial );
+        MemWrite( &item->gpuZoneEnd.cpuTime, Profiler::GetTime() );
+        MemWrite( &item->gpuZoneEnd.thread, GetThreadHandle() );
+        MemWrite( &item->gpuZoneEnd.queryId, uint16_t( queryId ) );
+        MemWrite( &item->gpuZoneEnd.context, m_ctx->GetId() );
+        Profiler::QueueSerialFinish();
+    }
+
+private:
+    const bool m_active;
+
+    VkCommandBuffer m_cmdbuf;
+    VkCtx* m_ctx;
+};
+
+#if defined TRACY_VK_USE_SYMBOL_TABLE
+void* CreateVkContext( VkInstance instance, VkPhysicalDevice physdev, VkDevice device, VkQueue queue, VkCommandBuffer cmdbuf, PFN_vkGetInstanceProcAddr instanceProcAddr, PFN_vkGetDeviceProcAddr getDeviceProcAddr, bool calibrated = false )
+#else
+void* CreateVkContext( VkPhysicalDevice physdev, VkDevice device, VkQueue queue, VkCommandBuffer cmdbuf, PFN_vkGetPhysicalDeviceCalibrateableTimeDomainsEXT gpdctd, PFN_vkGetCalibratedTimestampsEXT gct )
+#endif
+{
+    auto ctx = (VkCtx*)tracy_malloc( sizeof( VkCtx ) );
+#if defined TRACY_VK_USE_SYMBOL_TABLE
+    new(ctx) VkCtx( instance, physdev, device, queue, cmdbuf, instanceProcAddr, getDeviceProcAddr, calibrated );
+#else
+    new(ctx) VkCtx( physdev, device, queue, cmdbuf, gpdctd, gct );
+#endif
+    return ctx;
+}
+
+#if defined VK_EXT_host_query_reset
+#if defined TRACY_VK_USE_SYMBOL_TABLE
+void* CreateVkContext( VkInstance instance, VkPhysicalDevice physdev, VkDevice device, PFN_vkGetInstanceProcAddr instanceProcAddr, PFN_vkGetDeviceProcAddr getDeviceProcAddr )
+#else
+void* CreateVkContext( VkPhysicalDevice physdev, VkDevice device, PFN_vkResetQueryPoolEXT qpreset, PFN_vkGetPhysicalDeviceCalibrateableTimeDomainsEXT gpdctd, PFN_vkGetCalibratedTimestampsEXT gct )
+#endif
+{
+    auto ctx = (VkCtx*)tracy_malloc( sizeof( VkCtx ) );
+#if defined TRACY_VK_USE_SYMBOL_TABLE
+    new(ctx) VkCtx( instance, physdev, device, instanceProcAddr, getDeviceProcAddr );
+#else
+    new(ctx) VkCtx( physdev, device, qpreset, gpdctd, gct );
+#endif
+    ctx->Name("Vulkan", 6);
+    return ctx;
+}
+#endif
+
+void DestroyVkContext( void* ctx )
+{
+    ((VkCtx*)ctx)->~VkCtx();
+    tracy_free( ctx );
+}
+
+void CollectVkContext( void* ctx, VkCommandBuffer cmdbuf )
+{
+    ((VkCtx*)ctx)->Collect(cmdbuf);
+}
+
+void BeginVkZoneScope(void* zone, void* ctx, VkCommandBuffer cmdbuf, const char* name, size_t nameLen)
+{
+    new(zone) tracy::VkCtxScope{ (tracy::VkCtx*)ctx, 0, 0, 0, 0, 0, name, nameLen, cmdbuf, true };
+}
+
+void EndVkZoneScope(void* zone)
+{
+    ((tracy::VkCtxScope*)zone)->~VkCtxScope();
+}
+
+}
+
+using TracyVkCtx = tracy::VkCtx*;
+
+#if defined TRACY_VK_USE_SYMBOL_TABLE
+#define TracyVkContext( instance, physdev, device, queue, cmdbuf, instanceProcAddr, deviceProcAddr ) tracy::CreateVkContext( instance, physdev, device, queue, cmdbuf, instanceProcAddr, deviceProcAddr );
+#else
+#define TracyVkContext( physdev, device, queue, cmdbuf ) tracy::CreateVkContext( physdev, device, queue, cmdbuf, nullptr, nullptr );
+#endif
+#if defined TRACY_VK_USE_SYMBOL_TABLE
+#define TracyVkContextCalibrated( instance, physdev, device, queue, cmdbuf, instanceProcAddr, deviceProcAddr ) tracy::CreateVkContext( instance, physdev, device, queue, cmdbuf, instanceProcAddr, deviceProcAddr, true );
+#else
+#define TracyVkContextCalibrated( physdev, device, queue, cmdbuf, gpdctd, gct ) tracy::CreateVkContext( physdev, device, queue, cmdbuf, gpdctd, gct );
+#endif
+#if defined VK_EXT_host_query_reset
+#if defined TRACY_VK_USE_SYMBOL_TABLE
+#define TracyVkContextHostCalibrated( instance, physdev, device, instanceProcAddr, deviceProcAddr ) tracy::CreateVkContext( instance, physdev, device, instanceProcAddr, deviceProcAddr );
+#else
+#define TracyVkContextHostCalibrated( physdev, device, qpreset, gpdctd, gct ) tracy::CreateVkContext( physdev, device, qpreset, gpdctd, gct );
+#endif
+#endif
+#define TracyVkDestroy( ctx ) tracy::DestroyVkContext( ctx );
+#define TracyVkContextName( ctx, name, size ) ctx->Name( name, size );
+#if defined TRACY_HAS_CALLSTACK && defined TRACY_CALLSTACK
+#  define TracyVkNamedZone( ctx, varname, cmdbuf, name, active ) static constexpr tracy::SourceLocationData TracyConcat(__tracy_gpu_source_location,TracyLine) { name, TracyFunction,  TracyFile, (uint32_t)TracyLine, 0 }; tracy::VkCtxScope varname( ctx, &TracyConcat(__tracy_gpu_source_location,TracyLine), cmdbuf, TRACY_CALLSTACK, active );
+#  define TracyVkNamedZoneC( ctx, varname, cmdbuf, name, color, active ) static constexpr tracy::SourceLocationData TracyConcat(__tracy_gpu_source_location,TracyLine) { name, TracyFunction,  TracyFile, (uint32_t)TracyLine, color }; tracy::VkCtxScope varname( ctx, &TracyConcat(__tracy_gpu_source_location,TracyLine), cmdbuf, TRACY_CALLSTACK, active );
+#  define TracyVkZone( ctx, cmdbuf, name ) TracyVkNamedZoneS( ctx, ___tracy_gpu_zone, cmdbuf, name, TRACY_CALLSTACK, true )
+#  define TracyVkZoneC( ctx, cmdbuf, name, color ) TracyVkNamedZoneCS( ctx, ___tracy_gpu_zone, cmdbuf, name, color, TRACY_CALLSTACK, true )
+#  define TracyVkZoneTransient( ctx, varname, cmdbuf, name, active ) TracyVkZoneTransientS( ctx, varname, cmdbuf, name, TRACY_CALLSTACK, active )
+#else
+#  define TracyVkNamedZone( ctx, varname, cmdbuf, name, active ) static constexpr tracy::SourceLocationData TracyConcat(__tracy_gpu_source_location,TracyLine) { name, TracyFunction,  TracyFile, (uint32_t)TracyLine, 0 }; tracy::VkCtxScope varname( ctx, &TracyConcat(__tracy_gpu_source_location,TracyLine), cmdbuf, active );
+#  define TracyVkNamedZoneC( ctx, varname, cmdbuf, name, color, active ) static constexpr tracy::SourceLocationData TracyConcat(__tracy_gpu_source_location,TracyLine) { name, TracyFunction,  TracyFile, (uint32_t)TracyLine, color }; tracy::VkCtxScope varname( ctx, &TracyConcat(__tracy_gpu_source_location,TracyLine), cmdbuf, active );
+#  define TracyVkZone( ctx, cmdbuf, name ) TracyVkNamedZone( ctx, ___tracy_gpu_zone, cmdbuf, name, true )
+#  define TracyVkZoneC( ctx, cmdbuf, name, color ) TracyVkNamedZoneC( ctx, ___tracy_gpu_zone, cmdbuf, name, color, true )
+#  define TracyVkZoneTransient( ctx, varname, cmdbuf, name, active ) tracy::VkCtxScope varname( ctx, TracyLine, TracyFile, strlen( TracyFile ), TracyFunction, strlen( TracyFunction ), name, strlen( name ), cmdbuf, active );
+#endif
+#define TracyVkCollect( ctx, cmdbuf ) ctx->Collect( cmdbuf );
+
+#ifdef TRACY_HAS_CALLSTACK
+#  define TracyVkNamedZoneS( ctx, varname, cmdbuf, name, depth, active ) static constexpr tracy::SourceLocationData TracyConcat(__tracy_gpu_source_location,TracyLine) { name, TracyFunction,  TracyFile, (uint32_t)TracyLine, 0 }; tracy::VkCtxScope varname( ctx, &TracyConcat(__tracy_gpu_source_location,TracyLine), cmdbuf, depth, active );
+#  define TracyVkNamedZoneCS( ctx, varname, cmdbuf, name, color, depth, active ) static constexpr tracy::SourceLocationData TracyConcat(__tracy_gpu_source_location,TracyLine) { name, TracyFunction,  TracyFile, (uint32_t)TracyLine, color }; tracy::VkCtxScope varname( ctx, &TracyConcat(__tracy_gpu_source_location,TracyLine), cmdbuf, depth, active );
+#  define TracyVkZoneS( ctx, cmdbuf, name, depth ) TracyVkNamedZoneS( ctx, ___tracy_gpu_zone, cmdbuf, name, depth, true )
+#  define TracyVkZoneCS( ctx, cmdbuf, name, color, depth ) TracyVkNamedZoneCS( ctx, ___tracy_gpu_zone, cmdbuf, name, color, depth, true )
+#  define TracyVkZoneTransientS( ctx, varname, cmdbuf, name, depth, active ) tracy::VkCtxScope varname( ctx, TracyLine, TracyFile, strlen( TracyFile ), TracyFunction, strlen( TracyFunction ), name, strlen( name ), cmdbuf, depth, active );
+#else
+#  define TracyVkNamedZoneS( ctx, varname, cmdbuf, name, depth, active ) TracyVkNamedZone( ctx, varname, cmdbuf, name, active )
+#  define TracyVkNamedZoneCS( ctx, varname, cmdbuf, name, color, depth, active ) TracyVkNamedZoneC( ctx, varname, cmdbuf, name, color, active )
+#  define TracyVkZoneS( ctx, cmdbuf, name, depth ) TracyVkZone( ctx, cmdbuf, name )
+#  define TracyVkZoneCS( ctx, cmdbuf, name, color, depth ) TracyVkZoneC( ctx, cmdbuf, name, color )
+#  define TracyVkZoneTransientS( ctx, varname, cmdbuf, name, depth, active ) TracyVkZoneTransient( ctx, varname, cmdbuf, name, active )
+#endif
+
+#else
+
+// Forward declarations to be used in engine
+namespace tracy
+{
+#if defined TRACY_VK_USE_SYMBOL_TABLE
+    extern void* CreateVkContext(VkInstance instance, VkPhysicalDevice physdev, VkDevice device, VkQueue queue, VkCommandBuffer cmdbuf, PFN_vkGetInstanceProcAddr instanceProcAddr, PFN_vkGetDeviceProcAddr getDeviceProcAddr, bool calibrated = false);
+#else
+    extern void* CreateVkContext(VkPhysicalDevice physdev, VkDevice device, VkQueue queue, VkCommandBuffer cmdbuf, PFN_vkGetPhysicalDeviceCalibrateableTimeDomainsEXT gpdctd, PFN_vkGetCalibratedTimestampsEXT gct);
+#endif
+#if defined VK_EXT_host_query_reset
+#if defined TRACY_VK_USE_SYMBOL_TABLE
+    extern void* CreateVkContext(VkInstance instance, VkPhysicalDevice physdev, VkDevice device, PFN_vkGetInstanceProcAddr instanceProcAddr, PFN_vkGetDeviceProcAddr getDeviceProcAddr);
+#else
+    extern void* CreateVkContext(VkPhysicalDevice physdev, VkDevice device, PFN_vkResetQueryPoolEXT qpreset, PFN_vkGetPhysicalDeviceCalibrateableTimeDomainsEXT gpdctd, PFN_vkGetCalibratedTimestampsEXT gct);
+#endif
+#endif
+    extern void DestroyVkContext(void* ctx);
+    extern void CollectVkContext(void* ctx, VkCommandBuffer cmdbuf);
+    extern void BeginVkZoneScope(void* zone, void* ctx, VkCommandBuffer cmdbuf, const char* name, size_t nameLen);
+    extern void EndVkZoneScope(void* zone);
+}
+
+#endif
+
+#endif
--- a/Source/Tools/Flax.Build/Build/NativeCpp/BuildOptions.cs
+++ b/Source/Tools/Flax.Build/Build/NativeCpp/BuildOptions.cs
@@ -73,7 +73,7 @@ namespace Flax.Build.NativeCpp
    /// <summary>
    /// The native C++ module build settings container.
    /// </summary>
-    public sealed class BuildOptions
+    public sealed class BuildOptions : ICloneable
    {
        /// <summary>
        /// The target that builds this module.
@@ -442,5 +442,26 @@ namespace Flax.Build.NativeCpp
                SourcePaths.Clear();
            }
        }
+
+        /// <inheritdoc />
+        public object Clone()
+        {
+            var clone = new BuildOptions
+            {
+                Target = Target,
+                Platform = Platform,
+                Toolchain = Toolchain,
+                Architecture = Architecture,
+                Configuration = Configuration,
+                CompileEnv = (CompileEnvironment)CompileEnv.Clone(),
+                LinkEnv = (LinkEnvironment)LinkEnv.Clone(),
+                IntermediateFolder = IntermediateFolder,
+                OutputFolder = OutputFolder,
+                WorkingDirectory = WorkingDirectory,
+                HotReloadPostfix = HotReloadPostfix,
+                Flags = Flags,
+            };
+            return clone;
+        }
    }
 }