Rendering modes

Rendering modes

The default render paths supplied with Urho3D implement forward, light pre-pass and deferred rendering modes. Where they differ is how per-pixel lighting is calculated for opaque objects; transparent objects always use forward rendering. Note that on OpenGL ES 2.0 only forward rendering is available.

Forward rendering

Forward rendering begins with an ambient light pass for the objects; this also adds any per-vertex lights. Then, the objects are re-rendered for each per-pixel light affecting them (basic multipass rendering), up to the maximum per-pixel light count which is by default unlimited, but can be reduced with SetMaxLights(). The render operations are sorted by light, ie. render the effect of the first light on all affected objects first, then the second etc. If shadow maps are re-used (default on), a shadow casting light's shadow map will be updated immediately before rendering the lit objects. When shadow maps are not re-used, all shadow maps are updated first even before drawing the ambient pass.

Materials can also define an optimization pass, called "litbase", for forward rendering where the ambient light and the first per-pixel light are combined. This pass can not be used, however, if there are per-vertex lights affecting the object, or if the ambient light has a per-vertex gradient.

Light pre-pass rendering

Light pre-pass requires a minimum of two passes per object. First the normal, specular power, depth and lightmask (8 low bits only) of opaque objects are rendered to the following G-buffer:

  • RT0: World-space normal and specular power (D3DFMT_A8R8G8B8)
  • RT1: Linear depth (D3DFMT_R32F)
  • DS: Hardware depth and lightmask (D3DFMT_D24S8)

After the G-buffer is complete, light volumes (spot and point lights) or fullscreen quads (directional lights) will be rendered to a light accumulation buffer to calculate the diffuse and specular light at each opaque pixel. Specular light is stored as intensity only. Stencil compare (AND operation) with the 8 low bits of the light's lightmask will be used for light culling. Similarly to forward rendering, shadow maps will be updated before each light as necessary.

Finally the opaque objects are re-rendered during the material pass, which combines ambient and vertex lighting with per-pixel lighting from the light accumulation buffer. After this rendering proceeds to the post-opaque and refract passes, transparent object rendering pass, and the post-alpha pass, just like forward rendering.

Deferred rendering

Deferred rendering needs to render each opaque object only once to the G-buffer, but this rendering pass is much heavier than in light pre-pass rendering, as also ambient, emissive and diffuse albedo information is output at the same time. The G-buffer is the following:

  • RT0: Final rendertarget with ambient, per-vertex and emissive color (D3DFMT_X8R8G8B8)
  • RT1: Diffuse albedo and specular intensity (D3DFMT_A8R8G8B8)
  • RT2: World-space normal and specular power (D3DFMT_A8R8G8B8)
  • RT3: Linear depth (D3DFMT_R32F)
  • DS: Hardware depth and lightmask (D3DFMT_D24S8)

After the G-buffer has been rendered, light volumes will be rendered into the final rendertarget to accumulate per-pixel lighting. As the material albedo is available, all lighting calculations are final and output both the diffuse and specular color at the same time. After light accumulation rendering proceeds to post-opaque, refract, transparent, and post-alpha passes, as in other rendering modes.

Advantages and disadvantages

Whether using forward or deferred rendering modes is more advantageous depends on the scene and lighting complexity.

If the scene contains a large number of complex objects lit by multiple lights, forward rendering quickly increases the total draw call and vertex count due to re-rendering the objects for each light. However, light pre-pass and deferred rendering have a higher fixed cost due to the generation of the G-buffer. Also, in forward per-pixel lighting more calculations (such as light direction and shadow map coordinates) can be done at the vertex shader level, while in deferred all calculations need to happen per-pixel. This means that for a low light count, for example 1-2 per object, forward rendering will run faster based on the more efficient lighting calculations alone.

Forward rendering makes it possible to use hardware multisampling and different shading models in different materials if needed, while neither is possible in the deferred modes. Also, only forward rendering allows to calculate the material's diffuse and specular light response with the most accuracy. Light pre-pass rendering needs to reconstruct light specular color from the accumulated diffuse light color, which is inaccurate in case of overlapping lights. Deferred rendering on the other hand can not use the material's full specular color, it only stores a monochromatic intensity based on the green component into the G-buffer.

Light pre-pass rendering has a much more lightweight G-buffer pass, but it must render all opaque geometry twice. Light accumulation in pre-pass mode is slightly faster than in deferred. Despite this, unless there is significant overdraw, in vertex-heavy scenes deferred rendering will likely be faster than light pre-pass.

Finally note that due to OpenGL framebuffer object limitations an extra framebuffer blit has to happen at the end in both light pre-pass and deferred rendering, which costs some performance. Also, because multiple rendertargets on OpenGL must have the same format, an R32F texture can not be used for linear depth, but instead 24-bit depth is manually encoded and decoded into RGB channels.