Game Development Reference
should be fast and easy. The architecture should be simple and not get in
the way of creative expression.
We have devised a rendering pipeline that we believe meets these objectives well
and is a good match for modern GPU hardware going into the foreseeable future.
We refer to it as the Forward+ rendering pipeline [Harada et al. 11].
The Forward+ rendering pipeline requires three stages:
Z prepass. Z prepass is an option for forward rendering, but it is essential
for Forward+ to reduce the pixel overdraws of the final shading step. This
is especially expensive for Forward+ due to the possible traversal of many
lights per pixel, which we will detail later in this section.
Light culling. Light culling is a stage that calculates the list of lights affecting
Final shading. Final shading, which is an extension to the shading pass in
forward rendering, shades the entire surface. A required change is the way
to pass lights to shaders. In Forward+, any lights in a scene have to be
accessible from shaders rather than binding some subset of lights for each
objects as is typical of traditional forward rendering.
The light-culling stage is similar to the light-accumulation step of deferred light-
ing. Instead of calculating lighting components, light culling calculates a list of
light indices overlapping a pixel. The list of lights can be calculated for each
pixel, which is a better choice for final shading.
However, storing a per-pixel light list requires a large memory footprint and
significant computation at the light-culling stage. Instead, the screen is split into
tiling can add false positives to the list for a pixel in a tile, it reduces the overall
memory footprint and computation time necessary for generating the light lists.
Thus we are making a tradeoff between light-index buffer memory and final shader
By utilizing the computing capability of modern GPUs, light culling can be
implemented entirely on the GPU as detailed in Section 5.3. Therefore, the whole
lighting pipeline can be executed entirely on the GPU.