fix: ssgi outlines and terrain occlusion slow update (doodlum#350)

* fix: ssgi outlines * chore: doubling terrain shadow update length
FlayaN · Jul 23, 2024 · d7cd1d0 · d7cd1d0
1 parent c34fc04
commit d7cd1d0
Show file tree

Hide file tree

Showing 4 changed files with 10 additions and 6 deletions.
diff --git a/features/Screen Space GI/Shaders/ScreenSpaceGI/gi.cs.hlsl b/features/Screen Space GI/Shaders/ScreenSpaceGI/gi.cs.hlsl
@@ -350,7 +350,7 @@ void CalculateGI(
 	const static bool useHistory = false;
 #endif
 
-	float2 uv = (pxCoord + .5f) * RcpFrameDim;
+	float2 uv = (pxCoord + .5) * RcpFrameDim;
 	uint eyeIndex = GetEyeIndexFromTexCoord(uv);
 
 	float viewspaceZ = srcWorkingDepth[pxCoord];

diff --git a/features/Screen Space GI/Shaders/ScreenSpaceGI/prefilterDepths.cs.hlsl b/features/Screen Space GI/Shaders/ScreenSpaceGI/prefilterDepths.cs.hlsl
@@ -67,7 +67,7 @@ groupshared float g_scratchDepths[8][8];
 	const uint2 pixCoord = baseCoord * 2;
 	const float2 uv = (pixCoord + .5) * RcpFrameDim;
 
-	float4 depths4 = srcNDCDepth.GatherRed(samplerPointClamp, uv * frameScale, int2(1, 1));
+	float4 depths4 = srcNDCDepth.GatherRed(samplerPointClamp, uv * frameScale);
 	float depth0 = ClampDepth(ScreenToViewDepth(depths4.w));
 	float depth1 = ClampDepth(ScreenToViewDepth(depths4.z));
 	float depth2 = ClampDepth(ScreenToViewDepth(depths4.x));

diff --git a/features/Terrain Occlusion/Shaders/TerrainOcclusion/ShadowUpdate.cs.hlsl b/features/Terrain Occlusion/Shaders/TerrainOcclusion/ShadowUpdate.cs.hlsl
@@ -50,7 +50,7 @@ float2 GetInterpolatedHeightRW(float2 pxCoord, bool isVertical)
 		return heightA;
 }
 
-#define NTHREADS 64
+#define NTHREADS 128
 groupshared float2 g_shadowHeight[NTHREADS];
 
 [numthreads(NTHREADS, 1, 1)] void main(const uint gtid

diff --git a/src/Features/TerrainOcclusion.cpp b/src/Features/TerrainOcclusion.cpp
@@ -422,6 +422,10 @@ void TerrainOcclusion::UpdateShadow()
 	if (!IsHeightMapReady())
 		return;
 
+	// don't forget to change NTHREADS in shader!
+	constexpr uint updateLength = 128u;
+	constexpr uint logUpdateLength = std::bit_width(128u) - 1;  // integer log2, https://stackoverflow.com/questions/994593/how-to-do-an-integer-log2-in-c
+
 	auto& context = State::GetSingleton()->context;
 	auto accumulator = RE::BSGraphics::BSShaderAccumulator::GetCurrentAccumulator();
 	auto sunLight = skyrim_cast<RE::NiDirectionalLight*>(accumulator->GetRuntimeData().activeShadowSceneNode->GetRuntimeData().sunLight->light.get());
@@ -454,12 +458,12 @@ void TerrainOcclusion::UpdateShadow()
 			stepMult = 1.f / abs(dirLightPxDir.x);
 			edgePxCoord = dirLightPxDir.x > 0 ? 0 : (width - 1);
 			signDir = dirLightPxDir.x > 0 ? 1 : -1;
-			maxUpdates = ((width - 1) >> 6) + 1;
+			maxUpdates = (width + updateLength - 1) >> logUpdateLength;
 		} else {
 			stepMult = 1.f / abs(dirLightPxDir.y);
 			edgePxCoord = dirLightPxDir.y > 0 ? 0 : height - 1;
 			signDir = dirLightPxDir.y > 0 ? 1 : -1;
-			maxUpdates = ((height - 1) >> 6) + 1;
+			maxUpdates = (height + updateLength - 1) >> logUpdateLength;
 		}
 		dirLightPxDir *= stepMult;
 
@@ -474,7 +478,7 @@ void TerrainOcclusion::UpdateShadow()
 		shadowUpdateCBData.LightDeltaZ = -(lenUV / invScale.z * stepMult) * float2{ std::tan(upperAngle), std::tan(lowerAngle) };
 	}
 
-	shadowUpdateCBData.StartPxCoord = edgePxCoord + signDir * shadowUpdateIdx * 64u;
+	shadowUpdateCBData.StartPxCoord = edgePxCoord + signDir * shadowUpdateIdx * updateLength;
 	shadowUpdateCBData.PxSize = { 1.f / texNormalisedHeight->desc.Width, 1.f / texNormalisedHeight->desc.Height };
 
 	shadowUpdateCB->Update(shadowUpdateCBData);