UnrealEngineUWP/Engine/Source/Runtime/RenderCore/Private/RenderTargetPool.cpp

// Copyright 1998-2019 Epic Games, Inc. All Rights Reserved.

/*=============================================================================
	RenderTargetPool.cpp: Scene render target pool manager.
=============================================================================*/

#include "RenderTargetPool.h"
#include "RHIStaticStates.h"

#if !(UE_BUILD_SHIPPING || UE_BUILD_TEST) && 0
	#include "ClearQuad.h"
#endif

/** The global render targets pool. */
TGlobalResource<FRenderTargetPool> GRenderTargetPool;

DEFINE_LOG_CATEGORY_STATIC(LogRenderTargetPool, Warning, All);

static void DumpRenderTargetPoolMemory(FOutputDevice& OutputDevice)
{
	GRenderTargetPool.DumpMemoryUsage(OutputDevice);
}
static FAutoConsoleCommandWithOutputDevice GDumpRenderTargetPoolMemoryCmd(
	TEXT("r.DumpRenderTargetPoolMemory"),
	TEXT("Dump allocation information for the render target pool."),
	FConsoleCommandWithOutputDeviceDelegate::CreateStatic(DumpRenderTargetPoolMemory)
	);

void RenderTargetPoolEvents(const TArray<FString>& Args)
{
	uint32 SizeInKBThreshold = -1;
	if(Args.Num() && Args[0].IsNumeric())
	{
		SizeInKBThreshold = FCString::Atof(*Args[0]);
	}

	if(SizeInKBThreshold != -1)
	{
		UE_LOG(LogRenderTargetPool, Display, TEXT("r.DumpRenderTargetPoolEvents is now enabled, use r.DumpRenderTargetPoolEvents ? for help"));

		GRenderTargetPool.EventRecordingSizeThreshold = SizeInKBThreshold;
		GRenderTargetPool.bStartEventRecordingNextTick = true;
	}
	else
	{
		GRenderTargetPool.DisableEventDisplay();

		UE_LOG(LogRenderTargetPool, Display, TEXT("r.DumpRenderTargetPoolEvents is now disabled, use r.DumpRenderTargetPoolEvents <SizeInKB> to enable or r.DumpRenderTargetPoolEvents ? for help"));
	}
}

// CVars and commands
static FAutoConsoleCommand GRenderTargetPoolEventsCmd(
	TEXT("r.RenderTargetPool.Events"),
	TEXT("Visualize the render target pool events over time in one frame. Optional parameter defines threshold in KB.\n")
	TEXT("To disable the view use the command without any parameter"),
	FConsoleCommandWithArgsDelegate::CreateStatic(RenderTargetPoolEvents)
	);

#if !(UE_BUILD_SHIPPING || UE_BUILD_TEST)
static TAutoConsoleVariable<int32> CVarClobberRenderTarget(
	TEXT("r.Test.ClobberRenderRarget"),
	0,
	TEXT("Clears the texture returned by the rendertarget pool with a special color\n")
	TEXT("so we can see better which passes would need to clear. Doesn't work on volume textures and non rendertargets yet.\n")
	TEXT(" 0:off (default);\n")
	TEXT(" 1: 1000 on RGBA channels;\n")
	TEXT(" 2: NaN on RGBA channels;\n")
	TEXT(" 3: +INFINITY on RGBA channels.\n"),
	ECVF_Cheat | ECVF_RenderThreadSafe);
#endif

static TAutoConsoleVariable<int32> CVarAllowMultipleAliasingDiscardsPerFrame(
	TEXT("r.RenderTargetPool.AllowMultipleAliasingDiscardsPerFrame"),
	0,
	TEXT("If enabled, allows rendertargets to be discarded and reacquired in the same frame.\n")
	TEXT("This should give better aliasing efficiency, but carries some RHIthread/GPU performance overhead\n")
	TEXT("with some RHIs (due to additional commandlist flushes)\n")
	TEXT(" 0:off (default), 1:on"),
	ECVF_Cheat | ECVF_RenderThreadSafe);

static TAutoConsoleVariable<int32> CVarRtPoolTransientMode(
	TEXT("r.RenderTargetPool.TransientAliasingMode"),
	2,
	TEXT("Enables transient resource aliasing for rendertargets. Used only if GSupportsTransientResourceAliasing is true.\n")
	TEXT("0 : Disabled\n")
	TEXT("1 : enable transient resource aliasing for fastVRam rendertargets\n")
	TEXT("2 : enable transient resource aliasing for fastVRam rendertargets and those with a Transient hint. Best for memory usage - has some GPU cost (~0.2ms)\n")
	TEXT("3 : enable transient resource aliasing for ALL rendertargets (not recommended)\n"),
	ECVF_RenderThreadSafe);

bool FRenderTargetPool::IsEventRecordingEnabled() const
{
#if !(UE_BUILD_SHIPPING || UE_BUILD_TEST)
	return bEventRecordingStarted && bEventRecordingActive; 
#else
	return false;
#endif
}

IPooledRenderTarget* FRenderTargetPoolEvent::GetValidatedPointer() const
{
	int32 Index = GRenderTargetPool.FindIndex(Pointer);

	if(Index >= 0)
	{
		return Pointer;
	}

	return 0;
}

bool FRenderTargetPoolEvent::NeedsDeallocEvent()
{
	if(GetEventType() == ERTPE_Alloc)
	{
		if(Pointer)
		{
			IPooledRenderTarget* ValidPointer = GetValidatedPointer();
			if(!ValidPointer || ValidPointer->IsFree())
			{
				Pointer = 0;
				return true;
			}
		}
	}

	return false;
}


static uint32 ComputeSizeInKB(FPooledRenderTarget& Element)
{
	return (Element.ComputeMemorySize() + 1023) / 1024;
}

FRenderTargetPool::FRenderTargetPool()
	: AllocationLevelInKB(0)
	, bCurrentlyOverBudget(false)
	, bStartEventRecordingNextTick(false)
	, EventRecordingSizeThreshold(0)
	, bEventRecordingActive(false)
	, bEventRecordingStarted(false)
	, CurrentEventRecordingTime(0)
{
}

// Logic for determining whether to make a rendertarget transient
bool FRenderTargetPool::DoesTargetNeedTransienceOverride(const FPooledRenderTargetDesc& InputDesc, ERenderTargetTransience TransienceHint) const
{
	if (!GSupportsTransientResourceAliasing)
	{
		return false;
	}
	int32 AliasingMode = CVarRtPoolTransientMode.GetValueOnRenderThread();

	// We only override transience if aliasing is supported and enabled, the format is suitable, and the target is not already transient
	if (AliasingMode > 0 &&
	  	(InputDesc.TargetableFlags & (TexCreate_RenderTargetable | TexCreate_DepthStencilTargetable | TexCreate_UAV)) && 
		((InputDesc.Flags & TexCreate_Transient) == 0) )
	{
		if (AliasingMode == 1)
		{
			// Mode 1: Only make FastVRAM rendertargets transient
			if (InputDesc.Flags & TexCreate_FastVRAM)
			{
				return true;
			}
		}
		else if (AliasingMode == 2)
		{
			// Mode 2: Make fastvram and ERenderTargetTransience::Transient rendertargets transient
			if (InputDesc.Flags & TexCreate_FastVRAM || TransienceHint == ERenderTargetTransience::Transient)
			{
				return true;
			}
		}
		else if (AliasingMode == 3)
		{
			// Mode 3 : All rendertargets are transient
			return true;
		}
	}
	return false;
}

void FRenderTargetPool::TransitionTargetsWritable(FRHICommandListImmediate& RHICmdList)
{
	QUICK_SCOPE_CYCLE_COUNTER(STAT_RenderTargetPoolTransition);
	check(IsInRenderingThread());
	WaitForTransitionFence();
	
	TransitionTargets.Reset();	

	for (int32 i = 0; i < PooledRenderTargets.Num(); ++i)
	{
		FPooledRenderTarget* PooledRT = PooledRenderTargets[i];
		if (PooledRT && PooledRT->GetDesc().AutoWritable)
		{
			FTextureRHIParamRef RenderTarget = PooledRT->GetRenderTargetItem().TargetableTexture;
			if (RenderTarget)
			{				
				TransitionTargets.Add(RenderTarget);
			}
		}
	}

	if (TransitionTargets.Num() > 0)
	{
		RHICmdList.TransitionResourceArrayNoCopy(EResourceTransitionAccess::EWritable, TransitionTargets);
		if (IsRunningRHIInSeparateThread())
		{
			TransitionFence = RHICmdList.RHIThreadFence(false);
		}
	}
}

void FRenderTargetPool::WaitForTransitionFence()
{
	QUICK_SCOPE_CYCLE_COUNTER(STAT_RenderTargetPoolTransitionWait);
	check(IsInRenderingThread());
	if (TransitionFence)
	{		
		check(IsInRenderingThread());		
		FRHICommandListExecutor::WaitOnRHIThreadFence(TransitionFence);
		TransitionFence = nullptr;		
	}
	TransitionTargets.Reset();
	DeferredDeleteArray.Reset();
}

#if 1 // TODO(RDG): Needs dependency on UtilityShaders to reimplement ClobberAllocatedRenderTarget functionality, or just move the clearing shaders in RenderCore?

static void ClobberAllocatedRenderTarget(FRHICommandList& RHICmdList, const FPooledRenderTargetDesc& Desc, IPooledRenderTarget* Out)
{

}


#elif !(UE_BUILD_SHIPPING || UE_BUILD_TEST)

static void ClobberAllocatedRenderTarget(FRHICommandList& RHICmdList, const FPooledRenderTargetDesc& Desc, IPooledRenderTarget* Out)
{
	int32 ClearId = CVarClobberRenderTarget.GetValueOnRenderThread();

	if (!ClearId)
	{
		return;
	}
	if (Desc.bIsCubemap || Desc.bIsArray)
	{
		return;
	}

	FLinearColor Color = FLinearColor(1000, 1000, 1000, 1000);

	if (ClearId == 2)
	{
		Color = FLinearColor(NAN, NAN, NAN, NAN);
	}
	else if (ClearId == 3)
	{
		Color = FLinearColor(INFINITY, INFINITY, INFINITY, INFINITY);
	}

	// TODO(RDG): draw events will need to be rewritten with render graph builder instead of RHICmdList.
	//SCOPED_DRAW_EVENT(RHICmdList, ClobberAllocatedRenderTarget);

	if (Out->GetDesc().TargetableFlags & TexCreate_RenderTargetable)
	{
		SetRenderTarget(RHICmdList, Out->GetRenderTargetItem().TargetableTexture, FTextureRHIRef());
		DrawClearQuad(RHICmdList, Color);
	}
	else if (Out->GetDesc().TargetableFlags & TexCreate_UAV)
	{
		ClearUAV(RHICmdList, Out->GetRenderTargetItem(), Color);
	}

	if (Desc.TargetableFlags & TexCreate_DepthStencilTargetable)
	{
		SetRenderTarget(RHICmdList, FTextureRHIRef(), Out->GetRenderTargetItem().TargetableTexture);
		DrawClearQuad(RHICmdList, false, FLinearColor::Black, true, 0.0f, true, 0);
	}
}

#endif

bool FRenderTargetPool::FindFreeElement(FRHICommandList& RHICmdList, const FPooledRenderTargetDesc& InputDesc, TRefCountPtr<IPooledRenderTarget> &Out, const TCHAR* InDebugName, bool bDoWritableBarrier, ERenderTargetTransience TransienceHint)
{
	check(IsInRenderingThread());

	if(!InputDesc.IsValid())
	{
		// no need to do anything
		return true;
	}

	// Querying a render target that have no mip levels makes no sens.
	check(InputDesc.NumMips > 0);

	// Make sure if requesting a depth format that the clear value is correct
	ensure(!IsDepthOrStencilFormat(InputDesc.Format) || (InputDesc.ClearValue.ColorBinding == EClearBinding::ENoneBound || InputDesc.ClearValue.ColorBinding == EClearBinding::EDepthStencilBound));

	// If we're doing aliasing, we may need to override Transient flags, depending on the input format and mode
	FPooledRenderTargetDesc ModifiedDesc;
	bool bMakeTransient = DoesTargetNeedTransienceOverride(InputDesc, TransienceHint);
	if (bMakeTransient)
			{
				ModifiedDesc = InputDesc;
				ModifiedDesc.Flags |= TexCreate_Transient;
	}

	// Override the descriptor if necessary
	const FPooledRenderTargetDesc& Desc = bMakeTransient ? ModifiedDesc : InputDesc;


	// if we can keep the current one, do that
	if(Out)
	{
		FPooledRenderTarget* Current = (FPooledRenderTarget*)Out.GetReference();

		check(!Current->IsSnapshot());

		const bool bExactMatch = true;

		if(Out->GetDesc().Compare(Desc, bExactMatch))
		{
			// we can reuse the same, but the debug name might have changed
			Current->Desc.DebugName = InDebugName;
			RHIBindDebugLabelName(Current->GetRenderTargetItem().TargetableTexture, InDebugName);
			check(!Out->IsFree());
			#if !(UE_BUILD_SHIPPING || UE_BUILD_TEST)
				ClobberAllocatedRenderTarget(RHICmdList, Desc, Out);
			#endif
			return true;
		}
		else
		{
			// release old reference, it might free a RT we can use
			Out = 0;

			if(Current->IsFree())
			{
				AllocationLevelInKB -= ComputeSizeInKB(*Current);

				int32 Index = FindIndex(Current);

				check(Index >= 0);

				// we don't use Remove() to not shuffle around the elements for better transparency on RenderTargetPoolEvents
				PooledRenderTargets[Index] = 0;

				VerifyAllocationLevel();
			}
		}
	}

	int32 AliasingMode = CVarRtPoolTransientMode.GetValueOnRenderThread();
	FPooledRenderTarget* Found = 0;
	uint32 FoundIndex = -1;
	bool bReusingExistingTarget = false;
	// try to find a suitable element in the pool
	{
		//don't spend time doing 2 passes if the platform doesn't support fastvram
		uint32 PassCount = 1;
		if (AliasingMode == 0)
		{
			if ((Desc.Flags & TexCreate_FastVRAM) && FPlatformMemory::SupportsFastVRAMMemory() )
			{
				PassCount = 2;
			}
		}

		bool bAllowMultipleDiscards = ( CVarAllowMultipleAliasingDiscardsPerFrame.GetValueOnRenderThread() != 0 );
		// first we try exact, if that fails we try without TexCreate_FastVRAM
		// (easily we can run out of VRam, if this search becomes a performance problem we can optimize or we should use less TexCreate_FastVRAM)
		for(uint32 Pass = 0; Pass < PassCount; ++Pass)
		{
			bool bExactMatch = (Pass == 0); //-V547
    
			for(uint32 i = 0, Num = (uint32)PooledRenderTargets.Num(); i < Num; ++i)
			{
				FPooledRenderTarget* Element = PooledRenderTargets[i];
    
				if(Element && Element->IsFree() && Element->GetDesc().Compare(Desc, bExactMatch))
				{
					if ( ( Desc.Flags & TexCreate_Transient ) && bAllowMultipleDiscards == false && Element->HasBeenDiscardedThisFrame() )
					{
						// We can't re-use transient resources if they've already been discarded this frame
						continue;
					}
					check(!Element->IsSnapshot());
					Found = Element;
					FoundIndex = i;
					bReusingExistingTarget = true;
					goto Done;
				}
			}
		}		
	}
Done:

	if(!Found)
	{		
		UE_LOG(LogRenderTargetPool, Display, TEXT("%d MB, NewRT %s %s"), (AllocationLevelInKB + 1023) / 1024, *Desc.GenerateInfoString(), InDebugName);

		// not found in the pool, create a new element
		Found = new FPooledRenderTarget(Desc, this);

		PooledRenderTargets.Add(Found);
		
		// TexCreate_UAV should be used on Desc.TargetableFlags
		check(!(Desc.Flags & TexCreate_UAV));
		// TexCreate_FastVRAM should be used on Desc.Flags
		ensure(!(Desc.TargetableFlags & TexCreate_FastVRAM));

		FRHIResourceCreateInfo CreateInfo(Desc.ClearValue);
		CreateInfo.DebugName = InDebugName;

		if(Desc.TargetableFlags & (TexCreate_RenderTargetable | TexCreate_DepthStencilTargetable | TexCreate_UAV))
		{
			if(Desc.Is2DTexture())
			{
				if (!Desc.IsArray())
				{
					RHICreateTargetableShaderResource2D(
						Desc.Extent.X,
						Desc.Extent.Y,
						Desc.Format,
						Desc.NumMips,
						Desc.Flags,
						Desc.TargetableFlags,
						Desc.bForceSeparateTargetAndShaderResource,
						CreateInfo,
						(FTexture2DRHIRef&)Found->RenderTargetItem.TargetableTexture,
						(FTexture2DRHIRef&)Found->RenderTargetItem.ShaderResourceTexture,
						Desc.NumSamples
					);
				}
				else
				{
					RHICreateTargetableShaderResource2DArray(
						Desc.Extent.X,
						Desc.Extent.Y,
						Desc.ArraySize,
						Desc.Format,
						Desc.NumMips,
						Desc.Flags,
						Desc.TargetableFlags,
						CreateInfo,
						(FTexture2DArrayRHIRef&)Found->RenderTargetItem.TargetableTexture,
						(FTexture2DArrayRHIRef&)Found->RenderTargetItem.ShaderResourceTexture,
						Desc.NumSamples
					);
				}

				if (GSupportsRenderTargetWriteMask && Desc.bCreateRenderTargetWriteMask)
				{
					Found->RenderTargetItem.RTWriteMaskDataBufferRHI = RHICreateRTWriteMaskBuffer((FTexture2DRHIRef&)Found->RenderTargetItem.TargetableTexture);
					Found->RenderTargetItem.RTWriteMaskBufferRHI_SRV = RHICreateShaderResourceView(Found->RenderTargetItem.RTWriteMaskDataBufferRHI);
				}

				if( Desc.NumMips > 1 )
				{
					Found->RenderTargetItem.MipSRVs.SetNum( Desc.NumMips );
					for( uint16 i = 0; i < Desc.NumMips; i++ )
					{
						Found->RenderTargetItem.MipSRVs[i] = RHICreateShaderResourceView( (FTexture2DRHIRef&)Found->RenderTargetItem.ShaderResourceTexture, i );
					}
				}
			}
			else if(Desc.Is3DTexture())
			{
				Found->RenderTargetItem.ShaderResourceTexture = RHICreateTexture3D(
					Desc.Extent.X,
					Desc.Extent.Y,
					Desc.Depth,
					Desc.Format,
					Desc.NumMips,
					Desc.Flags | Desc.TargetableFlags,
					CreateInfo);

				// similar to RHICreateTargetableShaderResource2D
				Found->RenderTargetItem.TargetableTexture = Found->RenderTargetItem.ShaderResourceTexture;
			}
			else
			{
				check(Desc.IsCubemap());
				if(Desc.IsArray())
				{
					RHICreateTargetableShaderResourceCubeArray(
						Desc.Extent.X,
						Desc.ArraySize,
						Desc.Format,
						Desc.NumMips,
						Desc.Flags,
						Desc.TargetableFlags,
						false,
						CreateInfo,
						(FTextureCubeRHIRef&)Found->RenderTargetItem.TargetableTexture,
						(FTextureCubeRHIRef&)Found->RenderTargetItem.ShaderResourceTexture
						);
				}
				else
				{
					RHICreateTargetableShaderResourceCube(
						Desc.Extent.X,
						Desc.Format,
						Desc.NumMips,
						Desc.Flags,
						Desc.TargetableFlags,
						false,
						CreateInfo,
						(FTextureCubeRHIRef&)Found->RenderTargetItem.TargetableTexture,
						(FTextureCubeRHIRef&)Found->RenderTargetItem.ShaderResourceTexture
						);

					if( Desc.NumMips > 1 )
					{
						Found->RenderTargetItem.MipSRVs.SetNum( Desc.NumMips );
						for( uint16 i = 0; i < Desc.NumMips; i++ )
						{
							Found->RenderTargetItem.MipSRVs[i] = RHICreateShaderResourceView( (FTextureCubeRHIRef&)Found->RenderTargetItem.ShaderResourceTexture, i );
						}
					}
				}

			}

			RHIBindDebugLabelName(Found->RenderTargetItem.TargetableTexture, InDebugName);
		}
		else
		{
			if(Desc.Is2DTexture())
			{
				// this is useful to get a CPU lockable texture through the same interface
				Found->RenderTargetItem.ShaderResourceTexture = RHICreateTexture2D(
					Desc.Extent.X,
					Desc.Extent.Y,
					Desc.Format,
					Desc.NumMips,
					Desc.NumSamples,
					Desc.Flags,
					CreateInfo);
			}
			else if(Desc.Is3DTexture())
			{
				Found->RenderTargetItem.ShaderResourceTexture = RHICreateTexture3D(
					Desc.Extent.X,
					Desc.Extent.Y,
					Desc.Depth,
					Desc.Format,
					Desc.NumMips,
					Desc.Flags,
					CreateInfo);
			}
			else 
			{
				check(Desc.IsCubemap());
				if(Desc.IsArray())
				{
					FTextureCubeRHIRef CubeTexture = RHICreateTextureCubeArray(Desc.Extent.X,Desc.ArraySize,Desc.Format,Desc.NumMips,Desc.Flags | Desc.TargetableFlags | TexCreate_ShaderResource,CreateInfo);
					Found->RenderTargetItem.TargetableTexture = Found->RenderTargetItem.ShaderResourceTexture = CubeTexture;
				}
				else
				{
					FTextureCubeRHIRef CubeTexture = RHICreateTextureCube(Desc.Extent.X,Desc.Format,Desc.NumMips,Desc.Flags | Desc.TargetableFlags | TexCreate_ShaderResource,CreateInfo);
					Found->RenderTargetItem.TargetableTexture = Found->RenderTargetItem.ShaderResourceTexture = CubeTexture;
				}
			}

			RHIBindDebugLabelName(Found->RenderTargetItem.ShaderResourceTexture, InDebugName);
		}

		if(Desc.TargetableFlags & TexCreate_UAV)
		{
			// The render target desc is invalid if a UAV is requested with an RHI that doesn't support the high-end feature level.
			check(GMaxRHIFeatureLevel == ERHIFeatureLevel::SM5);
			Found->RenderTargetItem.MipUAVs.Reserve(Desc.NumMips);
			for (uint32 MipLevel = 0; MipLevel < Desc.NumMips; MipLevel++)
			{
				Found->RenderTargetItem.MipUAVs.Add(RHICreateUnorderedAccessView(Found->RenderTargetItem.TargetableTexture, MipLevel));
			}

			Found->RenderTargetItem.UAV = Found->RenderTargetItem.MipUAVs[0];
		}

		AllocationLevelInKB += ComputeSizeInKB(*Found);
		VerifyAllocationLevel();

		FoundIndex = PooledRenderTargets.Num() - 1;

		Found->Desc.DebugName = InDebugName;
	}

	#if !(UE_BUILD_SHIPPING || UE_BUILD_TEST)
		ClobberAllocatedRenderTarget(RHICmdList, Desc, Found);
	#endif

	check(Found->IsFree());
	check(!Found->IsSnapshot());

	Found->Desc.DebugName = InDebugName;
	Found->UnusedForNFrames = 0;

	AddAllocEvent(FoundIndex, Found);

	uint32 OriginalNumRefs = Found->GetRefCount();

	// assign to the reference counted variable
	Out = Found;

	check(!Found->IsFree());

	// Only referenced by the pool, map the physical pages
	if (OriginalNumRefs == 1 && Found->GetRenderTargetItem().TargetableTexture != nullptr)
	{
		RHIAcquireTransientResource(Found->GetRenderTargetItem().TargetableTexture);
	}

	if (bReusingExistingTarget)
	{
		if (bDoWritableBarrier)
		{
			RHICmdList.TransitionResource(EResourceTransitionAccess::EWritable, Found->GetRenderTargetItem().TargetableTexture);
		}

	}

	// Transient RTs have to be targettable
	check( ( Desc.Flags & TexCreate_Transient ) == 0 || Found->GetRenderTargetItem().TargetableTexture != nullptr );

	return false;
}

void FRenderTargetPool::CreateUntrackedElement(const FPooledRenderTargetDesc& Desc, TRefCountPtr<IPooledRenderTarget> &Out, const FSceneRenderTargetItem& Item)
{
	check(IsInRenderingThread());

	Out = 0;

	// not found in the pool, create a new element
	FPooledRenderTarget* Found = new FPooledRenderTarget(Desc, NULL);

	Found->RenderTargetItem = Item;
	check(!Found->IsSnapshot());

	// assign to the reference counted variable
	Out = Found;
}

IPooledRenderTarget* FRenderTargetPool::MakeSnapshot(const TRefCountPtr<IPooledRenderTarget>& In)
{
	check(IsInRenderingThread());
	FPooledRenderTarget* NewSnapshot = nullptr;
	if (In.GetReference())
	{
		NewSnapshot = new (FMemStack::Get()) FPooledRenderTarget(*static_cast<FPooledRenderTarget*>(In.GetReference()));
		PooledRenderTargetSnapshots.Add(NewSnapshot);
	}
	return NewSnapshot;
}


void FRenderTargetPool::GetStats(uint32& OutWholeCount, uint32& OutWholePoolInKB, uint32& OutUsedInKB) const
{
	OutWholeCount = (uint32)PooledRenderTargets.Num();
	OutUsedInKB = 0;
	OutWholePoolInKB = 0;
		
	for(uint32 i = 0; i < (uint32)PooledRenderTargets.Num(); ++i)
	{
		FPooledRenderTarget* Element = PooledRenderTargets[i];

		if(Element)
		{
			check(!Element->IsSnapshot());
			uint32 SizeInKB = ComputeSizeInKB(*Element);

			OutWholePoolInKB += SizeInKB;

			if(!Element->IsFree())
			{
				OutUsedInKB += SizeInKB;
			}
		}
	}

	// if this triggers uncomment the code in VerifyAllocationLevel() and debug the issue, we might leak memory or not release when we could
	ensure(AllocationLevelInKB == OutWholePoolInKB);
}

void FRenderTargetPool::AddPhaseEvent(const TCHAR *InPhaseName)
{
	if(IsEventRecordingEnabled())
	{
		AddDeallocEvents();

		const FString* LastName = GetLastEventPhaseName();

		if(!LastName || *LastName != InPhaseName)
		{
			if(CurrentEventRecordingTime)
			{
				// put a break to former data
				++CurrentEventRecordingTime;
			}

			FRenderTargetPoolEvent NewEvent(InPhaseName, CurrentEventRecordingTime);

			RenderTargetPoolEvents.Add(NewEvent);
		}
	}
}

const FString* FRenderTargetPool::GetLastEventPhaseName()
{
	// could be optimized but this is a debug view

	// start from the end for better performance
	for(int32 i = RenderTargetPoolEvents.Num() - 1; i >= 0; --i)
	{
		const FRenderTargetPoolEvent* Event = &RenderTargetPoolEvents[i];

		if(Event->GetEventType() == ERTPE_Phase)
		{
			return &Event->GetPhaseName();
		}
	}

	return 0;
}

FRenderTargetPool::SMemoryStats FRenderTargetPool::ComputeView()
{
	SMemoryStats MemoryStats;

#if !(UE_BUILD_SHIPPING || UE_BUILD_TEST)
	{
		struct FRTPColumn
		{
			// index into the column, -1 if this is no valid column
			uint32 PoolEntryId;
			// for sorting
			uint64 SizeInBytes;
			// for sorting
			bool bVRam;

			// default constructor
			FRTPColumn()
				: PoolEntryId(-1)
				, SizeInBytes(0)
			{
			}

			// constructor
			FRTPColumn(const FRenderTargetPoolEvent& Event)
				: PoolEntryId(Event.GetPoolEntryId())
				, bVRam((Event.GetDesc().Flags & TexCreate_FastVRAM) != 0)
			{
				 SizeInBytes = Event.GetSizeInBytes();
			}

			// sort criteria
			bool operator <(const FRTPColumn& rhs) const
			{
				// sort VRam first (only matters on XboxOne but nice to always see it)
// sorting only useful for XboxOne		if(bVRam != rhs.bVRam) return bVRam > rhs.bVRam;

				// we want the large ones first
				return SizeInBytes > rhs.SizeInBytes;
			}
		};

		TArray<FRTPColumn> Colums;

		// generate Colums
		for(int32 i = 0, Num = RenderTargetPoolEvents.Num(); i < Num; i++)
		{
			FRenderTargetPoolEvent* Event = &RenderTargetPoolEvents[i];

			if(Event->GetEventType() == ERTPE_Alloc)
			{
				uint32 PoolEntryId = Event->GetPoolEntryId();

				if(PoolEntryId >= (uint32)Colums.Num())
				{
					Colums.SetNum(PoolEntryId + 1);
				}

				Colums[PoolEntryId] = FRTPColumn(*Event);
			}
		}

		Colums.Sort();

		{
			uint32 ColumnX = 0;

			for(int32 ColumnIndex = 0, ColumnsNum = Colums.Num(); ColumnIndex < ColumnsNum; ++ColumnIndex)
			{
				const FRTPColumn& RTPColumn = Colums[ColumnIndex];

				uint32 ColumnSize = RTPColumn.SizeInBytes;

				// hide columns that are too small to make a difference (e.g. <1 MB)
				if(RTPColumn.SizeInBytes <= EventRecordingSizeThreshold * 1024)
				{
					ColumnSize = 0;
				}
				else
				{
					MemoryStats.DisplayedUsageInBytes += RTPColumn.SizeInBytes;

					// give an entry some size to be more UI friendly (if we get mouse UI for zooming in we might not want that any more)
					ColumnSize = FMath::Max((uint32)(1024 * 1024), ColumnSize);
				}

				MemoryStats.TotalColumnSize += ColumnSize;
				MemoryStats.TotalUsageInBytes += RTPColumn.SizeInBytes;
				
				for(int32 EventIndex = 0, PoolEventsNum = RenderTargetPoolEvents.Num(); EventIndex < PoolEventsNum; EventIndex++)
				{
					FRenderTargetPoolEvent* Event = &RenderTargetPoolEvents[EventIndex];

					if(Event->GetEventType() != ERTPE_Phase)
					{
						uint32 PoolEntryId = Event->GetPoolEntryId();

						if(RTPColumn.PoolEntryId == PoolEntryId)
						{
							Event->SetColumn(ColumnIndex, ColumnX, ColumnSize);
						}
					}
				}
				ColumnX += ColumnSize;
			}
		}
	}
#endif

	return MemoryStats;
}

void FRenderTargetPool::UpdateElementSize(const TRefCountPtr<IPooledRenderTarget>& Element, const uint32 OldElementSize)
{
	check(Element.IsValid() && FindIndex(&(*Element)) >= 0);
	AllocationLevelInKB -= (OldElementSize + 1023) / 1024;
	AllocationLevelInKB += (Element->ComputeMemorySize() + 1023) / 1024;
}

void FRenderTargetPool::AddDeallocEvents()
{
	check(IsInRenderingThread());

	bool bWorkWasDone = false;

	for(uint32 i = 0, Num = (uint32)RenderTargetPoolEvents.Num(); i < Num; ++i)
	{
		FRenderTargetPoolEvent& Event = RenderTargetPoolEvents[i];

		if(Event.NeedsDeallocEvent())
		{
			FRenderTargetPoolEvent NewEvent(Event.GetPoolEntryId(), CurrentEventRecordingTime);

			// for convenience - is actually redundant
			NewEvent.SetDesc(Event.GetDesc());

			RenderTargetPoolEvents.Add(NewEvent);
			bWorkWasDone = true;
		}
	}

	if(bWorkWasDone)
	{
		++CurrentEventRecordingTime;
	}
}

void FRenderTargetPool::AddAllocEvent(uint32 InPoolEntryId, FPooledRenderTarget* In)
{
	check(In);

	if(IsEventRecordingEnabled())
	{
		AddDeallocEvents();

		check(IsInRenderingThread());

		FRenderTargetPoolEvent NewEvent(InPoolEntryId, CurrentEventRecordingTime++, In);

		RenderTargetPoolEvents.Add(NewEvent);
	}
}


void FRenderTargetPool::AddAllocEventsFromCurrentState()
{
	if(!IsEventRecordingEnabled())
	{
		return;
	}

	check(IsInRenderingThread());

	bool bWorkWasDone = false;

	for(uint32 i = 0; i < (uint32)PooledRenderTargets.Num(); ++i)
	{
		FPooledRenderTarget* Element = PooledRenderTargets[i];

		if(Element && !Element->IsFree())
		{
			FRenderTargetPoolEvent NewEvent(i, CurrentEventRecordingTime, Element);

			RenderTargetPoolEvents.Add(NewEvent);
			bWorkWasDone = true;
		}
	}

	if(bWorkWasDone)
	{
		++CurrentEventRecordingTime;
	}
}

void FRenderTargetPool::TickPoolElements()
{
	check(IsInRenderingThread());
	WaitForTransitionFence();

	if(bStartEventRecordingNextTick)
	{
		bStartEventRecordingNextTick = false;
		bEventRecordingStarted = true;
	}

	uint32 MinimumPoolSizeInKB;
	{
		static const auto CVar = IConsoleManager::Get().FindTConsoleVariableDataInt(TEXT("r.RenderTargetPoolMin"));

		MinimumPoolSizeInKB = FMath::Clamp(CVar->GetValueOnRenderThread(), 0, 2000) * 1024;
	}

	CompactPool();

	for(uint32 i = 0; i < (uint32)PooledRenderTargets.Num(); ++i)
	{
		FPooledRenderTarget* Element = PooledRenderTargets[i];

		if(Element)
		{
			check(!Element->IsSnapshot());
			Element->OnFrameStart();
		}
	}

	// we need to release something, take the oldest ones first
	while(AllocationLevelInKB > MinimumPoolSizeInKB)
	{
		// -1: not set
		int32 OldestElementIndex = -1;

		// find oldest element we can remove
		for(uint32 i = 0, Num = (uint32)PooledRenderTargets.Num(); i < Num; ++i)
		{
			FPooledRenderTarget* Element = PooledRenderTargets[i];

			if(Element && Element->UnusedForNFrames > 2)
			{
				if(OldestElementIndex != -1)
				{
					if(PooledRenderTargets[OldestElementIndex]->UnusedForNFrames < Element->UnusedForNFrames)
					{
						OldestElementIndex = i;
					}
				}
				else
				{
					OldestElementIndex = i;
				}
			}
		}

		if(OldestElementIndex != -1)
		{
			AllocationLevelInKB -= ComputeSizeInKB(*PooledRenderTargets[OldestElementIndex]);

			// we assume because of reference counting the resource gets released when not needed any more
			// we don't use Remove() to not shuffle around the elements for better transparency on RenderTargetPoolEvents
			PooledRenderTargets[OldestElementIndex] = 0;

			VerifyAllocationLevel();
		}
		else
		{
			// There is no element we can remove but we are over budget, better we log that.
			// Options:
			//   * Increase the pool
			//   * Reduce rendering features or resolution
			//   * Investigate allocations, order or reusing other render targets can help
			//   * Ignore (editor case, might start using slow memory which can be ok)
			if(!bCurrentlyOverBudget)
			{
				UE_CLOG(IsRunningClientOnly(), LogRenderTargetPool, Warning, TEXT("r.RenderTargetPoolMin exceeded %d/%d MB (ok in editor, bad on fixed memory platform)"), (AllocationLevelInKB + 1023) / 1024, MinimumPoolSizeInKB / 1024);
				bCurrentlyOverBudget = true;
			}
			// at this point we need to give up
			break;
		}

/*	
	// confused more than it helps (often a name is used on two elements in the pool and some pool elements are not rendered to this frame)
	else
	{
		// initial state of a render target (e.g. Velocity@0)
		GVisualizeTexture.SetCheckPoint(Element);
*/	}

	if(AllocationLevelInKB <= MinimumPoolSizeInKB)
	{
		if(bCurrentlyOverBudget)
		{
			UE_LOG(LogRenderTargetPool, Display, TEXT("r.RenderTargetPoolMin resolved %d/%d MB"), (AllocationLevelInKB + 1023) / 1024, MinimumPoolSizeInKB / 1024);
			bCurrentlyOverBudget = false;
		}
	}

//	CompactEventArray();

	AddPhaseEvent(TEXT("FromLastFrame"));
	AddAllocEventsFromCurrentState();
	AddPhaseEvent(TEXT("Rendering"));

#if STATS
	uint32 Count, SizeKB, UsedKB;
	GetStats(Count, SizeKB, UsedKB);
	SET_MEMORY_STAT(STAT_RenderTargetPoolSize, int64(SizeKB) * 1024ll);
	SET_MEMORY_STAT(STAT_RenderTargetPoolUsed, int64(UsedKB) * 1024ll);
	SET_DWORD_STAT(STAT_RenderTargetPoolCount, Count);
#endif // STATS
}


int32 FRenderTargetPool::FindIndex(IPooledRenderTarget* In) const
{
	check(IsInRenderingThread());

	if(In)
	{
		for(uint32 i = 0, Num = (uint32)PooledRenderTargets.Num(); i < Num; ++i)
		{
			const FPooledRenderTarget* Element = PooledRenderTargets[i];

			if(Element == In)
			{
				check(!Element->IsSnapshot());
				return i;
			}
		}
	}

	// not found
	return -1;
}

void FRenderTargetPool::FreeUnusedResource(TRefCountPtr<IPooledRenderTarget>& In)
{
	check(IsInRenderingThread());
	
	int32 Index = FindIndex(In);

	if(Index != -1)
	{
		FPooledRenderTarget* Element = PooledRenderTargets[Index];

		if(Element)
		{
			check(!Element->IsSnapshot());
			AllocationLevelInKB -= ComputeSizeInKB(*Element);
			// we assume because of reference counting the resource gets released when not needed any more
			// we don't use Remove() to not shuffle around the elements for better transparency on RenderTargetPoolEvents
			DeferredDeleteArray.Add(PooledRenderTargets[Index]);
			PooledRenderTargets[Index] = 0;

			In.SafeRelease();

			VerifyAllocationLevel();
		}
	}
}

void FRenderTargetPool::FreeUnusedResources()
{
	check(IsInRenderingThread());

	for(uint32 i = 0, Num = (uint32)PooledRenderTargets.Num(); i < Num; ++i)
	{
		FPooledRenderTarget* Element = PooledRenderTargets[i];

		if(Element && Element->IsFree())
		{
			check(!Element->IsSnapshot());
			AllocationLevelInKB -= ComputeSizeInKB(*Element);
			// we assume because of reference counting the resource gets released when not needed any more
			// we don't use Remove() to not shuffle around the elements for better transparency on RenderTargetPoolEvents
			DeferredDeleteArray.Add(PooledRenderTargets[i]);
			PooledRenderTargets[i] = 0;
		}
	}

	VerifyAllocationLevel();
}


void FRenderTargetPool::DumpMemoryUsage(FOutputDevice& OutputDevice)
{
	OutputDevice.Logf(TEXT("Pooled Render Targets:"));
	for(int32 i = 0; i < PooledRenderTargets.Num(); ++i)
	{
		FPooledRenderTarget* Element = PooledRenderTargets[i];

		if(Element)
		{
			check(!Element->IsSnapshot());
			OutputDevice.Logf(
				TEXT("  %6.3fMB %4dx%4d%s%s %2dmip(s) %s (%s) %s %s"),
				ComputeSizeInKB(*Element) / 1024.0f,
				Element->Desc.Extent.X,
				Element->Desc.Extent.Y,
				Element->Desc.Depth > 1 ? *FString::Printf(TEXT("x%3d"), Element->Desc.Depth) : (Element->Desc.IsCubemap() ? TEXT("cube") : TEXT("    ")),
				Element->Desc.bIsArray ? *FString::Printf(TEXT("[%3d]"), Element->Desc.ArraySize) : TEXT("     "),
				Element->Desc.NumMips,
				Element->Desc.DebugName,
				GPixelFormats[Element->Desc.Format].Name,
				Element->IsTransient() ? TEXT("(transient)") : TEXT(""),
				GSupportsTransientResourceAliasing ? *FString::Printf(TEXT("Frames since last discard: %d"), GFrameNumberRenderThread - Element->FrameNumberLastDiscard) : TEXT("")
				);
		}
	}
	uint32 NumTargets=0;
	uint32 UsedKB=0;
	uint32 PoolKB=0;
	GetStats(NumTargets,PoolKB,UsedKB);
	OutputDevice.Logf(TEXT("%.3fMB total, %.3fMB used, %d render targets"), PoolKB / 1024.f, UsedKB / 1024.f, NumTargets);
}

uint32 FPooledRenderTarget::AddRef() const
{
	if (!bSnapshot)
	{
		check(IsInRenderingThread());
		return uint32(++NumRefs);
	}
	check(NumRefs == 1);
	return 1;
}

uint32 FPooledRenderTarget::Release()
{
	if (!bSnapshot)
	{
		check(IsInRenderingThread());
		uint32 Refs = uint32(--NumRefs);
		if(Refs == 0)
		{
			RenderTargetItem.SafeRelease();
			delete this;
		}
		else if (Refs == 1 && RenderTargetPool && IsTransient() )
		{
			// Discard the resource
			check(GetRenderTargetItem().TargetableTexture != nullptr);
			if (GetRenderTargetItem().TargetableTexture)
			{
				RHIDiscardTransientResource(GetRenderTargetItem().TargetableTexture);
			}
			FrameNumberLastDiscard = GFrameNumberRenderThread;
		}
		return Refs;
	}
	check(NumRefs == 1);
	return 1;
}

uint32 FPooledRenderTarget::GetRefCount() const
{
	return uint32(NumRefs);
}

void FPooledRenderTarget::SetDebugName(const TCHAR *InName)
{
	check(InName);

	Desc.DebugName = InName;
}

const FPooledRenderTargetDesc& FPooledRenderTarget::GetDesc() const
{
	return Desc;
}

void FRenderTargetPool::ReleaseDynamicRHI()
{
	check(IsInRenderingThread());
	WaitForTransitionFence();

	PooledRenderTargets.Empty();
	if (PooledRenderTargetSnapshots.Num())
	{
		DestructSnapshots();
	}
}

void FRenderTargetPool::DestructSnapshots()
{
	for (auto Snapshot : PooledRenderTargetSnapshots)
	{
		Snapshot->~FPooledRenderTarget();
	}
	PooledRenderTargetSnapshots.Reset();
}

// for debugging purpose
FPooledRenderTarget* FRenderTargetPool::GetElementById(uint32 Id) const
{
	// is used in game and render thread

	if(Id >= (uint32)PooledRenderTargets.Num())
	{
		return 0;
	}

	return PooledRenderTargets[Id];
}

void FRenderTargetPool::VerifyAllocationLevel() const
{
/*
	// uncomment to verify internal consistency
	uint32 OutWholeCount;
	uint32 OutWholePoolInKB;
	uint32 OutUsedInKB;

	GetStats(OutWholeCount, OutWholePoolInKB, OutUsedInKB);
*/
}

void FRenderTargetPool::CompactPool()
{
	for(uint32 i = 0, Num = (uint32)PooledRenderTargets.Num(); i < Num; ++i)
	{
		FPooledRenderTarget* Element = PooledRenderTargets[i];

		if(!Element)
		{
			PooledRenderTargets.RemoveAtSwap(i);
			--Num;
		}
	}
}

bool FPooledRenderTarget::OnFrameStart()
{
	check(IsInRenderingThread() && !bSnapshot);

	// If there are any references to the pooled render target other than the pool itself, then it may not be freed.
	if(!IsFree())
	{
		check(!UnusedForNFrames);
		return false;
	}

	++UnusedForNFrames;

	// this logic can be improved
	if(UnusedForNFrames > 10)
	{
		// release
		return true;
	}

	return false;
}

uint32 FPooledRenderTarget::ComputeMemorySize() const
{
	uint32 Size = 0;
	if (!bSnapshot)
	{
		if(Desc.Is2DTexture())
		{
			Size += RHIComputeMemorySize((const FTexture2DRHIRef&)RenderTargetItem.TargetableTexture);
			if(RenderTargetItem.ShaderResourceTexture != RenderTargetItem.TargetableTexture)
			{
				Size += RHIComputeMemorySize((const FTexture2DRHIRef&)RenderTargetItem.ShaderResourceTexture);
			}
		}
		else if(Desc.Is3DTexture())
		{
			Size += RHIComputeMemorySize((const FTexture3DRHIRef&)RenderTargetItem.TargetableTexture);
			if(RenderTargetItem.ShaderResourceTexture != RenderTargetItem.TargetableTexture)
			{
				Size += RHIComputeMemorySize((const FTexture3DRHIRef&)RenderTargetItem.ShaderResourceTexture);
			}
		}
		else
		{
			Size += RHIComputeMemorySize((const FTextureCubeRHIRef&)RenderTargetItem.TargetableTexture);
			if(RenderTargetItem.ShaderResourceTexture != RenderTargetItem.TargetableTexture)
			{
				Size += RHIComputeMemorySize((const FTextureCubeRHIRef&)RenderTargetItem.ShaderResourceTexture);
			}
		}
	}
	return Size;
}

bool FPooledRenderTarget::IsFree() const
{
	uint32 RefCount = GetRefCount();
	check(RefCount >= 1);

	// If the only reference to the pooled render target is from the pool, then it's unused.
	return !bSnapshot && RefCount == 1;
}