You've already forked UnrealEngineUWP
mirror of
https://github.com/izzy2lost/UnrealEngineUWP.git
synced 2026-03-26 18:15:20 -07:00
Fix Gather4 offsets in shadow map filtering (off-by half error)
#rb Andrew.Lauritzen [CL 14218017 by Ola Olsson in ue5-main branch]
This commit is contained in:
Ola Olsson
@@ -227,29 +227,32 @@ float3 CalculateOcclusion(float3 ShadowmapDepth, FPCFSamplerSettings Settings)
|
||||
}
|
||||
}
|
||||
|
||||
void FetchRowOfThree(float2 Sample00TexelCenter, float VerticalOffset, out float3 Values0, FPCFSamplerSettings Settings)
|
||||
float3 FetchRowOfThree(float2 Sample00TexelCenter, float VerticalOffset, FPCFSamplerSettings Settings)
|
||||
{
|
||||
Values0.x = Texture2DSampleLevel(Settings.ShadowDepthTexture, Settings.ShadowDepthTextureSampler, (Sample00TexelCenter + float2(0, VerticalOffset)) * Settings.ShadowBufferSize.zw, 0).r;
|
||||
Values0.y = Texture2DSampleLevel(Settings.ShadowDepthTexture, Settings.ShadowDepthTextureSampler, (Sample00TexelCenter + float2(1, VerticalOffset)) * Settings.ShadowBufferSize.zw, 0).r;
|
||||
Values0.z = Texture2DSampleLevel(Settings.ShadowDepthTexture, Settings.ShadowDepthTextureSampler, (Sample00TexelCenter + float2(2, VerticalOffset)) * Settings.ShadowBufferSize.zw, 0).r;
|
||||
Values0 = CalculateOcclusion(Values0, Settings);
|
||||
float3 Values;
|
||||
Values.x = Texture2DSampleLevel(Settings.ShadowDepthTexture, Settings.ShadowDepthTextureSampler, (Sample00TexelCenter + float2(0, VerticalOffset)) * Settings.ShadowBufferSize.zw, 0).r;
|
||||
Values.y = Texture2DSampleLevel(Settings.ShadowDepthTexture, Settings.ShadowDepthTextureSampler, (Sample00TexelCenter + float2(1, VerticalOffset)) * Settings.ShadowBufferSize.zw, 0).r;
|
||||
Values.z = Texture2DSampleLevel(Settings.ShadowDepthTexture, Settings.ShadowDepthTextureSampler, (Sample00TexelCenter + float2(2, VerticalOffset)) * Settings.ShadowBufferSize.zw, 0).r;
|
||||
return CalculateOcclusion(Values, Settings);
|
||||
}
|
||||
|
||||
void FetchRowOfFour(float2 Sample00TexelCenter, float VerticalOffset, out float4 Values0, FPCFSamplerSettings Settings)
|
||||
float4 FetchRowOfFour(float2 Sample00TexelCenter, float VerticalOffset, FPCFSamplerSettings Settings)
|
||||
{
|
||||
Values0.x = Texture2DSampleLevel(Settings.ShadowDepthTexture, Settings.ShadowDepthTextureSampler, (Sample00TexelCenter + float2(0, VerticalOffset)) * Settings.ShadowBufferSize.zw, 0).r;
|
||||
Values0.y = Texture2DSampleLevel(Settings.ShadowDepthTexture, Settings.ShadowDepthTextureSampler, (Sample00TexelCenter + float2(1, VerticalOffset)) * Settings.ShadowBufferSize.zw, 0).r;
|
||||
Values0.z = Texture2DSampleLevel(Settings.ShadowDepthTexture, Settings.ShadowDepthTextureSampler, (Sample00TexelCenter + float2(2, VerticalOffset)) * Settings.ShadowBufferSize.zw, 0).r;
|
||||
Values0.w = Texture2DSampleLevel(Settings.ShadowDepthTexture, Settings.ShadowDepthTextureSampler, (Sample00TexelCenter + float2(3, VerticalOffset)) * Settings.ShadowBufferSize.zw, 0).r;
|
||||
Values0 = CalculateOcclusion(Values0, Settings);
|
||||
float4 Values;
|
||||
Values.x = Texture2DSampleLevel(Settings.ShadowDepthTexture, Settings.ShadowDepthTextureSampler, (Sample00TexelCenter + float2(0, VerticalOffset)) * Settings.ShadowBufferSize.zw, 0).r;
|
||||
Values.y = Texture2DSampleLevel(Settings.ShadowDepthTexture, Settings.ShadowDepthTextureSampler, (Sample00TexelCenter + float2(1, VerticalOffset)) * Settings.ShadowBufferSize.zw, 0).r;
|
||||
Values.z = Texture2DSampleLevel(Settings.ShadowDepthTexture, Settings.ShadowDepthTextureSampler, (Sample00TexelCenter + float2(2, VerticalOffset)) * Settings.ShadowBufferSize.zw, 0).r;
|
||||
Values.w = Texture2DSampleLevel(Settings.ShadowDepthTexture, Settings.ShadowDepthTextureSampler, (Sample00TexelCenter + float2(3, VerticalOffset)) * Settings.ShadowBufferSize.zw, 0).r;
|
||||
return CalculateOcclusion(Values, Settings);
|
||||
}
|
||||
|
||||
void FetchRowOfThreeAfterFour(float2 Sample00TexelCenter, float VerticalOffset, out float3 Values1, FPCFSamplerSettings Settings)
|
||||
float3 FetchRowOfThreeAfterFour(float2 Sample00TexelCenter, float VerticalOffset, FPCFSamplerSettings Settings)
|
||||
{
|
||||
Values1.x = Texture2DSampleLevel(Settings.ShadowDepthTexture, Settings.ShadowDepthTextureSampler, (Sample00TexelCenter + float2(4, VerticalOffset)) * Settings.ShadowBufferSize.zw, 0).r;
|
||||
Values1.y = Texture2DSampleLevel(Settings.ShadowDepthTexture, Settings.ShadowDepthTextureSampler, (Sample00TexelCenter + float2(5, VerticalOffset)) * Settings.ShadowBufferSize.zw, 0).r;
|
||||
Values1.z = Texture2DSampleLevel(Settings.ShadowDepthTexture, Settings.ShadowDepthTextureSampler, (Sample00TexelCenter + float2(6, VerticalOffset)) * Settings.ShadowBufferSize.zw, 0).r;
|
||||
Values1 = CalculateOcclusion(Values1, Settings);
|
||||
float3 Values;
|
||||
Values.x = Texture2DSampleLevel(Settings.ShadowDepthTexture, Settings.ShadowDepthTextureSampler, (Sample00TexelCenter + float2(4, VerticalOffset)) * Settings.ShadowBufferSize.zw, 0).r;
|
||||
Values.y = Texture2DSampleLevel(Settings.ShadowDepthTexture, Settings.ShadowDepthTextureSampler, (Sample00TexelCenter + float2(5, VerticalOffset)) * Settings.ShadowBufferSize.zw, 0).r;
|
||||
Values.z = Texture2DSampleLevel(Settings.ShadowDepthTexture, Settings.ShadowDepthTextureSampler, (Sample00TexelCenter + float2(6, VerticalOffset)) * Settings.ShadowBufferSize.zw, 0).r;
|
||||
return CalculateOcclusion(Values, Settings);
|
||||
}
|
||||
|
||||
float Manual3x3PCF(float2 ShadowPosition, FPCFSamplerSettings Settings)
|
||||
@@ -258,24 +261,25 @@ float Manual3x3PCF(float2 ShadowPosition, FPCFSamplerSettings Settings)
|
||||
{
|
||||
float2 TexelPos = ShadowPosition * Settings.ShadowBufferSize.xy - 0.5f; // bias to be consistent with texture filtering hardware
|
||||
float2 Fraction = frac(TexelPos);
|
||||
float2 TexelCenter = floor(TexelPos) + 0.5f; // bias to get reliable texel center content
|
||||
float2 TexelCorner = floor(TexelPos); // bias to get reliable texel center content
|
||||
{
|
||||
float2 Sample00TexelCenter = TexelCenter - float2(1, 1);
|
||||
|
||||
float4 SampleValues0, SampleValues1, SampleValues2, SampleValues3;
|
||||
|
||||
#if FEATURE_GATHER4
|
||||
float2 SamplePos = TexelCenter * Settings.ShadowBufferSize.zw; // bias to get reliable texel center content
|
||||
SampleValues0 = CalculateOcclusion(Settings.ShadowDepthTexture.Gather(Settings.ShadowDepthTextureSampler, SamplePos, int2(-1, -1)), Settings);
|
||||
SampleValues1 = CalculateOcclusion(Settings.ShadowDepthTexture.Gather(Settings.ShadowDepthTextureSampler, SamplePos, int2(1, -1)), Settings);
|
||||
SampleValues2 = CalculateOcclusion(Settings.ShadowDepthTexture.Gather(Settings.ShadowDepthTextureSampler, SamplePos, int2(-1, 1)), Settings);
|
||||
SampleValues3 = CalculateOcclusion(Settings.ShadowDepthTexture.Gather(Settings.ShadowDepthTextureSampler, SamplePos, int2(1, 1)), Settings);
|
||||
// Gather4 samples "at the following locations: (-,+),(+,+),(+,-),(-,-), where the magnitude of the deltas are always half a texel" - DX11 Func. Spec.
|
||||
// So we need to offset to the centre of the 2x2 grid we want to sample.
|
||||
float2 SamplePos = (TexelCorner + 1.0f) * Settings.ShadowBufferSize.zw; // bias to get reliable texel center content
|
||||
float4 SampleValues0 = CalculateOcclusion(Settings.ShadowDepthTexture.Gather(Settings.ShadowDepthTextureSampler, SamplePos, int2(-1, -1)), Settings);
|
||||
float4 SampleValues1 = CalculateOcclusion(Settings.ShadowDepthTexture.Gather(Settings.ShadowDepthTextureSampler, SamplePos, int2(1, -1)), Settings);
|
||||
float4 SampleValues2 = CalculateOcclusion(Settings.ShadowDepthTexture.Gather(Settings.ShadowDepthTextureSampler, SamplePos, int2(-1, 1)), Settings);
|
||||
float4 SampleValues3 = CalculateOcclusion(Settings.ShadowDepthTexture.Gather(Settings.ShadowDepthTextureSampler, SamplePos, int2(1, 1)), Settings);
|
||||
return PCF3x3gather(Fraction, SampleValues0, SampleValues1, SampleValues2, SampleValues3);
|
||||
#else // FEATURE_GATHER4
|
||||
FetchRowOfFour(Sample00TexelCenter, 0, SampleValues0, Settings);
|
||||
FetchRowOfFour(Sample00TexelCenter, 1, SampleValues1, Settings);
|
||||
FetchRowOfFour(Sample00TexelCenter, 2, SampleValues2, Settings);
|
||||
FetchRowOfFour(Sample00TexelCenter, 3, SampleValues3, Settings);
|
||||
float2 Sample00TexelCenter = TexelCorner - 0.5f;
|
||||
float4 SampleValues0 = FetchRowOfFour(Sample00TexelCenter, 0, Settings);
|
||||
float4 SampleValues1 = FetchRowOfFour(Sample00TexelCenter, 1, Settings);
|
||||
float4 SampleValues2 = FetchRowOfFour(Sample00TexelCenter, 2, Settings);
|
||||
float4 SampleValues3 = FetchRowOfFour(Sample00TexelCenter, 3, Settings);
|
||||
return PCF3x3(Fraction, SampleValues0, SampleValues1, SampleValues2, SampleValues3);
|
||||
#endif // FEATURE_GATHER4
|
||||
}
|
||||
@@ -291,11 +295,9 @@ float Manual2x2PCF(float2 ShadowPosition, FPCFSamplerSettings Settings)
|
||||
|
||||
float2 Sample00TexelCenter = TexelCenter - float2(1, 1);
|
||||
|
||||
float3 SamplesValues0, SamplesValues1, SamplesValues2;
|
||||
|
||||
FetchRowOfThree(Sample00TexelCenter, 0, SamplesValues0, Settings);
|
||||
FetchRowOfThree(Sample00TexelCenter, 1, SamplesValues1, Settings);
|
||||
FetchRowOfThree(Sample00TexelCenter, 2, SamplesValues2, Settings);
|
||||
float3 SamplesValues0 = FetchRowOfThree(Sample00TexelCenter, 0, Settings);
|
||||
float3 SamplesValues1 = FetchRowOfThree(Sample00TexelCenter, 1, Settings);
|
||||
float3 SamplesValues2 = FetchRowOfThree(Sample00TexelCenter, 2, Settings);
|
||||
|
||||
return PCF2x2(Fraction, SamplesValues0, SamplesValues1, SamplesValues2);
|
||||
}
|
||||
@@ -309,11 +311,9 @@ float Manuax2PCF(float2 ShadowPosition, FPCFSamplerSettings Settings)
|
||||
|
||||
float2 Sample00TexelCenter = TexelCenter - float2(1, 1);
|
||||
|
||||
float3 SamplesValues0, SamplesValues1, SamplesValues2;
|
||||
|
||||
FetchRowOfThree(Sample00TexelCenter, 0, SamplesValues0, Settings);
|
||||
FetchRowOfThree(Sample00TexelCenter, 1, SamplesValues1, Settings);
|
||||
FetchRowOfThree(Sample00TexelCenter, 2, SamplesValues2, Settings);
|
||||
float3 SamplesValues0 = FetchRowOfThree(Sample00TexelCenter, 0, Settings);
|
||||
float3 SamplesValues1 = FetchRowOfThree(Sample00TexelCenter, 1, Settings);
|
||||
float3 SamplesValues2 = FetchRowOfThree(Sample00TexelCenter, 2, Settings);
|
||||
|
||||
return PCF2x2(Fraction, SamplesValues0, SamplesValues1, SamplesValues2);
|
||||
}
|
||||
@@ -328,14 +328,17 @@ float Manual1x1PCF(float2 ShadowPosition, FPCFSamplerSettings Settings)
|
||||
{
|
||||
float2 TexelPos = ShadowPosition * Settings.ShadowBufferSize.xy - 0.5f; // bias to be consistent with texture filtering hardware
|
||||
float2 Fraction = frac(TexelPos);
|
||||
float2 TexelCenter = floor(TexelPos) + 0.5f; // bias to get reliable texel center content
|
||||
|
||||
// using Gather: xyzw in counter clockwise order starting with the sample to the lower left of the queried location
|
||||
float4 Samples;
|
||||
|
||||
#if FEATURE_GATHER4
|
||||
Samples = Settings.ShadowDepthTexture.Gather(Settings.ShadowDepthTextureSampler, TexelCenter * Settings.ShadowBufferSize.zw);
|
||||
// Gather4 samples "at the following locations: (-,+),(+,+),(+,-),(-,-), where the magnitude of the deltas are always half a texel" - DX11 Func. Spec.
|
||||
// So we need to offset to the centre of the 2x2 grid we want to sample.
|
||||
float2 QuadCenter = floor(TexelPos) + 1.0f;
|
||||
Samples = Settings.ShadowDepthTexture.Gather(Settings.ShadowDepthTextureSampler, QuadCenter * Settings.ShadowBufferSize.zw);
|
||||
#else
|
||||
float2 TexelCenter = floor(TexelPos) + 0.5f; // bias to get reliable texel center content
|
||||
Samples.x = Texture2DSampleLevel(Settings.ShadowDepthTexture, Settings.ShadowDepthTextureSampler, (TexelCenter.xy + float2(0, 1)) * Settings.ShadowBufferSize.zw, 0).r;
|
||||
Samples.y = Texture2DSampleLevel(Settings.ShadowDepthTexture, Settings.ShadowDepthTextureSampler, (TexelCenter.xy + float2(1, 1)) * Settings.ShadowBufferSize.zw, 0).r;
|
||||
Samples.z = Texture2DSampleLevel(Settings.ShadowDepthTexture, Settings.ShadowDepthTextureSampler, (TexelCenter.xy + float2(1, 0)) * Settings.ShadowBufferSize.zw, 0).r;
|
||||
@@ -361,17 +364,16 @@ float ManualPCF(float2 ShadowPosition, FPCFSamplerSettings Settings)
|
||||
{
|
||||
float2 TexelPos = ShadowPosition * Settings.ShadowBufferSize.xy - 0.5f; // bias to be consistent with texture filtering hardware
|
||||
float2 Fraction = frac(TexelPos);
|
||||
float2 TexelCenter = floor(TexelPos);
|
||||
float2 SamplePos = (TexelCenter + 0.5f) * Settings.ShadowBufferSize.zw; // bias to get reliable texel center content
|
||||
|
||||
float Results;
|
||||
// Gather4 samples "at the following locations: (-,+),(+,+),(+,-),(-,-), where the magnitude of the deltas are always half a texel" - DX11 Func. Spec.
|
||||
// So we need to offset to the centre of the 2x2 grid we want to sample.
|
||||
float2 SamplePos = (floor(TexelPos) + 1.0f) * Settings.ShadowBufferSize.zw; // bias to get reliable texel center content
|
||||
|
||||
float4 Values00 = CalculateOcclusion(Settings.ShadowDepthTexture.Gather(Settings.ShadowDepthTextureSampler, SamplePos, int2(-2, -2)), Settings);
|
||||
float4 Values20 = CalculateOcclusion(Settings.ShadowDepthTexture.Gather(Settings.ShadowDepthTextureSampler, SamplePos, int2(0, -2)), Settings);
|
||||
float4 Values40 = CalculateOcclusion(Settings.ShadowDepthTexture.Gather(Settings.ShadowDepthTextureSampler, SamplePos, int2(2, -2)), Settings);
|
||||
|
||||
float2 Row0 = HorizontalPCF5x2(Fraction, Values00, Values20, Values40);
|
||||
Results = Row0.x * (1.0f - Fraction.y) + Row0.y;
|
||||
float Results = Row0.x * (1.0f - Fraction.y) + Row0.y;
|
||||
|
||||
float4 Values02 = CalculateOcclusion(Settings.ShadowDepthTexture.Gather(Settings.ShadowDepthTextureSampler, SamplePos, int2(-2, 0)), Settings);
|
||||
float4 Values22 = CalculateOcclusion(Settings.ShadowDepthTexture.Gather(Settings.ShadowDepthTextureSampler, SamplePos, int2(0, 0)), Settings);
|
||||
@@ -412,28 +414,25 @@ float ManualPCF(float2 ShadowPosition, FPCFSamplerSettings Settings)
|
||||
// Group work to minimize temporary registers and to split texture work with PCF ALU operations to hide texture latency
|
||||
// Without this layout (all texture lookups at the beginning, PCF ALU's at the end) this shader was 4x slower on Nvidia cards
|
||||
{
|
||||
float4 SampleValues00;
|
||||
FetchRowOfFour(Sample00TexelCenter, 0, SampleValues00, Settings);
|
||||
float4 SampleValues00 = FetchRowOfFour(Sample00TexelCenter, 0, Settings);
|
||||
SampleValues10.x = SampleValues00.w;
|
||||
|
||||
float4 SampleValues01;
|
||||
FetchRowOfFour(Sample00TexelCenter, 1, SampleValues01, Settings);
|
||||
float4 SampleValues01 = FetchRowOfFour(Sample00TexelCenter, 1, Settings);
|
||||
SampleValues11.x = SampleValues01.w;
|
||||
|
||||
float4 SampleValues02;
|
||||
FetchRowOfFour(Sample00TexelCenter, 2, SampleValues02, Settings);
|
||||
float4 SampleValues02 = FetchRowOfFour(Sample00TexelCenter, 2, Settings);
|
||||
SampleValues12.x = SampleValues02.w;
|
||||
|
||||
FetchRowOfFour(Sample00TexelCenter, 3, SampleValues03, Settings);
|
||||
SampleValues03 = FetchRowOfFour(Sample00TexelCenter, 3, Settings);
|
||||
SampleValues13.x = SampleValues03.w;
|
||||
Results.x = PCF3x3(Fraction, SampleValues00, SampleValues01, SampleValues02, SampleValues03);
|
||||
}
|
||||
|
||||
{
|
||||
FetchRowOfThreeAfterFour(Sample00TexelCenter, 0, SampleValues10.yzw, Settings);
|
||||
FetchRowOfThreeAfterFour(Sample00TexelCenter, 1, SampleValues11.yzw, Settings);
|
||||
FetchRowOfThreeAfterFour(Sample00TexelCenter, 2, SampleValues12.yzw, Settings);
|
||||
FetchRowOfThreeAfterFour(Sample00TexelCenter, 3, SampleValues13.yzw, Settings);
|
||||
SampleValues10.yzw = FetchRowOfThreeAfterFour(Sample00TexelCenter, 0, Settings);
|
||||
SampleValues11.yzw = FetchRowOfThreeAfterFour(Sample00TexelCenter, 1, Settings);
|
||||
SampleValues12.yzw = FetchRowOfThreeAfterFour(Sample00TexelCenter, 2, Settings);
|
||||
SampleValues13.yzw = FetchRowOfThreeAfterFour(Sample00TexelCenter, 3, Settings);
|
||||
Results.y = PCF3x3(Fraction, SampleValues10, SampleValues11, SampleValues12, SampleValues13);
|
||||
}
|
||||
}
|
||||
@@ -444,25 +443,22 @@ float ManualPCF(float2 ShadowPosition, FPCFSamplerSettings Settings)
|
||||
float4 SampleValues16;
|
||||
|
||||
{
|
||||
float4 SampleValues04;
|
||||
FetchRowOfFour(Sample00TexelCenter, 4, SampleValues04, Settings);
|
||||
float4 SampleValues04 = FetchRowOfFour(Sample00TexelCenter, 4, Settings);
|
||||
SampleValues14.x = SampleValues04.w;
|
||||
|
||||
float4 SampleValues05;
|
||||
FetchRowOfFour(Sample00TexelCenter, 5, SampleValues05, Settings);
|
||||
float4 SampleValues05 = FetchRowOfFour(Sample00TexelCenter, 5, Settings);
|
||||
SampleValues15.x = SampleValues05.w;
|
||||
|
||||
float4 SampleValues06;
|
||||
FetchRowOfFour(Sample00TexelCenter, 6, SampleValues06, Settings);
|
||||
float4 SampleValues06 = FetchRowOfFour(Sample00TexelCenter, 6, Settings);
|
||||
SampleValues16.x = SampleValues06.w;
|
||||
|
||||
Results.z = PCF3x3(Fraction, SampleValues03, SampleValues04, SampleValues05, SampleValues06);
|
||||
}
|
||||
|
||||
{
|
||||
FetchRowOfThreeAfterFour(Sample00TexelCenter, 4, SampleValues14.yzw, Settings);
|
||||
FetchRowOfThreeAfterFour(Sample00TexelCenter, 5, SampleValues15.yzw, Settings);
|
||||
FetchRowOfThreeAfterFour(Sample00TexelCenter, 6, SampleValues16.yzw, Settings);
|
||||
SampleValues14.yzw = FetchRowOfThreeAfterFour(Sample00TexelCenter, 4, Settings);
|
||||
SampleValues15.yzw = FetchRowOfThreeAfterFour(Sample00TexelCenter, 5, Settings);
|
||||
SampleValues16.yzw = FetchRowOfThreeAfterFour(Sample00TexelCenter, 6, Settings);
|
||||
Results.w = PCF3x3(Fraction, SampleValues13, SampleValues14, SampleValues15, SampleValues16);
|
||||
}
|
||||
}
|
||||
|
||||
Reference in New Issue
Block a user