An older technique (introduced with the Maxwell architecture), Multi-Res Shading splits the render into multiple viewports. This technique takes advantage of the face the VR pipeline requires warping to fit to the optics of a HMD--the user view is split up into 9 sections where those at the edges are rendered at a lower resolution.
While this technique certainly provides a performance improvement, the following technique for the most part replaces Multi-Res Shading and is therefore not recommended for project development unless limited to Maxwell architecture.
Lens-Matched Shading improves upon Multi-Res Shading by taking advantage of Simultaneous Multi-Projection resulting in a further performance increase. In essence, the image is rendered onto a surface that more closely approximates the VR lens.
This technique produces even better results than its predecessor with a lower performance overhead and is recommended for use in any VR project.
VRS is a technique that controls “shading rate” (sampling density) dynamically--ranges from sub-sampling to super-sampling. as little as once per 16 pixels or as often as 8 times per pixel (max Multisample Anti-aliasing (MSAA) rate)
There are several classes of VRS-based algorithms:
Content level of detail (Content Adaptive Shading): the further away, the lower the sampling density
Rate of content motion (Motion Adaptive Shading): the larger the magnitude of motion, the lower the sampling density
Targeted toward lens resolution and eye position (Foveated Rendering)
While this technique is a powerful abstraction for rendering in general, a VR application won't see maximal performance gains unless the headset is equipped with pupil tracking for foveated rendering.
Single-Pass Stereo simply allows shaders to output two x co-coordinates (one per eye). This eliminates the need to redraw all geometry for both eyes.
Single-Pass Stereo is essentially supplanted by Multi-View stereo and is recommended only for projects limited to older architecture.
Multi-View Rendering is the next evolution of Single-Pass stereo. This technique can render up to four projection views which can accelerate canted (non-coplanar) HMDs with extreme FOV (> 180°). Among many other things, the technique is useful for LOD techniques like Cascaded Shadow Maps.
This technique is recommended for any VR application with hardware that can take advantage of it.
This feature in VRWorks simply allows two GPUs to both contribute to the pipeline in an optimal way. Rather than split the view, VR lends itself naturally to parallelization per-eye. Each GPU renders these normally while sharing relevant data such as shadow maps and physics.
With available hardware, this feature is highly recommended for the dramatic almost two-fold performance increase.