Better shadows: soft shadows using percentage-closer filtering (PCF), tighter shadow frustum (from scene AABBs)

CHANGELOG.md

@@ -15,6 +15,9 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 - core/DepthAndShadowPass.h : handle multiple shadow maps using a texture atlas (https://github.com/Simple-Robotics/candlewick/pull/69)
 - multibody : handle two-light setup w/ shadow mapping (https://github.com/Simple-Robotics/candlewick/pull/69)
 - multibody/Visualizer : use `H` key to toggle GUI (https://github.com/Simple-Robotics/candlewick/pull/70)
+- shaders : Soft shadows in PBR using percentage-closer filtering (PCF) (https://github.com/Simple-Robotics/candlewick/pull/71)
+- core : Tighter shadow frustum around the world-scene AABB (https://github.com/Simple-Robotics/candlewick/pull/71)
+- core : add `getAABBCorners()` util function in `Collision.h` header (https://github.com/Simple-Robotics/candlewick/pull/71)
 
 ## [0.1.0] - 2025-05-21
 

shaders/compiled/PbrBasic.frag.msl

@@ -55,7 +55,6 @@ struct PbrMaterial
 struct ShadowAtlasInfo
 {
     int4 lightRegions[2];
-    uint2 atlasSize;
 };
 
 struct LightBlock
@@ -170,57 +169,57 @@ float3 calculatePbrLighting(thread const float3& normal, thread const float3& V,
 static inline __attribute__((always_inline))
 bool isCoordsInRange(thread const float3& uv)
 {
-    bool _69 = uv.x >= 0.0;
-    bool _76;
-    if (_69)
+    bool _72 = uv.x >= 0.0;
+    bool _79;
+    if (_72)
     {
-        _76 = uv.y >= 0.0;
+        _79 = uv.y >= 0.0;
     }
     else
     {
-        _76 = _69;
+        _79 = _72;
     }
-    bool _83;
-    if (_76)
+    bool _86;
+    if (_79)
     {
-        _83 = uv.x <= 1.0;
+        _86 = uv.x <= 1.0;
     }
     else
     {
-        _83 = _76;
+        _86 = _79;
     }
-    bool _89;
-    if (_83)
+    bool _92;
+    if (_86)
     {
-        _89 = uv.y <= 1.0;
+        _92 = uv.y <= 1.0;
     }
     else
     {
-        _89 = _83;
+        _92 = _86;
     }
-    bool _96;
-    if (_89)
+    bool _99;
+    if (_92)
     {
-        _96 = uv.z >= 0.0;
+        _99 = uv.z >= 0.0;
     }
     else
     {
-        _96 = _89;
+        _99 = _92;
     }
-    bool _102;
-    if (_96)
+    bool _105;
+    if (_99)
     {
-        _102 = uv.z <= 1.0;
+        _105 = uv.z <= 1.0;
     }
     else
     {
-        _102 = _96;
+        _105 = _99;
     }
-    return _102;
+    return _105;
 }
 
 static inline __attribute__((always_inline))
-float calcShadowmap(thread const int& lightIndex, thread const float& NdotL, thread spvUnsafeArray<float3, 2>& fragLightPos, constant ShadowAtlasInfo& _420, depth2d<float> shadowMap, sampler shadowMapSmplr)
+float calcShadowmap(thread const int& lightIndex, thread const float& NdotL, thread const int2& atlasSize, thread spvUnsafeArray<float3, 2>& fragLightPos, constant ShadowAtlasInfo& _422, depth2d<float> shadowMap, sampler shadowMapSmplr)
 {
     float bias0 = 0.004999999888241291046142578125;
     float3 lightSpacePos = fragLightPos[lightIndex];
@@ -233,11 +232,24 @@ float calcShadowmap(thread const int& lightIndex, thread const float& NdotL, thr
     {
         return 1.0;
     }
-    int4 region = _420.lightRegions[lightIndex];
+    int4 region = _422.lightRegions[lightIndex];
     uv = float2(region.xy) + (uv * float2(region.zw));
-    uv /= float2(_420.atlasSize);
-    float3 texCoords = float3(uv, depthRef);
-    return shadowMap.sample_compare(shadowMapSmplr, texCoords.xy, texCoords.z);
+    uv /= float2(atlasSize);
+    float2 regionMin = float2(region.xy) / float2(atlasSize);
+    float2 regionMax = float2(region.xy + region.zw) / float2(atlasSize);
+    float value = 0.0;
+    float2 offsets = float2(1.0) / float2(atlasSize);
+    for (int i = -1; i <= 1; i++)
+    {
+        for (int j = -1; j <= 1; j++)
+        {
+            float2 offUV = uv + (float2(float(i), float(j)) * offsets);
+            offUV = fast::clamp(offUV, regionMin, regionMax);
+            float3 texCoords = float3(offUV, depthRef);
+            value += (0.111111111938953399658203125 * shadowMap.sample_compare(shadowMapSmplr, texCoords.xy, texCoords.z));
+        }
+    }
+    return value;
 }
 
 static inline __attribute__((always_inline))
@@ -264,14 +276,15 @@ float3 uncharted2ToneMapping(thread const float3& color)
     return curr * white_scale;
 }
 
-fragment main0_out main0(main0_in in [[stage_in]], constant Material& _505 [[buffer(0)]], constant LightBlock& light [[buffer(1)]], constant EffectParams& params [[buffer(2)]], constant ShadowAtlasInfo& _420 [[buffer(3)]], depth2d<float> shadowMap [[texture(0)]], texture2d<float> ssaoTex [[texture(1)]], sampler shadowMapSmplr [[sampler(0)]], sampler ssaoTexSmplr [[sampler(1)]], bool gl_FrontFacing [[front_facing]], float4 gl_FragCoord [[position]])
+fragment main0_out main0(main0_in in [[stage_in]], constant Material& _571 [[buffer(0)]], constant LightBlock& light [[buffer(1)]], constant EffectParams& params [[buffer(2)]], constant ShadowAtlasInfo& _422 [[buffer(3)]], depth2d<float> shadowMap [[texture(0)]], texture2d<float> ssaoTex [[texture(1)]], sampler shadowMapSmplr [[sampler(0)]], sampler ssaoTexSmplr [[sampler(1)]], bool gl_FrontFacing [[front_facing]], float4 gl_FragCoord [[position]])
 {
     main0_out out = {};
     spvUnsafeArray<float3, 2> fragLightPos = {};
     fragLightPos[0] = in.fragLightPos_0;
     fragLightPos[1] = in.fragLightPos_1;
     float3 normal = fast::normalize(in.fragViewNormal);
     float3 V = fast::normalize(-in.fragViewPos);
+    int2 atlasSize = int2(shadowMap.get_width(), shadowMap.get_height());
     if (!gl_FrontFacing)
     {
         normal = -normal;
@@ -283,34 +296,35 @@ fragment main0_out main0(main0_in in [[stage_in]], constant Material& _505 [[buf
         float3 param = normal;
         float3 param_1 = V;
         float3 param_2 = lightDir;
-        PbrMaterial _518;
-        _518.baseColor = _505.material.baseColor;
-        _518.metalness = _505.material.metalness;
-        _518.roughness = _505.material.roughness;
-        _518.ao = _505.material.ao;
-        PbrMaterial param_3 = _518;
+        PbrMaterial _584;
+        _584.baseColor = _571.material.baseColor;
+        _584.metalness = _571.material.metalness;
+        _584.roughness = _571.material.roughness;
+        _584.ao = _571.material.ao;
+        PbrMaterial param_3 = _584;
         float3 param_4 = light.color[i];
         float param_5 = light.intensity[i].x;
         float3 _lo = calculatePbrLighting(param, param_1, param_2, param_3, param_4, param_5);
         float NdotL = fast::max(dot(normal, lightDir), 0.0);
         int param_6 = i;
         float param_7 = NdotL;
-        float shadowValue = calcShadowmap(param_6, param_7, fragLightPos, _420, shadowMap, shadowMapSmplr);
+        int2 param_8 = atlasSize;
+        float shadowValue = calcShadowmap(param_6, param_7, param_8, fragLightPos, _422, shadowMap, shadowMapSmplr);
         _lo *= shadowValue;
         Lo += _lo;
     }
-    float3 ambient = (float3(0.100000001490116119384765625) * _505.material.baseColor.xyz) * _505.material.ao;
+    float3 ambient = (float3(0.100000001490116119384765625) * _571.material.baseColor.xyz) * _571.material.ao;
     if (params.useSsao == 1u)
     {
         float2 ssaoTexSize = float2(int2(ssaoTex.get_width(), ssaoTex.get_height()));
         float2 ssaoUV = gl_FragCoord.xy / ssaoTexSize;
         ambient *= ssaoTex.sample(ssaoTexSmplr, ssaoUV).x;
     }
     float3 color = ambient + Lo;
-    float3 param_8 = color;
-    color = uncharted2ToneMapping(param_8);
+    float3 param_9 = color;
+    color = uncharted2ToneMapping(param_9);
     color = powr(color, float3(0.4545454680919647216796875));
-    out.fragColor = float4(color, _505.material.baseColor.w);
+    out.fragColor = float4(color, _571.material.baseColor.w);
     out.outNormal = in.fragViewNormal.xy;
     return out;
 }

shaders/src/PbrBasic.frag

@@ -32,7 +32,6 @@ layout(set=3, binding=2) uniform EffectParams {
 
 layout(set = 3, binding = 3) uniform ShadowAtlasInfo {
     ivec4 lightRegions[NUM_LIGHTS];
-    uvec2 atlasSize;
 };
 
 #ifdef HAS_SHADOW_MAPS
@@ -49,7 +48,7 @@ layout(location=0) out vec4 fragColor;
 #endif
 
 #ifdef HAS_SHADOW_MAPS
-float calcShadowmap(int lightIndex, float NdotL) {
+float calcShadowmap(int lightIndex, float NdotL, ivec2 atlasSize) {
     // float bias = max(0.05 * (1.0 - NdotL), 0.005);
     float bias = 0.005;
     vec3 lightSpacePos = fragLightPos[lightIndex];
@@ -62,17 +61,35 @@ float calcShadowmap(int lightIndex, float NdotL) {
     }
 
     ivec4 region = lightRegions[lightIndex];
+    vec2 regionMin = vec2(region.xy) / atlasSize;
+    vec2 regionMax = vec2(region.xy + region.zw) / atlasSize;
     uv = region.xy + uv * region.zw;
     uv = uv / atlasSize;
 
-    vec3 texCoords = vec3(uv, depthRef);
-    return texture(shadowMap, texCoords);
+    float value = 0.0;
+    const vec2 offsets = 1.0 / atlasSize;
+    // pcf loop
+    const int halfKernel = 1;
+    const float weight = 1.0 / pow(2 * halfKernel + 1, 2);
+    for (int i = -halfKernel; i <= halfKernel; i++) {
+        for (int j = -halfKernel; j <= halfKernel; j++) {
+            vec2 offUV = uv + vec2(i, j) * offsets;
+            offUV = clamp(offUV, regionMin, regionMax);
+            vec3 texCoords = vec3(offUV, depthRef);
+            value += weight * texture(shadowMap, texCoords);
+        }
+    }
+
+    return value;
 }
 #endif
 
 void main() {
     vec3 normal = normalize(fragViewNormal);
-    vec3 V = normalize(-fragViewPos);
+    const vec3 V = normalize(-fragViewPos);
+#ifdef HAS_SHADOW_MAPS
+    const ivec2 atlasSize = textureSize(shadowMap, 0).xy;
+#endif
 
     if (!gl_FrontFacing) {
         // Flip normal for back faces
@@ -91,8 +108,8 @@ void main() {
             light.intensity[i]
         );
 #ifdef HAS_SHADOW_MAPS
-        float NdotL = max(dot(normal, lightDir), 0.0);
-        float shadowValue = calcShadowmap(i, NdotL);
+        const float NdotL = max(dot(normal, lightDir), 0.0);
+        const float shadowValue = calcShadowmap(i, NdotL, atlasSize);
         _lo *= shadowValue;
 #endif
         Lo += _lo;

shaders/src/tone_mapping.glsl

@@ -5,10 +5,9 @@ vec3 reinhardToneMapping(vec3 color) {
     return color / (color + vec3(1.0));
 }
 
-// Exposure + Gamma (Simplified)
 vec3 exposureToneMapping(vec3 color, float exposure) {
-    // Exposure adjustment followed by gamma correction
-    return pow(1.0 - exp(-color * exposure), vec3(1.0/2.2));
+    // Exposure adjustment
+    return 1.0 - exp(-color * exposure);
 }
 
 // ACES Filmic Tone Mapping (Academy Color Encoding System)

src/candlewick/core/Collision.h

@@ -24,4 +24,20 @@ inline Mat4f toTransformationMatrix(const OBB &obb) {
   return T;
 }
 
+inline std::array<Float3, 8> getAABBCorners(const AABB &aabb) {
+  const Float3 min = aabb.min_.cast<float>();
+  const Float3 max = aabb.max_.cast<float>();
+
+  return {
+      Float3{min.x(), min.y(), min.z()}, // 000
+      Float3{max.x(), min.y(), min.z()}, // 100
+      Float3{min.x(), max.y(), min.z()}, // 010
+      Float3{max.x(), max.y(), min.z()}, // 110
+      Float3{min.x(), min.y(), max.z()}, // 001
+      Float3{max.x(), min.y(), max.z()}, // 101
+      Float3{min.x(), max.y(), max.z()}, // 011
+      Float3{max.x(), max.y(), max.z()}, // 111
+  };
+}
+
 } // namespace candlewick

src/candlewick/core/DepthAndShadowPass.cpp

@@ -273,23 +273,32 @@ void renderShadowPassFromFrustum(CommandBuffer &cmdBuf, ShadowMapPass &passInfo,
 void renderShadowPassFromAABB(CommandBuffer &cmdBuf, ShadowMapPass &passInfo,
                               std::span<const DirectionalLight> dirLight,
                               std::span<const OpaqueCastable> castables,
-                              const AABB &worldSceneBounds) {
+                              const AABB &worldAABB) {
   using Eigen::Vector3d;
 
-  Float3 center = worldSceneBounds.center().cast<float>();
-  float radius = 0.5f * float(worldSceneBounds.size());
-  radius = std::ceil(radius * 16.f) / 16.f;
+  Float3 center = worldAABB.center().cast<float>();
 
   for (size_t i = 0; i < passInfo.numLights(); i++) {
-    const Float3 eye = center - radius * dirLight[i].direction.normalized();
+    Float3 tmpEye = center - 100.0f * dirLight[i].direction.normalized();
+    Mat4f tmplightView = lookAt(tmpEye, center, Float3::UnitZ());
+
     auto &lightView = passInfo.cam[i].view;
     auto &lightProj = passInfo.cam[i].projection;
-    lightView = lookAt(eye, center, Float3::UnitZ());
 
-    AABB bounds{Vector3d::Constant(-radius), Vector3d::Constant(radius)};
+    Mat3f R = tmplightView.topLeftCorner<3, 3>();
+    Float3 t = tmplightView.topRightCorner<3, 1>();
+    AABB bounds = coal::translate(coal::rotate(worldAABB, R.cast<double>()),
+                                  t.cast<double>());
+
+    Float3 lightSpaceCenter = bounds.center().cast<float>();
+    float radius = float(bounds.max_.z());
+
+    Float3 finalEye = center - radius * dirLight[i].direction.normalized();
+
+    lightView = lookAt(finalEye, center, Float3::UnitZ());
     lightProj = shadowOrthographicMatrix({bounds.width(), bounds.height()},
-                                         float(bounds.min_.z()),
-                                         float(bounds.max_.z()));
+                                         float(bounds.max_.z()),
+                                         float(bounds.min_.z()));
   }
   passInfo.render(cmdBuf, castables);
 }

src/candlewick/multibody/RobotScene.cpp

@@ -33,7 +33,6 @@ struct alignas(16) LightSpaceMatricesUbo {
 struct alignas(16) ShadowAtlasInfoUbo {
   using Vec4u = Eigen::Matrix<Uint32, 4, 1, Eigen::DontAlign>;
   std::array<Vec4u, kNumLights> regions;
-  std::array<Uint32, 2> atlasSize;
 };
 
 void updateRobotTransforms(entt::registry &registry,
@@ -455,7 +454,6 @@ void RobotScene::renderPBRTriangleGeometry(CommandBuffer &command_buffer,
   const bool enable_shadows = m_config.enable_shadows;
   ShadowAtlasInfoUbo shadowAtlasUbo{
       .regions{},
-      .atlasSize{shadowPass.atlasDims()},
   };
   Mat4f lightViewProj[kNumLights];
   for (size_t i = 0; i < numLights; i++) {