Try a fix for nvidia and GI

data/shaders/slicedscreenquad.vert

@@ -3,22 +3,16 @@ in vec2 Texcoord;
 
 #ifndef VSLayer
 out int layer;
-out vec2 uv_in;
 #else
-out vec2 uv;
 flat out int slice;
 #endif
 
-
-
 void main() {
 #ifdef VSLayer
     gl_Layer = gl_InstanceID;
-    uv = Texcoord;
     slice = gl_InstanceID;
 #else
     layer = gl_InstanceID;
-    uv_in = Texcoord;
 #endif
     gl_Position = vec4(Position, 0., 1.);
 }

data/shaders/slicedscreenquad_nvworkaround.vert

@@ -0,0 +1,11 @@
+uniform int slice;
+
+in vec2 Position;
+in vec2 Texcoord;
+
+out int layer;
+
+void main() {
+    layer = slice;
+    gl_Position = vec4(Position, 0., 1.);
+}

src/graphics/irr_driver.cpp

@@ -466,11 +466,15 @@ void IrrDriver::initDevice()
         Log::info("IrrDriver", "OpenGL version string: %s", glGetString(GL_VERSION));
 
         m_need_ubo_workaround = false;
+        m_need_rh_workaround = false;
 #ifdef WIN32
         // Fix for Intel Sandy Bridge on Windows which supports GL up to 3.1 only
         if (strstr((const char *)glGetString(GL_VENDOR), "Intel") != nullptr && (GLMajorVersion == 3 && GLMinorVersion == 1))
             m_need_ubo_workaround = true;
 #endif
+        // Fix for Nvidia and instanced RH
+        if (strstr((const char *)glGetString(GL_VENDOR), "NVIDIA") != nullptr)
+            m_need_rh_workaround = true;
     }
     m_glsl = (GLMajorVersion > 3 || (GLMajorVersion == 3 && GLMinorVersion >= 1));
 

src/graphics/irr_driver.hpp

@@ -197,6 +197,7 @@ private:
     int GLMajorVersion, GLMinorVersion;
     bool hasVSLayer;
     bool m_need_ubo_workaround;
+    bool m_need_rh_workaround;
     /** The irrlicht device. */
     IrrlichtDevice             *m_device;
     /** Irrlicht scene manager. */
@@ -284,6 +285,11 @@ public:
         return m_need_ubo_workaround;
     }
 
+    bool needRHWorkaround() const
+    {
+        return m_need_ubo_workaround;
+    }
+
     bool hasVSLayerExtension() const
     {
         return hasVSLayer;

src/graphics/render_lighting.cpp

@@ -136,13 +136,28 @@ void IrrDriver::renderLights(unsigned pointlightcount)
         ScopedGPUTimer timer(irr_driver->getGPUTimer(Q_RH));
         glDisable(GL_BLEND);
         m_rtts->getRH().Bind();
-        glUseProgram(FullScreenShader::RadianceHintsConstructionShader::getInstance()->Program);
         glBindVertexArray(SharedObject::FullScreenQuadVAO);
-        setTexture(FullScreenShader::RadianceHintsConstructionShader::getInstance()->TU_ctex, m_rtts->getRSM().getRTT()[0], GL_LINEAR, GL_LINEAR);
-        setTexture(FullScreenShader::RadianceHintsConstructionShader::getInstance()->TU_ntex, m_rtts->getRSM().getRTT()[1], GL_LINEAR, GL_LINEAR);
-        setTexture(FullScreenShader::RadianceHintsConstructionShader::getInstance()->TU_dtex, m_rtts->getRSM().getDepthTexture(), GL_LINEAR, GL_LINEAR);
-        FullScreenShader::RadianceHintsConstructionShader::getInstance()->setUniforms(rsm_matrix, rh_matrix, rh_extend);
-        glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, 32);
+        if (irr_driver->needRHWorkaround())
+        {
+            glUseProgram(FullScreenShader::NVWorkaroundRadianceHintsConstructionShader::getInstance()->Program);
+            setTexture(FullScreenShader::NVWorkaroundRadianceHintsConstructionShader::getInstance()->TU_ctex, m_rtts->getRSM().getRTT()[0], GL_LINEAR, GL_LINEAR);
+            setTexture(FullScreenShader::NVWorkaroundRadianceHintsConstructionShader::getInstance()->TU_ntex, m_rtts->getRSM().getRTT()[1], GL_LINEAR, GL_LINEAR);
+            setTexture(FullScreenShader::NVWorkaroundRadianceHintsConstructionShader::getInstance()->TU_dtex, m_rtts->getRSM().getDepthTexture(), GL_LINEAR, GL_LINEAR);
+            for (unsigned i = 0; i < 32; i++)
+            {
+                FullScreenShader::NVWorkaroundRadianceHintsConstructionShader::getInstance()->setUniforms(rsm_matrix, rh_matrix, rh_extend, i);
+                glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
+            }
+        }
+        else
+        {
+            glUseProgram(FullScreenShader::RadianceHintsConstructionShader::getInstance()->Program);
+            setTexture(FullScreenShader::RadianceHintsConstructionShader::getInstance()->TU_ctex, m_rtts->getRSM().getRTT()[0], GL_LINEAR, GL_LINEAR);
+            setTexture(FullScreenShader::RadianceHintsConstructionShader::getInstance()->TU_ntex, m_rtts->getRSM().getRTT()[1], GL_LINEAR, GL_LINEAR);
+            setTexture(FullScreenShader::RadianceHintsConstructionShader::getInstance()->TU_dtex, m_rtts->getRSM().getDepthTexture(), GL_LINEAR, GL_LINEAR);
+            FullScreenShader::RadianceHintsConstructionShader::getInstance()->setUniforms(rsm_matrix, rh_matrix, rh_extend);
+            glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, 32);
+        }
     }
 
     for (unsigned i = 0; i < sun_ortho_matrix.size(); i++)

src/graphics/shaders.cpp

@@ -1584,6 +1584,22 @@ namespace FullScreenShader
         AssignTextureUnit(Program, TexUnit(TU_ctex, "ctex"), TexUnit(TU_ntex, "ntex"), TexUnit(TU_dtex, "dtex"));
     }
 
+    NVWorkaroundRadianceHintsConstructionShader::NVWorkaroundRadianceHintsConstructionShader()
+    {
+        if (irr_driver->getGLSLVersion() < 150)
+            return;
+        Program = LoadProgram(
+            GL_VERTEX_SHADER, file_manager->getAsset("shaders/slicedscreenquad_nvworkaround.vert").c_str(),
+            GL_GEOMETRY_SHADER, file_manager->getAsset("shaders/rhpassthrough.geom").c_str(),
+            GL_FRAGMENT_SHADER, file_manager->getAsset("shaders/rh.frag").c_str());
+
+        AssignUniforms("RSMMatrix", "RHMatrix", "extents", "slice");
+        TU_ctex = 0;
+        TU_ntex = 1;
+        TU_dtex = 2;
+        AssignTextureUnit(Program, TexUnit(TU_ctex, "ctex"), TexUnit(TU_ntex, "ntex"), TexUnit(TU_dtex, "dtex"));
+    }
+
     RHDebug::RHDebug()
     {
         Program = LoadProgram(

src/graphics/shaders.hpp

@@ -588,6 +588,15 @@ public:
     RadianceHintsConstructionShader();
 };
 
+// Workaround for a bug found in kepler nvidia linux and fermi nvidia windows
+class NVWorkaroundRadianceHintsConstructionShader : public ShaderHelperSingleton<NVWorkaroundRadianceHintsConstructionShader, core::matrix4, core::matrix4, core::vector3df, int>
+{
+public:
+    GLuint TU_ctex, TU_ntex, TU_dtex;
+
+    NVWorkaroundRadianceHintsConstructionShader();
+};
+
 class RHDebug : public ShaderHelperSingleton<RHDebug, core::matrix4, core::vector3df>
 {
 public: