Much better SSAO

supertuxkart · Jan 16, 2025 · 6db816e · 6db816e
1 parent f8c6c66
commit 6db816e
Show file tree

Hide file tree

Showing 11 changed files with 130 additions and 104 deletions.
diff --git a/data/shaders/IBL.frag b/data/shaders/IBL.frag
@@ -1,6 +1,7 @@
 uniform sampler2D ntex;
 uniform sampler2D dtex;
 uniform sampler2D albedo;
+uniform sampler2D ssao;
 
 #ifdef GL_ES
 layout (location = 0) out vec4 Diff;
@@ -76,8 +77,9 @@ void main(void)
 {
     vec2 uv = gl_FragCoord.xy / u_screen;
     vec3 normal = DecodeNormal(texture(ntex, uv).xy);
+    float ao = texture(ssao, uv).x;
 
-    Diff = vec4(0.25 * DiffuseIBL(normal), 1.);
+    Diff = vec4(0.25 * DiffuseIBL(normal) * ao, 1.);
 
     float z = texture(dtex, uv).x;
 
@@ -86,8 +88,11 @@ void main(void)
     // Extract roughness
     float specval = texture(ntex, uv).z;
 
+    // Lagarde and de Rousiers 2014, "Moving Frostbite to PBR"
+    float ao_spec = clamp(pow(max(dot(normal, eyedir), 0.) + ao, exp2(-16.0 * (1.0 - specval) - 1.0)) - 1.0 + ao);
+
 #ifdef GL_ES
-    Spec = vec4(.25 * SpecularIBL(normal, eyedir, specval), 1.);
+    Spec = vec4(.25 * SpecularIBL(normal, eyedir, specval) * ao_spec, 1.);
 #else
     // :::::::: Compute Space Screen Reflection ::::::::::::::::::::::::::::::::::::
 
@@ -123,7 +128,7 @@ void main(void)
         outColor = fallback;
     }
 
-    Spec = vec4(outColor.rgb, 1.0);
+    Spec = vec4(outColor.rgb * ao_spec, 1.0);
 #endif
 
 }
diff --git a/data/shaders/bilateralH.frag b/data/shaders/bilateralH.frag
@@ -8,7 +8,7 @@ out vec4 FragColor;
 
 void main()
 {
-    float sigma = 5.;
+    float sigma = 2.;
 
     vec2 uv = gl_FragCoord.xy * pixel;
     float X = uv.x;
@@ -23,7 +23,7 @@ void main()
     g0 *= g1;
     g1 *= g2;
     float tmp_weight, total_weight = g0;
-    for (int i = 1; i < 9; i++) {
+    for (int i = 1; i < 3; i++) {
         tmp_weight = max(0.0, 1.0 - .001 * abs(texture(depth, vec2(X - float(i) * pixel.x, Y)).x - pixel_depth));
         sum += texture(tex, vec2(X - float(i) * pixel.x, Y)) * g0 * tmp_weight;
         total_weight += g0 * tmp_weight;

diff --git a/data/shaders/bilateralV.frag b/data/shaders/bilateralV.frag
@@ -8,7 +8,7 @@ out vec4 FragColor;
 
 void main()
 {
-    float sigma = 5.;
+    float sigma = 2.;
 
     vec2 uv = gl_FragCoord.xy * pixel;
     float X = uv.x;
@@ -23,7 +23,7 @@ void main()
     g0 *= g1;
     g1 *= g2;
     float tmp_weight, total_weight = g0;
-    for (int i = 1; i < 9; i++) {
+    for (int i = 1; i < 3; i++) {
         tmp_weight = max(0.0, 1.0 - .001 * abs(texture(depth, vec2(X, Y - float(i) * pixel.y)).x - pixel_depth));
         sum += texture(tex, vec2(X, Y - float(i) * pixel.y)) * g0 * tmp_weight;
         total_weight += g0 * tmp_weight;

diff --git a/data/shaders/combine_diffuse_color.frag b/data/shaders/combine_diffuse_color.frag
@@ -1,6 +1,5 @@
 uniform sampler2D diffuse_map;
 uniform sampler2D specular_map;
-uniform sampler2D ssao_tex;
 uniform sampler2D normal_color;
 uniform sampler2D diffuse_color;
 #if defined(GL_ES) && defined(GL_FRAGMENT_PRECISION_HIGH)
@@ -28,7 +27,6 @@ void main()
     float specMapValue = texture(normal_color, tc).z;
     float emitMapValue = diffuseMatColor.w;
 
-    float ao = texture(ssao_tex, tc).x;
     vec3 DiffuseComponent = texture(diffuse_map, tc).xyz;
     vec3 SpecularComponent = texture(specular_map, tc).xyz;
 
@@ -42,7 +40,7 @@ void main()
     vec3 tmp = DiffuseComponent * mix(diffuseMatColor.xyz, vec3(0.0), metallicMapValue) + (metallicMatColor * SpecularComponent);
 
     vec3 emitCol = diffuseMatColor.xyz + (diffuseMatColor.xyz * diffuseMatColor.xyz * emitMapValue * emitMapValue * 10.0);
-    vec4 color_1 = vec4(tmp * ao + (emitMapValue * emitCol), 1.0);
+    vec4 color_1 = vec4(tmp + (emitMapValue * emitCol), 1.0);
 
     // Fog
     float depth = texture(depth_stencil, tc).x;

diff --git a/data/shaders/degraded_ibl.frag b/data/shaders/degraded_ibl.frag
@@ -1,4 +1,5 @@
 uniform sampler2D ntex;
+uniform sampler2D ssao;
 
 #ifdef GL_ES
 layout (location = 0) out vec4 Diff;
@@ -17,7 +18,9 @@ void main(void)
 {
     vec2 uv = gl_FragCoord.xy / u_screen;
     vec3 normal = DecodeNormal(texture(ntex, uv).xy);
+    vec3 spec_color = vec3(0.031, 0.106, 0.173);
+    float ao = texture(ssao, uv).x;
 
-    Diff = vec4(0.25 * DiffuseIBL(normal), 1.);
-    Spec = vec4(0.031, 0.106, 0.173, 1.);
+    Diff = vec4(0.25 * DiffuseIBL(normal) * ao, 1.);
+    Spec = vec4(spec_color * ao, 1.);
 }
diff --git a/data/shaders/ssao.frag b/data/shaders/ssao.frag
@@ -1,18 +1,18 @@
 // From paper http://graphics.cs.williams.edu/papers/AlchemyHPG11/
 // and improvements here http://graphics.cs.williams.edu/papers/SAOHPG12/
+// and implementations here https://github.com/google/filament/blob/026b985c07b7eec4f678e0e5130d0a4e742e9c61/filament/src/materials/ssao/saoImpl.fs
 
 uniform sampler2D dtex;
 uniform float radius;
 uniform float k;
 uniform float sigma;
 out float AO;
 
-const float tau = 7.;
-const float beta = 0.002;
-const float epsilon = .00001;
+const float bias = .0005;
+const float thickness = 10.0;
 
-#define SAMPLES 16
-const float invSamples = 0.0625; // 1. / SAMPLES
+#define SAMPLES 8
+const float invSamples = 0.125; // 1. / SAMPLES
 
 vec3 getXcYcZc(int x, int y, float zC)
 {
@@ -22,6 +22,12 @@ vec3 getXcYcZc(int x, int y, float zC)
     return vec3(xC, yC, zC);
 }
 
+float interleavedGradientNoise(highp vec2 w)
+{
+    const vec3 m = vec3(0.06711056, 0.00583715, 52.9829189);
+    return fract(m.z * fract(dot(w, m.xy)));
+}
+
 void main(void)
 {
     vec2 uv = gl_FragCoord.xy / u_screen;
@@ -34,32 +40,38 @@ void main(void)
     vec3 ddy = dFdy(FragPos);
     vec3 norm = normalize(cross(ddy, ddx));
 
-    float r = radius / FragPos.z;
-    float phi = 3. * float((x ^ y) + x * y);
+    float r = radius / FragPos.z * 0.5;
+    float phi = interleavedGradientNoise(vec2(gl_FragCoord.x, gl_FragCoord.y));
     float bl = 0.0;
     float m = log2(r) + 6. + log2(invSamples);
 
-    float theta = mod(2. * 3.14 * tau * .5 * invSamples + phi, 6.283185307179586);
+    float peak = 0.1 * radius;
+    float peak2 = peak * peak;
+    float intensity = 2.0 * 3.14159 * sigma * peak * invSamples;
+
+    float theta = phi * 2.0 * 2.4 * 3.14159;
     vec2 rotations = vec2(cos(theta), sin(theta)) * u_screen;
     vec2 offset = vec2(cos(invSamples), sin(invSamples));
 
-    for(int i = 0; i < SAMPLES; ++i) {
+    for(int i = 0; i < SAMPLES; ++i)
+    {
         float alpha = (float(i) + .5) * invSamples;
         rotations = vec2(rotations.x * offset.x - rotations.y * offset.y, rotations.x * offset.y + rotations.y * offset.x);
         float h = r * alpha;
         vec2 localoffset = h * rotations;
 
-        m = m + .5;
-        ivec2 ioccluder_uv = ivec2(x, y) + ivec2(localoffset);
-
-        if (ioccluder_uv.x < 0 || ioccluder_uv.x > int(u_screen.x) || ioccluder_uv.y < 0 || ioccluder_uv.y > int(u_screen.y)) continue;
+        ivec2 ioccluder_uv = clamp(ivec2(x, y) + ivec2(localoffset), ivec2(0), ivec2(u_screen));
 
         float LinearoccluderFragmentDepth = textureLod(dtex, vec2(ioccluder_uv) / u_screen, max(m, 0.)).x;
         vec3 OccluderPos = getXcYcZc(ioccluder_uv.x, ioccluder_uv.y, LinearoccluderFragmentDepth);
 
         vec3 vi = OccluderPos - FragPos;
-        bl += max(0., dot(vi, norm) - FragPos.z * beta) / (dot(vi, vi) + epsilon);
+        float vv = dot(vi, vi);
+        float vn = dot(vi, norm);
+        float w = max(0.0, 1.0 - vv / thickness / thickness);
+        w = w * w;
+        bl += w * max(0., vn - FragPos.z * bias) / (vv + peak);
     }
 
-    AO = max(pow(1.0 - min(2. * sigma * bl * invSamples, 0.99), k), 0.);
+    AO = pow(max(1.0 - sqrt(bl * intensity), 0.), k);
 }
diff --git a/src/graphics/lighting_passes.cpp b/src/graphics/lighting_passes.cpp
@@ -170,7 +170,7 @@ class PointLightScatterShader : public TextureShader<PointLightScatterShader,
 };
 
 // ============================================================================
-class IBLShader : public TextureShader<IBLShader, 4>
+class IBLShader : public TextureShader<IBLShader, 5>
 {
 public:
     IBLShader()
@@ -181,20 +181,22 @@ class IBLShader : public TextureShader<IBLShader, 4>
         assignSamplerNames(0, "ntex",  ST_NEAREST_FILTERED,
                            1, "dtex",  ST_NEAREST_FILTERED,
                            2, "probe", ST_TRILINEAR_CUBEMAP,
-                           3, "albedo",ST_NEAREST_FILTERED);
+                           3, "albedo",ST_NEAREST_FILTERED,
+                           4, "ssao",  ST_NEAREST_FILTERED);
     }   // IBLShader
 };   // IBLShader
 
 // ============================================================================
-class DegradedIBLShader : public TextureShader<DegradedIBLShader, 1>
+class DegradedIBLShader : public TextureShader<DegradedIBLShader, 2>
 {
 public:
     DegradedIBLShader()
     {
         loadProgram(OBJECT, GL_VERTEX_SHADER, "screenquad.vert",
                             GL_FRAGMENT_SHADER, "degraded_ibl.frag");
         assignUniforms();
-        assignSamplerNames(0, "ntex", ST_NEAREST_FILTERED);
+        assignSamplerNames(0, "ntex", ST_NEAREST_FILTERED,
+                           1, "ssao", ST_NEAREST_FILTERED);
     }   // DegradedIBLShader
 };   // DegradedIBLShader
 
@@ -298,7 +300,8 @@ static void renderPointLights(unsigned count,
 void LightingPasses::renderEnvMap(GLuint normal_depth_texture,
                                   GLuint depth_stencil_texture,
                                   GLuint specular_probe,
-                                  GLuint albedo_buffer)
+                                  GLuint albedo_buffer,
+                                  GLuint ssao_texture)
 {
     glDisable(GL_DEPTH_TEST);
     glEnable(GL_BLEND);
@@ -311,7 +314,7 @@ void LightingPasses::renderEnvMap(GLuint normal_depth_texture,
         glBindVertexArray(SharedGPUObjects::getFullScreenQuadVAO());
 
         DegradedIBLShader::getInstance()
-            ->setTextureUnits(normal_depth_texture);
+            ->setTextureUnits(normal_depth_texture, ssao_texture);
         DegradedIBLShader::getInstance()->setUniforms();
     }
     else
@@ -323,7 +326,8 @@ void LightingPasses::renderEnvMap(GLuint normal_depth_texture,
             normal_depth_texture,
             depth_stencil_texture,
             specular_probe,
-            albedo_buffer);
+            albedo_buffer,
+            ssao_texture);
         IBLShader::getInstance()->setUniforms();
     }
 
@@ -434,14 +438,16 @@ void LightingPasses::renderLights(  bool has_shadow,
                                     GLuint depth_stencil_texture,
                                     GLuint albedo_texture,
                                     const FrameBuffer* shadow_framebuffer,
+                                    GLuint ssao_texture,
                                     GLuint specular_probe)
 {
     {
         ScopedGPUTimer timer(irr_driver->getGPUTimer(Q_ENVMAP));
         renderEnvMap(normal_depth_texture,
                      depth_stencil_texture,
                      specular_probe,
-                     albedo_texture);
+                     albedo_texture,
+                     ssao_texture);
     }
 
     // Render sunlight if and only if track supports shadow

diff --git a/src/graphics/lighting_passes.hpp b/src/graphics/lighting_passes.hpp
@@ -47,7 +47,8 @@ class LightingPasses
     void renderEnvMap(GLuint normal_depth_texture,
                       GLuint depth_stencil_texture,
                       GLuint specular_probe,
-                      GLuint albedo_buffer);
+                      GLuint albedo_buffer,
+                      GLuint ssao_buffer);
 
     /** Generate diffuse and specular map */
     void         renderSunlight(const core::vector3df &direction,
@@ -65,6 +66,7 @@ class LightingPasses
                         GLuint depth_stencil_texture,
                         GLuint albedo_texture,
                         const FrameBuffer* shadow_framebuffer,
+                        GLuint ssao_texture,
                         GLuint specular_probe);
     void renderLightsScatter(GLuint depth_stencil_texture,
                              const FrameBuffer& half1_framebuffer,