Files
UnrealEngineUWP/Engine/Shaders/Private/VirtualShadowMaps/VirtualShadowMapCacheInvalidation.ush
wouter dek 8f310241b2 Use DoubleFloat instead of tile+offset in various systems on GPU to improve numerical precision.
#jira UE-171685
#rb Charles.deRousiers, tim.doerries

[CL 30763421 by wouter dek in ue5-main branch]
2024-01-21 21:02:35 -05:00

139 lines
5.8 KiB
Plaintext

// Copyright Epic Games, Inc. All Rights Reserved.
/*=============================================================================
VirtualShadowMapCacheInvalidation.ush:
=============================================================================*/
#pragma once
#include "../Common.ush"
#include "../SceneData.ush"
#include "VirtualShadowMapProjectionCommon.ush"
#include "VirtualShadowMapPageOverlap.ush"
#include "VirtualShadowMapPageCacheCommon.ush"
#include "../ScreenSpaceDenoise/SSDDefinitions.ush" // For LIGHT_TYPE's
#include "../Nanite/NaniteHZBCull.ush"
RWStructuredBuffer<FPhysicalPageMetaData> PhysicalPageMetaDataOut;
bool IsValidInstanceAndCastShadows(uint InstanceId)
{
FInstanceSceneData InstanceSceneData = GetInstanceSceneData(InstanceId, Scene.GPUScene.InstanceDataSOAStride);
if (!InstanceSceneData.ValidInstance)
{
return false;
}
const bool bCastShadows = (GetPrimitiveData(InstanceSceneData.PrimitiveId).Flags & PRIMITIVE_SCENE_DATA_FLAG_CAST_SHADOWS) != 0U;
// TODO: test the flag on the instance instead once it is updated correctly InstanceSceneData.CastShadows
return bCastShadows;
}
void InvalidateInstancePages(int VirtualShadowMapId, uint InstanceId, bool bForceInvalidateStaticPage)
{
FInstanceSceneData InstanceSceneData = GetInstanceSceneData(InstanceId, Scene.GPUScene.InstanceDataSOAStride);
const bool bIsNanite = InstanceSceneData.NaniteRuntimeResourceID != 0xFFFFFFFFu;
FVirtualShadowMapProjectionShaderData ProjectionData = GetVirtualShadowMapProjectionData(VirtualShadowMapId);
const bool bDirectionalLight = (ProjectionData.LightType == LIGHT_TYPE_DIRECTIONAL);
// NOTE: This is the *shadow view*'s translated world, not primary view
float4x4 LocalToTranslatedWorld = DFMultiplyTranslationDemote(InstanceSceneData.LocalToWorld, ProjectionData.PreViewTranslation);
// distance cull invalidations for local lights
if (!bDirectionalLight)
{
float InstanceRadius = length(InstanceSceneData.LocalBoundsExtent * InstanceSceneData.NonUniformScale.xyz);
float3 TranslatedWorldCenter = mul(float4(InstanceSceneData.LocalBoundsCenter, 1.0f), LocalToTranslatedWorld).xyz;
if (length2(TranslatedWorldCenter) > Square(ProjectionData.LightRadius + InstanceRadius))
{
return;
}
}
// Go back to clip space
float4x4 UVToClip;
UVToClip[0] = float4(2, 0, 0, 0);
UVToClip[1] = float4(0, -2, 0, 0);
UVToClip[2] = float4(0, 0, 1, 0);
UVToClip[3] = float4(-1, 1, 0, 1);
const bool bIsOrtho = bDirectionalLight;
const bool bNearClip = !bDirectionalLight;
FFrustumCullData Cull = BoxCullFrustum(
InstanceSceneData.LocalBoundsCenter,
InstanceSceneData.LocalBoundsExtent,
LocalToTranslatedWorld,
mul(ProjectionData.TranslatedWorldToShadowUVMatrix, UVToClip),
ProjectionData.ShadowViewToClipMatrix, bIsOrtho, bNearClip, false);
float PixelEstRadius = CalcClipSpaceRadiusEstimate(bIsOrtho, InstanceSceneData, LocalToTranslatedWorld, ProjectionData.ShadowViewToClipMatrix) * float(VSM_VIRTUAL_MAX_RESOLUTION_XY);
if (Cull.bIsVisible)
{
bool bAllowWPO = VirtualShadowMapIsWPOAllowed(GetPrimitiveData(InstanceSceneData.PrimitiveId), VirtualShadowMapId);
bool bShouldCacheAsStatic = ShouldCacheInstanceAsStatic(InstanceId, ProjectionData.bUnCached, bAllowWPO);
bool bInvalidateStaticPage = bShouldCacheAsStatic || bForceInvalidateStaticPage;
uint InvalidationFlags = (bInvalidateStaticPage ? VSM_STATIC_UNCACHED_FLAG : VSM_DYNAMIC_UNCACHED_FLAG) << VSM_PHYSICAL_PAGE_INVALIDATION_FLAGS_SHIFT;
// 2. figure out overlap and all that
// case #1 mip-map VSM - loop all mip levels, case #2 clipmap, just one 'mip level'
int NumMipLevels = (ProjectionData.ClipmapLevelCountRemaining <= 0) ? VSM_MAX_MIP_LEVELS : 1;
{
for (int MipLevel = 0; MipLevel < NumMipLevels; ++MipLevel)
{
int ViewDim = int(uint(VSM_VIRTUAL_MAX_RESOLUTION_XY) >> MipLevel);
FScreenRect Rect = GetScreenRect(int4(0, 0, ViewDim, ViewDim), Cull, 4);
// Add a small epsilon to the HZB depth test
// This is to handle the rare case where an object that is fully parallel to the
// light's near plane might self-occlude the HZB test due to minor precision differences
// in the computation. While rare, this can come up with things like point lights and
// axis aligned boxes.
Rect.Depth += 1e-8f;
uint4 RectPages = GetPageRect(Rect, ProjectionData.VirtualShadowMapId, MipLevel);
uint FlagMask = VSM_ALLOCATED_FLAG | GetDetailGeometryFlagMaskForCulling(bInvalidateStaticPage, bIsNanite, PixelEstRadius);
PixelEstRadius *= 0.5f;
// Use Hierarchical mip test to speed up (allows skipping invalidating areas that don't have any flags anyway)
if (OverlapsAnyValidPage(ProjectionData.VirtualShadowMapId, MipLevel, RectPages, FlagMask))
{
#if USE_HZB_OCCLUSION
FPageTestScreenRect HZBTestRect = SetupPageHZBRect(Rect, ProjectionData.VirtualShadowMapId, MipLevel);
#endif // USE_HZB_OCCLUSION
// 3. do invalidation
FVirtualSMLevelOffset PageTableLevelOffset = CalcPageTableLevelOffset(ProjectionData.VirtualShadowMapId, MipLevel);
for (uint y = RectPages.y; y <= RectPages.w; y++)
{
for (uint x = RectPages.x; x <= RectPages.z; x++)
{
uint PageFlagOffset = CalcPageOffset(PageTableLevelOffset, MipLevel, uint2(x, y));
uint PageFlag = VirtualShadowMap.PageFlags[PageFlagOffset];
if ((PageFlag & FlagMask) == FlagMask)
{
#if USE_HZB_OCCLUSION
if (!IsPageVisibleHZB(uint2(x, y), PageFlagOffset, HZBTestRect))
{
continue;
}
#endif // USE_HZB_OCCLUSION
// Accumulate static/dynamic invalidation flags
// TODO: Wave version
uint2 PhysicalAddress = ShadowGetPhysicalPage(PageFlagOffset).PhysicalAddress;
uint PhysicalPageIndex = VSMPhysicalPageAddressToIndex(PhysicalAddress);
InterlockedOr(PhysicalPageMetaDataOut[PhysicalPageIndex].Flags, InvalidationFlags);
}
}
}
}
}
}
}
}