Tech Interview: Halo: Reach

Halo 3 and Halo 3: ODST used a 'semi'-deferred two-pass rendering approach, except for small decoration objects like grass and pebbles which were forward-rendered in a single pass for speed. Semi-deferred rendering allowed us to have cheap decals, but we didn't use it for deferred lighting; the lighting was rendered in the second pass over the geometry so we could have complex light map lighting and nice-looking metallic shininess (i.e. something better than Phong specular). For Halo: Reach, we rebuilt the deferred buffers so they could better approximate our specular models, which let us use fast deferred dynamic lights everywhere without losing shininess.

On top of that, we also built a system to determine when objects were not taking advantage of the deferred path (ie. they had no decals or complex deferred lights touching them) and switch those objects on the fly to the faster one-pass forward rendering. Yaohua Hu also spent a lot of time researching an improved light-map representation (it's better than spherical harmonics!) that gives us the same support for area light sources, improved contrast, fewer artifacts, a smaller memory footprint and much better performance. This helped to free up a lot of GPU time to use for the dynamic deferred lights and other graphical goodies.

We built a particle system to handle the specific case of numerous small transient particles - basically rock chips, dirt puffs, rain drops, splashes, sparks and that kind of thing. I am presenting it in more detail at the next GDC, but the neat part is that it can handle tens of thousands of collisions/bounces each frame by reading the depth and normal buffers, and the whole thing takes less than 0.3 ms (about 1/100th of a frame); which looks pretty good compared to the seven (7) standard particle collisions per frame allowed by the effects budget.

The new particle system allowed the effect artists to use huge numbers of these small colliding particles in their effects without worrying about performance at all. Oh and it's used for the rain too: if you watch the rain in slow-motion in the theatre mode you can follow a single rain drop as it falls until it splashes onto something!

We use a single 7e3 buffer for our final render target in Reach. This results in a more limited HDR (about 8x over the white point, as opposed to 128x in Halo 3) but is much faster for transparents and post-processing. In practice, the difference between 8x and 128x HDR is slight - the main thing you may notice is that the bloom around bright areas loses its color more often, desaturating to white.

And yes, a single 7e3 buffer gives us more available EDRAM for the final lighting pass, but the render resolution is still limited by the three buffers used in the main deferred pass. The resolution in Halo 3 was more limited because we save some EDRAM for dynamic shadows during the lighting pass, alongside the 2 HDR buffers and a depth buffer. But with a single 7e3 buffer, we have plenty of extra room available for the shadows, and it's only limited by the 3 buffers used during the deferred pass.

Multiple tiles are problematic: they either add too much controller latency (because predicated tiling delays the GPU start), or they result in too many passes over the geometry, eating up large amounts of CPU (basically rendering everything twice). Another factor is the 360's DAC, which has the super fancy up-sampling filter that hides any artifacts - we actually ran user tests on various resolutions and no one could tell the difference! So we chose to take the additional performance and reduced controller latency over the nearly imperceptible resolution increase.