Shadow: Use 4 cascades.

Now that we use GS and textures array it's easier to add a cascade.
4 is still not optimal but I'd like to avoid cascade count inflation as
much as possible ; 4 x 1024x1024 has the bandwidth requirement of the
single 2048x2048 shadowmaps we previously had.

data/shaders/shadow.geom

@@ -1,8 +1,8 @@
 #version 330 core
-uniform mat4 ModelViewProjectionMatrix[3];
+uniform mat4 ModelViewProjectionMatrix[4];
 
 layout(triangles) in;
-layout(triangle_strip, max_vertices=9) out;
+layout(triangle_strip, max_vertices=12) out;
 
 in vec2 tc[3];
 
@@ -10,7 +10,7 @@ out vec2 uv;
 
 void main(void)
 {
-  for (int j = 0; j<3; j++)
+  for (int j = 0; j<4; j++)
   {
     gl_Layer = j;
     for(int i=0; i<3; i++)

data/shaders/sunlightshadow.frag

@@ -8,7 +8,7 @@ uniform sampler2DArrayShadow shadowtex;
 uniform vec3 direction;
 uniform vec3 col;
 uniform mat4 invproj;
-uniform mat4 shadowmat[3];
+uniform mat4 shadowmat[4];
 //uniform int hasclouds;
 //uniform vec2 wind;
 //uniform float shadowoffset;
@@ -26,7 +26,7 @@ vec3 DecodeNormal(vec2 n)
 
 float getShadowFactor(vec3 pos, float bias)
 {
-	if (pos.z < 10.)
+	if (pos.z < 5.)
 	{
 		vec4 shadowcoord = (shadowmat[0] * vec4(pos, 1.0));
 		shadowcoord /= shadowcoord.w;
@@ -43,19 +43,26 @@ float getShadowFactor(vec3 pos, float bias)
 	//	bias += smoothstep(0.001, 0.1, moved) * 0.014; // According to the warping
 		return texture(shadowtex, vec4(shadowtexcoord, 0., 0.5 * (shadowcoord.z + bias * 0.001) + 0.5));
 	}
-	else if (pos.z < 60)
+	else if (pos.z < 10.)
 	{
 		vec4 shadowcoord = (shadowmat[1] * vec4(pos, 1.0));
 		shadowcoord /= shadowcoord.w;
 		vec2 shadowtexcoord = shadowcoord.xy * 0.5 + 0.5;
 		return texture(shadowtex, vec4(shadowtexcoord, 1., 0.5 * (shadowcoord.z + bias * 0.001) + 0.5));
 	}
-	else
+	else if (pos.z < 75.)
 	{
 		vec4 shadowcoord = (shadowmat[2] * vec4(pos, 1.0));
 		shadowcoord /= shadowcoord.w;
 		vec2 shadowtexcoord = shadowcoord.xy * 0.5 + 0.5;
-		return texture(shadowtex, vec4(shadowtexcoord, 2., 0.5 * (shadowcoord.z + bias) + 0.5));
+		return texture(shadowtex, vec4(shadowtexcoord, 2., 0.5 * (shadowcoord.z + bias * 0.001) + 0.5));
+	}
+	else
+	{
+		vec4 shadowcoord = (shadowmat[3] * vec4(pos, 1.0));
+		shadowcoord /= shadowcoord.w;
+		vec2 shadowtexcoord = shadowcoord.xy * 0.5 + 0.5;
+		return texture(shadowtex, vec4(shadowtexcoord, 3., 0.5 * (shadowcoord.z + bias) + 0.5));
 	}
 }
 

src/graphics/render.cpp

@@ -484,22 +484,24 @@ void IrrDriver::renderShadows(//ShadowImportanceProvider * const sicb,
     const float oldnear = camnode->getNearValue();
     float FarValues[] =
     {
+        5.,
         10.,
-        60.,
+        75.,
         oldfar,
     };
     float NearValues[] =
     {
         oldnear,
+        5.,
         10.,
-        60.,
+        75.,
     };
 
     const core::matrix4 &SunCamViewMatrix = m_suncam->getViewMatrix();
     sun_ortho_matrix.clear();
 
     // Build the 3 ortho projection (for the 3 shadow resolution levels)
-    for (unsigned i = 0; i < 3; i++)
+    for (unsigned i = 0; i < 4; i++)
     {
         camnode->setFarValue(FarValues[i]);
         camnode->setNearValue(NearValues[i]);
@@ -554,7 +556,7 @@ void IrrDriver::renderShadows(//ShadowImportanceProvider * const sicb,
 
         sun_ortho_matrix.push_back(getVideoDriver()->getTransform(video::ETS_PROJECTION) * getVideoDriver()->getTransform(video::ETS_VIEW));
     }
-    assert(sun_ortho_matrix.size() == 3);
+    assert(sun_ortho_matrix.size() == 4);
 
     irr_driver->setPhase(SHADOW_PASS);
     glDisable(GL_BLEND);

src/graphics/rtts.cpp

@@ -145,10 +145,10 @@ RTT::RTT()
     glBindFramebuffer(GL_FRAMEBUFFER, shadowFBO);
     glGenTextures(1, &shadowColorTex);
     glBindTexture(GL_TEXTURE_2D_ARRAY, shadowColorTex);
-    glTexImage3D(GL_TEXTURE_2D_ARRAY, 0, GL_R8, 1024, 1024, 3, 0, GL_RED, GL_UNSIGNED_BYTE, 0);
+    glTexImage3D(GL_TEXTURE_2D_ARRAY, 0, GL_R8, 1024, 1024, 4, 0, GL_RED, GL_UNSIGNED_BYTE, 0);
     glGenTextures(1, &shadowDepthTex);
     glBindTexture(GL_TEXTURE_2D_ARRAY, shadowDepthTex);
-    glTexImage3D(GL_TEXTURE_2D_ARRAY, 0, GL_DEPTH_COMPONENT24, 1024, 1024, 3, 0, GL_DEPTH_COMPONENT, GL_FLOAT, 0);
+    glTexImage3D(GL_TEXTURE_2D_ARRAY, 0, GL_DEPTH_COMPONENT24, 1024, 1024, 4, 0, GL_DEPTH_COMPONENT, GL_FLOAT, 0);
     glFramebufferTexture(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, shadowColorTex, 0);
     glFramebufferTexture(GL_FRAMEBUFFER, GL_DEPTH_ATTACHMENT, shadowDepthTex, 0);
     GLenum result = glCheckFramebufferStatus(GL_FRAMEBUFFER);