This part of the DDGI article explains the algorithm in detail for
industry programmers, researchers, and scientists seeking to implement
it. This focuses on the mathematical concepts and opportunities for
making your own design decisions and potential improvements for a
minimal prototype implementation.
The following part gives a checklist for steps for a full production
DDGI implementation when integrating into an existing rendering
system, with recommendations for the best practices we use with NVIDIA
partners.
## Sampling
DDGI supports efficient sampling of the irradiance field at _any_
point $X$ in a scene relative to a surface normal $\n$ by a single
shader function[^shader]:
`Color3 sampleDDGI(Point3 X, Vector3 n, DDGIVolume ddgi);`
The `sampleDDGI` function can be called from anywhere. For example, a
deferred-shading compute or full-screen pixel shader, rasterization
forward shader, volumetric ray marcher, texture-space shading routine,
or ray hit shader.
Because DDGI encodes incoming light, shading simply involes modulating
the sampled result by the scattering function (BSDF). Combined with a
glossy term computed from an environment-map or by ray tracing, the
net _outgoing_ light
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ GLSL
Radiance3 L =
lerp(glossyLight,
lambertianReflectivity * sampleDDGI(X, n, ddgi),
fresnel);
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
[^shader]: I use macros to rename `vec3`/`float3` shader types to be
self-documenting, so imagine `#define Color3 float3` and so on in some
header file. All of the support code for our reference implementation
is directly available in the [G3D Innovation Engine](https://casual-effects.com).
## Radial Gaussian Distance
DDGI contains three terms that manage the visibility
* * * *
I wrote this blog article in collaboration with my colleagues [Zander Majercik](https://research.nvidia.com/person/zander-majercik)
and [Adam Marrs](http://www.visualextract.com/) at NVIDIA.