|
|
@@ -27,7 +27,7 @@
|
|
|
[[vk::binding(4)]] SamplerState g_linearAnyClampSampler;
|
|
|
|
|
|
# if ANKI_COMPUTE_SHADER
|
|
|
-[[vk::binding(5)]] RWTexture2D<Vec4> g_outUav;
|
|
|
+[[vk::binding(5)]] RWTexture2D<RVec4> g_bentNormalsAndSsaoUav;
|
|
|
# endif
|
|
|
|
|
|
[[vk::push_constant]] ConstantBuffer<SsaoConstants> g_consts;
|
|
|
@@ -52,7 +52,7 @@ RF32 computeFalloff(RF32 len)
|
|
|
# if ANKI_COMPUTE_SHADER
|
|
|
[numthreads(8, 8, 1)] void main(UVec2 svDispatchThreadId : SV_DISPATCHTHREADID)
|
|
|
# else
|
|
|
-Vec2 main([[vk::location(0)]] Vec2 uv : TEXCOORD, Vec4 svPosition : SV_POSITION) : SV_TARGET0
|
|
|
+RVec4 main([[vk::location(0)]] Vec2 uv : TEXCOORD, Vec4 svPosition : SV_POSITION) : SV_TARGET0
|
|
|
# endif
|
|
|
{
|
|
|
# if ANKI_COMPUTE_SHADER
|
|
|
@@ -148,12 +148,18 @@ Vec2 main([[vk::location(0)]] Vec2 uv : TEXCOORD, Vec4 svPosition : SV_POSITION)
|
|
|
// Apply power
|
|
|
Vd = pow(Vd, g_consts.m_ssaoPower);
|
|
|
|
|
|
- const F32 linearDepth = linearizeDepthOptimal(depth, g_consts.m_linearizeDepthParams.x, g_consts.m_linearizeDepthParams.y);
|
|
|
+ // Compute bent normal: see "Algorithm 2 Extension that computes bent normals b."
|
|
|
+ const RF32 t0 =
|
|
|
+ (6.0f * sin(h1 - n) - sin(3.0f * h1 - n) + 6.0f * sin(h2 - n) - sin(3.0f * h2 - n) + 16.0f * sin(n) - 3.0f * (sin(h1 + n) + sin(h2 + n)))
|
|
|
+ / 12.0f;
|
|
|
+ const RF32 t1 = (-cos(3.0f * h1 - n) - cos(3.0f * h2 - n) + 8.0f * cos(n) - 3.0f * (cos(h1 + n) + cos(h2 + n))) / 12.0f;
|
|
|
+ RVec3 bentNormal = RVec3(-dir2d.x * t0, -dir2d.y * t0, t1);
|
|
|
+ bentNormal = normalize(bentNormal);
|
|
|
|
|
|
# if ANKI_COMPUTE_SHADER
|
|
|
- g_outUav[svDispatchThreadId] = Vec4(Vd, linearDepth, 0.0f, 0.0f);
|
|
|
+ g_bentNormalsAndSsaoUav[svDispatchThreadId] = RVec4(bentNormal, Vd);
|
|
|
# else
|
|
|
- return Vec2(Vd, linearDepth);
|
|
|
+ return RVec4(bentNormal, Vd);
|
|
|
# endif
|
|
|
}
|
|
|
#endif // ANKI_TECHNIQUE_Ssao && (ANKI_COMPUTE_SHADER || ANKI_FRAGMENT_SHADER)
|
|
|
@@ -163,76 +169,91 @@ Vec2 main([[vk::location(0)]] Vec2 uv : TEXCOORD, Vec4 svPosition : SV_POSITION)
|
|
|
// ===========================================================================
|
|
|
#if ANKI_TECHNIQUE_SsaoSpatialDenoise && (ANKI_COMPUTE_SHADER || ANKI_FRAGMENT_SHADER)
|
|
|
# include <AnKi/Shaders/BilateralFilter.hlsl>
|
|
|
+# include <AnKi/Shaders/Include/MiscRendererTypes.h>
|
|
|
+# include <AnKi/Shaders/Functions.hlsl>
|
|
|
|
|
|
[[vk::binding(0)]] SamplerState g_linearAnyClampSampler;
|
|
|
-[[vk::binding(1)]] Texture2D<Vec4> g_inTex;
|
|
|
+[[vk::binding(1)]] Texture2D<RVec4> g_bentNormalsAndSsaoTex;
|
|
|
+[[vk::binding(2)]] Texture2D<Vec4> g_depthTex;
|
|
|
|
|
|
# if ANKI_COMPUTE_SHADER
|
|
|
-[[vk::binding(2)]] RWTexture2D<RVec4> g_outImg;
|
|
|
+[[vk::binding(3)]] RWTexture2D<RVec4> g_bentNormalsAndSsaoUav;
|
|
|
# endif
|
|
|
|
|
|
+[[vk::push_constant]] ConstantBuffer<SsaoSpatialDenoiseConstants> g_consts;
|
|
|
+
|
|
|
F32 computeWeight(F32 depth, F32 refDepth)
|
|
|
{
|
|
|
const F32 diff = abs(depth - refDepth);
|
|
|
return sqrt(1.0 / (0.0003 + diff));
|
|
|
}
|
|
|
|
|
|
-void sampleTex(Vec2 uv, IVec2 offset, F32 refDepth, inout RF32 col, inout F32 weight)
|
|
|
+void sampleTex(Vec2 uv, IVec2 offset, F32 refDepth, inout RF32 ssao, inout RVec3 bentNormal, inout F32 weight)
|
|
|
{
|
|
|
- const Vec2 ssaoWithDepth = g_inTex.SampleLevel(g_linearAnyClampSampler, uv, 0.0, offset).xy;
|
|
|
- const F32 w = computeWeight(refDepth, ssaoWithDepth.y);
|
|
|
- col += ssaoWithDepth.x * w;
|
|
|
+ const F32 linearDepth = linearizeDepthOptimal(g_depthTex.SampleLevel(g_linearAnyClampSampler, uv, 0.0, offset).x,
|
|
|
+ g_consts.m_linearizeDepthParams.x, g_consts.m_linearizeDepthParams.y);
|
|
|
+ const RVec4 bentNormalAndSsao = g_bentNormalsAndSsaoTex.SampleLevel(g_linearAnyClampSampler, uv, 0.0, offset);
|
|
|
+ const F32 w = computeWeight(refDepth, linearDepth);
|
|
|
+ ssao += bentNormalAndSsao.w * w;
|
|
|
+ bentNormal += bentNormalAndSsao.xyz * w;
|
|
|
weight += w;
|
|
|
}
|
|
|
|
|
|
# if ANKI_COMPUTE_SHADER
|
|
|
[numthreads(8, 8, 1)] void main(UVec2 svDispatchThreadId : SV_DISPATCHTHREADID)
|
|
|
# else
|
|
|
-RF32 main([[vk::location(0)]] Vec2 uv : TEXCOORD) : SV_TARGET0
|
|
|
+RVec4 main([[vk::location(0)]] Vec2 uv : TEXCOORD) : SV_TARGET0
|
|
|
# endif
|
|
|
{
|
|
|
// Set UVs
|
|
|
# if ANKI_COMPUTE_SHADER
|
|
|
Vec2 textureSize;
|
|
|
U32 mipCount;
|
|
|
- g_inTex.GetDimensions(0, textureSize.x, textureSize.y, mipCount);
|
|
|
+ g_bentNormalsAndSsaoTex.GetDimensions(0, textureSize.x, textureSize.y, mipCount);
|
|
|
const Vec2 uv = (Vec2(svDispatchThreadId) + 0.5f) / textureSize;
|
|
|
# endif
|
|
|
|
|
|
// Sample ref
|
|
|
- const Vec2 refSsaoAndDepth = g_inTex.SampleLevel(g_linearAnyClampSampler, uv, 0.0).xy;
|
|
|
- RF32 color = refSsaoAndDepth.x;
|
|
|
- const F32 refDepth = refSsaoAndDepth.y;
|
|
|
+ const RVec4 refBentNormalAndSsao = g_bentNormalsAndSsaoTex.SampleLevel(g_linearAnyClampSampler, uv, 0.0);
|
|
|
+ RF32 ssao = refBentNormalAndSsao.w;
|
|
|
+ RVec3 bentNormal = refBentNormalAndSsao.xyz;
|
|
|
+ const F32 refDepth = linearizeDepthOptimal(g_depthTex.SampleLevel(g_linearAnyClampSampler, uv, 0.0).x, g_consts.m_linearizeDepthParams.x,
|
|
|
+ g_consts.m_linearizeDepthParams.y);
|
|
|
F32 weight = computeWeight(0.0f, 0.0f); // Highest weight that this function can give
|
|
|
|
|
|
// Sample taps
|
|
|
- sampleTex(uv, IVec2(1, 1), refDepth, color, weight);
|
|
|
- sampleTex(uv, IVec2(0, 1), refDepth, color, weight);
|
|
|
- sampleTex(uv, IVec2(-1, 1), refDepth, color, weight);
|
|
|
- sampleTex(uv, IVec2(-1, 0), refDepth, color, weight);
|
|
|
- sampleTex(uv, IVec2(-1, -1), refDepth, color, weight);
|
|
|
- sampleTex(uv, IVec2(0, -1), refDepth, color, weight);
|
|
|
- sampleTex(uv, IVec2(1, -1), refDepth, color, weight);
|
|
|
- sampleTex(uv, IVec2(1, 0), refDepth, color, weight);
|
|
|
+ sampleTex(uv, IVec2(1, 1), refDepth, ssao, bentNormal, weight);
|
|
|
+ sampleTex(uv, IVec2(0, 1), refDepth, ssao, bentNormal, weight);
|
|
|
+ sampleTex(uv, IVec2(-1, 1), refDepth, ssao, bentNormal, weight);
|
|
|
+ sampleTex(uv, IVec2(-1, 0), refDepth, ssao, bentNormal, weight);
|
|
|
+ sampleTex(uv, IVec2(-1, -1), refDepth, ssao, bentNormal, weight);
|
|
|
+ sampleTex(uv, IVec2(0, -1), refDepth, ssao, bentNormal, weight);
|
|
|
+ sampleTex(uv, IVec2(1, -1), refDepth, ssao, bentNormal, weight);
|
|
|
+ sampleTex(uv, IVec2(1, 0), refDepth, ssao, bentNormal, weight);
|
|
|
|
|
|
# if SPATIAL_DENOISE_QUALITY == 1
|
|
|
- sampleTex(uv, IVec2(2, 2), refDepth, color, weight);
|
|
|
- sampleTex(uv, IVec2(0, 2), refDepth, color, weight);
|
|
|
- sampleTex(uv, IVec2(-2, 2), refDepth, color, weight);
|
|
|
- sampleTex(uv, IVec2(-2, 0), refDepth, color, weight);
|
|
|
- sampleTex(uv, IVec2(-2, -2), refDepth, color, weight);
|
|
|
- sampleTex(uv, IVec2(0, -2), refDepth, color, weight);
|
|
|
- sampleTex(uv, IVec2(2, -2), refDepth, color, weight);
|
|
|
- sampleTex(uv, IVec2(2, 0), refDepth, color, weight);
|
|
|
+ sampleTex(uv, IVec2(2, 2), refDepth, ssao, bentNormal, weight);
|
|
|
+ sampleTex(uv, IVec2(0, 2), refDepth, ssao, bentNormal, weight);
|
|
|
+ sampleTex(uv, IVec2(-2, 2), refDepth, ssao, bentNormal, weight);
|
|
|
+ sampleTex(uv, IVec2(-2, 0), refDepth, ssao, bentNormal, weight);
|
|
|
+ sampleTex(uv, IVec2(-2, -2), refDepth, ssao, bentNormal, weight);
|
|
|
+ sampleTex(uv, IVec2(0, -2), refDepth, ssao, bentNormal, weight);
|
|
|
+ sampleTex(uv, IVec2(2, -2), refDepth, ssao, bentNormal, weight);
|
|
|
+ sampleTex(uv, IVec2(2, 0), refDepth, ssao, bentNormal, weight);
|
|
|
# endif
|
|
|
|
|
|
- color /= weight;
|
|
|
+ ssao /= weight;
|
|
|
+ ssao = saturate(ssao);
|
|
|
+
|
|
|
+ bentNormal /= weight;
|
|
|
+ bentNormal = normalize(bentNormal);
|
|
|
+ bentNormal = mul(g_consts.m_viewToWorldMat, Vec4(bentNormal, 0.0f));
|
|
|
|
|
|
// Write value
|
|
|
# if ANKI_COMPUTE_SHADER
|
|
|
- g_outImg[svDispatchThreadId] = color;
|
|
|
+ g_bentNormalsAndSsaoUav[svDispatchThreadId] = RVec4(bentNormal, ssao);
|
|
|
# else
|
|
|
- return color;
|
|
|
+ return RVec4(bentNormal, ssao);
|
|
|
# endif
|
|
|
}
|
|
|
#endif
|
|
|
@@ -242,54 +263,55 @@ RF32 main([[vk::location(0)]] Vec2 uv : TEXCOORD) : SV_TARGET0
|
|
|
// ===========================================================================
|
|
|
#if ANKI_TECHNIQUE_SsaoTemporalDenoise && (ANKI_COMPUTE_SHADER || ANKI_FRAGMENT_SHADER)
|
|
|
# include <AnKi/Shaders/Functions.hlsl>
|
|
|
+# include <AnKi/Shaders/TonemappingFunctions.hlsl>
|
|
|
|
|
|
[[vk::binding(0)]] SamplerState g_linearAnyClampSampler;
|
|
|
-[[vk::binding(1)]] Texture2D<RVec4> g_ssaoTex;
|
|
|
-[[vk::binding(2)]] Texture2D<RVec4> g_historySsaoTex;
|
|
|
+[[vk::binding(1)]] Texture2D<RVec4> g_bentNormalsAndSsaoTex;
|
|
|
+[[vk::binding(2)]] Texture2D<RVec4> g_historyBentNormalsAndSsaoTex;
|
|
|
[[vk::binding(3)]] Texture2D<Vec4> g_motionVectorsTex;
|
|
|
|
|
|
# if ANKI_COMPUTE_SHADER
|
|
|
-[[vk::binding(4)]] RWTexture2D<RVec4> g_outImg;
|
|
|
+[[vk::binding(4)]] RWTexture2D<RVec4> g_bentNormalsAndSsaoUav;
|
|
|
# endif
|
|
|
|
|
|
# if ANKI_COMPUTE_SHADER
|
|
|
[numthreads(8, 8, 1)] void main(UVec2 svDispatchThreadId : SV_DISPATCHTHREADID)
|
|
|
# else
|
|
|
-RF32 main([[vk::location(0)]] Vec2 uv : TEXCOORD) : SV_TARGET0
|
|
|
+RVec4 main([[vk::location(0)]] Vec2 uv : TEXCOORD) : SV_TARGET0
|
|
|
# endif
|
|
|
{
|
|
|
# if ANKI_COMPUTE_SHADER
|
|
|
Vec2 textureSize;
|
|
|
U32 mipCount;
|
|
|
- g_ssaoTex.GetDimensions(0, textureSize.x, textureSize.y, mipCount);
|
|
|
+ g_bentNormalsAndSsaoTex.GetDimensions(0, textureSize.x, textureSize.y, mipCount);
|
|
|
const Vec2 uv = (Vec2(svDispatchThreadId) + 0.5f) / textureSize;
|
|
|
# endif
|
|
|
|
|
|
const Vec2 historyUv = uv + g_motionVectorsTex.SampleLevel(g_linearAnyClampSampler, uv, 0.0f).xy;
|
|
|
|
|
|
// Read textures
|
|
|
- RF32 history = g_historySsaoTex.SampleLevel(g_linearAnyClampSampler, historyUv, 0.0f).x;
|
|
|
- RF32 current = g_ssaoTex.SampleLevel(g_linearAnyClampSampler, uv, 0.0f).x;
|
|
|
+ RVec4 history = g_historyBentNormalsAndSsaoTex.SampleLevel(g_linearAnyClampSampler, historyUv, 0.0f);
|
|
|
+ RVec4 current = g_bentNormalsAndSsaoTex.SampleLevel(g_linearAnyClampSampler, uv, 0.0f);
|
|
|
|
|
|
// Remove ghosting by clamping the history color to neighbour's AABB
|
|
|
- const RF32 near0 = g_ssaoTex.SampleLevel(g_linearAnyClampSampler, uv, 0, IVec2(1, 0)).x;
|
|
|
- const RF32 near1 = g_ssaoTex.SampleLevel(g_linearAnyClampSampler, uv, 0, IVec2(0, 1)).x;
|
|
|
- const RF32 near2 = g_ssaoTex.SampleLevel(g_linearAnyClampSampler, uv, 0, IVec2(-1, 0)).x;
|
|
|
- const RF32 near3 = g_ssaoTex.SampleLevel(g_linearAnyClampSampler, uv, 0, IVec2(0, -1)).x;
|
|
|
+ const RVec4 near0 = g_bentNormalsAndSsaoTex.SampleLevel(g_linearAnyClampSampler, uv, 0, IVec2(1, 0));
|
|
|
+ const RVec4 near1 = g_bentNormalsAndSsaoTex.SampleLevel(g_linearAnyClampSampler, uv, 0, IVec2(0, 1));
|
|
|
+ const RVec4 near2 = g_bentNormalsAndSsaoTex.SampleLevel(g_linearAnyClampSampler, uv, 0, IVec2(-1, 0));
|
|
|
+ const RVec4 near3 = g_bentNormalsAndSsaoTex.SampleLevel(g_linearAnyClampSampler, uv, 0, IVec2(0, -1));
|
|
|
|
|
|
# if 0
|
|
|
- const RF32 boxMin = min(current, min4(near0, near1, near2, near3));
|
|
|
- const RF32 boxMax = max(current, max4(near0, near1, near2, near3));
|
|
|
+ const RVec4 boxMin = min(current, min4(near0, near1, near2, near3));
|
|
|
+ const RVec4 boxMax = max(current, max4(near0, near1, near2, near3));
|
|
|
# else
|
|
|
- const RF32 m1 = current + near0 + near1 + near2 + near3;
|
|
|
- const RF32 m2 = current * current + near0 * near0 + near1 * near1 + near2 * near2 + near3 * near3;
|
|
|
+ const RVec4 m1 = current + near0 + near1 + near2 + near3;
|
|
|
+ const RVec4 m2 = current * current + near0 * near0 + near1 * near1 + near2 * near2 + near3 * near3;
|
|
|
|
|
|
- const RF32 mu = m1 / 5.0;
|
|
|
- const RF32 sigma = sqrt(m2 / 5.0 - mu * mu);
|
|
|
+ const RVec4 mu = m1 / 5.0;
|
|
|
+ const RVec4 sigma = sqrt(m2 / 5.0 - mu * mu);
|
|
|
|
|
|
const F32 varianceClippingGamma = 1.2f;
|
|
|
- const RF32 boxMin = mu - varianceClippingGamma * sigma;
|
|
|
- const RF32 boxMax = mu + varianceClippingGamma * sigma;
|
|
|
+ const RVec4 boxMin = mu - varianceClippingGamma * sigma;
|
|
|
+ const RVec4 boxMax = mu + varianceClippingGamma * sigma;
|
|
|
# endif
|
|
|
|
|
|
history = clamp(history, boxMin, boxMax);
|
|
|
@@ -297,8 +319,8 @@ RF32 main([[vk::location(0)]] Vec2 uv : TEXCOORD) : SV_TARGET0
|
|
|
// Final
|
|
|
const RF32 kBlendFactor = 0.1f;
|
|
|
|
|
|
- const F32 lum0 = current;
|
|
|
- const F32 lum1 = history;
|
|
|
+ const F32 lum0 = computeLuminance(current.xyz) * current.w;
|
|
|
+ const F32 lum1 = computeLuminance(history.xyz) * history.w;
|
|
|
const F32 maxLum = 1.0;
|
|
|
|
|
|
RF32 diff = abs(lum0 - lum1) / max(lum0, max(lum1, maxLum + kEpsilonF32));
|
|
|
@@ -306,13 +328,14 @@ RF32 main([[vk::location(0)]] Vec2 uv : TEXCOORD) : SV_TARGET0
|
|
|
diff = diff * diff;
|
|
|
const RF32 feedback = lerp(0.0, kBlendFactor, diff);
|
|
|
|
|
|
- const RF32 ssao = lerp(history, current, feedback);
|
|
|
+ RVec4 finalVal = lerp(history, current, feedback);
|
|
|
+ finalVal.xyz = normalize(finalVal.xyz);
|
|
|
|
|
|
// Write value
|
|
|
# if ANKI_COMPUTE_SHADER
|
|
|
- g_outImg[svDispatchThreadId] = ssao;
|
|
|
+ g_bentNormalsAndSsaoUav[svDispatchThreadId] = finalVal;
|
|
|
# else
|
|
|
- return ssao;
|
|
|
+ return finalVal;
|
|
|
# endif
|
|
|
}
|
|
|
#endif
|
Panagiotis Christopoulos Charitos