Implement specular IBL properly

data/shaders/IBL.frag

@@ -50,9 +50,8 @@ void main(void)
     sampleDirection = (InverseViewMatrix * vec4(sampleDirection, 0.)).xyz;
 
     float specval = texture(ntex, uv).z;
-    // From http://graphics.cs.williams.edu/papers/EnvMipReport2013/
-    int texSize = textureSize(tex, 0).x;
-    float lodval = clamp(log2(texSize * sqrt(3.)) - (5. * specval + 1.), 0., 10.);
+     // Assume 8 level of lod (ie 256x256 texture)
+    float lodval = 8. * (1. - specval);
     vec4 specular = textureLod(tex, sampleDirection, lodval);
     Spec = max(specular, vec4(0.));
 }

data/shaders/importance_sampling_specular.frag

@@ -0,0 +1,30 @@
+uniform samplerCube tex;
+uniform float samples[2048];
+uniform float ViewportSize;
+
+uniform mat4 PermutationMatrix;
+
+out vec4 FragColor;
+
+void main(void)
+{
+    vec2 uv = gl_FragCoord.xy / ViewportSize;
+    vec3 RayDir = 2. * vec3(uv, 1.) - 1.;
+    RayDir = normalize((PermutationMatrix * vec4(RayDir, 0.)).xyz);
+
+    vec4 FinalColor = vec4(0.);
+    vec3 up = (RayDir.y < .99) ? vec3(0., 1., 0.) : vec3(0., 0., 1.);
+    vec3 Tangent = normalize(cross(up, RayDir));
+    vec3 Bitangent = cross(RayDir, Tangent);
+
+    for (int i = 0; i < 1024; i++)
+    {
+        float Theta = samples[2 * i];
+        float Phi = samples[2 * i + 1];
+
+        vec3 sampleDir = cos(Theta) * RayDir + sin(Theta) * cos(Phi) * Tangent + sin(Theta) * sin(Phi) * Bitangent;
+        FinalColor += textureLod(tex, sampleDir, 0.);
+    }
+
+    FragColor = FinalColor / 1024.;
+}

src/graphics/IBL.cpp

@@ -1,6 +1,8 @@
 #include "IBL.hpp"
 #include "gl_headers.hpp"
+#include "shaders.hpp"
 #include <cmath>
+#include <set>
 
 static void getXYZ(GLenum face, float i, float j, float &x, float &y, float &z)
 {
@@ -204,3 +206,107 @@ void SphericalHarmonics(Color *CubemapFace[6], size_t edge_size, float *blueSHCo
         delete[] Y22[face];
     }
 }
+
+// From http://http.developer.nvidia.com/GPUGems3/gpugems3_ch20.html
+/** Returns the index-th pair from Hammersley set of pseudo random set.
+    Hammersley set is a uniform distribution between 0 and 1 for 2 components.
+    We use the natural indexation on the set to avoid storing the whole set.
+    \param index of the pair
+    \param size of the set. */
+std::pair<float, float> HammersleySequence(int index, int samples)
+{
+    float InvertedBinaryRepresentation = 0.;
+    for (size_t i = 0; i < 32; i++)
+    {
+        InvertedBinaryRepresentation += ((index >> i) & 0x1) * powf(.5, (float) (i + 1.));
+    }
+    return std::make_pair(float(index) / float(samples), InvertedBinaryRepresentation);
+}
+
+
+/** Returns a pseudo random (theta, phi) generated from a probability density function modeled after Phong function.
+    \param a pseudo random float pair from a uniform density function between 0 and 1.
+    \param exponent from the Phong formula. */
+std::pair<float, float> ImportanceSamplingPhong(std::pair<float, float> Seeds, float exponent)
+{
+    return std::make_pair(acosf(powf(Seeds.first, 1.f / (exponent + 1.f))), 2.f * 3.14f * Seeds.second);
+}
+
+static
+core::matrix4 getPermutationMatrix(size_t indexX, float valX, size_t indexY, float valY, size_t indexZ, float valZ)
+{
+    core::matrix4 resultMat;
+    float *M = resultMat.pointer();
+    memset(M, 0, 16 * sizeof(float));
+    assert(indexX < 4);
+    assert(indexY < 4);
+    assert(indexZ < 4);
+    M[indexX] = valX;
+    M[4 + indexY] = valY;
+    M[8 + indexZ] = valZ;
+    return resultMat;
+}
+
+GLuint generateSpecularCubemap(GLuint probe)
+{
+    GLuint cubemap_texture;
+
+    glGenTextures(1, &cubemap_texture);
+    glBindTexture(GL_TEXTURE_CUBE_MAP, cubemap_texture);
+    size_t cubemap_size = 256;
+    for (int i = 0; i < 6; i++)
+        glTexImage2D(GL_TEXTURE_CUBE_MAP_POSITIVE_X + i, 0, GL_RGBA16F, cubemap_size, cubemap_size, 0, GL_BGRA, GL_FLOAT, 0);
+    glGenerateMipmap(GL_TEXTURE_CUBE_MAP);
+
+    GLuint fbo;
+    glGenFramebuffers(1, &fbo);
+    glBindFramebuffer(GL_FRAMEBUFFER, fbo);
+    glViewport(0, 0, cubemap_size, cubemap_size);
+    GLenum bufs[] = { GL_COLOR_ATTACHMENT0 };
+    glDrawBuffers(1, bufs);
+    glUseProgram(UtilShader::SpecularIBLGenerator::getInstance()->Program);
+    glBindVertexArray(SharedObject::FullScreenQuadVAO);
+
+    glDisable(GL_BLEND);
+    glDisable(GL_DEPTH_TEST);
+    glDisable(GL_CULL_FACE);
+
+    core::matrix4 M[6] = {
+        getPermutationMatrix(2, -1., 1, -1., 0, 1.),
+        getPermutationMatrix(2, 1., 1, -1., 0, -1.),
+        getPermutationMatrix(0, 1., 2, 1., 1, 1.),
+        getPermutationMatrix(0, 1., 2, -1., 1, -1.),
+        getPermutationMatrix(0, 1., 1, -1., 2, 1.),
+        getPermutationMatrix(0, -1., 1, -1., 2, -1.),
+    };
+
+    for (unsigned level = 0; level < 8; level++)
+    {
+        // Blinn Phong can be approximated by Phong with 4x the specular coefficient
+        // See http://seblagarde.wordpress.com/2012/03/29/relationship-between-phong-and-blinn-lighting-model/
+        float roughness = (8 - level) * 4 * pow(2., 10.) / 8.;
+        float viewportSize = 1 << (8 - level);
+
+        std::vector<float> Samples;
+        for (unsigned i = 0; i < 1024; i++)
+        {
+            std::pair<float, float> sample = ImportanceSamplingPhong(HammersleySequence(i, 1024), roughness);
+            Samples.push_back(sample.first);
+            Samples.push_back(sample.second);
+        }
+
+        for (unsigned face = 0; face < 6; face++)
+        {
+            glFramebufferTexture2D(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, GL_TEXTURE_CUBE_MAP_POSITIVE_X + face, cubemap_texture, level);
+            GLuint status = glCheckFramebufferStatus(GL_FRAMEBUFFER);
+            assert(status == GL_FRAMEBUFFER_COMPLETE);
+
+            UtilShader::SpecularIBLGenerator::getInstance()->SetTextureUnits(probe);
+            UtilShader::SpecularIBLGenerator::getInstance()->setUniforms(M[face], Samples, viewportSize);
+            glDrawArrays(GL_TRIANGLES, 0, 3);
+        }
+    }
+
+    glDeleteFramebuffers(1, &fbo);
+    return cubemap_texture;
+}

src/graphics/IBL.hpp

@@ -1,6 +1,8 @@
 #ifndef IBL_HPP
 #define IBL_HPP
 
+#include "gl_headers.hpp"
+
 struct Color
 {
     float Red;
@@ -14,4 +16,6 @@ using the cubemap provided by CubemapFace.
 *  \param row/columns count of textures.
 */
 void SphericalHarmonics(Color *CubemapFace[6], size_t edge_size, float *blueSHCoeff, float *greenSHCoeff, float *redSHCoeff);
+
+GLuint generateSpecularCubemap(GLuint probe);
 #endif

src/graphics/irr_driver.cpp

@@ -1365,6 +1365,7 @@ scene::ISceneNode *IrrDriver::addSkyBox(const std::vector<video::ITexture*> &tex
     SkyboxTextures = texture;
     SphericalHarmonicsTextures = sphericalHarmonics;
     SkyboxCubeMap = 0;
+    SkyboxSpecularProbe = 0;
     m_SH_dirty = true;
     return m_scene_manager->addSkyBoxSceneNode(texture[0], texture[1],
                                                texture[2], texture[3],
@@ -1377,8 +1378,12 @@ void IrrDriver::suppressSkyBox()
     SphericalHarmonicsTextures.clear();
     m_SH_dirty = true;
     if ((SkyboxCubeMap) && (!ProfileWorld::isNoGraphics()))
+    {
         glDeleteTextures(1, &SkyboxCubeMap);
+        glDeleteTextures(1, &SkyboxSpecularProbe);
+    }
     SkyboxCubeMap = 0;
+    SkyboxSpecularProbe = 0;
 }
 
 // ----------------------------------------------------------------------------

src/graphics/irr_driver.hpp

@@ -254,6 +254,7 @@ private:
 
 public:
     GLuint SkyboxCubeMap;
+    GLuint SkyboxSpecularProbe;
     /** A simple class to store video resolutions. */
     class VideoMode
     {

src/graphics/render_lighting.cpp

@@ -155,7 +155,7 @@ void IrrDriver::renderLights(unsigned pointlightcount, bool hasShadow)
 
     {
         ScopedGPUTimer timer(irr_driver->getGPUTimer(Q_ENVMAP));
-        m_post_processing->renderEnvMap(blueSHCoeff, greenSHCoeff, redSHCoeff, SkyboxCubeMap);
+        m_post_processing->renderEnvMap(blueSHCoeff, greenSHCoeff, redSHCoeff, SkyboxSpecularProbe);
     }
 
     // Render sunlight if and only if track supports shadow

src/graphics/render_skybox.cpp

@@ -280,6 +280,7 @@ void IrrDriver::generateSkyboxCubemap()
 
     assert(SkyboxTextures.size() == 6);
     SkyboxCubeMap = generateCubeMapFromTextures(SkyboxTextures);
+    SkyboxSpecularProbe = generateSpecularCubemap(SkyboxCubeMap);
 }
 
 void IrrDriver::generateDiffuseCoefficients()
@@ -322,8 +323,8 @@ void IrrDriver::generateDiffuseCoefficients()
     }
     else
     {
-        int sh_w = 16;
-        int sh_h = 16;
+        sh_w = 16;
+        sh_h = 16;
 
         video::SColor ambient = m_scene_manager->getAmbientLight().toSColor();
 

src/graphics/shaders.cpp

@@ -889,6 +889,18 @@ unsigned getGLSLVersion()
     return irr_driver->getGLSLVersion();
 }
 
+namespace UtilShader
+{
+    SpecularIBLGenerator::SpecularIBLGenerator()
+    {
+        Program = LoadProgram(OBJECT,
+            GL_VERTEX_SHADER, file_manager->getAsset("shaders/screenquad.vert").c_str(),
+            GL_FRAGMENT_SHADER, file_manager->getAsset("shaders/importance_sampling_specular.frag").c_str());
+        AssignUniforms("PermutationMatrix", "samples[0]", "ViewportSize");
+        AssignSamplerNames(Program, 0, "tex");
+    }
+}
+
 namespace MeshShader
 {
     // Solid Normal and depth pass shaders

src/graphics/shaders.hpp

@@ -48,6 +48,12 @@ public:
     static void init();
     static void setUniforms(const irr::video::SColor &);
 };
+
+class SpecularIBLGenerator : public ShaderHelperSingleton<SpecularIBLGenerator, core::matrix4, std::vector<float>, float >, public TextureRead<Trilinear_cubemap>
+{
+public:
+    SpecularIBLGenerator();
+};
 }