bool CharacterDisplay::Init(CCharacterFX* pParent, HRECORD hDisplay, const LTVector2n& sz, ClientDisplayData* pData )
{
//if it's the same we shouldn't have to do anything...
if ( m_pParent == pParent &&
m_hDisplayRecord == hDisplay &&
m_hRenderTarget == pData->hNameTarget &&
m_vSize.x == (float)sz.x &&
m_vSize.y == (float)sz.y &&
m_hMaterial == pData->hNameMaterial
)
{
return true;
}
m_pParent = pParent;
m_vSize = LTVector2((float)sz.x,(float)sz.y);
if (!m_pParent || !m_pParent->GetModel())
{
return false;
}
m_hDisplayRecord = hDisplay;
if (!m_hDisplayRecord)
{
LTERROR( "Invalid CharacterDisplay" );
return false;
}
// Get the layout record from the localized database.
m_hLayoutRecord = DATABASE_CATEGORY( CharacterDisplayDB ).GetRecordLinkToLocalizedDB( g_pLTDatabase->GetAttribute( m_hDisplayRecord, "Layout" ),
0, DATABASE_CATEGORY( CharacterDisplayLayoutDB ).GetCategory( ));
if( !m_hLayoutRecord )
{
LTERROR_PARAM1( "Invalid CharacterDisplayLayout %s", g_pLTDatabase->GetRecordName( m_hDisplayRecord ));
return false;
}
ReleaseRenderTarget();
char const* pszMaterial = DATABASE_CATEGORY( CharacterDisplayDB ).GETRECORDATTRIB( m_hDisplayRecord, Material );
//check to see if we're using the same base material
if (pData->hNameMaterial && !LTStrIEquals(pData->sNameMaterialFile.c_str(),pszMaterial) )
{
g_pLTClient->GetRenderer()->ReleaseMaterialInstance(pData->hNameMaterial);
pData->hNameMaterial = NULL;
}
//hNameMaterial is a reference to the material stored in a client info record, if this is null
// we need to clone a material for this client to use
if (!pData->hNameMaterial)
{
HMATERIAL hBaseMat = g_pLTClient->GetRenderer()->CreateMaterialInstance( pszMaterial );
if (!hBaseMat) return false;
pData->hNameMaterial = g_pLTClient->GetRenderer()->CloneMaterialInstance(hBaseMat);
pData->sNameMaterialFile = pszMaterial;
g_pLTClient->GetRenderer()->ReleaseMaterialInstance(hBaseMat);
}
if (m_hMaterial != pData->hNameMaterial)
{
if (m_hMaterial)
{
g_pLTClient->GetRenderer()->ReleaseMaterialInstance(m_hMaterial);
}
m_hMaterial = pData->hNameMaterial;
g_pLTClient->GetRenderer()->AddRefMaterialInstance(m_hMaterial);
}
if (m_hMaterial == NULL)
{
return false;
}
if (!SetRenderTarget(pData->hNameTarget))
{
return false;
}
LTVector2 vPos = DATABASE_CATEGORY( CharacterDisplayLayoutDB ).GETRECORDATTRIB( m_hLayoutRecord, TextOffset );
m_Text.SetPos(LTVector2n(uint32(vPos.x),uint32(vPos.y)));
m_Text.SetAlignment(kCenter);
m_Text.SetDropShadow(1);
for (uint8 n = 0; n < 3; ++n)
{
m_nTextSize[n] = DATABASE_CATEGORY( CharacterDisplayLayoutDB ).GETRECORDATTRIBINDEX( m_hLayoutRecord, TextSize, n);
}
HRECORD hFont = DATABASE_CATEGORY( CharacterDisplayLayoutDB ).GETRECORDATTRIB( m_hLayoutRecord, Font );
char const* pszFont = DATABASE_CATEGORY( CharacterDisplayLayoutDB ).GetString( hFont, "Face" );
CFontInfo textFont(pszFont,m_nTextSize[2]);
m_Text.SetFont(textFont);
m_Text.SetColor(DATABASE_CATEGORY( CharacterDisplayDB ).GETRECORDATTRIB( m_hDisplayRecord, TextColor));
m_cTextBackColor = DATABASE_CATEGORY( CharacterDisplayDB ).GETRECORDATTRIB( m_hDisplayRecord, TextBackColor);
LTVector4 vTextBackRect = DATABASE_CATEGORY( CharacterDisplayDB ).GETRECORDATTRIB( m_hDisplayRecord, TextBackRect );
m_rTextBackRect.Init(( int32 )vTextBackRect.x, ( int32 )vTextBackRect.y, ( int32 )vTextBackRect.z, ( int32 )vTextBackRect.w );
return true;
}
void FRCPassPostProcessVisualizeBuffer::Process(FRenderingCompositePassContext& Context)
{
SCOPED_DRAW_EVENT(Context.RHICmdList, VisualizeBuffer);
const FPooledRenderTargetDesc* InputDesc = GetInputDesc(ePId_Input0);
if(!InputDesc)
{
// input is not hooked up correctly
return;
}
const FSceneView& View = Context.View;
const FSceneViewFamily& ViewFamily = *(View.Family);
FIntRect SrcRect = View.ViewRect;
FIntRect DestRect = View.ViewRect;
FIntPoint SrcSize = InputDesc->Extent;
const FSceneRenderTargetItem& DestRenderTarget = PassOutputs[0].RequestSurface(Context);
// Set the view family's render target/viewport.
SetRenderTarget(Context.RHICmdList, DestRenderTarget.TargetableTexture, FTextureRHIRef());
Context.SetViewportAndCallRHI(DestRect);
// set the state
Context.RHICmdList.SetBlendState(TStaticBlendState<>::GetRHI());
Context.RHICmdList.SetRasterizerState(TStaticRasterizerState<>::GetRHI());
Context.RHICmdList.SetDepthStencilState(TStaticDepthStencilState<false, CF_Always>::GetRHI());
{
FShader* VertexShader = SetShaderTempl<false>(Context);
// Draw a quad mapping scene color to the view's render target
DrawRectangle(
Context.RHICmdList,
0, 0,
DestRect.Width(), DestRect.Height(),
SrcRect.Min.X, SrcRect.Min.Y,
SrcRect.Width(), SrcRect.Height(),
DestRect.Size(),
SrcSize,
VertexShader,
EDRF_UseTriangleOptimization);
}
Context.RHICmdList.SetBlendState(TStaticBlendState<CW_RGB, BO_Add, BF_SourceAlpha, BF_InverseSourceAlpha>::GetRHI());
TShaderMapRef<FPostProcessVS> VertexShader(Context.GetShaderMap());
TShaderMapRef<FPostProcessVisualizeBufferPS<true> > PixelShader(Context.GetShaderMap());
static FGlobalBoundShaderState BoundShaderState;
SetGlobalBoundShaderState(Context.RHICmdList, Context.GetFeatureLevel(), BoundShaderState, GFilterVertexDeclaration.VertexDeclarationRHI, *VertexShader, *PixelShader);
PixelShader->SetPS(Context);
// Track the name and position of each tile we draw so we can write text labels over them
struct LabelRecord
{
FString Label;
int32 LocationX;
int32 LocationY;
};
TArray<LabelRecord> Labels;
const int32 MaxTilesX = 4;
const int32 MaxTilesY = 4;
const int32 TileWidth = DestRect.Width() / MaxTilesX;
const int32 TileHeight = DestRect.Height() / MaxTilesY;
int32 CurrentTileIndex = 0;
for (TArray<TileData>::TConstIterator It = Tiles.CreateConstIterator(); It; ++It, ++CurrentTileIndex)
{
FRenderingCompositeOutputRef Tile = It->Source;
if (Tile.IsValid())
{
FTextureRHIRef Texture = Tile.GetOutput()->PooledRenderTarget->GetRenderTargetItem().TargetableTexture;
int32 TileX = CurrentTileIndex % MaxTilesX;
int32 TileY = CurrentTileIndex / MaxTilesX;
PixelShader->SetSourceTexture(Context.RHICmdList, Texture);
DrawRectangle(
Context.RHICmdList,
TileX * TileWidth, TileY * TileHeight,
TileWidth, TileHeight,
SrcRect.Min.X, SrcRect.Min.Y,
SrcRect.Width(), SrcRect.Height(),
DestRect.Size(),
SrcSize,
*VertexShader,
EDRF_Default
);
Labels.Add(LabelRecord());
Labels.Last().Label = It->Name;
Labels.Last().LocationX = 8 + TileX * TileWidth;
Labels.Last().LocationY = (TileY + 1) * TileHeight - 19;
}
}
// Draw tile labels
// this is a helper class for FCanvas to be able to get screen size
class FRenderTargetTemp : public FRenderTarget
{
public:
const FSceneView& View;
const FTexture2DRHIRef Texture;
FRenderTargetTemp(const FSceneView& InView, const FTexture2DRHIRef InTexture)
: View(InView), Texture(InTexture)
{
}
virtual FIntPoint GetSizeXY() const
{
return View.ViewRect.Size();
};
virtual const FTexture2DRHIRef& GetRenderTargetTexture() const
{
return Texture;
}
} TempRenderTarget(View, (const FTexture2DRHIRef&)DestRenderTarget.TargetableTexture);
FCanvas Canvas(&TempRenderTarget, NULL, ViewFamily.CurrentRealTime, ViewFamily.CurrentWorldTime, ViewFamily.DeltaWorldTime, Context.GetFeatureLevel());
FLinearColor LabelColor(1, 1, 0);
for (auto It = Labels.CreateConstIterator(); It; ++It)
{
Canvas.DrawShadowedString(It->LocationX, It->LocationY, *It->Label, GetStatsFont(), LabelColor);
}
Canvas.Flush_RenderThread(Context.RHICmdList);
Context.RHICmdList.CopyToResolveTarget(DestRenderTarget.TargetableTexture, DestRenderTarget.ShaderResourceTexture, false, FResolveParams());
}
void FRCPassPostProcessDeferredDecals::Process(FRenderingCompositePassContext& Context)
{
FRHICommandListImmediate& RHICmdList = Context.RHICmdList;
FSceneRenderTargets& SceneContext = FSceneRenderTargets::Get(RHICmdList);
const bool bShaderComplexity = Context.View.Family->EngineShowFlags.ShaderComplexity;
const bool bDBuffer = IsDBufferEnabled();
const bool bStencilSizeThreshold = CVarStencilSizeThreshold.GetValueOnRenderThread() >= 0;
SCOPED_DRAW_EVENT(RHICmdList, PostProcessDeferredDecals);
enum EDecalResolveBufferIndex
{
SceneColorIndex,
GBufferAIndex,
GBufferBIndex,
GBufferCIndex,
DBufferAIndex,
DBufferBIndex,
DBufferCIndex,
ResolveBufferMax,
};
FTextureRHIParamRef TargetsToResolve[ResolveBufferMax] = { nullptr };
if(DecalRenderStage == DRS_BeforeBasePass)
{
// before BasePass, only if DBuffer is enabled
check(bDBuffer);
// DBuffer: Decal buffer
FPooledRenderTargetDesc Desc(FPooledRenderTargetDesc::Create2DDesc(SceneContext.GBufferA->GetDesc().Extent,
PF_B8G8R8A8,
FClearValueBinding::None,
TexCreate_None,
TexCreate_ShaderResource | TexCreate_RenderTargetable,
false));
if(!SceneContext.DBufferA)
{
Desc.ClearValue = FClearValueBinding::Black;
GRenderTargetPool.FindFreeElement(Desc, SceneContext.DBufferA, TEXT("DBufferA"));
}
if(!SceneContext.DBufferB)
{
Desc.ClearValue = FClearValueBinding(FLinearColor(128.0f / 255.0f, 128.0f / 255.0f, 128.0f / 255.0f, 1));
GRenderTargetPool.FindFreeElement(Desc, SceneContext.DBufferB, TEXT("DBufferB"));
}
Desc.Format = PF_R8G8;
if(!SceneContext.DBufferC)
{
Desc.ClearValue = FClearValueBinding(FLinearColor(0, 1, 0, 1));
GRenderTargetPool.FindFreeElement(Desc, SceneContext.DBufferC, TEXT("DBufferC"));
}
// we assume views are non overlapping, then we need to clear only once in the beginning, otherwise we would need to set scissor rects
// and don't get FastClear any more.
bool bFirstView = Context.View.Family->Views[0] == &Context.View;
if(bFirstView)
{
SCOPED_DRAW_EVENT(RHICmdList, DBufferClear);
FRHIRenderTargetView RenderTargets[3];
RenderTargets[0] = FRHIRenderTargetView(SceneContext.DBufferA->GetRenderTargetItem().TargetableTexture, 0, -1, ERenderTargetLoadAction::EClear, ERenderTargetStoreAction::EStore);
RenderTargets[1] = FRHIRenderTargetView(SceneContext.DBufferB->GetRenderTargetItem().TargetableTexture, 0, -1, ERenderTargetLoadAction::EClear, ERenderTargetStoreAction::EStore);
RenderTargets[2] = FRHIRenderTargetView(SceneContext.DBufferC->GetRenderTargetItem().TargetableTexture, 0, -1, ERenderTargetLoadAction::EClear, ERenderTargetStoreAction::EStore);
FRHIDepthRenderTargetView DepthView(SceneContext.GetSceneDepthSurface(), ERenderTargetLoadAction::ELoad, ERenderTargetStoreAction::ENoAction, ERenderTargetLoadAction::ELoad, ERenderTargetStoreAction::ENoAction, FExclusiveDepthStencil(FExclusiveDepthStencil::DepthRead_StencilWrite));
FRHISetRenderTargetsInfo Info(3, RenderTargets, DepthView);
RHICmdList.SetRenderTargetsAndClear(Info);
TargetsToResolve[DBufferAIndex] = SceneContext.DBufferA->GetRenderTargetItem().TargetableTexture;
TargetsToResolve[DBufferBIndex] = SceneContext.DBufferB->GetRenderTargetItem().TargetableTexture;
TargetsToResolve[DBufferCIndex] = SceneContext.DBufferC->GetRenderTargetItem().TargetableTexture;
}
}
// this cast is safe as only the dedicated server implements this differently and this pass should not be executed on the dedicated server
const FViewInfo& View = Context.View;
const FSceneViewFamily& ViewFamily = *(View.Family);
FScene& Scene = *(FScene*)ViewFamily.Scene;
//don't early return. Resolves must be run for fast clears to work.
bool bRenderDecal = Scene.Decals.Num() && ViewFamily.EngineShowFlags.Decals;
if (bRenderDecal)
{
// Build a list of decals that need to be rendered for this view
FTransientDecalRenderDataList SortedDecals;
FDecalRendering::BuildVisibleDecalList(Scene, View, DecalRenderStage, SortedDecals);
if (SortedDecals.Num() > 0)
{
FIntRect SrcRect = View.ViewRect;
FIntRect DestRect = View.ViewRect;
bool bStencilDecalsInThisStage = true;
#if DBUFFER_DONT_USE_STENCIL_YET
if (DecalRenderStage != DRS_BeforeLighting)
{
bStencilDecalsInThisStage = false;
}
#endif
// Setup a stencil mask to prevent certain pixels from receiving deferred decals
if (bStencilDecalsInThisStage)
{
StencilDecalMask(RHICmdList, View, Context.HasHmdMesh());
}
// optimization to have less state changes
EDecalRasterizerState LastDecalRasterizerState = DRS_Undefined;
FDecalDepthState LastDecalDepthState;
int32 LastDecalBlendMode = -1;
int32 LastDecalHasNormal = -1; // Decal state can change based on its normal property.(SM5)
FDecalRendering::ERenderTargetMode LastRenderTargetMode = FDecalRendering::RTM_Unknown;
const ERHIFeatureLevel::Type SMFeatureLevel = Context.GetFeatureLevel();
SCOPED_DRAW_EVENT(RHICmdList, Decals);
INC_DWORD_STAT_BY(STAT_Decals, SortedDecals.Num());
for (int32 DecalIndex = 0, DecalCount = SortedDecals.Num(); DecalIndex < DecalCount; DecalIndex++)
{
const FTransientDecalRenderData& DecalData = SortedDecals[DecalIndex];
const FDeferredDecalProxy& DecalProxy = *DecalData.DecalProxy;
const FMatrix ComponentToWorldMatrix = DecalProxy.ComponentTrans.ToMatrixWithScale();
const FMatrix FrustumComponentToClip = FDecalRendering::ComputeComponentToClipMatrix(View, ComponentToWorldMatrix);
EDecalBlendMode DecalBlendMode = DecalData.DecalBlendMode;
bool bStencilThisDecal = bStencilDecalsInThisStage;
#if DBUFFER_DONT_USE_STENCIL_YET
if (FDecalRendering::ComputeRenderStage(View.GetShaderPlatform(), DecalBlendMode) != DRS_BeforeLighting)
{
bStencilThisDecal = false;
}
#endif
FDecalRendering::ERenderTargetMode CurrentRenderTargetMode = FDecalRendering::ComputeRenderTargetMode(View.GetShaderPlatform(), DecalBlendMode);
if (bShaderComplexity)
{
CurrentRenderTargetMode = FDecalRendering::RTM_SceneColor;
// we want additive blending for the ShaderComplexity mode
DecalBlendMode = DBM_Emissive;
}
// fewer rendertarget switches if possible
if (CurrentRenderTargetMode != LastRenderTargetMode)
{
LastRenderTargetMode = CurrentRenderTargetMode;
switch (CurrentRenderTargetMode)
{
case FDecalRendering::RTM_SceneColorAndGBuffer:
{
TargetsToResolve[SceneColorIndex] = SceneContext.GetSceneColor()->GetRenderTargetItem().TargetableTexture;
TargetsToResolve[GBufferAIndex] = SceneContext.GBufferA->GetRenderTargetItem().TargetableTexture;
TargetsToResolve[GBufferBIndex] = SceneContext.GBufferB->GetRenderTargetItem().TargetableTexture;
TargetsToResolve[GBufferCIndex] = SceneContext.GBufferC->GetRenderTargetItem().TargetableTexture;
SetRenderTargets(RHICmdList, 4, TargetsToResolve, SceneContext.GetSceneDepthSurface(), ESimpleRenderTargetMode::EExistingColorAndDepth, FExclusiveDepthStencil::DepthRead_StencilWrite);
}
break;
case FDecalRendering::RTM_SceneColorAndGBufferDepthWrite:
{
TargetsToResolve[SceneColorIndex] = SceneContext.GetSceneColor()->GetRenderTargetItem().TargetableTexture;
TargetsToResolve[GBufferAIndex] = SceneContext.GBufferA->GetRenderTargetItem().TargetableTexture;
TargetsToResolve[GBufferBIndex] = SceneContext.GBufferB->GetRenderTargetItem().TargetableTexture;
TargetsToResolve[GBufferCIndex] = SceneContext.GBufferC->GetRenderTargetItem().TargetableTexture;
SetRenderTargets(RHICmdList, 4, TargetsToResolve, SceneContext.GetSceneDepthSurface(), ESimpleRenderTargetMode::EExistingColorAndDepth, FExclusiveDepthStencil::DepthWrite_StencilWrite);
}
break;
case FDecalRendering::RTM_GBufferNormal:
TargetsToResolve[GBufferAIndex] = SceneContext.GBufferA->GetRenderTargetItem().TargetableTexture;
SetRenderTarget(RHICmdList, TargetsToResolve[GBufferAIndex], SceneContext.GetSceneDepthSurface(), ESimpleRenderTargetMode::EExistingColorAndDepth, FExclusiveDepthStencil::DepthRead_StencilWrite);
break;
case FDecalRendering::RTM_SceneColor:
TargetsToResolve[SceneColorIndex] = SceneContext.GetSceneColor()->GetRenderTargetItem().TargetableTexture;
SetRenderTarget(RHICmdList, TargetsToResolve[SceneColorIndex], SceneContext.GetSceneDepthSurface(), ESimpleRenderTargetMode::EExistingColorAndDepth, FExclusiveDepthStencil::DepthRead_StencilWrite);
break;
case FDecalRendering::RTM_DBuffer:
{
TargetsToResolve[DBufferAIndex] = SceneContext.DBufferA->GetRenderTargetItem().TargetableTexture;
TargetsToResolve[DBufferBIndex] = SceneContext.DBufferB->GetRenderTargetItem().TargetableTexture;
TargetsToResolve[DBufferCIndex] = SceneContext.DBufferC->GetRenderTargetItem().TargetableTexture;
SetRenderTargets(RHICmdList, 3, &TargetsToResolve[DBufferAIndex], SceneContext.GetSceneDepthSurface(), ESimpleRenderTargetMode::EExistingColorAndDepth, FExclusiveDepthStencil::DepthRead_StencilWrite);
}
break;
default:
check(0);
break;
}
Context.SetViewportAndCallRHI(DestRect);
// we need to reset the stream source after any call to SetRenderTarget (at least for Metal, which doesn't queue up VB assignments)
RHICmdList.SetStreamSource(0, GetUnitCubeVertexBuffer(), sizeof(FVector4), 0);
}
bool bThisDecalUsesStencil = false;
if (bStencilThisDecal)
{
if (bStencilSizeThreshold)
{
// note this is after a SetStreamSource (in if CurrentRenderTargetMode != LastRenderTargetMode) call as it needs to get the VB input
bThisDecalUsesStencil = RenderPreStencil(Context, ComponentToWorldMatrix, FrustumComponentToClip);
LastDecalRasterizerState = DRS_Undefined;
LastDecalDepthState = FDecalDepthState();
LastDecalBlendMode = -1;
}
}
const bool bBlendStateChange = DecalBlendMode != LastDecalBlendMode;// Has decal mode changed.
const bool bDecalNormalChanged = GSupportsSeparateRenderTargetBlendState && // has normal changed for SM5 stain/translucent decals?
(DecalBlendMode == DBM_Translucent || DecalBlendMode == DBM_Stain) &&
(int32)DecalData.bHasNormal != LastDecalHasNormal;
// fewer blend state changes if possible
if (bBlendStateChange || bDecalNormalChanged)
{
LastDecalBlendMode = DecalBlendMode;
LastDecalHasNormal = (int32)DecalData.bHasNormal;
SetDecalBlendState(RHICmdList, SMFeatureLevel, DecalRenderStage, (EDecalBlendMode)LastDecalBlendMode, DecalData.bHasNormal);
}
// todo
const float ConservativeRadius = DecalData.ConservativeRadius;
// const int32 IsInsideDecal = ((FVector)View.ViewMatrices.ViewOrigin - ComponentToWorldMatrix.GetOrigin()).SizeSquared() < FMath::Square(ConservativeRadius * 1.05f + View.NearClippingDistance * 2.0f) + ( bThisDecalUsesStencil ) ? 2 : 0;
const bool bInsideDecal = ((FVector)View.ViewMatrices.ViewOrigin - ComponentToWorldMatrix.GetOrigin()).SizeSquared() < FMath::Square(ConservativeRadius * 1.05f + View.NearClippingDistance * 2.0f);
// const bool bInsideDecal = !(IsInsideDecal & 1);
// update rasterizer state if needed
{
EDecalRasterizerState DecalRasterizerState = ComputeDecalRasterizerState(bInsideDecal, View);
if (LastDecalRasterizerState != DecalRasterizerState)
{
LastDecalRasterizerState = DecalRasterizerState;
SetDecalRasterizerState(DecalRasterizerState, RHICmdList);
}
}
// update DepthStencil state if needed
{
FDecalDepthState DecalDepthState = ComputeDecalDepthState(DecalBlendMode, bInsideDecal, bStencilDecalsInThisStage, bThisDecalUsesStencil);
if (LastDecalDepthState != DecalDepthState)
{
LastDecalDepthState = DecalDepthState;
SetDecalDepthState(DecalDepthState, RHICmdList);
}
}
FDecalRendering::SetShader(RHICmdList, View, bShaderComplexity, DecalData, FrustumComponentToClip);
RHICmdList.DrawIndexedPrimitive(GetUnitCubeIndexBuffer(), PT_TriangleList, 0, 0, 8, 0, ARRAY_COUNT(GCubeIndices) / 3, 1);
}
// we don't modify stencil but if out input was having stencil for us (after base pass - we need to clear)
// Clear stencil to 0, which is the assumed default by other passes
RHICmdList.Clear(false, FLinearColor::White, false, (float)ERHIZBuffer::FarPlane, true, 0, FIntRect());
if (DecalRenderStage == DRS_BeforeBasePass)
{
// before BasePass
GRenderTargetPool.VisualizeTexture.SetCheckPoint(RHICmdList, SceneContext.DBufferA);
GRenderTargetPool.VisualizeTexture.SetCheckPoint(RHICmdList, SceneContext.DBufferB);
GRenderTargetPool.VisualizeTexture.SetCheckPoint(RHICmdList, SceneContext.DBufferC);
}
}
}
// resolve the targets we wrote to.
FResolveParams ResolveParams;
for (int32 i = 0; i < ResolveBufferMax; ++i)
{
if (TargetsToResolve[i])
{
RHICmdList.CopyToResolveTarget(TargetsToResolve[i], TargetsToResolve[i], true, ResolveParams);
}
}
}
void FRCPassPostProcessGBufferHints::Process(FRenderingCompositePassContext& Context)
{
SCOPED_DRAW_EVENT(Context.RHICmdList, GBufferHints);
const FPooledRenderTargetDesc* InputDesc = GetInputDesc(ePId_Input0);
if(!InputDesc)
{
// input is not hooked up correctly
return;
}
const FSceneView& View = Context.View;
const FSceneViewFamily& ViewFamily = *(View.Family);
FIntRect SrcRect = View.ViewRect;
FIntRect DestRect = View.ViewRect;
FIntPoint SrcSize = InputDesc->Extent;
const FSceneRenderTargetItem& DestRenderTarget = PassOutputs[0].RequestSurface(Context);
// Set the view family's render target/viewport.
SetRenderTarget(Context.RHICmdList, DestRenderTarget.TargetableTexture, FTextureRHIRef());
Context.SetViewportAndCallRHI(DestRect);
// set the state
Context.RHICmdList.SetBlendState(TStaticBlendState<>::GetRHI());
Context.RHICmdList.SetRasterizerState(TStaticRasterizerState<>::GetRHI());
Context.RHICmdList.SetDepthStencilState(TStaticDepthStencilState<false, CF_Always>::GetRHI());
TShaderMapRef<FPostProcessVS> VertexShader(Context.GetShaderMap());
TShaderMapRef<FPostProcessGBufferHintsPS> PixelShader(Context.GetShaderMap());
static FGlobalBoundShaderState BoundShaderState;
SetGlobalBoundShaderState(Context.RHICmdList, Context.GetFeatureLevel(), BoundShaderState, GFilterVertexDeclaration.VertexDeclarationRHI, *VertexShader, *PixelShader);
PixelShader->SetPS(Context);
// Draw a quad mapping scene color to the view's render target
DrawRectangle(
Context.RHICmdList,
0, 0,
DestRect.Width(), DestRect.Height(),
SrcRect.Min.X, SrcRect.Min.Y,
SrcRect.Width(), SrcRect.Height(),
DestRect.Size(),
SrcSize,
*VertexShader,
EDRF_UseTriangleOptimization);
// this is a helper class for FCanvas to be able to get screen size
class FRenderTargetTemp : public FRenderTarget
{
public:
const FSceneView& View;
const FTexture2DRHIRef Texture;
FRenderTargetTemp(const FSceneView& InView, const FTexture2DRHIRef InTexture)
: View(InView), Texture(InTexture)
{
}
virtual FIntPoint GetSizeXY() const
{
return View.ViewRect.Size();
};
virtual const FTexture2DRHIRef& GetRenderTargetTexture() const
{
return Texture;
}
} TempRenderTarget(View, (const FTexture2DRHIRef&)DestRenderTarget.TargetableTexture);
FCanvas Canvas(&TempRenderTarget, NULL, ViewFamily.CurrentRealTime, ViewFamily.CurrentWorldTime, ViewFamily.DeltaWorldTime, View.GetFeatureLevel());
float X = 30;
float Y = 8;
const float YStep = 14;
const float ColumnWidth = 250;
FString Line;
Line = FString::Printf(TEXT("GBufferHints"));
Canvas.DrawShadowedString( X, Y += YStep, *Line, GetStatsFont(), FLinearColor(1, 1, 1));
Y += YStep;
Line = FString::Printf(TEXT("Yellow: Unrealistic material (In nature even black materials reflect quite some light)"));
Canvas.DrawShadowedString( X, Y += YStep, *Line, GetStatsFont(), FLinearColor(0.8f, 0.8f, 0));
Line = FString::Printf(TEXT("Red: Impossive material (this material emits more light than it receives)"));
Canvas.DrawShadowedString( X, Y += YStep, *Line, GetStatsFont(), FLinearColor(1, 0, 0));
Canvas.Flush_RenderThread(Context.RHICmdList);
Context.RHICmdList.CopyToResolveTarget(DestRenderTarget.TargetableTexture, DestRenderTarget.ShaderResourceTexture, false, FResolveParams());
// AdjustGBufferRefCount(1) call is done in constructor
FSceneRenderTargets::Get(Context.RHICmdList).AdjustGBufferRefCount(-1);
}
//todo: this function is an abomination, this is just disgusting. fix it.
//...seriously, this is really, really horrible. I mean this is amazingly bad.
void OBS::MainCaptureLoop()
{
int curRenderTarget = 0, curYUVTexture = 0, curCopyTexture = 0;
int copyWait = NUM_RENDER_BUFFERS-1;
bSentHeaders = false;
bFirstAudioPacket = true;
bool bLogLongFramesProfile = GlobalConfig->GetInt(TEXT("General"), TEXT("LogLongFramesProfile"), LOGLONGFRAMESDEFAULT) != 0;
float logLongFramesProfilePercentage = GlobalConfig->GetFloat(TEXT("General"), TEXT("LogLongFramesProfilePercentage"), 10.f);
Vect2 baseSize = Vect2(float(baseCX), float(baseCY));
Vect2 outputSize = Vect2(float(outputCX), float(outputCY));
Vect2 scaleSize = Vect2(float(scaleCX), float(scaleCY));
HANDLE hScaleVal = yuvScalePixelShader->GetParameterByName(TEXT("baseDimensionI"));
//----------------------------------------
// x264 input buffers
int curOutBuffer = 0;
bool bUsingQSV = videoEncoder->isQSV();//GlobalConfig->GetInt(TEXT("Video Encoding"), TEXT("UseQSV")) != 0;
if(bUsingQSV)
bUsing444 = false;
EncoderPicture lastPic;
EncoderPicture outPics[NUM_OUT_BUFFERS];
DWORD outTimes[NUM_OUT_BUFFERS] = {0, 0, 0};
for(int i=0; i<NUM_OUT_BUFFERS; i++)
{
if(bUsingQSV)
{
outPics[i].mfxOut = new mfxFrameSurface1;
memset(outPics[i].mfxOut, 0, sizeof(mfxFrameSurface1));
mfxFrameData& data = outPics[i].mfxOut->Data;
videoEncoder->RequestBuffers(&data);
}
else
{
outPics[i].picOut = new x264_picture_t;
x264_picture_init(outPics[i].picOut);
}
}
if(bUsing444)
{
for(int i=0; i<NUM_OUT_BUFFERS; i++)
{
outPics[i].picOut->img.i_csp = X264_CSP_BGRA; //although the x264 input says BGR, x264 actually will expect packed UYV
outPics[i].picOut->img.i_plane = 1;
}
}
else
{
if(!bUsingQSV)
for(int i=0; i<NUM_OUT_BUFFERS; i++)
x264_picture_alloc(outPics[i].picOut, X264_CSP_NV12, outputCX, outputCY);
}
int bCongestionControl = AppConfig->GetInt (TEXT("Video Encoding"), TEXT("CongestionControl"), 0);
bool bDynamicBitrateSupported = App->GetVideoEncoder()->DynamicBitrateSupported();
int defaultBitRate = AppConfig->GetInt(TEXT("Video Encoding"), TEXT("MaxBitrate"), 1000);
int currentBitRate = defaultBitRate;
QWORD lastAdjustmentTime = 0;
UINT adjustmentStreamId = 0;
//----------------------------------------
// time/timestamp stuff
bufferedTimes.Clear();
ctsOffsets.Clear();
int bufferedFrames = 1; //to avoid constantly polling number of frames
#ifdef USE_100NS_TIME
QWORD streamTimeStart = GetQPCTime100NS();
QWORD frameTime100ns = 10000000/fps;
QWORD sleepTargetTime = 0;
bool bWasLaggedFrame = false;
#else
DWORD streamTimeStart = OSGetTime();
DWORD fpsTimeAdjust = 0;
#endif
totalStreamTime = 0;
lastAudioTimestamp = 0;
latestVideoTime = firstSceneTimestamp = GetQPCTimeMS();
DWORD fpsTimeNumerator = 1000-(frameTime*fps);
DWORD fpsTimeDenominator = fps;
DWORD cfrTime = 0;
DWORD cfrTimeAdjust = 0;
//----------------------------------------
// start audio capture streams
desktopAudio->StartCapture();
if(micAudio) micAudio->StartCapture();
//----------------------------------------
// status bar/statistics stuff
DWORD fpsCounter = 0;
int numLongFrames = 0;
int numTotalFrames = 0;
int numTotalDuplicatedFrames = 0;
bytesPerSec = 0;
captureFPS = 0;
curFramesDropped = 0;
curStrain = 0.0;
PostMessage(hwndMain, OBS_UPDATESTATUSBAR, 0, 0);
QWORD lastBytesSent[3] = {0, 0, 0};
DWORD lastFramesDropped = 0;
#ifdef USE_100NS_TIME
double bpsTime = 0.0;
#else
float bpsTime = 0.0f;
#endif
double lastStrain = 0.0f;
DWORD numSecondsWaited = 0;
//----------------------------------------
// 444->420 thread data
int numThreads = MAX(OSGetTotalCores()-2, 1);
HANDLE *h420Threads = (HANDLE*)Allocate(sizeof(HANDLE)*numThreads);
Convert444Data *convertInfo = (Convert444Data*)Allocate(sizeof(Convert444Data)*numThreads);
zero(h420Threads, sizeof(HANDLE)*numThreads);
zero(convertInfo, sizeof(Convert444Data)*numThreads);
for(int i=0; i<numThreads; i++)
{
convertInfo[i].width = outputCX;
convertInfo[i].height = outputCY;
convertInfo[i].hSignalConvert = CreateEvent(NULL, FALSE, FALSE, NULL);
convertInfo[i].hSignalComplete = CreateEvent(NULL, FALSE, FALSE, NULL);
convertInfo[i].bNV12 = bUsingQSV;
if(i == 0)
convertInfo[i].startY = 0;
else
convertInfo[i].startY = convertInfo[i-1].endY;
if(i == (numThreads-1))
convertInfo[i].endY = outputCY;
else
convertInfo[i].endY = ((outputCY/numThreads)*(i+1)) & 0xFFFFFFFE;
}
bool bFirstFrame = true;
bool bFirstImage = true;
bool bFirst420Encode = true;
bool bUseThreaded420 = bUseMultithreadedOptimizations && (OSGetTotalCores() > 1) && !bUsing444;
List<HANDLE> completeEvents;
if(bUseThreaded420)
{
for(int i=0; i<numThreads; i++)
{
h420Threads[i] = OSCreateThread((XTHREAD)Convert444Thread, convertInfo+i);
completeEvents << convertInfo[i].hSignalComplete;
}
}
//----------------------------------------
QWORD curStreamTime = 0, lastStreamTime, firstFrameTime = GetQPCTimeMS();
#ifdef USE_100NS_TIME
lastStreamTime = GetQPCTime100NS()-frameTime100ns;
#else
lastStreamTime = firstFrameTime-frameTime;
#endif
//bool bFirstAudioPacket = true;
List<ProfilerNode> threadedProfilers;
bool bUsingThreadedProfilers = false;
while(bRunning || bufferedFrames)
{
#ifdef USE_100NS_TIME
QWORD renderStartTime = GetQPCTime100NS();
totalStreamTime = DWORD((renderStartTime-streamTimeStart)/10000);
if(sleepTargetTime == 0 || bWasLaggedFrame)
sleepTargetTime = renderStartTime;
#else
DWORD renderStartTime = OSGetTime();
totalStreamTime = renderStartTime-streamTimeStart;
DWORD frameTimeAdjust = frameTime;
fpsTimeAdjust += fpsTimeNumerator;
if(fpsTimeAdjust > fpsTimeDenominator)
{
fpsTimeAdjust -= fpsTimeDenominator;
++frameTimeAdjust;
}
#endif
bool bRenderView = !IsIconic(hwndMain) && bRenderViewEnabled;
profileIn("frame");
#ifdef USE_100NS_TIME
QWORD qwTime = renderStartTime/10000;
latestVideoTime = qwTime;
QWORD frameDelta = renderStartTime-lastStreamTime;
double fSeconds = double(frameDelta)*0.0000001;
//Log(TEXT("frameDelta: %f"), fSeconds);
lastStreamTime = renderStartTime;
#else
QWORD qwTime = GetQPCTimeMS();
latestVideoTime = qwTime;
QWORD frameDelta = qwTime-lastStreamTime;
float fSeconds = float(frameDelta)*0.001f;
//Log(TEXT("frameDelta: %llu"), frameDelta);
lastStreamTime = qwTime;
#endif
bool bUpdateBPS = false;
profileIn("frame preprocessing and rendering");
//------------------------------------
if(bRequestKeyframe && keyframeWait > 0)
{
keyframeWait -= int(frameDelta);
if(keyframeWait <= 0)
{
GetVideoEncoder()->RequestKeyframe();
bRequestKeyframe = false;
}
}
if(!bPushToTalkDown && pushToTalkTimeLeft > 0)
{
pushToTalkTimeLeft -= int(frameDelta);
OSDebugOut(TEXT("time left: %d\r\n"), pushToTalkTimeLeft);
if(pushToTalkTimeLeft <= 0)
{
pushToTalkTimeLeft = 0;
bPushToTalkOn = false;
}
}
//------------------------------------
OSEnterMutex(hSceneMutex);
if(bResizeRenderView)
{
GS->ResizeView();
bResizeRenderView = false;
}
//------------------------------------
if(scene)
{
profileIn("scene->Preprocess");
scene->Preprocess();
for(UINT i=0; i<globalSources.Num(); i++)
globalSources[i].source->Preprocess();
profileOut;
scene->Tick(float(fSeconds));
for(UINT i=0; i<globalSources.Num(); i++)
globalSources[i].source->Tick(float(fSeconds));
}
//------------------------------------
QWORD curBytesSent = network->GetCurrentSentBytes();
curFramesDropped = network->NumDroppedFrames();
bpsTime += fSeconds;
if(bpsTime > 1.0f)
{
if(numSecondsWaited < 3)
++numSecondsWaited;
//bytesPerSec = DWORD(curBytesSent - lastBytesSent);
bytesPerSec = DWORD(curBytesSent - lastBytesSent[0]) / numSecondsWaited;
if(bpsTime > 2.0)
bpsTime = 0.0f;
else
bpsTime -= 1.0;
if(numSecondsWaited == 3)
{
lastBytesSent[0] = lastBytesSent[1];
lastBytesSent[1] = lastBytesSent[2];
lastBytesSent[2] = curBytesSent;
}
else
lastBytesSent[numSecondsWaited] = curBytesSent;
captureFPS = fpsCounter;
fpsCounter = 0;
bUpdateBPS = true;
}
fpsCounter++;
curStrain = network->GetPacketStrain();
EnableBlending(TRUE);
BlendFunction(GS_BLEND_SRCALPHA, GS_BLEND_INVSRCALPHA);
//------------------------------------
// render the mini render texture
LoadVertexShader(mainVertexShader);
LoadPixelShader(mainPixelShader);
SetRenderTarget(mainRenderTextures[curRenderTarget]);
Ortho(0.0f, baseSize.x, baseSize.y, 0.0f, -100.0f, 100.0f);
SetViewport(0, 0, baseSize.x, baseSize.y);
if(scene)
scene->Render();
//------------------------------------
if(bTransitioning)
{
if(!transitionTexture)
{
transitionTexture = CreateTexture(baseCX, baseCY, GS_BGRA, NULL, FALSE, TRUE);
if(transitionTexture)
{
D3D10Texture *d3dTransitionTex = static_cast<D3D10Texture*>(transitionTexture);
D3D10Texture *d3dSceneTex = static_cast<D3D10Texture*>(mainRenderTextures[lastRenderTarget]);
GetD3D()->CopyResource(d3dTransitionTex->texture, d3dSceneTex->texture);
}
else
bTransitioning = false;
}
else if(transitionAlpha >= 1.0f)
{
delete transitionTexture;
transitionTexture = NULL;
bTransitioning = false;
}
}
if(bTransitioning)
{
EnableBlending(TRUE);
transitionAlpha += float(fSeconds)*5.0f;
if(transitionAlpha > 1.0f)
transitionAlpha = 1.0f;
}
else
EnableBlending(FALSE);
//------------------------------------
// render the mini view thingy
if(bRenderView)
{
// Cache
const Vect2 renderFrameSize = GetRenderFrameSize();
const Vect2 renderFrameOffset = GetRenderFrameOffset();
const Vect2 renderFrameCtrlSize = GetRenderFrameControlSize();
SetRenderTarget(NULL);
LoadVertexShader(mainVertexShader);
LoadPixelShader(mainPixelShader);
Ortho(0.0f, renderFrameCtrlSize.x, renderFrameCtrlSize.y, 0.0f, -100.0f, 100.0f);
if(renderFrameCtrlSize.x != oldRenderFrameCtrlWidth || renderFrameCtrlSize.y != oldRenderFrameCtrlHeight)
{
// User is drag resizing the window. We don't recreate the swap chains so our coordinates are wrong
SetViewport(0.0f, 0.0f, (float)oldRenderFrameCtrlWidth, (float)oldRenderFrameCtrlHeight);
}
else
SetViewport(0.0f, 0.0f, renderFrameCtrlSize.x, renderFrameCtrlSize.y);
// Draw background (Black if fullscreen, window colour otherwise)
if(bFullscreenMode)
ClearColorBuffer(0x000000);
else
ClearColorBuffer(GetSysColor(COLOR_BTNFACE));
if(bTransitioning)
{
BlendFunction(GS_BLEND_ONE, GS_BLEND_ZERO);
DrawSprite(transitionTexture, 0xFFFFFFFF,
renderFrameOffset.x, renderFrameOffset.y,
renderFrameOffset.x + renderFrameSize.x, renderFrameOffset.y + renderFrameSize.y);
BlendFunction(GS_BLEND_FACTOR, GS_BLEND_INVFACTOR, transitionAlpha);
}
DrawSprite(mainRenderTextures[curRenderTarget], 0xFFFFFFFF,
renderFrameOffset.x, renderFrameOffset.y,
renderFrameOffset.x + renderFrameSize.x, renderFrameOffset.y + renderFrameSize.y);
//draw selections if in edit mode
if(bEditMode && !bSizeChanging)
{
if(scene) {
LoadVertexShader(solidVertexShader);
LoadPixelShader(solidPixelShader);
solidPixelShader->SetColor(solidPixelShader->GetParameter(0), 0xFF0000);
scene->RenderSelections(solidPixelShader);
}
}
}
else if(bForceRenderViewErase)
{
InvalidateRect(hwndRenderFrame, NULL, TRUE);
UpdateWindow(hwndRenderFrame);
bForceRenderViewErase = false;
}
//------------------------------------
// actual stream output
LoadVertexShader(mainVertexShader);
LoadPixelShader(yuvScalePixelShader);
Texture *yuvRenderTexture = yuvRenderTextures[curRenderTarget];
SetRenderTarget(yuvRenderTexture);
if(downscale < 2.01)
yuvScalePixelShader->SetVector2(hScaleVal, 1.0f/baseSize);
else if(downscale < 3.01)
yuvScalePixelShader->SetVector2(hScaleVal, 1.0f/(outputSize*3.0f));
Ortho(0.0f, outputSize.x, outputSize.y, 0.0f, -100.0f, 100.0f);
SetViewport(0.0f, 0.0f, outputSize.x, outputSize.y);
//why am I using scaleSize instead of outputSize for the texture?
//because outputSize can be trimmed by up to three pixels due to 128-bit alignment.
//using the scale function with outputSize can cause slightly inaccurate scaled images
if(bTransitioning)
{
BlendFunction(GS_BLEND_ONE, GS_BLEND_ZERO);
DrawSpriteEx(transitionTexture, 0xFFFFFFFF, 0.0f, 0.0f, scaleSize.x, scaleSize.y, 0.0f, 0.0f, 1.0f, 1.0f);
BlendFunction(GS_BLEND_FACTOR, GS_BLEND_INVFACTOR, transitionAlpha);
}
DrawSpriteEx(mainRenderTextures[curRenderTarget], 0xFFFFFFFF, 0.0f, 0.0f, outputSize.x, outputSize.y, 0.0f, 0.0f, 1.0f, 1.0f);
//------------------------------------
if(bRenderView && !copyWait)
static_cast<D3D10System*>(GS)->swap->Present(0, 0);
OSLeaveMutex(hSceneMutex);
profileOut;
//------------------------------------
// present/upload
profileIn("video encoding and uploading");
bool bEncode = true;
if(copyWait)
{
copyWait--;
bEncode = false;
}
else
{
//audio sometimes takes a bit to start -- do not start processing frames until audio has started capturing
if(!bRecievedFirstAudioFrame)
{
static bool bWarnedAboutNoAudio = false;
if (qwTime-firstFrameTime > 10000 && !bWarnedAboutNoAudio)
{
bWarnedAboutNoAudio = true;
//AddStreamInfo (TEXT ("WARNING: OBS is not receiving audio frames. Please check your audio devices."), StreamInfoPriority_Critical);
}
bEncode = false;
}
else if(bFirstFrame)
{
firstFrameTime = qwTime;
bFirstFrame = false;
}
if(!bEncode)
{
if(curYUVTexture == (NUM_RENDER_BUFFERS-1))
curYUVTexture = 0;
else
curYUVTexture++;
}
}
if(bEncode)
{
curStreamTime = qwTime-firstFrameTime;
UINT prevCopyTexture = (curCopyTexture == 0) ? NUM_RENDER_BUFFERS-1 : curCopyTexture-1;
ID3D10Texture2D *copyTexture = copyTextures[curCopyTexture];
profileIn("CopyResource");
if(!bFirst420Encode && bUseThreaded420)
{
WaitForMultipleObjects(completeEvents.Num(), completeEvents.Array(), TRUE, INFINITE);
copyTexture->Unmap(0);
}
D3D10Texture *d3dYUV = static_cast<D3D10Texture*>(yuvRenderTextures[curYUVTexture]);
GetD3D()->CopyResource(copyTexture, d3dYUV->texture);
profileOut;
ID3D10Texture2D *prevTexture = copyTextures[prevCopyTexture];
if(bFirstImage) //ignore the first frame
bFirstImage = false;
else
{
HRESULT result;
D3D10_MAPPED_TEXTURE2D map;
if(SUCCEEDED(result = prevTexture->Map(0, D3D10_MAP_READ, 0, &map)))
{
int prevOutBuffer = (curOutBuffer == 0) ? NUM_OUT_BUFFERS-1 : curOutBuffer-1;
int nextOutBuffer = (curOutBuffer == NUM_OUT_BUFFERS-1) ? 0 : curOutBuffer+1;
EncoderPicture &prevPicOut = outPics[prevOutBuffer];
EncoderPicture &picOut = outPics[curOutBuffer];
EncoderPicture &nextPicOut = outPics[nextOutBuffer];
if(!bUsing444)
{
profileIn("conversion to 4:2:0");
if(bUseThreaded420)
{
outTimes[nextOutBuffer] = (DWORD)curStreamTime;
bool firstRun = threadedProfilers.Num() == 0;
if(firstRun)
threadedProfilers.SetSize(numThreads);
for(int i=0; i<numThreads; i++)
{
convertInfo[i].input = (LPBYTE)map.pData;
convertInfo[i].inPitch = map.RowPitch;
if(bUsingQSV)
{
mfxFrameData& data = nextPicOut.mfxOut->Data;
videoEncoder->RequestBuffers(&data);
convertInfo[i].outPitch = data.Pitch;
convertInfo[i].output[0] = data.Y;
convertInfo[i].output[1] = data.UV;
}
else
{
convertInfo[i].output[0] = nextPicOut.picOut->img.plane[0];
convertInfo[i].output[1] = nextPicOut.picOut->img.plane[1];
convertInfo[i].output[2] = nextPicOut.picOut->img.plane[2];
}
if(!firstRun)
threadedProfilers[i].~ProfilerNode();
::new (&threadedProfilers[i]) ProfilerNode(TEXT("Convert444Threads"), true);
threadedProfilers[i].MonitorThread(h420Threads[i]);
bUsingThreadedProfilers = true;
SetEvent(convertInfo[i].hSignalConvert);
}
if(bFirst420Encode)
bFirst420Encode = bEncode = false;
}
else
{
outTimes[curOutBuffer] = (DWORD)curStreamTime;
if(bUsingQSV)
{
mfxFrameData& data = picOut.mfxOut->Data;
videoEncoder->RequestBuffers(&data);
LPBYTE output[] = {data.Y, data.UV};
Convert444toNV12((LPBYTE)map.pData, outputCX, map.RowPitch, data.Pitch, outputCY, 0, outputCY, output);
}
else
Convert444toNV12((LPBYTE)map.pData, outputCX, map.RowPitch, outputCX, outputCY, 0, outputCY, picOut.picOut->img.plane);
prevTexture->Unmap(0);
}
profileOut;
}
else
{
outTimes[curOutBuffer] = (DWORD)curStreamTime;
picOut.picOut->img.i_stride[0] = map.RowPitch;
picOut.picOut->img.plane[0] = (uint8_t*)map.pData;
}
if(bEncode)
{
DWORD curFrameTimestamp = outTimes[prevOutBuffer];
//Log(TEXT("curFrameTimestamp: %u"), curFrameTimestamp);
//------------------------------------
FrameProcessInfo frameInfo;
frameInfo.firstFrameTime = firstFrameTime;
frameInfo.prevTexture = prevTexture;
if(bDupeFrames)
{
while(cfrTime < curFrameTimestamp)
{
DWORD frameTimeAdjust = frameTime;
cfrTimeAdjust += fpsTimeNumerator;
if(cfrTimeAdjust > fpsTimeDenominator)
{
cfrTimeAdjust -= fpsTimeDenominator;
++frameTimeAdjust;
}
DWORD halfTime = (frameTimeAdjust+1)/2;
EncoderPicture &nextPic = (curFrameTimestamp-cfrTime <= halfTime) ? picOut : prevPicOut;
//Log(TEXT("cfrTime: %u, time: %u"), cfrTime, curFrameTimestamp);
//these lines are just for counting duped frames
if(nextPic == lastPic)
++numTotalDuplicatedFrames;
else
lastPic = nextPic;
frameInfo.pic = &nextPic;
if(bUsingQSV)
frameInfo.pic->mfxOut->Data.TimeStamp = cfrTime;
else
frameInfo.pic->picOut->i_pts = cfrTime;
frameInfo.frameTimestamp = cfrTime;
ProcessFrame(frameInfo);
cfrTime += frameTimeAdjust;
//Log(TEXT("cfrTime: %u, chi frame: %u"), cfrTime, (curFrameTimestamp-cfrTime <= halfTime));
}
}
else
{
if(bUsingQSV)
picOut.mfxOut->Data.TimeStamp = curFrameTimestamp;
else
picOut.picOut->i_pts = curFrameTimestamp;
frameInfo.pic = &picOut;
frameInfo.frameTimestamp = curFrameTimestamp;
ProcessFrame(frameInfo);
}
if (!bRunning)
bufferedFrames = videoEncoder->GetBufferedFrames ();
}
if(bUsing444)
{
prevTexture->Unmap(0);
}
curOutBuffer = nextOutBuffer;
}
else
{
//We have to crash, or we end up deadlocking the thread when the convert threads are never signalled
if (result == DXGI_ERROR_DEVICE_REMOVED)
{
String message;
HRESULT reason = GetD3D()->GetDeviceRemovedReason();
switch (reason)
{
case DXGI_ERROR_DEVICE_RESET:
case DXGI_ERROR_DEVICE_HUNG:
message = TEXT("Your video card or driver froze and was reset. Please check for possible hardware / driver issues.");
break;
case DXGI_ERROR_DEVICE_REMOVED:
message = TEXT("Your video card disappeared from the system. Please check for possible hardware / driver issues.");
break;
case DXGI_ERROR_DRIVER_INTERNAL_ERROR:
message = TEXT("Your video driver reported an internal error. Please check for possible hardware / driver issues.");
break;
case DXGI_ERROR_INVALID_CALL:
message = TEXT("Your video driver reported an invalid call. Please check for possible driver issues.");
break;
default:
message = TEXT("DXGI_ERROR_DEVICE_REMOVED");
break;
}
CrashError (TEXT("Texture->Map failed: 0x%08x 0x%08x\r\n\r\n%s"), result, reason, message.Array());
}
else
CrashError (TEXT("Texture->Map failed: 0x%08x"), result);
}
}
if(curCopyTexture == (NUM_RENDER_BUFFERS-1))
curCopyTexture = 0;
else
curCopyTexture++;
if(curYUVTexture == (NUM_RENDER_BUFFERS-1))
curYUVTexture = 0;
else
curYUVTexture++;
if (bCongestionControl && bDynamicBitrateSupported && !bTestStream)
{
if (curStrain > 25)
{
if (qwTime - lastAdjustmentTime > 1500)
{
if (currentBitRate > 100)
{
currentBitRate = (int)(currentBitRate * (1.0 - (curStrain / 400)));
App->GetVideoEncoder()->SetBitRate(currentBitRate, -1);
if (!adjustmentStreamId)
adjustmentStreamId = App->AddStreamInfo (FormattedString(TEXT("Congestion detected, dropping bitrate to %d kbps"), currentBitRate).Array(), StreamInfoPriority_Low);
else
App->SetStreamInfo(adjustmentStreamId, FormattedString(TEXT("Congestion detected, dropping bitrate to %d kbps"), currentBitRate).Array());
bUpdateBPS = true;
}
lastAdjustmentTime = qwTime;
}
}
else if (currentBitRate < defaultBitRate && curStrain < 5 && lastStrain < 5)
{
if (qwTime - lastAdjustmentTime > 5000)
{
if (currentBitRate < defaultBitRate)
{
currentBitRate += (int)(defaultBitRate * 0.05);
if (currentBitRate > defaultBitRate)
currentBitRate = defaultBitRate;
}
App->GetVideoEncoder()->SetBitRate(currentBitRate, -1);
/*if (!adjustmentStreamId)
App->AddStreamInfo (FormattedString(TEXT("Congestion clearing, raising bitrate to %d kbps"), currentBitRate).Array(), StreamInfoPriority_Low);
else
App->SetStreamInfo(adjustmentStreamId, FormattedString(TEXT("Congestion clearing, raising bitrate to %d kbps"), currentBitRate).Array());*/
bUpdateBPS = true;
lastAdjustmentTime = qwTime;
}
}
}
}
lastRenderTarget = curRenderTarget;
if(curRenderTarget == (NUM_RENDER_BUFFERS-1))
curRenderTarget = 0;
else
curRenderTarget++;
if(bUpdateBPS || !CloseDouble(curStrain, lastStrain) || curFramesDropped != lastFramesDropped)
{
PostMessage(hwndMain, OBS_UPDATESTATUSBAR, 0, 0);
lastStrain = curStrain;
lastFramesDropped = curFramesDropped;
}
//------------------------------------
// we're about to sleep so we should flush the d3d command queue
profileIn("flush");
GetD3D()->Flush();
profileOut;
profileOut; //video encoding and uploading
profileOut; //frame
//------------------------------------
// frame sync
#ifdef USE_100NS_TIME
QWORD renderStopTime = GetQPCTime100NS();
sleepTargetTime += frameTime100ns;
if(bWasLaggedFrame = (sleepTargetTime <= renderStopTime))
{
numLongFrames++;
if(bLogLongFramesProfile && (numLongFrames/float(max(1, numTotalFrames)) * 100.) > logLongFramesProfilePercentage)
DumpLastProfileData();
}
else
SleepTo(sleepTargetTime);
#else
DWORD renderStopTime = OSGetTime();
DWORD totalTime = renderStopTime-renderStartTime;
if(totalTime > frameTimeAdjust)
{
numLongFrames++;
if(bLogLongFramesProfile && (numLongFrames/float(max(1, numTotalFrames)) * 100.) > logLongFramesProfilePercentage)
DumpLastProfileData();
}
else if(totalTime < frameTimeAdjust)
OSSleep(frameTimeAdjust-totalTime);
#endif
//OSDebugOut(TEXT("Frame adjust time: %d, "), frameTimeAdjust-totalTime);
numTotalFrames++;
}
if(!bUsing444)
{
if(bUseThreaded420)
{
for(int i=0; i<numThreads; i++)
{
if(h420Threads[i])
{
convertInfo[i].bKillThread = true;
SetEvent(convertInfo[i].hSignalConvert);
if(bUsingThreadedProfilers)
threadedProfilers[i].~ProfilerNode();
OSTerminateThread(h420Threads[i], 10000);
h420Threads[i] = NULL;
}
if(convertInfo[i].hSignalConvert)
{
CloseHandle(convertInfo[i].hSignalConvert);
convertInfo[i].hSignalConvert = NULL;
}
if(convertInfo[i].hSignalComplete)
{
CloseHandle(convertInfo[i].hSignalComplete);
convertInfo[i].hSignalComplete = NULL;
}
}
if(!bFirst420Encode)
{
ID3D10Texture2D *copyTexture = copyTextures[curCopyTexture];
copyTexture->Unmap(0);
}
}
if(bUsingQSV)
for(int i = 0; i < NUM_OUT_BUFFERS; i++)
delete outPics[i].mfxOut;
else
for(int i=0; i<NUM_OUT_BUFFERS; i++)
{
x264_picture_clean(outPics[i].picOut);
delete outPics[i].picOut;
}
}
Free(h420Threads);
Free(convertInfo);
Log(TEXT("Total frames rendered: %d, number of frames that lagged: %d (%0.2f%%) (it's okay for some frames to lag)"), numTotalFrames, numLongFrames, (double(numLongFrames)/double(numTotalFrames))*100.0);
if(bDupeFrames)
Log(TEXT("Total duplicated frames: %d (%0.2f%%)"), numTotalDuplicatedFrames, (double(numTotalDuplicatedFrames)/double(numTotalFrames))*100.0);
}
void
BasicCompositor::BeginFrame(const nsIntRegion& aInvalidRegion,
const gfx::Rect *aClipRectIn,
const gfx::Rect& aRenderBounds,
gfx::Rect *aClipRectOut /* = nullptr */,
gfx::Rect *aRenderBoundsOut /* = nullptr */)
{
mWidgetSize = mWidget->GetClientSize();
IntRect intRect = gfx::IntRect(IntPoint(), mWidgetSize);
Rect rect = Rect(0, 0, intRect.width, intRect.height);
// Sometimes the invalid region is larger than we want to draw.
nsIntRegion invalidRegionSafe;
invalidRegionSafe.And(aInvalidRegion, gfx::ThebesIntRect(intRect));
nsIntRect invalidRect = invalidRegionSafe.GetBounds();
mInvalidRect = IntRect(invalidRect.x, invalidRect.y, invalidRect.width, invalidRect.height);
mInvalidRegion = invalidRegionSafe;
if (aRenderBoundsOut) {
*aRenderBoundsOut = Rect();
}
if (mInvalidRect.width <= 0 || mInvalidRect.height <= 0) {
return;
}
if (mTarget) {
// If we have a copy target, then we don't have a widget-provided mDrawTarget (currently). Create a dummy
// placeholder so that CreateRenderTarget() works.
mDrawTarget = gfxPlatform::GetPlatform()->CreateOffscreenContentDrawTarget(IntSize(1,1), SurfaceFormat::B8G8R8A8);
} else {
mDrawTarget = mWidget->StartRemoteDrawing();
}
if (!mDrawTarget) {
return;
}
// Setup an intermediate render target to buffer all compositing. We will
// copy this into mDrawTarget (the widget), and/or mTarget in EndFrame()
RefPtr<CompositingRenderTarget> target = CreateRenderTarget(mInvalidRect, INIT_MODE_CLEAR);
if (!target) {
if (!mTarget) {
mWidget->EndRemoteDrawing();
}
return;
}
SetRenderTarget(target);
// We only allocate a surface sized to the invalidated region, so we need to
// translate future coordinates.
mRenderTarget->mDrawTarget->SetTransform(Matrix::Translation(-invalidRect.x,
-invalidRect.y));
gfxUtils::ClipToRegion(mRenderTarget->mDrawTarget, invalidRegionSafe);
if (aRenderBoundsOut) {
*aRenderBoundsOut = rect;
}
if (aClipRectIn) {
mRenderTarget->mDrawTarget->PushClipRect(*aClipRectIn);
} else {
mRenderTarget->mDrawTarget->PushClipRect(rect);
if (aClipRectOut) {
*aClipRectOut = rect;
}
}
}
void FRCPassPostProcessDOFRecombine::Process(FRenderingCompositePassContext& Context)
{
SCOPED_DRAW_EVENT(Context.RHICmdList, DOFRecombine);
const FPooledRenderTargetDesc* InputDesc = GetInputDesc(ePId_Input1);
if(!InputDesc)
{
// input is not hooked up correctly
return;
}
const FSceneView& View = Context.View;
const auto FeatureLevel = Context.GetFeatureLevel();
auto ShaderMap = Context.GetShaderMap();
FIntPoint TexSize = InputDesc->Extent;
// usually 1, 2, 4 or 8
uint32 ScaleToFullRes = GSceneRenderTargets.GetBufferSizeXY().X / TexSize.X;
FIntRect HalfResViewRect = View.ViewRect / ScaleToFullRes;
const FSceneRenderTargetItem& DestRenderTarget = PassOutputs[0].RequestSurface(Context);
// Set the view family's render target/viewport.
SetRenderTarget(Context.RHICmdList, DestRenderTarget.TargetableTexture, FTextureRHIRef());
// is optimized away if possible (RT size=view size, )
Context.RHICmdList.Clear(true, FLinearColor::Black, false, 1.0f, false, 0, View.ViewRect);
Context.SetViewportAndCallRHI(View.ViewRect);
// set the state
Context.RHICmdList.SetBlendState(TStaticBlendState<>::GetRHI());
Context.RHICmdList.SetRasterizerState(TStaticRasterizerState<>::GetRHI());
Context.RHICmdList.SetDepthStencilState(TStaticDepthStencilState<false, CF_Always>::GetRHI());
TShaderMapRef<FPostProcessVS> VertexShader(ShaderMap);
if (bNearBlurEnabled)
{
static FGlobalBoundShaderState BoundShaderState;
TShaderMapRef< FPostProcessDOFRecombinePS<1> > PixelShader(ShaderMap);
SetGlobalBoundShaderState(Context.RHICmdList, FeatureLevel, BoundShaderState, GFilterVertexDeclaration.VertexDeclarationRHI, *VertexShader, *PixelShader);
PixelShader->SetParameters(Context);
}
else
{
static FGlobalBoundShaderState BoundShaderState;
TShaderMapRef< FPostProcessDOFRecombinePS<0> > PixelShader(ShaderMap);
SetGlobalBoundShaderState(Context.RHICmdList, FeatureLevel, BoundShaderState, GFilterVertexDeclaration.VertexDeclarationRHI, *VertexShader, *PixelShader);
PixelShader->SetParameters(Context);
}
VertexShader->SetParameters(Context);
// Draw a quad mapping scene color to the view's render target
DrawRectangle(
Context.RHICmdList,
0, 0,
View.ViewRect.Width(), View.ViewRect.Height(),
HalfResViewRect.Min.X, HalfResViewRect.Min.Y,
HalfResViewRect.Width(), HalfResViewRect.Height(),
View.ViewRect.Size(),
TexSize,
*VertexShader,
EDRF_UseTriangleOptimization);
Context.RHICmdList.CopyToResolveTarget(DestRenderTarget.TargetableTexture, DestRenderTarget.ShaderResourceTexture, false, FResolveParams());
}
void ComputeDiffuseIrradiance(FRHICommandListImmediate& RHICmdList, ERHIFeatureLevel::Type FeatureLevel, FTextureRHIRef LightingSource, int32 LightingSourceMipIndex, FSHVectorRGB3* OutIrradianceEnvironmentMap)
{
auto ShaderMap = GetGlobalShaderMap(FeatureLevel);
FSceneRenderTargets& SceneContext = FSceneRenderTargets::Get(RHICmdList);
for (int32 CoefficientIndex = 0; CoefficientIndex < FSHVector3::MaxSHBasis; CoefficientIndex++)
{
// Copy the starting mip from the lighting texture, apply texel area weighting and appropriate SH coefficient
{
const int32 MipIndex = 0;
const int32 MipSize = GDiffuseIrradianceCubemapSize;
FSceneRenderTargetItem& EffectiveRT = GetEffectiveDiffuseIrradianceRenderTarget(SceneContext, MipIndex);
for (int32 CubeFace = 0; CubeFace < CubeFace_MAX; CubeFace++)
{
SetRenderTarget(RHICmdList, EffectiveRT.TargetableTexture, 0, CubeFace, NULL);
const FIntRect ViewRect(0, 0, MipSize, MipSize);
RHICmdList.SetViewport(0, 0, 0.0f, MipSize, MipSize, 1.0f);
RHICmdList.SetRasterizerState(TStaticRasterizerState<FM_Solid, CM_None>::GetRHI());
RHICmdList.SetDepthStencilState(TStaticDepthStencilState<false, CF_Always>::GetRHI());
RHICmdList.SetBlendState(TStaticBlendState<>::GetRHI());
TShaderMapRef<FCopyDiffuseIrradiancePS> PixelShader(ShaderMap);
TShaderMapRef<FScreenVS> VertexShader(GetGlobalShaderMap(FeatureLevel));
SetGlobalBoundShaderState(RHICmdList, FeatureLevel, CopyDiffuseIrradianceShaderState, GFilterVertexDeclaration.VertexDeclarationRHI, *VertexShader, *PixelShader);
PixelShader->SetParameters(RHICmdList, CubeFace, LightingSourceMipIndex, CoefficientIndex, MipSize, LightingSource);
DrawRectangle(
RHICmdList,
ViewRect.Min.X, ViewRect.Min.Y,
ViewRect.Width(), ViewRect.Height(),
ViewRect.Min.X, ViewRect.Min.Y,
ViewRect.Width(), ViewRect.Height(),
FIntPoint(ViewRect.Width(), ViewRect.Height()),
FIntPoint(MipSize, MipSize),
*VertexShader);
RHICmdList.CopyToResolveTarget(EffectiveRT.TargetableTexture, EffectiveRT.ShaderResourceTexture, true, FResolveParams(FResolveRect(), (ECubeFace)CubeFace, MipIndex));
}
}
const int32 NumMips = FMath::CeilLogTwo(GDiffuseIrradianceCubemapSize) + 1;
{
// Accumulate all the texel values through downsampling to 1x1 mip
for (int32 MipIndex = 1; MipIndex < NumMips; MipIndex++)
{
const int32 SourceMipIndex = FMath::Max(MipIndex - 1, 0);
const int32 MipSize = 1 << (NumMips - MipIndex - 1);
FSceneRenderTargetItem& EffectiveRT = GetEffectiveDiffuseIrradianceRenderTarget(SceneContext, MipIndex);
FSceneRenderTargetItem& EffectiveSource = GetEffectiveDiffuseIrradianceSourceTexture(SceneContext, MipIndex);
check(EffectiveRT.TargetableTexture != EffectiveSource.ShaderResourceTexture);
for (int32 CubeFace = 0; CubeFace < CubeFace_MAX; CubeFace++)
{
SetRenderTarget(RHICmdList, EffectiveRT.TargetableTexture, MipIndex, CubeFace, NULL);
const FIntRect ViewRect(0, 0, MipSize, MipSize);
RHICmdList.SetViewport(0, 0, 0.0f, MipSize, MipSize, 1.0f);
RHICmdList.SetRasterizerState(TStaticRasterizerState<FM_Solid, CM_None>::GetRHI());
RHICmdList.SetDepthStencilState(TStaticDepthStencilState<false, CF_Always>::GetRHI());
RHICmdList.SetBlendState(TStaticBlendState<>::GetRHI());
TShaderMapRef<FAccumulateDiffuseIrradiancePS> PixelShader(ShaderMap);
TShaderMapRef<FScreenVS> VertexShader(GetGlobalShaderMap(FeatureLevel));
SetGlobalBoundShaderState(RHICmdList, FeatureLevel, DiffuseIrradianceAccumulateShaderState, GFilterVertexDeclaration.VertexDeclarationRHI, *VertexShader, *PixelShader);
PixelShader->SetParameters(RHICmdList, CubeFace, NumMips, SourceMipIndex, CoefficientIndex, EffectiveSource.ShaderResourceTexture);
DrawRectangle(
RHICmdList,
ViewRect.Min.X, ViewRect.Min.Y,
ViewRect.Width(), ViewRect.Height(),
ViewRect.Min.X, ViewRect.Min.Y,
ViewRect.Width(), ViewRect.Height(),
FIntPoint(ViewRect.Width(), ViewRect.Height()),
FIntPoint(MipSize, MipSize),
*VertexShader);
RHICmdList.CopyToResolveTarget(EffectiveRT.TargetableTexture, EffectiveRT.ShaderResourceTexture, true, FResolveParams(FResolveRect(), (ECubeFace)CubeFace, MipIndex));
}
}
}
{
// Gather the cubemap face results and normalize, copy this coefficient to FSceneRenderTargets::Get(RHICmdList).SkySHIrradianceMap
FSceneRenderTargetItem& EffectiveRT = FSceneRenderTargets::Get(RHICmdList).SkySHIrradianceMap->GetRenderTargetItem();
//load/store actions so we don't lose results as we render one pixel at a time on tile renderers.
FRHIRenderTargetView RTV(EffectiveRT.TargetableTexture, 0, -1, ERenderTargetLoadAction::ELoad, ERenderTargetStoreAction::EStore);
RHICmdList.SetRenderTargets(1, &RTV, nullptr, 0, nullptr);
const FIntRect ViewRect(CoefficientIndex, 0, CoefficientIndex + 1, 1);
RHICmdList.SetViewport(0, 0, 0.0f, FSHVector3::MaxSHBasis, 1, 1.0f);
RHICmdList.SetRasterizerState(TStaticRasterizerState<FM_Solid, CM_None>::GetRHI());
RHICmdList.SetDepthStencilState(TStaticDepthStencilState<false, CF_Always>::GetRHI());
RHICmdList.SetBlendState(TStaticBlendState<>::GetRHI());
TShaderMapRef<FScreenVS> VertexShader(ShaderMap);
TShaderMapRef<FAccumulateCubeFacesPS> PixelShader(ShaderMap);
SetGlobalBoundShaderState(RHICmdList, FeatureLevel, AccumulateCubeFacesBoundShaderState, GFilterVertexDeclaration.VertexDeclarationRHI, *VertexShader, *PixelShader);
const int32 SourceMipIndex = NumMips - 1;
const int32 MipSize = 1;
FSceneRenderTargetItem& EffectiveSource = GetEffectiveDiffuseIrradianceRenderTarget(SceneContext, SourceMipIndex);
PixelShader->SetParameters(RHICmdList, SourceMipIndex, EffectiveSource.ShaderResourceTexture);
DrawRectangle(
RHICmdList,
ViewRect.Min.X, ViewRect.Min.Y,
ViewRect.Width(), ViewRect.Height(),
0, 0,
MipSize, MipSize,
FIntPoint(FSHVector3::MaxSHBasis, 1),
FIntPoint(MipSize, MipSize),
*VertexShader);
RHICmdList.CopyToResolveTarget(EffectiveRT.TargetableTexture, EffectiveRT.ShaderResourceTexture, true, FResolveParams());
}
}
{
// Read back the completed SH environment map
FSceneRenderTargetItem& EffectiveRT = FSceneRenderTargets::Get(RHICmdList).SkySHIrradianceMap->GetRenderTargetItem();
check(EffectiveRT.ShaderResourceTexture->GetFormat() == PF_FloatRGBA);
TArray<FFloat16Color> SurfaceData;
RHICmdList.ReadSurfaceFloatData(EffectiveRT.ShaderResourceTexture, FIntRect(0, 0, FSHVector3::MaxSHBasis, 1), SurfaceData, CubeFace_PosX, 0, 0);
check(SurfaceData.Num() == FSHVector3::MaxSHBasis);
for (int32 CoefficientIndex = 0; CoefficientIndex < FSHVector3::MaxSHBasis; CoefficientIndex++)
{
const FLinearColor CoefficientValue(SurfaceData[CoefficientIndex]);
OutIrradianceEnvironmentMap->R.V[CoefficientIndex] = CoefficientValue.R;
OutIrradianceEnvironmentMap->G.V[CoefficientIndex] = CoefficientValue.G;
OutIrradianceEnvironmentMap->B.V[CoefficientIndex] = CoefficientValue.B;
}
}
}
void FRCPassPostProcessScreenSpaceReflections::Process(FRenderingCompositePassContext& Context)
{
SCOPED_DRAW_EVENT(Context.RHICmdList, ScreenSpaceReflections);
const FSceneView& View = Context.View;
const auto FeatureLevel = Context.GetFeatureLevel();
const FSceneRenderTargetItem& DestRenderTarget = PassOutputs[0].RequestSurface(Context);
// Set the view family's render target/viewport.
SetRenderTarget(Context.RHICmdList, DestRenderTarget.TargetableTexture, FTextureRHIRef());
Context.RHICmdList.Clear(true, FLinearColor(0, 0, 0, 0), false, 1.0f, false, 0, FIntRect());
Context.SetViewportAndCallRHI(View.ViewRect);
// set the state
Context.RHICmdList.SetBlendState(TStaticBlendState<>::GetRHI());
Context.RHICmdList.SetRasterizerState(TStaticRasterizerState<>::GetRHI());
Context.RHICmdList.SetDepthStencilState(TStaticDepthStencilState<false, CF_Always>::GetRHI());
int SSRQuality = ComputeSSRQuality(View.FinalPostProcessSettings.ScreenSpaceReflectionQuality);
SSRQuality = FMath::Clamp(SSRQuality, 1, 4);
uint32 iPreFrame = bPrevFrame ? 1 : 0;
if (View.Family->EngineShowFlags.VisualizeSSR)
{
iPreFrame = 0;
SSRQuality = 0;
}
TShaderMapRef< FPostProcessVS > VertexShader(Context.GetShaderMap());
#define CASE(A, B) \
case (A + 2 * (B + 3 * 0 )): \
{ \
TShaderMapRef< FPostProcessScreenSpaceReflectionsPS<A, B> > PixelShader(Context.GetShaderMap()); \
static FGlobalBoundShaderState BoundShaderState; \
SetGlobalBoundShaderState(Context.RHICmdList, FeatureLevel, BoundShaderState, GFilterVertexDeclaration.VertexDeclarationRHI, *VertexShader, *PixelShader); \
VertexShader->SetParameters(Context); \
PixelShader->SetParameters(Context); \
}; \
break
switch (iPreFrame + 2 * (SSRQuality + 3 * 0))
{
CASE(0,0);
CASE(0,1); CASE(1,1);
CASE(0,2); CASE(1,2);
CASE(0,3); CASE(1,3);
CASE(0,4); CASE(1,4);
default:
check(!"Missing case in FRCPassPostProcessScreenSpaceReflections");
}
#undef CASE
// Draw a quad mapping scene color to the view's render target
DrawRectangle(
Context.RHICmdList,
0, 0,
View.ViewRect.Width(), View.ViewRect.Height(),
View.ViewRect.Min.X, View.ViewRect.Min.Y,
View.ViewRect.Width(), View.ViewRect.Height(),
View.ViewRect.Size(),
GSceneRenderTargets.GetBufferSizeXY(),
*VertexShader,
EDRF_UseTriangleOptimization);
Context.RHICmdList.CopyToResolveTarget(DestRenderTarget.TargetableTexture, DestRenderTarget.ShaderResourceTexture, false, FResolveParams());
}
void CRender::BeginRender()
{
SetRenderTarget(NULL);
_pCoreRenderer->Clear();
}
void CRender::EndRender()
{
SetRenderTarget(NULL);
_pCoreRenderer->Present();
}
void BuildHZB( FRHICommandListImmediate& RHICmdList, FViewInfo& View )
{
QUICK_SCOPE_CYCLE_COUNTER(STAT_BuildHZB);
// View.ViewRect.{Width,Height}() are most likely to be < 2^24, so the float
// conversion won't loss any precision (assuming float have 23bits for mantissa)
const int32 NumMipsX = FMath::Max(FPlatformMath::CeilToInt(FMath::Log2(float(View.ViewRect.Width()))) - 1, 1);
const int32 NumMipsY = FMath::Max(FPlatformMath::CeilToInt(FMath::Log2(float(View.ViewRect.Height()))) - 1, 1);
const uint32 NumMips = FMath::Max(NumMipsX, NumMipsY);
// Must be power of 2
const FIntPoint HZBSize( 1 << NumMipsX, 1 << NumMipsY );
View.HZBMipmap0Size = HZBSize;
FPooledRenderTargetDesc Desc(FPooledRenderTargetDesc::Create2DDesc(HZBSize, PF_R16F, FClearValueBinding::None, TexCreate_None, TexCreate_RenderTargetable | TexCreate_ShaderResource | TexCreate_NoFastClear, false, NumMips));
Desc.Flags |= TexCreate_FastVRAM;
GRenderTargetPool.FindFreeElement(RHICmdList, Desc, View.HZB, TEXT("HZB") );
FSceneRenderTargetItem& HZBRenderTarget = View.HZB->GetRenderTargetItem();
FTextureRHIParamRef HZBRenderTargetRef = HZBRenderTarget.TargetableTexture.GetReference();
// Mip 0
{
SCOPED_DRAW_EVENTF(RHICmdList, BuildHZB, TEXT("HZB SetupMip 0 %dx%d"), HZBSize.X, HZBSize.Y);
SetRenderTarget(RHICmdList, HZBRenderTarget.TargetableTexture, 0, NULL);
RHICmdList.SetBlendState(TStaticBlendState<>::GetRHI());
RHICmdList.SetRasterizerState(TStaticRasterizerState<>::GetRHI());
RHICmdList.SetDepthStencilState(TStaticDepthStencilState< false, CF_Always >::GetRHI());
TShaderMapRef< FPostProcessVS > VertexShader(View.ShaderMap);
TShaderMapRef< THZBBuildPS<0> > PixelShader(View.ShaderMap);
static FGlobalBoundShaderState BoundShaderState;
SetGlobalBoundShaderState(RHICmdList, View.GetFeatureLevel(), BoundShaderState, GFilterVertexDeclaration.VertexDeclarationRHI, *VertexShader, *PixelShader);
// Imperfect sampling, doesn't matter too much
PixelShader->SetParameters( RHICmdList, View );
RHICmdList.SetViewport(0, 0, 0.0f, HZBSize.X, HZBSize.Y, 1.0f);
DrawRectangle(
RHICmdList,
0, 0,
HZBSize.X, HZBSize.Y,
View.ViewRect.Min.X, View.ViewRect.Min.Y,
View.ViewRect.Width(), View.ViewRect.Height(),
HZBSize,
FSceneRenderTargets::Get(RHICmdList).GetBufferSizeXY(),
*VertexShader,
EDRF_UseTriangleOptimization);
//Use RWBarrier since we don't transition individual subresources. Basically treat the whole texture as R/W as we walk down the mip chain.
RHICmdList.TransitionResources(EResourceTransitionAccess::ERWSubResBarrier, &HZBRenderTargetRef, 1);
}
FIntPoint SrcSize = HZBSize;
FIntPoint DstSize = SrcSize / 2;
SCOPED_DRAW_EVENTF(RHICmdList, BuildHZB, TEXT("HZB SetupMips 1..%d %dx%d Mips:%d"), NumMips - 1, DstSize.X, DstSize.Y);
// Downsampling...
for( uint8 MipIndex = 1; MipIndex < NumMips; MipIndex++ )
{
DstSize.X = FMath::Max(DstSize.X, 1);
DstSize.Y = FMath::Max(DstSize.Y, 1);
SetRenderTarget(RHICmdList, HZBRenderTarget.TargetableTexture, MipIndex, NULL);
RHICmdList.SetBlendState(TStaticBlendState<>::GetRHI());
RHICmdList.SetRasterizerState(TStaticRasterizerState<>::GetRHI());
RHICmdList.SetDepthStencilState(TStaticDepthStencilState< false, CF_Always >::GetRHI());
TShaderMapRef< FPostProcessVS > VertexShader(View.ShaderMap);
TShaderMapRef< THZBBuildPS<1> > PixelShader(View.ShaderMap);
static FGlobalBoundShaderState BoundShaderState;
SetGlobalBoundShaderState(RHICmdList, View.GetFeatureLevel(), BoundShaderState, GFilterVertexDeclaration.VertexDeclarationRHI, *VertexShader, *PixelShader);
PixelShader->SetParameters(RHICmdList, View, SrcSize, HZBRenderTarget.MipSRVs[ MipIndex - 1 ] );
RHICmdList.SetViewport(0, 0, 0.0f, DstSize.X, DstSize.Y, 1.0f);
DrawRectangle(
RHICmdList,
0, 0,
DstSize.X, DstSize.Y,
0, 0,
SrcSize.X, SrcSize.Y,
DstSize,
SrcSize,
*VertexShader,
EDRF_UseTriangleOptimization);
SrcSize /= 2;
DstSize /= 2;
//Use ERWSubResBarrier since we don't transition individual subresources. Basically treat the whole texture as R/W as we walk down the mip chain.
RHICmdList.TransitionResources(EResourceTransitionAccess::ERWSubResBarrier, &HZBRenderTargetRef, 1);
}
GRenderTargetPool.VisualizeTexture.SetCheckPoint( RHICmdList, View.HZB );
}
void FHZBOcclusionTester::Submit(FRHICommandListImmediate& RHICmdList, const FViewInfo& View)
{
SCOPED_DRAW_EVENT(RHICmdList, SubmitHZB);
FSceneViewState* ViewState = (FSceneViewState*)View.State;
if( !ViewState )
{
return;
}
TRefCountPtr< IPooledRenderTarget > BoundsCenterTexture;
TRefCountPtr< IPooledRenderTarget > BoundsExtentTexture;
{
uint32 Flags = TexCreate_ShaderResource | TexCreate_Dynamic;
FPooledRenderTargetDesc Desc( FPooledRenderTargetDesc::Create2DDesc( FIntPoint( SizeX, SizeY ), PF_A32B32G32R32F, FClearValueBinding::None, Flags, TexCreate_None, false ) );
GRenderTargetPool.FindFreeElement(RHICmdList, Desc, BoundsCenterTexture, TEXT("HZBBoundsCenter") );
GRenderTargetPool.FindFreeElement(RHICmdList, Desc, BoundsExtentTexture, TEXT("HZBBoundsExtent") );
}
TRefCountPtr< IPooledRenderTarget > ResultsTextureGPU;
{
FPooledRenderTargetDesc Desc( FPooledRenderTargetDesc::Create2DDesc( FIntPoint( SizeX, SizeY ), PF_B8G8R8A8, FClearValueBinding::None, TexCreate_None, TexCreate_RenderTargetable, false ) );
GRenderTargetPool.FindFreeElement(RHICmdList, Desc, ResultsTextureGPU, TEXT("HZBResultsGPU") );
}
{
#if 0
static float CenterBuffer[ SizeX * SizeY ][4];
static float ExtentBuffer[ SizeX * SizeY ][4];
FMemory::Memset( CenterBuffer, 0, sizeof( CenterBuffer ) );
FMemory::Memset( ExtentBuffer, 0, sizeof( ExtentBuffer ) );
const uint32 NumPrimitives = Primitives.Num();
for( uint32 i = 0; i < NumPrimitives; i++ )
{
const FOcclusionPrimitive& Primitive = Primitives[i];
CenterBuffer[i][0] = Primitive.Center.X;
CenterBuffer[i][1] = Primitive.Center.Y;
CenterBuffer[i][2] = Primitive.Center.Z;
CenterBuffer[i][3] = 0.0f;
ExtentBuffer[i][0] = Primitive.Extent.X;
ExtentBuffer[i][1] = Primitive.Extent.Y;
ExtentBuffer[i][2] = Primitive.Extent.Z;
ExtentBuffer[i][3] = 1.0f;
}
FUpdateTextureRegion2D Region( 0, 0, 0, 0, SizeX, SizeY );
RHIUpdateTexture2D( (FTexture2DRHIRef&)BoundsCenterTexture->GetRenderTargetItem().ShaderResourceTexture, 0, Region, SizeX * 4 * sizeof( float ), (uint8*)CenterBuffer );
RHIUpdateTexture2D( (FTexture2DRHIRef&)BoundsExtentTexture->GetRenderTargetItem().ShaderResourceTexture, 0, Region, SizeX * 4 * sizeof( float ), (uint8*)ExtentBuffer );
#elif 0
static float CenterBuffer[ SizeX * SizeY ][4];
static float ExtentBuffer[ SizeX * SizeY ][4];
{
QUICK_SCOPE_CYCLE_COUNTER(STAT_HZBPackPrimitiveData);
FMemory::Memset( CenterBuffer, 0, sizeof( CenterBuffer ) );
FMemory::Memset( ExtentBuffer, 0, sizeof( ExtentBuffer ) );
const uint32 NumPrimitives = Primitives.Num();
for( uint32 i = 0; i < NumPrimitives; i++ )
{
const FOcclusionPrimitive& Primitive = Primitives[i];
uint32 x = FMath::ReverseMortonCode2( i >> 0 );
uint32 y = FMath::ReverseMortonCode2( i >> 1 );
uint32 m = x + y * SizeX;
CenterBuffer[m][0] = Primitive.Center.X;
CenterBuffer[m][1] = Primitive.Center.Y;
CenterBuffer[m][2] = Primitive.Center.Z;
CenterBuffer[m][3] = 0.0f;
ExtentBuffer[m][0] = Primitive.Extent.X;
ExtentBuffer[m][1] = Primitive.Extent.Y;
ExtentBuffer[m][2] = Primitive.Extent.Z;
ExtentBuffer[m][3] = 1.0f;
}
}
QUICK_SCOPE_CYCLE_COUNTER(STAT_HZBUpdateTextures);
FUpdateTextureRegion2D Region( 0, 0, 0, 0, SizeX, SizeY );
RHIUpdateTexture2D( (FTexture2DRHIRef&)BoundsCenterTexture->GetRenderTargetItem().ShaderResourceTexture, 0, Region, SizeX * 4 * sizeof( float ), (uint8*)CenterBuffer );
RHIUpdateTexture2D( (FTexture2DRHIRef&)BoundsExtentTexture->GetRenderTargetItem().ShaderResourceTexture, 0, Region, SizeX * 4 * sizeof( float ), (uint8*)ExtentBuffer );
#else
// Update in blocks to avoid large update
const uint32 BlockSize = 8;
const uint32 SizeInBlocksX = SizeX / BlockSize;
const uint32 SizeInBlocksY = SizeY / BlockSize;
const uint32 BlockStride = BlockSize * 4 * sizeof( float );
float CenterBuffer[ BlockSize * BlockSize ][4];
float ExtentBuffer[ BlockSize * BlockSize ][4];
const uint32 NumPrimitives = Primitives.Num();
for( uint32 i = 0; i < NumPrimitives; i += BlockSize * BlockSize )
{
const uint32 BlockEnd = FMath::Min( BlockSize * BlockSize, NumPrimitives - i );
for( uint32 b = 0; b < BlockEnd; b++ )
{
const FOcclusionPrimitive& Primitive = Primitives[ i + b ];
CenterBuffer[b][0] = Primitive.Center.X;
CenterBuffer[b][1] = Primitive.Center.Y;
CenterBuffer[b][2] = Primitive.Center.Z;
CenterBuffer[b][3] = 0.0f;
ExtentBuffer[b][0] = Primitive.Extent.X;
ExtentBuffer[b][1] = Primitive.Extent.Y;
ExtentBuffer[b][2] = Primitive.Extent.Z;
ExtentBuffer[b][3] = 1.0f;
}
// Clear rest of block
if( BlockEnd < BlockSize * BlockSize )
{
FMemory::Memset( (float*)CenterBuffer + BlockEnd * 4, 0, sizeof( CenterBuffer ) - BlockEnd * 4 * sizeof(float) );
FMemory::Memset( (float*)ExtentBuffer + BlockEnd * 4, 0, sizeof( ExtentBuffer ) - BlockEnd * 4 * sizeof(float) );
}
const int32 BlockIndex = i / (BlockSize * BlockSize);
const int32 BlockX = BlockIndex % SizeInBlocksX;
const int32 BlockY = BlockIndex / SizeInBlocksY;
FUpdateTextureRegion2D Region( BlockX * BlockSize, BlockY * BlockSize, 0, 0, BlockSize, BlockSize );
RHIUpdateTexture2D( (FTexture2DRHIRef&)BoundsCenterTexture->GetRenderTargetItem().ShaderResourceTexture, 0, Region, BlockStride, (uint8*)CenterBuffer );
RHIUpdateTexture2D( (FTexture2DRHIRef&)BoundsExtentTexture->GetRenderTargetItem().ShaderResourceTexture, 0, Region, BlockStride, (uint8*)ExtentBuffer );
}
#endif
Primitives.Empty();
}
// Draw test
{
SCOPED_DRAW_EVENT(RHICmdList, TestHZB);
SetRenderTarget(RHICmdList, ResultsTextureGPU->GetRenderTargetItem().TargetableTexture, NULL);
RHICmdList.SetBlendState(TStaticBlendState<>::GetRHI());
RHICmdList.SetRasterizerState(TStaticRasterizerState<>::GetRHI());
RHICmdList.SetDepthStencilState(TStaticDepthStencilState< false, CF_Always >::GetRHI());
TShaderMapRef< FScreenVS > VertexShader(View.ShaderMap);
TShaderMapRef< FHZBTestPS > PixelShader(View.ShaderMap);
static FGlobalBoundShaderState BoundShaderState;
SetGlobalBoundShaderState(RHICmdList, View.GetFeatureLevel(), BoundShaderState, GFilterVertexDeclaration.VertexDeclarationRHI, *VertexShader, *PixelShader);
PixelShader->SetParameters(RHICmdList, View, BoundsCenterTexture->GetRenderTargetItem().ShaderResourceTexture, BoundsExtentTexture->GetRenderTargetItem().ShaderResourceTexture );
RHICmdList.SetViewport(0, 0, 0.0f, SizeX, SizeY, 1.0f);
// TODO draw quads covering blocks added above
DrawRectangle(
RHICmdList,
0, 0,
SizeX, SizeY,
0, 0,
SizeX, SizeY,
FIntPoint( SizeX, SizeY ),
FIntPoint( SizeX, SizeY ),
*VertexShader,
EDRF_UseTriangleOptimization);
}
GRenderTargetPool.VisualizeTexture.SetCheckPoint(RHICmdList, ResultsTextureGPU);
// Transfer memory GPU -> CPU
RHICmdList.CopyToResolveTarget(ResultsTextureGPU->GetRenderTargetItem().TargetableTexture, ResultsTextureCPU->GetRenderTargetItem().ShaderResourceTexture, false, FResolveParams());
}
void FDeferredShadingSceneRenderer::BeginOcclusionTests(FRHICommandListImmediate& RHICmdList, bool bRenderQueries)
{
SCOPE_CYCLE_COUNTER(STAT_BeginOcclusionTestsTime);
int32 NumBufferedFrames = FOcclusionQueryHelpers::GetNumBufferedFrames();
FSceneRenderTargets& SceneContext = FSceneRenderTargets::Get(RHICmdList);
const bool bUseDownsampledDepth = SceneContext.UseDownsizedOcclusionQueries() && IsValidRef(SceneContext.SmallDepthZ) && IsValidRef(SceneContext.GetSmallDepthSurface());
if (bUseDownsampledDepth)
{
SetRenderTarget(RHICmdList, NULL, SceneContext.GetSmallDepthSurface(), ESimpleRenderTargetMode::EExistingColorAndDepth, FExclusiveDepthStencil::DepthRead_StencilWrite);
}
else
{
SetRenderTarget(RHICmdList, NULL, SceneContext.GetSceneDepthSurface(), ESimpleRenderTargetMode::EExistingColorAndDepth, FExclusiveDepthStencil::DepthRead_StencilWrite);
}
if (bRenderQueries)
{
RHICmdList.BeginOcclusionQueryBatch();
// Perform occlusion queries for each view
for (int32 ViewIndex = 0; ViewIndex < Views.Num(); ViewIndex++)
{
SCOPED_DRAW_EVENT(RHICmdList, BeginOcclusionTests);
FViewInfo& View = Views[ViewIndex];
if (bUseDownsampledDepth)
{
const uint32 DownsampledX = FMath::TruncToInt(View.ViewRect.Min.X / SceneContext.GetSmallColorDepthDownsampleFactor());
const uint32 DownsampledY = FMath::TruncToInt(View.ViewRect.Min.Y / SceneContext.GetSmallColorDepthDownsampleFactor());
const uint32 DownsampledSizeX = FMath::TruncToInt(View.ViewRect.Width() / SceneContext.GetSmallColorDepthDownsampleFactor());
const uint32 DownsampledSizeY = FMath::TruncToInt(View.ViewRect.Height() / SceneContext.GetSmallColorDepthDownsampleFactor());
// Setup the viewport for rendering to the downsampled depth buffer
RHICmdList.SetViewport(DownsampledX, DownsampledY, 0.0f, DownsampledX + DownsampledSizeX, DownsampledY + DownsampledSizeY, 1.0f);
}
else
{
RHICmdList.SetViewport(View.ViewRect.Min.X, View.ViewRect.Min.Y, 0.0f, View.ViewRect.Max.X, View.ViewRect.Max.Y, 1.0f);
}
FSceneViewState* ViewState = (FSceneViewState*)View.State;
if (ViewState && !View.bDisableQuerySubmissions)
{
// Depth tests, no depth writes, no color writes, opaque, solid rasterization wo/ backface culling.
RHICmdList.SetDepthStencilState(TStaticDepthStencilState<false, CF_DepthNearOrEqual>::GetRHI());
// We only need to render the front-faces of the culling geometry (this halves the amount of pixels we touch)
RHICmdList.SetRasterizerState(View.bReverseCulling ? TStaticRasterizerState<FM_Solid, CM_CCW>::GetRHI() : TStaticRasterizerState<FM_Solid, CM_CW>::GetRHI());
RHICmdList.SetBlendState(TStaticBlendState<CW_NONE>::GetRHI());
// Lookup the vertex shader.
TShaderMapRef<FOcclusionQueryVS> VertexShader(View.ShaderMap);
SetGlobalBoundShaderState(RHICmdList, View.GetFeatureLevel(), OcclusionTestBoundShaderState, GetVertexDeclarationFVector3(), *VertexShader, NULL);
VertexShader->SetParameters(RHICmdList, View);
// Issue this frame's occlusion queries (occlusion queries from last frame may still be in flight)
const uint32 QueryIndex = FOcclusionQueryHelpers::GetQueryIssueIndex(ViewState->PendingPrevFrameNumber, NumBufferedFrames);
FSceneViewState::ShadowKeyOcclusionQueryMap& ShadowOcclusionQueryMap = ViewState->ShadowOcclusionQueryMaps[QueryIndex];
// Clear primitives which haven't been visible recently out of the occlusion history, and reset old pending occlusion queries.
ViewState->TrimOcclusionHistory(RHICmdList, ViewFamily.CurrentRealTime, ViewFamily.CurrentRealTime - GEngine->PrimitiveProbablyVisibleTime, ViewFamily.CurrentRealTime, ViewState->OcclusionFrameCounter);
// Give back all these occlusion queries to the pool.
for (TMap<FSceneViewState::FProjectedShadowKey, FRenderQueryRHIRef>::TIterator QueryIt(ShadowOcclusionQueryMap); QueryIt; ++QueryIt)
{
//FRenderQueryRHIParamRef Query = QueryIt.Value();
//check( Query.GetRefCount() == 1 );
ViewState->OcclusionQueryPool.ReleaseQuery(QueryIt.Value());
}
ShadowOcclusionQueryMap.Reset();
{
SCOPED_DRAW_EVENT(RHICmdList, ShadowFrustumQueries);
for (TSparseArray<FLightSceneInfoCompact>::TConstIterator LightIt(Scene->Lights); LightIt; ++LightIt)
{
const FVisibleLightInfo& VisibleLightInfo = VisibleLightInfos[LightIt.GetIndex()];
for (int32 ShadowIndex = 0; ShadowIndex < VisibleLightInfo.AllProjectedShadows.Num(); ShadowIndex++)
{
const FProjectedShadowInfo& ProjectedShadowInfo = *VisibleLightInfo.AllProjectedShadows[ShadowIndex];
if (ProjectedShadowInfo.DependentView && ProjectedShadowInfo.DependentView != &View)
{
continue;
}
if (ProjectedShadowInfo.CascadeSettings.bOnePassPointLightShadow)
{
FLightSceneProxy& LightProxy = *(ProjectedShadowInfo.GetLightSceneInfo().Proxy);
// Query one pass point light shadows separately because they don't have a shadow frustum, they have a bounding sphere instead.
FSphere LightBounds = LightProxy.GetBoundingSphere();
const bool bCameraInsideLightGeometry = ((FVector)View.ViewMatrices.ViewOrigin - LightBounds.Center).SizeSquared() < FMath::Square(LightBounds.W * 1.05f + View.NearClippingDistance * 2.0f);
if (!bCameraInsideLightGeometry)
{
const FRenderQueryRHIRef ShadowOcclusionQuery = ViewState->OcclusionQueryPool.AllocateQuery();
RHICmdList.BeginRenderQuery(ShadowOcclusionQuery);
FSceneViewState::FProjectedShadowKey Key(ProjectedShadowInfo);
checkSlow(ShadowOcclusionQueryMap.Find(Key) == NULL);
ShadowOcclusionQueryMap.Add(Key, ShadowOcclusionQuery);
// Draw bounding sphere
VertexShader->SetParametersWithBoundingSphere(RHICmdList, View, LightBounds);
StencilingGeometry::DrawVectorSphere(RHICmdList);
RHICmdList.EndRenderQuery(ShadowOcclusionQuery);
}
}
// Don't query preshadows, since they are culled if their subject is occluded.
// Don't query if any subjects are visible because the shadow frustum will be definitely unoccluded
else if (!ProjectedShadowInfo.IsWholeSceneDirectionalShadow() && !ProjectedShadowInfo.bPreShadow && !ProjectedShadowInfo.SubjectsVisible(View))
{
IssueProjectedShadowOcclusionQuery(RHICmdList, View, ProjectedShadowInfo, *VertexShader, NumBufferedFrames);
}
}
// Issue occlusion queries for all per-object projected shadows that we would have rendered but were occluded last frame.
for (int32 ShadowIndex = 0; ShadowIndex < VisibleLightInfo.OccludedPerObjectShadows.Num(); ShadowIndex++)
{
const FProjectedShadowInfo& ProjectedShadowInfo = *VisibleLightInfo.OccludedPerObjectShadows[ShadowIndex];
IssueProjectedShadowOcclusionQuery(RHICmdList, View, ProjectedShadowInfo, *VertexShader, NumBufferedFrames);
}
}
}
// Don't do primitive occlusion if we have a view parent or are frozen.
#if !(UE_BUILD_SHIPPING || UE_BUILD_TEST)
if (!ViewState->HasViewParent() && !ViewState->bIsFrozen)
#endif
{
VertexShader->SetParameters(RHICmdList, View);
{
SCOPED_DRAW_EVENT(RHICmdList, IndividualQueries);
View.IndividualOcclusionQueries.Flush(RHICmdList);
}
{
SCOPED_DRAW_EVENT(RHICmdList, GroupedQueries);
View.GroupedOcclusionQueries.Flush(RHICmdList);
}
}
}
}
RHICmdList.EndOcclusionQueryBatch();
}
if (bUseDownsampledDepth)
{
// Restore default render target
SceneContext.BeginRenderingSceneColor(RHICmdList, ESimpleRenderTargetMode::EUninitializedColorExistingDepth, FExclusiveDepthStencil::DepthRead_StencilWrite);
}
}
void FRCPassPostProcessBloomSetup::Process(FRenderingCompositePassContext& Context)
{
SCOPED_DRAW_EVENT(Context.RHICmdList, PostProcessBloomSetup);
const FPooledRenderTargetDesc* InputDesc = GetInputDesc(ePId_Input0);
if(!InputDesc)
{
// input is not hooked up correctly
return;
}
const FSceneView& View = Context.View;
const FSceneViewFamily& ViewFamily = *(View.Family);
FIntPoint SrcSize = InputDesc->Extent;
FIntPoint DestSize = PassOutputs[0].RenderTargetDesc.Extent;
// e.g. 4 means the input texture is 4x smaller than the buffer size
uint32 ScaleFactor = GSceneRenderTargets.GetBufferSizeXY().X / SrcSize.X;
FIntRect SrcRect = View.ViewRect / ScaleFactor;
FIntRect DestRect = SrcRect;
const FSceneRenderTargetItem& DestRenderTarget = PassOutputs[0].RequestSurface(Context);
// Set the view family's render target/viewport.
SetRenderTarget(Context.RHICmdList, DestRenderTarget.TargetableTexture, FTextureRHIRef());
// is optimized away if possible (RT size=view size, )
Context.RHICmdList.Clear(true, FLinearColor::Black, false, 1.0f, false, 0, DestRect);
Context.SetViewportAndCallRHI(0, 0, 0.0f, DestSize.X, DestSize.Y, 1.0f );
// set the state
Context.RHICmdList.SetBlendState(TStaticBlendState<>::GetRHI());
Context.RHICmdList.SetRasterizerState(TStaticRasterizerState<>::GetRHI());
Context.RHICmdList.SetDepthStencilState(TStaticDepthStencilState<false, CF_Always>::GetRHI());
TShaderMapRef<FPostProcessBloomSetupVS> VertexShader(Context.GetShaderMap());
TShaderMapRef<FPostProcessBloomSetupPS> PixelShader(Context.GetShaderMap());
static FGlobalBoundShaderState BoundShaderState;
SetGlobalBoundShaderState(Context.RHICmdList, Context.GetFeatureLevel(), BoundShaderState, GFilterVertexDeclaration.VertexDeclarationRHI, *VertexShader, *PixelShader);
VertexShader->SetVS(Context);
PixelShader->SetPS(Context);
// Draw a quad mapping scene color to the view's render target
DrawRectangle(
Context.RHICmdList,
DestRect.Min.X, DestRect.Min.Y,
DestRect.Width(), DestRect.Height(),
SrcRect.Min.X, SrcRect.Min.Y,
SrcRect.Width(), SrcRect.Height(),
DestSize,
SrcSize,
*VertexShader,
EDRF_UseTriangleOptimization);
Context.RHICmdList.CopyToResolveTarget(DestRenderTarget.TargetableTexture, DestRenderTarget.ShaderResourceTexture, false, FResolveParams());
}
void FSteamVRHMD::RenderTexture_RenderThread(FRHICommandListImmediate& RHICmdList, FTexture2DRHIParamRef BackBuffer, FTexture2DRHIParamRef SrcTexture) const
{
check(IsInRenderingThread());
UpdateLayerTextures();
if (bSplashIsShown)
{
SetRenderTarget(RHICmdList, SrcTexture, FTextureRHIRef());
RHICmdList.Clear(true, FLinearColor(0, 0, 0, 0), false, (float)ERHIZBuffer::FarPlane, false, 0, FIntRect());
}
if (WindowMirrorMode != 0)
{
const uint32 ViewportWidth = BackBuffer->GetSizeX();
const uint32 ViewportHeight = BackBuffer->GetSizeY();
SetRenderTarget(RHICmdList, BackBuffer, FTextureRHIRef());
RHICmdList.SetViewport(0, 0, 0, ViewportWidth, ViewportHeight, 1.0f);
RHICmdList.SetBlendState(TStaticBlendState<>::GetRHI());
RHICmdList.SetRasterizerState(TStaticRasterizerState<>::GetRHI());
RHICmdList.SetDepthStencilState(TStaticDepthStencilState<false, CF_Always>::GetRHI());
const auto FeatureLevel = GMaxRHIFeatureLevel;
auto ShaderMap = GetGlobalShaderMap(FeatureLevel);
TShaderMapRef<FScreenVS> VertexShader(ShaderMap);
TShaderMapRef<FScreenPS> PixelShader(ShaderMap);
static FGlobalBoundShaderState BoundShaderState;
SetGlobalBoundShaderState(RHICmdList, FeatureLevel, BoundShaderState, RendererModule->GetFilterVertexDeclaration().VertexDeclarationRHI, *VertexShader, *PixelShader);
PixelShader->SetParameters(RHICmdList, TStaticSamplerState<SF_Bilinear>::GetRHI(), SrcTexture);
if (WindowMirrorMode == 1)
{
// need to clear when rendering only one eye since the borders won't be touched by the DrawRect below
RHICmdList.Clear(true, FLinearColor::Black, false, 0, false, 0, FIntRect());
RendererModule->DrawRectangle(
RHICmdList,
ViewportWidth / 4, 0,
ViewportWidth / 2, ViewportHeight,
0.1f, 0.2f,
0.3f, 0.6f,
FIntPoint(ViewportWidth, ViewportHeight),
FIntPoint(1, 1),
*VertexShader,
EDRF_Default);
}
else if (WindowMirrorMode == 2)
{
RendererModule->DrawRectangle(
RHICmdList,
0, 0,
ViewportWidth, ViewportHeight,
0.0f, 0.0f,
1.0f, 1.0f,
FIntPoint(ViewportWidth, ViewportHeight),
FIntPoint(1, 1),
*VertexShader,
EDRF_Default);
}
}
}
void FRCPassPostProcessVisualizeLPV::Process(FRenderingCompositePassContext& Context)
{
SCOPED_DRAW_EVENT(Context.RHICmdList, VisualizeLPV);
const FSceneView& View = Context.View;
const FSceneViewFamily& ViewFamily = *(View.Family);
// const FSceneRenderTargetItem& DestRenderTarget = PassOutputs[0].RequestSurface(Context);
const TRefCountPtr<IPooledRenderTarget> RenderTarget = GetInput(ePId_Input0)->GetOutput()->PooledRenderTarget;
const FSceneRenderTargetItem& DestRenderTarget = RenderTarget->GetRenderTargetItem();
// Set the view family's render target/viewport.
SetRenderTarget(Context.RHICmdList, DestRenderTarget.TargetableTexture, FTextureRHIRef());
{
// this is a helper class for FCanvas to be able to get screen size
class FRenderTargetTemp : public FRenderTarget
{
public:
const FSceneView& View;
const FTexture2DRHIRef Texture;
FRenderTargetTemp(const FSceneView& InView, const FTexture2DRHIRef InTexture)
: View(InView), Texture(InTexture)
{
}
virtual FIntPoint GetSizeXY() const
{
return View.ViewRect.Size();
};
virtual const FTexture2DRHIRef& GetRenderTargetTexture() const
{
return Texture;
}
} TempRenderTarget(View, (const FTexture2DRHIRef&)DestRenderTarget.TargetableTexture);
FCanvas Canvas(&TempRenderTarget, NULL, ViewFamily.CurrentRealTime, ViewFamily.CurrentWorldTime, ViewFamily.DeltaWorldTime, View.GetFeatureLevel());
float X = 30;
float Y = 28;
const float YStep = 14;
const float ColumnWidth = 250;
Canvas.DrawShadowedString( X, Y += YStep, TEXT("VisualizeLightPropagationVolume"), GetStatsFont(), FLinearColor(0.2f, 0.2f, 1));
Y += YStep;
const FLightPropagationVolumeSettings& Dest = View.FinalPostProcessSettings.BlendableManager.GetSingleFinalDataConst<FLightPropagationVolumeSettings>();
#define ENTRY(name)\
Canvas.DrawShadowedString( X, Y += YStep, TEXT(#name) TEXT(":"), GetStatsFont(), FLinearColor(1, 1, 1));\
Canvas.DrawShadowedString( X + ColumnWidth, Y, *FString::Printf(TEXT("%g"), Dest.name), GetStatsFont(), FLinearColor(1, 1, 1));
ENTRY(LPVIntensity)
ENTRY(LPVVplInjectionBias)
ENTRY(LPVSize)
ENTRY(LPVSecondaryOcclusionIntensity)
ENTRY(LPVSecondaryBounceIntensity)
ENTRY(LPVGeometryVolumeBias)
ENTRY(LPVEmissiveInjectionIntensity)
ENTRY(LPVDirectionalOcclusionIntensity)
ENTRY(LPVDirectionalOcclusionRadius)
ENTRY(LPVDiffuseOcclusionExponent)
ENTRY(LPVSpecularOcclusionExponent)
ENTRY(LPVDiffuseOcclusionIntensity)
ENTRY(LPVSpecularOcclusionIntensity)
#undef ENTRY
Canvas.Flush_RenderThread(Context.RHICmdList);
}
Context.RHICmdList.CopyToResolveTarget(DestRenderTarget.TargetableTexture, DestRenderTarget.ShaderResourceTexture, false, FResolveParams());
// to satify following passws
FRenderingCompositeOutput* Output = GetOutput(ePId_Output0);
Output->PooledRenderTarget = RenderTarget;
}
static void UpdatePlanarReflectionContents_RenderThread(
FRHICommandListImmediate& RHICmdList,
FSceneRenderer* MainSceneRenderer,
FSceneRenderer* SceneRenderer,
const FPlanarReflectionSceneProxy* SceneProxy,
FRenderTarget* RenderTarget,
FTexture* RenderTargetTexture,
const FPlane& MirrorPlane,
const FName OwnerName,
const FResolveParams& ResolveParams,
bool bUseSceneColorTexture)
{
QUICK_SCOPE_CYCLE_COUNTER(STAT_RenderPlanarReflection);
FBox PlanarReflectionBounds = SceneProxy->WorldBounds;
bool bIsInAnyFrustum = false;
for (int32 ViewIndex = 0; ViewIndex < SceneRenderer->Views.Num(); ++ViewIndex)
{
FViewInfo& View = SceneRenderer->Views[ViewIndex];
if (View.ViewFrustum.IntersectBox(PlanarReflectionBounds.GetCenter(), PlanarReflectionBounds.GetExtent()))
{
bIsInAnyFrustum = true;
break;
}
}
if (bIsInAnyFrustum)
{
bool bIsVisibleInAnyView = true;
for (int32 ViewIndex = 0; ViewIndex < SceneRenderer->Views.Num(); ++ViewIndex)
{
FViewInfo& View = SceneRenderer->Views[ViewIndex];
FSceneViewState* ViewState = View.ViewState;
if (ViewState)
{
FIndividualOcclusionHistory& OcclusionHistory = ViewState->PlanarReflectionOcclusionHistories.FindOrAdd(SceneProxy->PlanarReflectionId);
// +1 to buffered frames because the query is submitted late into the main frame, but read at the beginning of a reflection capture frame
const int32 NumBufferedFrames = FOcclusionQueryHelpers::GetNumBufferedFrames() + 1;
// +1 to frame counter because we are operating before the main view's InitViews, which is where OcclusionFrameCounter is incremented
uint32 OcclusionFrameCounter = ViewState->OcclusionFrameCounter + 1;
FRenderQueryRHIRef& PastQuery = OcclusionHistory.GetPastQuery(OcclusionFrameCounter, NumBufferedFrames);
if (IsValidRef(PastQuery))
{
uint64 NumSamples = 0;
QUICK_SCOPE_CYCLE_COUNTER(STAT_PlanarReflectionOcclusionQueryResults);
if (RHIGetRenderQueryResult(PastQuery.GetReference(), NumSamples, true))
{
bIsVisibleInAnyView = NumSamples > 0;
if (bIsVisibleInAnyView)
{
break;
}
}
}
}
}
if (bIsVisibleInAnyView)
{
FMemMark MemStackMark(FMemStack::Get());
// update any resources that needed a deferred update
FDeferredUpdateResource::UpdateResources(RHICmdList);
{
#if WANTS_DRAW_MESH_EVENTS
FString EventName;
OwnerName.ToString(EventName);
SCOPED_DRAW_EVENTF(RHICmdList, SceneCapture, TEXT("PlanarReflection %s"), *EventName);
#else
SCOPED_DRAW_EVENT(RHICmdList, UpdatePlanarReflectionContent_RenderThread);
#endif
const FRenderTarget* Target = SceneRenderer->ViewFamily.RenderTarget;
SetRenderTarget(RHICmdList, Target->GetRenderTargetTexture(), NULL, true);
// Note: relying on GBuffer SceneColor alpha being cleared to 1 in the main scene rendering
check(GetSceneColorClearAlpha() == 1.0f);
RHICmdList.Clear(true, FLinearColor(0, 0, 0, 1), false, (float)ERHIZBuffer::FarPlane, false, 0, FIntRect());
// Reflection view late update
if (SceneRenderer->Views.Num() > 1)
{
const FMirrorMatrix MirrorMatrix(MirrorPlane);
for (int32 ViewIndex = 0; ViewIndex < SceneRenderer->Views.Num(); ++ViewIndex)
{
FViewInfo& ReflectionViewToUpdate = SceneRenderer->Views[ViewIndex];
const FViewInfo& UpdatedParentView = MainSceneRenderer->Views[ViewIndex];
ReflectionViewToUpdate.UpdatePlanarReflectionViewMatrix(UpdatedParentView, MirrorMatrix);
}
}
// Render the scene normally
{
SCOPED_DRAW_EVENT(RHICmdList, RenderScene);
SceneRenderer->Render(RHICmdList);
}
for (int32 ViewIndex = 0; ViewIndex < SceneRenderer->Views.Num(); ++ViewIndex)
{
FViewInfo& View = SceneRenderer->Views[ViewIndex];
if (MainSceneRenderer->Scene->GetShadingPath() == EShadingPath::Deferred)
{
PrefilterPlanarReflection<true>(RHICmdList, View, SceneProxy, Target);
}
else
{
PrefilterPlanarReflection<false>(RHICmdList, View, SceneProxy, Target);
}
}
RHICmdList.CopyToResolveTarget(RenderTarget->GetRenderTargetTexture(), RenderTargetTexture->TextureRHI, false, ResolveParams);
}
FSceneRenderer::WaitForTasksClearSnapshotsAndDeleteSceneRenderer(RHICmdList, SceneRenderer);
}
}
}
void FRCPassPostProcessMaterial::Process(FRenderingCompositePassContext& Context)
{
FMaterialRenderProxy* Proxy = MaterialInterface->GetRenderProxy(false);
check(Proxy);
const FMaterial* Material = Proxy->GetMaterial(Context.View.GetFeatureLevel());
check(Material);
SCOPED_DRAW_EVENTF(Context.RHICmdList, PostProcessMaterial, TEXT("PostProcessMaterial Material=%s"), *Material->GetFriendlyName());
const FPooledRenderTargetDesc* InputDesc = GetInputDesc(ePId_Input0);
if(!InputDesc)
{
// input is not hooked up correctly
return;
}
const FSceneView& View = Context.View;
const FSceneViewFamily& ViewFamily = *(View.Family);
// hacky cast
FRenderingCompositePassContext RenderingCompositePassContext(Context.RHICmdList, (FViewInfo&)View);
RenderingCompositePassContext.Pass = this;
FIntRect SrcRect = View.ViewRect;
FIntRect DestRect = View.ViewRect;
FIntPoint SrcSize = InputDesc->Extent;
const FSceneRenderTargetItem& DestRenderTarget = PassOutputs[0].RequestSurface(Context);
// Set the view family's render target/viewport.
SetRenderTarget(Context.RHICmdList, DestRenderTarget.TargetableTexture, FTextureRHIParamRef());
if( ViewFamily.RenderTarget->GetRenderTargetTexture() != DestRenderTarget.TargetableTexture )
{
Context.RHICmdList.Clear(true, FLinearColor::Black, false, 1.0f, false, 0, View.ViewRect);
}
Context.SetViewportAndCallRHI(View.ViewRect);
// set the state
Context.RHICmdList.SetBlendState(TStaticBlendState<>::GetRHI());
Context.RHICmdList.SetRasterizerState(TStaticRasterizerState<>::GetRHI());
Context.RHICmdList.SetDepthStencilState(TStaticDepthStencilState<false, CF_Always>::GetRHI());
float ScaleX = 1.0f / InputDesc->Extent.X;
float ScaleY = 1.0f / InputDesc->Extent.Y;
const FMaterialShaderMap* MaterialShaderMap = Material->GetRenderingThreadShaderMap();
FPostProcessMaterialPS* PixelShader = MaterialShaderMap->GetShader<FPostProcessMaterialPS>();
FPostProcessMaterialVS* VertexShader = MaterialShaderMap->GetShader<FPostProcessMaterialVS>();
Context.RHICmdList.SetLocalBoundShaderState(Context.RHICmdList.BuildLocalBoundShaderState(GetVertexDeclarationFVector4(), VertexShader->GetVertexShader(), FHullShaderRHIRef(), FDomainShaderRHIRef(), PixelShader->GetPixelShader(), FGeometryShaderRHIRef()));
VertexShader->SetParameters(Context.RHICmdList, Context);
PixelShader->SetParameters(Context.RHICmdList, Context, MaterialInterface->GetRenderProxy(false));
DrawRectangle(
Context.RHICmdList,
0, 0,
DestRect.Width(), DestRect.Height(),
SrcRect.Min.X, SrcRect.Min.Y,
SrcRect.Width(), SrcRect.Height(),
DestRect.Size(),
SrcSize,
VertexShader,
EDRF_UseTriangleOptimization);
Context.RHICmdList.CopyToResolveTarget(DestRenderTarget.TargetableTexture, DestRenderTarget.ShaderResourceTexture, false, FResolveParams());
if(Material->NeedsGBuffer())
{
GSceneRenderTargets.AdjustGBufferRefCount(-1);
}
}
void FRCPassPostProcessUpscale::Process(FRenderingCompositePassContext& Context)
{
SCOPED_DRAW_EVENT(Context.RHICmdList, PostProcessUpscale);
const FPooledRenderTargetDesc* InputDesc = GetInputDesc(ePId_Input0);
if(!InputDesc)
{
// input is not hooked up correctly
return;
}
const FSceneView& View = Context.View;
const FSceneViewFamily& ViewFamily = *(View.Family);
FIntRect SrcRect = View.ViewRect;
FIntRect DestRect = View.UnscaledViewRect;
FIntPoint SrcSize = InputDesc->Extent;
const FSceneRenderTargetItem& DestRenderTarget = PassOutputs[0].RequestSurface(Context);
// Set the view family's render target/viewport.
SetRenderTarget(Context.RHICmdList, DestRenderTarget.TargetableTexture, FTextureRHIRef());
Context.SetViewportAndCallRHI(DestRect);
bool bTessellatedQuad = PaniniConfig.D >= 0.01f;
// with distortion (bTessellatedQuad) we need to clear the background
FIntRect ExcludeRect = bTessellatedQuad ? FIntRect() : DestRect;
Context.RHICmdList.Clear(true, FLinearColor::Black, false, 1.0f, false, 0, ExcludeRect);
// set the state
Context.RHICmdList.SetBlendState(TStaticBlendState<>::GetRHI());
Context.RHICmdList.SetRasterizerState(TStaticRasterizerState<>::GetRHI());
Context.RHICmdList.SetDepthStencilState(TStaticDepthStencilState<false, CF_Always>::GetRHI());
FShader* VertexShader = 0;
if(bTessellatedQuad)
{
switch (UpscaleQuality)
{
case 0:
VertexShader = SetShader<0, 1>(Context, PaniniConfig);
break;
case 1:
VertexShader = SetShader<1, 1>(Context, PaniniConfig);
break;
case 2:
VertexShader = SetShader<2, 1>(Context, PaniniConfig);
break;
case 3:
VertexShader = SetShader<3, 1>(Context, PaniniConfig);
break;
default:
checkNoEntry();
break;
}
}
else
{
switch (UpscaleQuality)
{
case 0:
VertexShader = SetShader<0, 0>(Context, PaniniParams::Default);
break;
case 1:
VertexShader = SetShader<1, 0>(Context, PaniniParams::Default);
break;
case 2:
VertexShader = SetShader<2, 0>(Context, PaniniParams::Default);
break;
case 3:
VertexShader = SetShader<3, 0>(Context, PaniniParams::Default);
break;
default:
checkNoEntry();
break;
}
}
// Draw a quad, a triangle or a tessellated quad
DrawRectangle(
Context.RHICmdList,
0, 0,
DestRect.Width(), DestRect.Height(),
SrcRect.Min.X, SrcRect.Min.Y,
SrcRect.Width(), SrcRect.Height(),
DestRect.Size(),
SrcSize,
VertexShader,
bTessellatedQuad ? EDRF_UseTesselatedIndexBuffer: EDRF_UseTriangleOptimization);
Context.RHICmdList.CopyToResolveTarget(DestRenderTarget.TargetableTexture, DestRenderTarget.ShaderResourceTexture, false, FResolveParams());
}
void FRCPassPostProcessDownsample::Process(FRenderingCompositePassContext& Context)
{
const FPooledRenderTargetDesc* InputDesc = GetInputDesc(ePId_Input0);
if(!InputDesc)
{
// input is not hooked up correctly
return;
}
const FSceneView& View = Context.View;
const FSceneViewFamily& ViewFamily = *(View.Family);
FIntPoint SrcSize = InputDesc->Extent;
FIntPoint DestSize = PassOutputs[0].RenderTargetDesc.Extent;
// e.g. 4 means the input texture is 4x smaller than the buffer size
uint32 ScaleFactor = FMath::DivideAndRoundUp(FSceneRenderTargets::Get(Context.RHICmdList).GetBufferSizeXY().Y, SrcSize.Y);
FIntRect SrcRect = View.ViewRect / ScaleFactor;
FIntRect DestRect = FIntRect::DivideAndRoundUp(SrcRect, 2);
SrcRect = DestRect * 2;
SCOPED_DRAW_EVENTF(Context.RHICmdList, Downsample, TEXT("Downsample %dx%d"), DestRect.Width(), DestRect.Height());
const FSceneRenderTargetItem& DestRenderTarget = PassOutputs[0].RequestSurface(Context);
// check if we have to clear the whole surface.
// Otherwise perform the clear when the dest rectangle has been computed.
auto FeatureLevel = Context.View.GetFeatureLevel();
if (FeatureLevel == ERHIFeatureLevel::ES2 || FeatureLevel == ERHIFeatureLevel::ES3_1)
{
// Set the view family's render target/viewport.
SetRenderTarget(Context.RHICmdList, DestRenderTarget.TargetableTexture, FTextureRHIRef(), ESimpleRenderTargetMode::EClearColorAndDepth);
Context.SetViewportAndCallRHI(0, 0, 0.0f, DestSize.X, DestSize.Y, 1.0f );
}
else
{
// Set the view family's render target/viewport.
SetRenderTarget(Context.RHICmdList, DestRenderTarget.TargetableTexture, FTextureRHIRef(), ESimpleRenderTargetMode::EExistingColorAndDepth);
Context.SetViewportAndCallRHI(0, 0, 0.0f, DestSize.X, DestSize.Y, 1.0f );
DrawClearQuad(Context.RHICmdList, Context.GetFeatureLevel(), true, FLinearColor(0, 0, 0, 0), false, 1.0f, false, 0, DestSize, DestRect);
}
// set the state
Context.RHICmdList.SetBlendState(TStaticBlendState<>::GetRHI());
Context.RHICmdList.SetRasterizerState(TStaticRasterizerState<>::GetRHI());
Context.RHICmdList.SetDepthStencilState(TStaticDepthStencilState<false, CF_Always>::GetRHI());
// InflateSize increases the size of the source/dest rectangle to compensate for bilinear reads and UIBlur pass requirements.
int32 InflateSize;
// if second input is hooked up
if (IsDepthInputAvailable())
{
// also put depth in alpha
InflateSize = 2;
SetShader<2>(Context, InputDesc);
}
else
{
if (Quality == 0)
{
SetShader<0>(Context, InputDesc);
InflateSize = 1;
}
else
{
SetShader<1>(Context, InputDesc);
InflateSize = 2;
}
}
TShaderMapRef<FPostProcessDownsampleVS> VertexShader(Context.GetShaderMap());
DrawPostProcessPass(
Context.RHICmdList,
DestRect.Min.X, DestRect.Min.Y,
DestRect.Width(), DestRect.Height(),
SrcRect.Min.X, SrcRect.Min.Y,
SrcRect.Width(), SrcRect.Height(),
DestSize,
SrcSize,
*VertexShader,
View.StereoPass,
false, // This pass is input for passes that can't use the hmd mask, so we need to disable it to ensure valid input data
EDRF_UseTriangleOptimization);
Context.RHICmdList.CopyToResolveTarget(DestRenderTarget.TargetableTexture, DestRenderTarget.ShaderResourceTexture, false, FResolveParams());
}
// TODO: REMOVE if no longer needed:
void FSceneRenderer::GammaCorrectToViewportRenderTarget(FRHICommandList& RHICmdList, const FViewInfo* View, float OverrideGamma)
{
// Set the view family's render target/viewport.
SetRenderTarget(RHICmdList, ViewFamily.RenderTarget->GetRenderTargetTexture(), FTextureRHIRef());
// Deferred the clear until here so the garbage left in the non rendered regions by the post process effects do not show up
if( ViewFamily.bDeferClear )
{
RHICmdList.Clear(true, FLinearColor::Black, false, 0.0f, false, 0, FIntRect());
ViewFamily.bDeferClear = false;
}
SCOPED_DRAW_EVENT(RHICmdList, GammaCorrection);
// turn off culling and blending
RHICmdList.SetRasterizerState(TStaticRasterizerState<FM_Solid, CM_None>::GetRHI());
RHICmdList.SetBlendState(TStaticBlendState<>::GetRHI());
// turn off depth reads/writes
RHICmdList.SetDepthStencilState(TStaticDepthStencilState<false, CF_Always>::GetRHI());
TShaderMapRef<FGammaCorrectionVS> VertexShader(View->ShaderMap);
TShaderMapRef<FGammaCorrectionPS> PixelShader(View->ShaderMap);
static FGlobalBoundShaderState PostProcessBoundShaderState;
SetGlobalBoundShaderState(RHICmdList, View->GetFeatureLevel(), PostProcessBoundShaderState, GFilterVertexDeclaration.VertexDeclarationRHI, *VertexShader, *PixelShader);
float InvDisplayGamma = 1.0f / ViewFamily.RenderTarget->GetDisplayGamma();
if (OverrideGamma != 0)
{
InvDisplayGamma = 1 / OverrideGamma;
}
const FPixelShaderRHIParamRef ShaderRHI = PixelShader->GetPixelShader();
SetShaderValue(
RHICmdList,
ShaderRHI,
PixelShader->InverseGamma,
InvDisplayGamma
);
SetShaderValue(RHICmdList, ShaderRHI,PixelShader->ColorScale,View->ColorScale);
SetShaderValue(RHICmdList, ShaderRHI,PixelShader->OverlayColor,View->OverlayColor);
const FTextureRHIRef DesiredSceneColorTexture = GSceneRenderTargets.GetSceneColorTexture();
SetTextureParameter(
RHICmdList,
ShaderRHI,
PixelShader->SceneTexture,
PixelShader->SceneTextureSampler,
TStaticSamplerState<SF_Bilinear>::GetRHI(),
DesiredSceneColorTexture
);
// Draw a quad mapping scene color to the view's render target
DrawRectangle(
RHICmdList,
View->UnscaledViewRect.Min.X,View->UnscaledViewRect.Min.Y,
View->UnscaledViewRect.Width(),View->UnscaledViewRect.Height(),
View->ViewRect.Min.X,View->ViewRect.Min.Y,
View->ViewRect.Width(),View->ViewRect.Height(),
ViewFamily.RenderTarget->GetSizeXY(),
GSceneRenderTargets.GetBufferSizeXY(),
*VertexShader,
EDRF_UseTriangleOptimization);
}
void FRCPassPostProcessNoiseBlur::Process(FRenderingCompositePassContext& Context)
{
SCOPED_DRAW_EVENT(Context.RHICmdList, NoiseBlur);
const FPooledRenderTargetDesc* InputDesc = GetInputDesc(ePId_Input0);
if(!InputDesc)
{
// input is not hooked up correctly
return;
}
const FSceneView& View = Context.View;
const FSceneViewFamily& ViewFamily = *(View.Family);
FIntPoint SrcSize = InputDesc->Extent;
FIntPoint DestSize = PassOutputs[0].RenderTargetDesc.Extent;
// e.g. 4 means the input texture is 4x smaller than the buffer size
uint32 ScaleFactor = FSceneRenderTargets::Get(Context.RHICmdList).GetBufferSizeXY().X / SrcSize.X;
FIntRect SrcRect = View.ViewRect / ScaleFactor;
FIntRect DestRect = SrcRect;
const FSceneRenderTargetItem& DestRenderTarget = PassOutputs[0].RequestSurface(Context);
// Set the view family's render target/viewport.
SetRenderTarget(Context.RHICmdList, DestRenderTarget.TargetableTexture, FTextureRHIRef());
// is optimized away if possible (RT size=view size, )
Context.RHICmdList.Clear(true, FLinearColor(0, 0, 0, 0), false, 1.0f, false, 0, DestRect);
Context.SetViewportAndCallRHI(0, 0, 0.0f, DestSize.X, DestSize.Y, 1.0f );
// set the state
Context.RHICmdList.SetBlendState(TStaticBlendState<>::GetRHI());
Context.RHICmdList.SetRasterizerState(TStaticRasterizerState<>::GetRHI());
Context.RHICmdList.SetDepthStencilState(TStaticDepthStencilState<false, CF_Always>::GetRHI());
if(Quality == 0)
{
SetNoiseBlurShader<0>(Context, Radius);
}
else if(Quality == 1)
{
SetNoiseBlurShader<1>(Context, Radius);
}
else
{
SetNoiseBlurShader<2>(Context, Radius);
}
// Draw a quad mapping scene color to the view's render target
TShaderMapRef<FPostProcessVS> VertexShader(Context.GetShaderMap());
DrawRectangle(
Context.RHICmdList,
DestRect.Min.X, DestRect.Min.Y,
DestRect.Width(), DestRect.Height(),
SrcRect.Min.X, SrcRect.Min.Y,
SrcRect.Width(), SrcRect.Height(),
DestSize,
SrcSize,
*VertexShader,
EDRF_UseTriangleOptimization);
Context.RHICmdList.CopyToResolveTarget(DestRenderTarget.TargetableTexture, DestRenderTarget.ShaderResourceTexture, false, FResolveParams());
}
void FRCPassPostProcessLensBlur::Process(FRenderingCompositePassContext& Context)
{
SCOPED_DRAW_EVENT(Context.RHICmdList, PassPostProcessLensBlur, DEC_SCENE_ITEMS);
const FPooledRenderTargetDesc* InputDesc = GetInputDesc(ePId_Input0);
if(!InputDesc)
{
// input is not hooked up correctly
return;
}
const FSceneView& View = Context.View;
FIntPoint TexSize = InputDesc->Extent;
// usually 1, 2, 4 or 8
uint32 ScaleToFullRes = GSceneRenderTargets.GetBufferSizeXY().X / TexSize.X;
FIntRect ViewRect = FIntRect::DivideAndRoundUp(View.ViewRect, ScaleToFullRes);
FIntPoint ViewSize = ViewRect.Size();
const FSceneRenderTargetItem& DestRenderTarget = PassOutputs[0].RequestSurface(Context);
// Set the view family's render target/viewport.
SetRenderTarget(Context.RHICmdList, DestRenderTarget.TargetableTexture, FTextureRHIRef());
Context.RHICmdList.Clear(true, FLinearColor::Black, false, 1.0f, false, 0, FIntRect());
Context.SetViewportAndCallRHI(ViewRect);
// set the state (additive blending)
Context.RHICmdList.SetBlendState(TStaticBlendState<CW_RGB, BO_Add, BF_One, BF_One>::GetRHI());
Context.RHICmdList.SetRasterizerState(TStaticRasterizerState<>::GetRHI());
Context.RHICmdList.SetDepthStencilState(TStaticDepthStencilState<false, CF_Always>::GetRHI());
TShaderMapRef<FPostProcessLensBlurVS> VertexShader(Context.GetShaderMap());
TShaderMapRef<FPostProcessLensBlurPS> PixelShader(Context.GetShaderMap());
static FGlobalBoundShaderState BoundShaderState;
SetGlobalBoundShaderState(Context.RHICmdList, Context.GetFeatureLevel(), BoundShaderState, GFilterVertexDeclaration.VertexDeclarationRHI, *VertexShader, *PixelShader);
uint32 TileSize = 1;
FIntPoint TileCount = ViewSize / TileSize;
float PixelKernelSize = PercentKernelSize / 100.0f * ViewSize.X;
VertexShader->SetParameters(Context, TileCount, TileSize, PixelKernelSize, Threshold);
PixelShader->SetParameters(Context, PixelKernelSize);
Context.RHICmdList.SetStreamSource(0, NULL, 0, 0);
// needs to be the same on shader side (faster on NVIDIA and AMD)
int32 QuadsPerInstance = 4;
Context.RHICmdList.DrawPrimitive(PT_TriangleList, 0, 2, FMath::DivideAndRoundUp(TileCount.X * TileCount.Y, QuadsPerInstance));
Context.RHICmdList.CopyToResolveTarget(DestRenderTarget.TargetableTexture, DestRenderTarget.ShaderResourceTexture, false, FResolveParams());
}
void RendererGPUBenchmark(FRHICommandListImmediate& RHICmdList, FSynthBenchmarkResults& InOut, const FSceneView& View, float WorkScale, bool bDebugOut)
{
check(IsInRenderingThread());
// two RT to ping pong so we force the GPU to flush it's pipeline
TRefCountPtr<IPooledRenderTarget> RTItems[3];
{
FPooledRenderTargetDesc Desc(FPooledRenderTargetDesc::Create2DDesc(FIntPoint(GBenchmarkResolution, GBenchmarkResolution), PF_B8G8R8A8, FClearValueBinding::None, TexCreate_None, TexCreate_RenderTargetable | TexCreate_ShaderResource, false));
GRenderTargetPool.FindFreeElement(Desc, RTItems[0], TEXT("Benchmark0"));
GRenderTargetPool.FindFreeElement(Desc, RTItems[1], TEXT("Benchmark1"));
Desc.Extent = FIntPoint(1, 1);
Desc.Flags = TexCreate_CPUReadback; // needs TexCreate_ResolveTargetable?
Desc.TargetableFlags = TexCreate_None;
GRenderTargetPool.FindFreeElement(Desc, RTItems[2], TEXT("BenchmarkReadback"));
}
// set the state
RHICmdList.SetBlendState(TStaticBlendState<>::GetRHI());
RHICmdList.SetRasterizerState(TStaticRasterizerState<>::GetRHI());
RHICmdList.SetDepthStencilState(TStaticDepthStencilState<false,CF_Always>::GetRHI());
{
// larger number means more accuracy but slower, some slower GPUs might timeout with a number to large
const uint32 IterationCount = 70;
const uint32 MethodCount = ARRAY_COUNT(InOut.GPUStats);
// 0 / 1
uint32 DestRTIndex = 0;
const uint32 TimerSampleCount = IterationCount * MethodCount + 1;
static FRenderQueryRHIRef TimerQueries[TimerSampleCount];
static float LocalWorkScale[IterationCount];
for(uint32 i = 0; i < TimerSampleCount; ++i)
{
TimerQueries[i] = GTimerQueryPool.AllocateQuery();
}
const bool bSupportsTimerQueries = (TimerQueries[0] != NULL);
if(!bSupportsTimerQueries)
{
UE_LOG(LogSynthBenchmark, Warning, TEXT("GPU driver does not support timer queries."));
// Temporary workaround for GL_TIMESTAMP being unavailable and GL_TIME_ELAPSED workaround breaking drivers
#if PLATFORM_MAC
GLint RendererID = 0;
float PerfScale = 1.0f;
[[NSOpenGLContext currentContext] getValues:&RendererID forParameter:NSOpenGLCPCurrentRendererID];
{
switch((RendererID & kCGLRendererIDMatchingMask))
{
case kCGLRendererATIRadeonX4000ID: // AMD 7xx0 & Dx00 series - should be pretty beefy
PerfScale = 1.2f;
break;
case kCGLRendererATIRadeonX3000ID: // AMD 5xx0, 6xx0 series - mostly OK
case kCGLRendererGeForceID: // Nvidia 6x0 & 7x0 series - mostly OK
PerfScale = 2.0f;
break;
case kCGLRendererIntelHD5000ID: // Intel HD 5000, Iris, Iris Pro - not dreadful
PerfScale = 4.2f;
break;
case kCGLRendererIntelHD4000ID: // Intel HD 4000 - quite slow
PerfScale = 7.5f;
break;
case kCGLRendererATIRadeonX2000ID: // ATi 4xx0, 3xx0, 2xx0 - almost all very slow and drivers are now very buggy
case kCGLRendererGeForce8xxxID: // Nvidia 3x0, 2x0, 1x0, 9xx0, 8xx0 - almost all very slow
case kCGLRendererIntelHDID: // Intel HD 3000 - very, very slow and very buggy driver
default:
PerfScale = 10.0f;
break;
}
}
InOut.GPUStats[0] = FSynthBenchmarkStat(TEXT("ALUHeavyNoise"), 1.0f / 4.601f, TEXT("s/GigaPix"));
InOut.GPUStats[1] = FSynthBenchmarkStat(TEXT("TexHeavy"), 1.0f / 7.447f, TEXT("s/GigaPix"));
InOut.GPUStats[2] = FSynthBenchmarkStat(TEXT("DepTexHeavy"), 1.0f / 3.847f, TEXT("s/GigaPix"));
InOut.GPUStats[3] = FSynthBenchmarkStat(TEXT("FillOnly"), 1.0f / 25.463f, TEXT("s/GigaPix"));
InOut.GPUStats[4] = FSynthBenchmarkStat(TEXT("Bandwidth"), 1.0f / 1.072f, TEXT("s/GigaPix"));
InOut.GPUStats[0].SetMeasuredTime( FTimeSample(PerfScale, PerfScale * (1.0f / 4.601f)) );
InOut.GPUStats[1].SetMeasuredTime( FTimeSample(PerfScale, PerfScale * (1.0f / 7.447f)) );
InOut.GPUStats[2].SetMeasuredTime( FTimeSample(PerfScale, PerfScale * (1.0f / 3.847f)) );
InOut.GPUStats[3].SetMeasuredTime( FTimeSample(PerfScale, PerfScale * (1.0f / 25.463f)) );
InOut.GPUStats[4].SetMeasuredTime( FTimeSample(PerfScale, PerfScale * (1.0f / 1.072f)) );
#endif
return;
}
// TimingValues are in Seconds
FTimingSeries TimingSeries[MethodCount];
// in 1/1000000 Seconds
uint64 TotalTimes[MethodCount];
for(uint32 MethodIterator = 0; MethodIterator < MethodCount; ++MethodIterator)
{
TotalTimes[MethodIterator] = 0;
TimingSeries[MethodIterator].Init(IterationCount);
}
check(MethodCount == 5);
InOut.GPUStats[0] = FSynthBenchmarkStat(TEXT("ALUHeavyNoise"), 1.0f / 4.601f, TEXT("s/GigaPix"));
InOut.GPUStats[1] = FSynthBenchmarkStat(TEXT("TexHeavy"), 1.0f / 7.447f, TEXT("s/GigaPix"));
InOut.GPUStats[2] = FSynthBenchmarkStat(TEXT("DepTexHeavy"), 1.0f / 3.847f, TEXT("s/GigaPix"));
InOut.GPUStats[3] = FSynthBenchmarkStat(TEXT("FillOnly"), 1.0f / 25.463f, TEXT("s/GigaPix"));
InOut.GPUStats[4] = FSynthBenchmarkStat(TEXT("Bandwidth"), 1.0f / 1.072f, TEXT("s/GigaPix"));
// e.g. on NV670: Method3 (mostly fill rate )-> 26GP/s (seems realistic)
// reference: http://en.wikipedia.org/wiki/Comparison_of_Nvidia_graphics_processing_units theoretical: 29.3G/s
RHICmdList.EndRenderQuery(TimerQueries[0]);
// multiple iterations to see how trust able the values are
for(uint32 Iteration = 0; Iteration < IterationCount; ++Iteration)
{
for(uint32 MethodIterator = 0; MethodIterator < MethodCount; ++MethodIterator)
{
// alternate between forward and backward (should give the same number)
// uint32 MethodId = (Iteration % 2) ? MethodIterator : (MethodCount - 1 - MethodIterator);
uint32 MethodId = MethodIterator;
uint32 QueryIndex = 1 + Iteration * MethodCount + MethodId;
// 0 / 1
const uint32 SrcRTIndex = 1 - DestRTIndex;
GRenderTargetPool.VisualizeTexture.SetCheckPoint(RHICmdList, RTItems[DestRTIndex]);
SetRenderTarget(RHICmdList, RTItems[DestRTIndex]->GetRenderTargetItem().TargetableTexture, FTextureRHIRef());
// decide how much work we do in this pass
LocalWorkScale[Iteration] = (Iteration / 10.f + 1.f) * WorkScale;
RunBenchmarkShader(RHICmdList, View, MethodId, RTItems[SrcRTIndex], LocalWorkScale[Iteration]);
RHICmdList.CopyToResolveTarget(RTItems[DestRTIndex]->GetRenderTargetItem().TargetableTexture, RTItems[DestRTIndex]->GetRenderTargetItem().ShaderResourceTexture, false, FResolveParams());
/*if(bGPUCPUSync)
{
// more consistent timing but strangely much faster to the level that is unrealistic
FResolveParams Param;
Param.Rect = FResolveRect(0, 0, 1, 1);
RHICmdList.CopyToResolveTarget(
RTItems[DestRTIndex]->GetRenderTargetItem().TargetableTexture,
RTItems[2]->GetRenderTargetItem().ShaderResourceTexture,
false,
Param);
void* Data = 0;
int Width = 0;
int Height = 0;
RHIMapStagingSurface(RTItems[2]->GetRenderTargetItem().ShaderResourceTexture, Data, Width, Height);
RHIUnmapStagingSurface(RTItems[2]->GetRenderTargetItem().ShaderResourceTexture);
}*/
RHICmdList.EndRenderQuery(TimerQueries[QueryIndex]);
// ping pong
DestRTIndex = 1 - DestRTIndex;
}
}
{
uint64 OldAbsTime = 0;
// flushes the RHI thread to make sure all RHICmdList.EndRenderQuery() commands got executed.
RHICmdList.ImmediateFlush(EImmediateFlushType::FlushRHIThread);
RHICmdList.GetRenderQueryResult(TimerQueries[0], OldAbsTime, true);
GTimerQueryPool.ReleaseQuery(RHICmdList, TimerQueries[0]);
for(uint32 Iteration = 0; Iteration < IterationCount; ++Iteration)
{
uint32 Results[MethodCount];
for(uint32 MethodId = 0; MethodId < MethodCount; ++MethodId)
{
uint32 QueryIndex = 1 + Iteration * MethodCount + MethodId;
uint64 AbsTime;
RHICmdList.GetRenderQueryResult(TimerQueries[QueryIndex], AbsTime, true);
GTimerQueryPool.ReleaseQuery(RHICmdList, TimerQueries[QueryIndex]);
uint64 RelTime = AbsTime - OldAbsTime;
TotalTimes[MethodId] += RelTime;
Results[MethodId] = RelTime;
OldAbsTime = AbsTime;
}
for(uint32 MethodId = 0; MethodId < MethodCount; ++MethodId)
{
float TimeInSec = Results[MethodId] / 1000000.0f;
// to normalize from seconds to seconds per GPixel
float SamplesInGPix = LocalWorkScale[Iteration] * GBenchmarkResolution * GBenchmarkResolution / 1000000000.0f;
// TimingValue in Seconds per GPixel
TimingSeries[MethodId].SetEntry(Iteration, TimeInSec / SamplesInGPix);
}
}
if(bSupportsTimerQueries)
{
for(uint32 MethodId = 0; MethodId < MethodCount; ++MethodId)
{
float Confidence = 0.0f;
// in seconds per GPixel
float NormalizedTime = TimingSeries[MethodId].ComputeValue(Confidence);
if(Confidence > 0)
{
FTimeSample TimeSample(TotalTimes[MethodId] / 1000000.0f, NormalizedTime);
InOut.GPUStats[MethodId].SetMeasuredTime(TimeSample, Confidence);
}
}
}
}
}
void FRCPassPostProcessMotionBlur::Process(FRenderingCompositePassContext& Context)
{
SCOPED_DRAW_EVENT(Context.RHICmdList, MotionBlur);
const FPooledRenderTargetDesc* InputDesc = GetInputDesc(ePId_Input0);
if(!InputDesc)
{
// input is not hooked up correctly
return;
}
const FSceneView& View = Context.View;
FIntPoint TexSize = InputDesc->Extent;
// we assume the input and output is full resolution
FIntPoint SrcSize = InputDesc->Extent;
FIntPoint DestSize = PassOutputs[0].RenderTargetDesc.Extent;
// e.g. 4 means the input texture is 4x smaller than the buffer size
uint32 ScaleFactor = GSceneRenderTargets.GetBufferSizeXY().X / SrcSize.X;
FIntRect SrcRect = FIntRect::DivideAndRoundUp(View.ViewRect, ScaleFactor);
FIntRect DestRect = SrcRect;
const FSceneRenderTargetItem& DestRenderTarget = PassOutputs[0].RequestSurface(Context);
// Set the view family's render target/viewport.
SetRenderTarget(Context.RHICmdList, DestRenderTarget.TargetableTexture, FTextureRHIRef());
// is optimized away if possible (RT size=view size, )
Context.RHICmdList.Clear(true, FLinearColor::Black, false, 1.0f, false, 0, SrcRect);
Context.SetViewportAndCallRHI(SrcRect);
// set the state
Context.RHICmdList.SetBlendState(TStaticBlendState<>::GetRHI());
Context.RHICmdList.SetRasterizerState(TStaticRasterizerState<>::GetRHI());
Context.RHICmdList.SetDepthStencilState(TStaticDepthStencilState<false, CF_Always>::GetRHI());
if(Quality == 1)
{
SetMotionBlurShaderTempl<1>(Context);
}
else if(Quality == 2)
{
SetMotionBlurShaderTempl<2>(Context);
}
else if(Quality == 3)
{
SetMotionBlurShaderTempl<3>(Context);
}
else
{
check(Quality == 4);
SetMotionBlurShaderTempl<4>(Context);
}
TShaderMapRef<FPostProcessVS> VertexShader(Context.GetShaderMap());
// Draw a quad mapping scene color to the view's render target
DrawRectangle(
Context.RHICmdList,
0, 0,
SrcRect.Width(), SrcRect.Height(),
SrcRect.Min.X, SrcRect.Min.Y,
SrcRect.Width(), SrcRect.Height(),
SrcRect.Size(),
SrcSize,
*VertexShader,
EDRF_UseTriangleOptimization);
Context.RHICmdList.CopyToResolveTarget(DestRenderTarget.TargetableTexture, DestRenderTarget.ShaderResourceTexture, false, FResolveParams());
}
//todo: this function is an abomination, this is just disgusting. fix it.
//...seriously, this is really, really horrible. I mean this is amazingly bad.
void OBS::MainCaptureLoop()
{
int curRenderTarget = 0, curYUVTexture = 0, curCopyTexture = 0;
int copyWait = NUM_RENDER_BUFFERS-1;
bSentHeaders = false;
bFirstAudioPacket = true;
bool bLogLongFramesProfile = GlobalConfig->GetInt(TEXT("General"), TEXT("LogLongFramesProfile"), LOGLONGFRAMESDEFAULT) != 0;
float logLongFramesProfilePercentage = GlobalConfig->GetFloat(TEXT("General"), TEXT("LogLongFramesProfilePercentage"), 10.f);
Vect2 baseSize = Vect2(float(baseCX), float(baseCY));
Vect2 outputSize = Vect2(float(outputCX), float(outputCY));
Vect2 scaleSize = Vect2(float(scaleCX), float(scaleCY));
HANDLE hMatrix = yuvScalePixelShader->GetParameterByName(TEXT("yuvMat"));
HANDLE hScaleVal = yuvScalePixelShader->GetParameterByName(TEXT("baseDimensionI"));
//----------------------------------------
// x264 input buffers
int curOutBuffer = 0;
bool bUsingQSV = videoEncoder->isQSV();//GlobalConfig->GetInt(TEXT("Video Encoding"), TEXT("UseQSV")) != 0;
bUsing444 = false;
EncoderPicture lastPic;
EncoderPicture outPics[NUM_OUT_BUFFERS];
for(int i=0; i<NUM_OUT_BUFFERS; i++)
{
if(bUsingQSV)
{
outPics[i].mfxOut = new mfxFrameSurface1;
memset(outPics[i].mfxOut, 0, sizeof(mfxFrameSurface1));
mfxFrameData& data = outPics[i].mfxOut->Data;
videoEncoder->RequestBuffers(&data);
}
else
{
outPics[i].picOut = new x264_picture_t;
x264_picture_init(outPics[i].picOut);
}
}
if(bUsing444)
{
for(int i=0; i<NUM_OUT_BUFFERS; i++)
{
outPics[i].picOut->img.i_csp = X264_CSP_BGRA; //although the x264 input says BGR, x264 actually will expect packed UYV
outPics[i].picOut->img.i_plane = 1;
}
}
else
{
if(!bUsingQSV)
for(int i=0; i<NUM_OUT_BUFFERS; i++)
x264_picture_alloc(outPics[i].picOut, X264_CSP_NV12, outputCX, outputCY);
}
int bCongestionControl = AppConfig->GetInt (TEXT("Video Encoding"), TEXT("CongestionControl"), 0);
bool bDynamicBitrateSupported = App->GetVideoEncoder()->DynamicBitrateSupported();
int defaultBitRate = AppConfig->GetInt(TEXT("Video Encoding"), TEXT("MaxBitrate"), 1000);
int currentBitRate = defaultBitRate;
QWORD lastAdjustmentTime = 0;
UINT adjustmentStreamId = 0;
//std::unique_ptr<ProfilerNode> encodeThreadProfiler;
//----------------------------------------
// time/timestamp stuff
bool bWasLaggedFrame = false;
totalStreamTime = 0;
lastAudioTimestamp = 0;
//----------------------------------------
// start audio capture streams
desktopAudio->StartCapture();
if(micAudio) micAudio->StartCapture();
//----------------------------------------
// status bar/statistics stuff
DWORD fpsCounter = 0;
int numLongFrames = 0;
int numTotalFrames = 0;
bytesPerSec = 0;
captureFPS = 0;
curFramesDropped = 0;
curStrain = 0.0;
PostMessage(hwndMain, OBS_UPDATESTATUSBAR, 0, 0);
QWORD lastBytesSent[3] = {0, 0, 0};
DWORD lastFramesDropped = 0;
double bpsTime = 0.0;
double lastStrain = 0.0f;
DWORD numSecondsWaited = 0;
//----------------------------------------
// 444->420 thread data
int numThreads = MAX(OSGetTotalCores()-2, 1);
HANDLE *h420Threads = (HANDLE*)Allocate(sizeof(HANDLE)*numThreads);
Convert444Data *convertInfo = (Convert444Data*)Allocate(sizeof(Convert444Data)*numThreads);
zero(h420Threads, sizeof(HANDLE)*numThreads);
zero(convertInfo, sizeof(Convert444Data)*numThreads);
for(int i=0; i<numThreads; i++)
{
convertInfo[i].width = outputCX;
convertInfo[i].height = outputCY;
convertInfo[i].hSignalConvert = CreateEvent(NULL, FALSE, FALSE, NULL);
convertInfo[i].hSignalComplete = CreateEvent(NULL, FALSE, FALSE, NULL);
convertInfo[i].bNV12 = bUsingQSV;
convertInfo[i].numThreads = numThreads;
if(i == 0)
convertInfo[i].startY = 0;
else
convertInfo[i].startY = convertInfo[i-1].endY;
if(i == (numThreads-1))
convertInfo[i].endY = outputCY;
else
convertInfo[i].endY = ((outputCY/numThreads)*(i+1)) & 0xFFFFFFFE;
}
bool bEncode;
bool bFirstFrame = true;
bool bFirstImage = true;
bool bFirstEncode = true;
bool bUseThreaded420 = bUseMultithreadedOptimizations && (OSGetTotalCores() > 1) && !bUsing444;
List<HANDLE> completeEvents;
if(bUseThreaded420)
{
for(int i=0; i<numThreads; i++)
{
h420Threads[i] = OSCreateThread((XTHREAD)Convert444Thread, convertInfo+i);
completeEvents << convertInfo[i].hSignalComplete;
}
}
//----------------------------------------
QWORD streamTimeStart = GetQPCTimeNS();
QWORD lastStreamTime = 0;
QWORD firstFrameTimeMS = streamTimeStart/1000000;
QWORD frameLengthNS = 1000000000/fps;
while(WaitForSingleObject(hVideoEvent, INFINITE) == WAIT_OBJECT_0)
{
if (bShutdownVideoThread)
break;
QWORD renderStartTime = GetQPCTimeNS();
totalStreamTime = DWORD((renderStartTime-streamTimeStart)/1000000);
bool bRenderView = !IsIconic(hwndMain) && bRenderViewEnabled;
QWORD renderStartTimeMS = renderStartTime/1000000;
QWORD curStreamTime = latestVideoTimeNS;
if (!lastStreamTime)
lastStreamTime = curStreamTime-frameLengthNS;
QWORD frameDelta = curStreamTime-lastStreamTime;
//if (!lastStreamTime)
// lastStreamTime = renderStartTime-frameLengthNS;
//QWORD frameDelta = renderStartTime-lastStreamTime;
double fSeconds = double(frameDelta)*0.000000001;
//lastStreamTime = renderStartTime;
bool bUpdateBPS = false;
profileIn("video thread frame");
//Log(TEXT("Stream Time: %llu"), curStreamTime);
//Log(TEXT("frameDelta: %lf"), fSeconds);
//------------------------------------
if(bRequestKeyframe && keyframeWait > 0)
{
keyframeWait -= int(frameDelta);
if(keyframeWait <= 0)
{
GetVideoEncoder()->RequestKeyframe();
bRequestKeyframe = false;
}
}
if(!bPushToTalkDown && pushToTalkTimeLeft > 0)
{
pushToTalkTimeLeft -= int(frameDelta);
OSDebugOut(TEXT("time left: %d\r\n"), pushToTalkTimeLeft);
if(pushToTalkTimeLeft <= 0)
{
pushToTalkTimeLeft = 0;
bPushToTalkOn = false;
}
}
//------------------------------------
OSEnterMutex(hSceneMutex);
if (bPleaseEnableProjector)
ActuallyEnableProjector();
else if(bPleaseDisableProjector)
DisableProjector();
if(bResizeRenderView)
{
GS->ResizeView();
bResizeRenderView = false;
}
//------------------------------------
if(scene)
{
profileIn("scene->Preprocess");
scene->Preprocess();
for(UINT i=0; i<globalSources.Num(); i++)
globalSources[i].source->Preprocess();
profileOut;
scene->Tick(float(fSeconds));
for(UINT i=0; i<globalSources.Num(); i++)
globalSources[i].source->Tick(float(fSeconds));
}
//------------------------------------
QWORD curBytesSent = 0;
if(network) {
curBytesSent = network->GetCurrentSentBytes();
curFramesDropped = network->NumDroppedFrames();
}
bpsTime += fSeconds;
if(bpsTime > 1.0f)
{
if(numSecondsWaited < 3)
++numSecondsWaited;
//bytesPerSec = DWORD(curBytesSent - lastBytesSent);
bytesPerSec = DWORD(curBytesSent - lastBytesSent[0]) / numSecondsWaited;
if(bpsTime > 2.0)
bpsTime = 0.0f;
else
bpsTime -= 1.0;
if(numSecondsWaited == 3)
{
lastBytesSent[0] = lastBytesSent[1];
lastBytesSent[1] = lastBytesSent[2];
lastBytesSent[2] = curBytesSent;
}
else
lastBytesSent[numSecondsWaited] = curBytesSent;
captureFPS = fpsCounter;
fpsCounter = 0;
bUpdateBPS = true;
}
fpsCounter++;
if(network) curStrain = network->GetPacketStrain();
EnableBlending(TRUE);
BlendFunction(GS_BLEND_SRCALPHA, GS_BLEND_INVSRCALPHA);
//------------------------------------
// render the mini render texture
LoadVertexShader(mainVertexShader);
LoadPixelShader(mainPixelShader);
SetRenderTarget(mainRenderTextures[curRenderTarget]);
Ortho(0.0f, baseSize.x, baseSize.y, 0.0f, -100.0f, 100.0f);
SetViewport(0, 0, baseSize.x, baseSize.y);
if(scene)
scene->Render();
//------------------------------------
if(bTransitioning)
{
if(!transitionTexture)
{
transitionTexture = CreateTexture(baseCX, baseCY, GS_BGRA, NULL, FALSE, TRUE);
if(transitionTexture)
{
D3D10Texture *d3dTransitionTex = static_cast<D3D10Texture*>(transitionTexture);
D3D10Texture *d3dSceneTex = static_cast<D3D10Texture*>(mainRenderTextures[lastRenderTarget]);
GetD3D()->CopyResource(d3dTransitionTex->texture, d3dSceneTex->texture);
}
else
bTransitioning = false;
}
else if(transitionAlpha >= 1.0f)
{
delete transitionTexture;
transitionTexture = NULL;
bTransitioning = false;
}
}
if(bTransitioning)
{
EnableBlending(TRUE);
transitionAlpha += float(fSeconds)*5.0f;
if(transitionAlpha > 1.0f)
transitionAlpha = 1.0f;
}
else
EnableBlending(FALSE);
//------------------------------------
// render the mini view thingy
if (bProjector) {
SetRenderTarget(projectorTexture);
Vect2 renderFrameSize, renderFrameOffset;
Vect2 projectorSize = Vect2(float(projectorWidth), float(projectorHeight));
float projectorAspect = (projectorSize.x / projectorSize.y);
float baseAspect = (baseSize.x / baseSize.y);
if (projectorAspect < baseAspect) {
float fProjectorWidth = float(projectorWidth);
renderFrameSize = Vect2(fProjectorWidth, fProjectorWidth / baseAspect);
renderFrameOffset = Vect2(0.0f, (projectorSize.y-renderFrameSize.y) * 0.5f);
} else {
float fProjectorHeight = float(projectorHeight);
renderFrameSize = Vect2(fProjectorHeight * baseAspect, fProjectorHeight);
renderFrameOffset = Vect2((projectorSize.x-renderFrameSize.x) * 0.5f, 0.0f);
}
DrawPreview(renderFrameSize, renderFrameOffset, projectorSize, curRenderTarget, Preview_Projector);
SetRenderTarget(NULL);
}
if(bRenderView)
{
// Cache
const Vect2 renderFrameSize = GetRenderFrameSize();
const Vect2 renderFrameOffset = GetRenderFrameOffset();
const Vect2 renderFrameCtrlSize = GetRenderFrameControlSize();
SetRenderTarget(NULL);
DrawPreview(renderFrameSize, renderFrameOffset, renderFrameCtrlSize, curRenderTarget,
bFullscreenMode ? Preview_Fullscreen : Preview_Standard);
//draw selections if in edit mode
if(bEditMode && !bSizeChanging)
{
if(scene) {
LoadVertexShader(solidVertexShader);
LoadPixelShader(solidPixelShader);
solidPixelShader->SetColor(solidPixelShader->GetParameter(0), 0xFF0000);
scene->RenderSelections(solidPixelShader);
}
}
}
else if(bForceRenderViewErase)
{
InvalidateRect(hwndRenderFrame, NULL, TRUE);
UpdateWindow(hwndRenderFrame);
bForceRenderViewErase = false;
}
//------------------------------------
// actual stream output
LoadVertexShader(mainVertexShader);
LoadPixelShader(yuvScalePixelShader);
Texture *yuvRenderTexture = yuvRenderTextures[curRenderTarget];
SetRenderTarget(yuvRenderTexture);
switch(colorDesc.matrix)
{
case ColorMatrix_GBR:
yuvScalePixelShader->SetMatrix(hMatrix, colorDesc.fullRange ? (float*)yuvFullMat[0] : (float*)yuvMat[0]);
break;
case ColorMatrix_YCgCo:
yuvScalePixelShader->SetMatrix(hMatrix, colorDesc.fullRange ? (float*)yuvFullMat[1] : (float*)yuvMat[1]);
break;
case ColorMatrix_BT2020NCL:
yuvScalePixelShader->SetMatrix(hMatrix, colorDesc.fullRange ? (float*)yuvFullMat[2] : (float*)yuvMat[2]);
break;
case ColorMatrix_BT709:
yuvScalePixelShader->SetMatrix(hMatrix, colorDesc.fullRange ? (float*)yuvFullMat[3] : (float*)yuvMat[3]);
break;
case ColorMatrix_SMPTE240M:
yuvScalePixelShader->SetMatrix(hMatrix, colorDesc.fullRange ? (float*)yuvFullMat[4] : (float*)yuvMat[4]);
break;
default:
yuvScalePixelShader->SetMatrix(hMatrix, colorDesc.fullRange ? (float*)yuvFullMat[5] : (float*)yuvMat[5]);
}
if(downscale < 2.01)
yuvScalePixelShader->SetVector2(hScaleVal, 1.0f/baseSize);
else if(downscale < 3.01)
yuvScalePixelShader->SetVector2(hScaleVal, 1.0f/(outputSize*3.0f));
Ortho(0.0f, outputSize.x, outputSize.y, 0.0f, -100.0f, 100.0f);
SetViewport(0.0f, 0.0f, outputSize.x, outputSize.y);
//why am I using scaleSize instead of outputSize for the texture?
//because outputSize can be trimmed by up to three pixels due to 128-bit alignment.
//using the scale function with outputSize can cause slightly inaccurate scaled images
if(bTransitioning)
{
BlendFunction(GS_BLEND_ONE, GS_BLEND_ZERO);
DrawSpriteEx(transitionTexture, 0xFFFFFFFF, 0.0f, 0.0f, scaleSize.x, scaleSize.y, 0.0f, 0.0f, 1.0f, 1.0f);
BlendFunction(GS_BLEND_FACTOR, GS_BLEND_INVFACTOR, transitionAlpha);
}
DrawSpriteEx(mainRenderTextures[curRenderTarget], 0xFFFFFFFF, 0.0f, 0.0f, outputSize.x, outputSize.y, 0.0f, 0.0f, 1.0f, 1.0f);
//------------------------------------
if (bProjector && !copyWait)
projectorSwap->Present(0, 0);
if(bRenderView && !copyWait)
static_cast<D3D10System*>(GS)->swap->Present(0, 0);
OSLeaveMutex(hSceneMutex);
//------------------------------------
// present/upload
profileIn("GPU download and conversion");
bEncode = true;
if(copyWait)
{
copyWait--;
bEncode = false;
}
else
{
//audio sometimes takes a bit to start -- do not start processing frames until audio has started capturing
if(!bRecievedFirstAudioFrame)
{
static bool bWarnedAboutNoAudio = false;
if (renderStartTimeMS-firstFrameTimeMS > 10000 && !bWarnedAboutNoAudio)
{
bWarnedAboutNoAudio = true;
//AddStreamInfo (TEXT ("WARNING: OBS is not receiving audio frames. Please check your audio devices."), StreamInfoPriority_Critical);
}
bEncode = false;
}
else if(bFirstFrame)
{
firstFrameTimestamp = lastStreamTime/1000000;
bFirstFrame = false;
}
if(!bEncode)
{
if(curYUVTexture == (NUM_RENDER_BUFFERS-1))
curYUVTexture = 0;
else
curYUVTexture++;
}
}
lastStreamTime = curStreamTime;
if(bEncode)
{
UINT prevCopyTexture = (curCopyTexture == 0) ? NUM_RENDER_BUFFERS-1 : curCopyTexture-1;
ID3D10Texture2D *copyTexture = copyTextures[curCopyTexture];
profileIn("CopyResource");
if(!bFirstEncode && bUseThreaded420)
{
WaitForMultipleObjects(completeEvents.Num(), completeEvents.Array(), TRUE, INFINITE);
copyTexture->Unmap(0);
}
D3D10Texture *d3dYUV = static_cast<D3D10Texture*>(yuvRenderTextures[curYUVTexture]);
GetD3D()->CopyResource(copyTexture, d3dYUV->texture);
profileOut;
ID3D10Texture2D *prevTexture = copyTextures[prevCopyTexture];
if(bFirstImage) //ignore the first frame
bFirstImage = false;
else
{
HRESULT result;
D3D10_MAPPED_TEXTURE2D map;
if(SUCCEEDED(result = prevTexture->Map(0, D3D10_MAP_READ, 0, &map)))
{
int prevOutBuffer = (curOutBuffer == 0) ? NUM_OUT_BUFFERS-1 : curOutBuffer-1;
int nextOutBuffer = (curOutBuffer == NUM_OUT_BUFFERS-1) ? 0 : curOutBuffer+1;
EncoderPicture &prevPicOut = outPics[prevOutBuffer];
EncoderPicture &picOut = outPics[curOutBuffer];
EncoderPicture &nextPicOut = outPics[nextOutBuffer];
if(!bUsing444)
{
profileIn("conversion to 4:2:0");
if(bUseThreaded420)
{
for(int i=0; i<numThreads; i++)
{
convertInfo[i].input = (LPBYTE)map.pData;
convertInfo[i].inPitch = map.RowPitch;
if(bUsingQSV)
{
mfxFrameData& data = nextPicOut.mfxOut->Data;
videoEncoder->RequestBuffers(&data);
convertInfo[i].outPitch = data.Pitch;
convertInfo[i].output[0] = data.Y;
convertInfo[i].output[1] = data.UV;
}
else
{
convertInfo[i].output[0] = nextPicOut.picOut->img.plane[0];
convertInfo[i].output[1] = nextPicOut.picOut->img.plane[1];
convertInfo[i].output[2] = nextPicOut.picOut->img.plane[2];
}
SetEvent(convertInfo[i].hSignalConvert);
}
if(bFirstEncode)
bFirstEncode = bEncode = false;
}
else
{
if(bUsingQSV)
{
mfxFrameData& data = picOut.mfxOut->Data;
videoEncoder->RequestBuffers(&data);
LPBYTE output[] = {data.Y, data.UV};
Convert444toNV12((LPBYTE)map.pData, outputCX, map.RowPitch, data.Pitch, outputCY, 0, outputCY, output);
}
else
Convert444toNV12((LPBYTE)map.pData, outputCX, map.RowPitch, outputCX, outputCY, 0, outputCY, picOut.picOut->img.plane);
prevTexture->Unmap(0);
}
profileOut;
}
if(bEncode)
{
//encodeThreadProfiler.reset(::new ProfilerNode(TEXT("EncodeThread"), true));
//encodeThreadProfiler->MonitorThread(hEncodeThread);
curFramePic = &picOut;
}
curOutBuffer = nextOutBuffer;
}
else
{
//We have to crash, or we end up deadlocking the thread when the convert threads are never signalled
if (result == DXGI_ERROR_DEVICE_REMOVED)
{
String message;
HRESULT reason = GetD3D()->GetDeviceRemovedReason();
switch (reason)
{
case DXGI_ERROR_DEVICE_RESET:
case DXGI_ERROR_DEVICE_HUNG:
message = TEXT("Your video card or driver froze and was reset. Please check for possible hardware / driver issues.");
break;
case DXGI_ERROR_DEVICE_REMOVED:
message = TEXT("Your video card disappeared from the system. Please check for possible hardware / driver issues.");
break;
case DXGI_ERROR_DRIVER_INTERNAL_ERROR:
message = TEXT("Your video driver reported an internal error. Please check for possible hardware / driver issues.");
break;
case DXGI_ERROR_INVALID_CALL:
message = TEXT("Your video driver reported an invalid call. Please check for possible driver issues.");
break;
default:
message = TEXT("DXGI_ERROR_DEVICE_REMOVED");
break;
}
message << TEXT(" This error can also occur if you have enabled opencl in x264 custom settings.");
CrashError (TEXT("Texture->Map failed: 0x%08x 0x%08x\r\n\r\n%s"), result, reason, message.Array());
}
else
CrashError (TEXT("Texture->Map failed: 0x%08x"), result);
}
}
if(curCopyTexture == (NUM_RENDER_BUFFERS-1))
curCopyTexture = 0;
else
curCopyTexture++;
if(curYUVTexture == (NUM_RENDER_BUFFERS-1))
curYUVTexture = 0;
else
curYUVTexture++;
if (bCongestionControl && bDynamicBitrateSupported && !bTestStream && totalStreamTime > 15000)
{
if (curStrain > 25)
{
if (renderStartTimeMS - lastAdjustmentTime > 1500)
{
if (currentBitRate > 100)
{
currentBitRate = (int)(currentBitRate * (1.0 - (curStrain / 400)));
App->GetVideoEncoder()->SetBitRate(currentBitRate, -1);
if (!adjustmentStreamId)
adjustmentStreamId = App->AddStreamInfo (FormattedString(TEXT("Congestion detected, dropping bitrate to %d kbps"), currentBitRate).Array(), StreamInfoPriority_Low);
else
App->SetStreamInfo(adjustmentStreamId, FormattedString(TEXT("Congestion detected, dropping bitrate to %d kbps"), currentBitRate).Array());
bUpdateBPS = true;
}
lastAdjustmentTime = renderStartTimeMS;
}
}
else if (currentBitRate < defaultBitRate && curStrain < 5 && lastStrain < 5)
{
if (renderStartTimeMS - lastAdjustmentTime > 5000)
{
if (currentBitRate < defaultBitRate)
{
currentBitRate += (int)(defaultBitRate * 0.05);
if (currentBitRate > defaultBitRate)
currentBitRate = defaultBitRate;
}
App->GetVideoEncoder()->SetBitRate(currentBitRate, -1);
/*if (!adjustmentStreamId)
App->AddStreamInfo (FormattedString(TEXT("Congestion clearing, raising bitrate to %d kbps"), currentBitRate).Array(), StreamInfoPriority_Low);
else
App->SetStreamInfo(adjustmentStreamId, FormattedString(TEXT("Congestion clearing, raising bitrate to %d kbps"), currentBitRate).Array());*/
App->RemoveStreamInfo(adjustmentStreamId);
adjustmentStreamId = 0;
bUpdateBPS = true;
lastAdjustmentTime = renderStartTimeMS;
}
}
}
}
lastRenderTarget = curRenderTarget;
if(curRenderTarget == (NUM_RENDER_BUFFERS-1))
curRenderTarget = 0;
else
curRenderTarget++;
if(bUpdateBPS || !CloseDouble(curStrain, lastStrain) || curFramesDropped != lastFramesDropped)
{
PostMessage(hwndMain, OBS_UPDATESTATUSBAR, 0, 0);
lastStrain = curStrain;
lastFramesDropped = curFramesDropped;
}
//------------------------------------
// we're about to sleep so we should flush the d3d command queue
profileIn("flush");
GetD3D()->Flush();
profileOut;
profileOut;
profileOut; //frame
//------------------------------------
// frame sync
QWORD renderStopTime = GetQPCTimeNS();
if(bWasLaggedFrame = (frameDelta > frameLengthNS))
{
numLongFrames++;
if(bLogLongFramesProfile && (numLongFrames/float(max(1, numTotalFrames)) * 100.) > logLongFramesProfilePercentage)
DumpLastProfileData();
}
//OSDebugOut(TEXT("Frame adjust time: %d, "), frameTimeAdjust-totalTime);
numTotalFrames++;
}
DisableProjector();
//encodeThreadProfiler.reset();
if(!bUsing444)
{
if(bUseThreaded420)
{
for(int i=0; i<numThreads; i++)
{
if(h420Threads[i])
{
convertInfo[i].bKillThread = true;
SetEvent(convertInfo[i].hSignalConvert);
OSTerminateThread(h420Threads[i], 10000);
h420Threads[i] = NULL;
}
if(convertInfo[i].hSignalConvert)
{
CloseHandle(convertInfo[i].hSignalConvert);
convertInfo[i].hSignalConvert = NULL;
}
if(convertInfo[i].hSignalComplete)
{
CloseHandle(convertInfo[i].hSignalComplete);
convertInfo[i].hSignalComplete = NULL;
}
}
if(!bFirstEncode)
{
ID3D10Texture2D *copyTexture = copyTextures[curCopyTexture];
copyTexture->Unmap(0);
}
}
if(bUsingQSV)
for(int i = 0; i < NUM_OUT_BUFFERS; i++)
delete outPics[i].mfxOut;
else
for(int i=0; i<NUM_OUT_BUFFERS; i++)
{
x264_picture_clean(outPics[i].picOut);
delete outPics[i].picOut;
}
}
Free(h420Threads);
Free(convertInfo);
Log(TEXT("Total frames rendered: %d, number of late frames: %d (%0.2f%%) (it's okay for some frames to be late)"), numTotalFrames, numLongFrames, (double(numLongFrames)/double(numTotalFrames))*100.0);
}
void FRCPassPostProcessVelocityScatter::Process(FRenderingCompositePassContext& Context)
{
SCOPED_DRAW_EVENT(Context.RHICmdList, PassPostProcessVelocityScatter);
const FPooledRenderTargetDesc* InputDesc = GetInputDesc(ePId_Input0);
if(!InputDesc)
{
// input is not hooked up correctly
return;
}
const FSceneView& View = Context.View;
FIntPoint SrcSize = InputDesc->Extent;
FIntPoint DestSize = PassOutputs[0].RenderTargetDesc.Extent;
// e.g. 4 means the input texture is 4x smaller than the buffer size
uint32 ScaleFactor = GSceneRenderTargets.GetBufferSizeXY().X / SrcSize.X;
FIntRect SrcRect = FIntRect::DivideAndRoundUp(View.ViewRect, ScaleFactor);
FIntRect DestRect = SrcRect;
const FSceneRenderTargetItem& DestRenderTarget = PassOutputs[0].RequestSurface(Context);
TRefCountPtr<IPooledRenderTarget> DepthTarget;
FPooledRenderTargetDesc Desc( FPooledRenderTargetDesc::Create2DDesc( DestRect.Size(), PF_ShadowDepth, TexCreate_None, TexCreate_DepthStencilTargetable, false ) );
GRenderTargetPool.FindFreeElement( Desc, DepthTarget, TEXT("VelocityScatterDepth") );
// Set the view family's render target/viewport.
SetRenderTarget( Context.RHICmdList, DestRenderTarget.TargetableTexture, DepthTarget->GetRenderTargetItem().TargetableTexture );
Context.RHICmdList.Clear( true, FLinearColor::Black, true, 0.0f, false, 0, FIntRect() );
Context.SetViewportAndCallRHI(SrcRect);
// set the state
Context.RHICmdList.SetBlendState(TStaticBlendState<>::GetRHI());
Context.RHICmdList.SetRasterizerState(TStaticRasterizerState<>::GetRHI());
Context.RHICmdList.SetDepthStencilState(TStaticDepthStencilState<true, CF_Greater>::GetRHI());
TShaderMapRef< FPostProcessVelocityScatterVS > VertexShader(Context.GetShaderMap());
TShaderMapRef< FPostProcessVelocityScatterPS > PixelShader(Context.GetShaderMap());
static FGlobalBoundShaderState BoundShaderState;
SetGlobalBoundShaderState(Context.RHICmdList, Context.GetFeatureLevel(), BoundShaderState, GFilterVertexDeclaration.VertexDeclarationRHI, *VertexShader, *PixelShader);
FIntPoint TileCount = SrcRect.Size();
VertexShader->SetParameters( Context, TileCount );
PixelShader->SetParameters( Context );
// needs to be the same on shader side (faster on NVIDIA and AMD)
int32 QuadsPerInstance = 8;
Context.RHICmdList.SetStreamSource(0, NULL, 0, 0);
Context.RHICmdList.DrawIndexedPrimitive(GScatterQuadIndexBuffer.IndexBufferRHI, PT_TriangleList, 0, 0, 32, 0, 2 * QuadsPerInstance, FMath::DivideAndRoundUp(TileCount.X * TileCount.Y, QuadsPerInstance));
Context.RHICmdList.CopyToResolveTarget(DestRenderTarget.TargetableTexture, DestRenderTarget.ShaderResourceTexture, false, FResolveParams());
}
/**
* Renders the view family.
*/
void FForwardShadingSceneRenderer::Render(FRHICommandListImmediate& RHICmdList)
{
if(!ViewFamily.EngineShowFlags.Rendering)
{
return;
}
auto FeatureLevel = ViewFamily.GetFeatureLevel();
// Initialize global system textures (pass-through if already initialized).
GSystemTextures.InitializeTextures(RHICmdList, FeatureLevel);
// Allocate the maximum scene render target space for the current view family.
GSceneRenderTargets.Allocate(ViewFamily);
// Find the visible primitives.
InitViews(RHICmdList);
RenderShadowDepthMaps(RHICmdList);
// Notify the FX system that the scene is about to be rendered.
if (Scene->FXSystem)
{
Scene->FXSystem->PreRender(RHICmdList);
}
GRenderTargetPool.VisualizeTexture.OnStartFrame(Views[0]);
// Dynamic vertex and index buffers need to be committed before rendering.
FGlobalDynamicVertexBuffer::Get().Commit();
FGlobalDynamicIndexBuffer::Get().Commit();
// This might eventually be a problem with multiple views.
// Using only view 0 to check to do on-chip transform of alpha.
FViewInfo& View = Views[0];
const bool bGammaSpace = !IsMobileHDR();
const bool bRequiresUpscale = ((uint32)ViewFamily.RenderTarget->GetSizeXY().X > ViewFamily.FamilySizeX || (uint32)ViewFamily.RenderTarget->GetSizeXY().Y > ViewFamily.FamilySizeY);
const bool bRenderToScene = bRequiresUpscale || FSceneRenderer::ShouldCompositeEditorPrimitives(View);
if (bGammaSpace && !bRenderToScene)
{
SetRenderTarget(RHICmdList, ViewFamily.RenderTarget->GetRenderTargetTexture(), GSceneRenderTargets.GetSceneDepthTexture(), ESimpleRenderTargetMode::EClearToDefault);
}
else
{
// Begin rendering to scene color
GSceneRenderTargets.BeginRenderingSceneColor(RHICmdList, ESimpleRenderTargetMode::EClearToDefault);
}
if (GIsEditor)
{
RHICmdList.Clear(true, Views[0].BackgroundColor, false, 0, false, 0, FIntRect());
}
RenderForwardShadingBasePass(RHICmdList);
// Make a copy of the scene depth if the current hardware doesn't support reading and writing to the same depth buffer
GSceneRenderTargets.ResolveSceneDepthToAuxiliaryTexture(RHICmdList);
// Notify the FX system that opaque primitives have been rendered.
if (Scene->FXSystem)
{
Scene->FXSystem->PostRenderOpaque(RHICmdList);
}
// Draw translucency.
if (ViewFamily.EngineShowFlags.Translucency)
{
SCOPE_CYCLE_COUNTER(STAT_TranslucencyDrawTime);
// Note: Forward pass has no SeparateTranslucency, so refraction effect order with Transluency is different.
// Having the distortion applied between two different translucency passes would make it consistent with the deferred pass.
// This is not done yet.
if (ViewFamily.EngineShowFlags.Refraction)
{
// to apply refraction effect by distorting the scene color
RenderDistortion(RHICmdList);
}
RenderTranslucency(RHICmdList);
}
static const auto CVarMobileMSAA = IConsoleManager::Get().FindTConsoleVariableDataInt(TEXT("r.MobileMSAA"));
bool bOnChipSunMask =
GSupportsRenderTargetFormat_PF_FloatRGBA &&
GSupportsShaderFramebufferFetch &&
ViewFamily.EngineShowFlags.PostProcessing &&
((View.bLightShaftUse) || (View.FinalPostProcessSettings.DepthOfFieldScale > 0.0) ||
((ViewFamily.GetShaderPlatform() == SP_METAL) && (CVarMobileMSAA ? CVarMobileMSAA->GetValueOnAnyThread() > 1 : false))
);
if (!bGammaSpace && bOnChipSunMask)
{
// Convert alpha from depth to circle of confusion with sunshaft intensity.
// This is done before resolve on hardware with framebuffer fetch.
// This will break when PrePostSourceViewportSize is not full size.
FIntPoint PrePostSourceViewportSize = GSceneRenderTargets.GetBufferSizeXY();
FMemMark Mark(FMemStack::Get());
FRenderingCompositePassContext CompositeContext(RHICmdList, View);
FRenderingCompositePass* PostProcessSunMask = CompositeContext.Graph.RegisterPass(new(FMemStack::Get()) FRCPassPostProcessSunMaskES2(PrePostSourceViewportSize, true));
CompositeContext.Root->AddDependency(FRenderingCompositeOutputRef(PostProcessSunMask));
CompositeContext.Process(TEXT("OnChipAlphaTransform"));
}
if (!bGammaSpace || bRenderToScene)
{
// Resolve the scene color for post processing.
GSceneRenderTargets.ResolveSceneColor(RHICmdList, FResolveRect(0, 0, ViewFamily.FamilySizeX, ViewFamily.FamilySizeY));
// Drop depth and stencil before post processing to avoid export.
RHICmdList.DiscardRenderTargets(true, true, 0);
}
if (!bGammaSpace)
{
// Finish rendering for each view, or the full stereo buffer if enabled
if (ViewFamily.bResolveScene)
{
if (ViewFamily.EngineShowFlags.StereoRendering)
{
check(Views.Num() > 1);
//@todo ES2 stereo post: until we get proper stereo postprocessing for ES2, process the stereo buffer as one view
FIntPoint OriginalMax0 = Views[0].ViewRect.Max;
Views[0].ViewRect.Max = Views[1].ViewRect.Max;
GPostProcessing.ProcessES2(RHICmdList, Views[0], bOnChipSunMask);
Views[0].ViewRect.Max = OriginalMax0;
}
else
{
SCOPED_DRAW_EVENT(RHICmdList, PostProcessing);
SCOPE_CYCLE_COUNTER(STAT_FinishRenderViewTargetTime);
for(int32 ViewIndex = 0;ViewIndex < Views.Num();ViewIndex++)
{
SCOPED_CONDITIONAL_DRAW_EVENTF(RHICmdList, EventView, Views.Num() > 1, TEXT("View%d"), ViewIndex);
GPostProcessing.ProcessES2(RHICmdList, Views[ViewIndex], bOnChipSunMask);
}
}
}
}
else if (bRenderToScene)
{
BasicPostProcess(RHICmdList, View, bRequiresUpscale, FSceneRenderer::ShouldCompositeEditorPrimitives(View));
}
RenderFinish(RHICmdList);
}
void Graphics_Mesh(bool isOpenGLMode)
{
StartGraphicsTest();
SetGLEnable(isOpenGLMode);
ace::Log* log = ace::Log_Imp::Create(L"graphics.html", L"メッシュ");
auto window = ace::Window_Imp::Create(640, 480, ace::ToAString(L"メッシュ").c_str());
ASSERT_TRUE(window != nullptr);
auto graphics = ace::Graphics_Imp::Create(window, isOpenGLMode, log, false);
ASSERT_TRUE(graphics != nullptr);
auto renderer3d = new ace::Renderer3D(graphics);
ASSERT_TRUE(renderer3d != nullptr);
renderer3d->SetWindowSize(ace::Vector2DI(640, 480));
auto mesh1 = CreateMesh(graphics);
auto mesh2 = CreateMesh(graphics);
auto deformer = CreateDeformer(graphics);
auto animation = CreateAnimation();
auto cameraObject = new ace::RenderedCameraObject3D(graphics);
cameraObject->SetPosition(ace::Vector3DF(0, 0, 10));
cameraObject->SetFocus(ace::Vector3DF(0, 0, 0));
cameraObject->SetFieldOfView(20.0f);
cameraObject->SetZNear(1.0f);
cameraObject->SetZFar(20.0f);
cameraObject->SetWindowSize(ace::Vector2DI(800, 600));
auto meshObject1 = new ace::RenderedModelObject3D(graphics);
meshObject1->AddMeshGroup();
meshObject1->AddMesh(0, mesh1);
meshObject1->SetPosition(ace::Vector3DF(1, 0, 0));
meshObject1->SetRotation(ace::Vector3DF(20.0f, 20.0f, 0.0f));
auto meshObject2 = new ace::RenderedModelObject3D(graphics);
meshObject2->AddMeshGroup();
meshObject2->AddMesh(0, mesh2);
meshObject2->SetDeformer(0, deformer.get());
meshObject2->SetPosition(ace::Vector3DF(-1, 0, 0));
meshObject2->SetRotation(ace::Vector3DF(20.0f, 20.0f, 0.0f));
meshObject2->AddAnimationClip(ace::ToAString("anime1").c_str(), animation.get());
meshObject2->PlayAnimation(ace::ToAString("anime1").c_str());
auto lightObject = new ace::RenderedDirectionalLightObject3D(graphics);
lightObject->SetRotation(ace::Vector3DF(30, 160, 0));
renderer3d->AddObject(cameraObject);
renderer3d->AddObject(meshObject1);
renderer3d->AddObject(meshObject2);
renderer3d->AddObject(lightObject);
auto renderer2d = new ace::Renderer2D_Imp(graphics, log, window->GetSize());
int32_t time = 0;
while (window->DoEvent())
{
graphics->Begin();
graphics->Clear(true, false, ace::Color(0, 0, 0, 255));
renderer3d->Flip();
renderer3d->BeginRendering();
ace::Sleep(100);
renderer3d->EndRendering();
graphics->SetRenderTarget(nullptr, nullptr);
ace::Vector2DF positions[4];
ace::Color colors[4];
ace::Vector2DF uvs[4];
colors[0] = ace::Color(255, 255, 255, 255);
colors[1] = ace::Color(255, 255, 255, 255);
colors[2] = ace::Color(255, 255, 255, 255);
colors[3] = ace::Color(255, 255, 255, 255);
positions[0].X = 0;
positions[0].Y = 0;
positions[1].X = 640;
positions[1].Y = 0;
positions[2].X = 640;
positions[2].Y = 480;
positions[3].X = 0;
positions[3].Y = 480;
uvs[0].X = 0;
uvs[0].Y = 0;
uvs[1].X = 1;
uvs[1].Y = 0;
uvs[2].X = 1;
uvs[2].Y = 1;
uvs[3].X = 0;
uvs[3].Y = 1;
renderer2d->AddSprite(positions, colors, uvs, renderer3d->GetRenderTarget(), ace::eAlphaBlend::ALPHA_BLEND_BLEND, 0);
renderer2d->DrawCache();
renderer2d->ClearCache();
graphics->Present();
graphics->End();
if (time == 10)
{
SAVE_SCREEN_SHOT(graphics, 0);
}
if (time == 11)
{
window->Close();
}
time++;
}
meshObject1->Release();
meshObject2->Release();
cameraObject->Release();
lightObject->Release();
delete renderer2d;
delete renderer3d;
graphics->Release();
window->Release();
delete log;
}
