Some of the core ideas were described at SIGGRAPH 2017 in the Advances in Real-Time Rendering course (course page link). Since this initial publication we have been actively working to extend the feature set, which includes innovations in the following areas.
It is well known that ocean shape can be well approximated by summing Gerstner waves. Dozens of these are required to obtain an interesting shape. In previous implementations this has been prohibitively expensive and shape is either generated using an online FFT, or precomputed and baked into textures.
We generate shape from Gerstner waves efficiently at runtime by rendering at multiple scales, and ensure that waves are never over-sampled (inefficient) or under-sampled (bad quality, aliasing). This is highly performant and gives detail close to the viewer, and shape to the horizon. This gives considerable flexibility in shape and opens possibilities such as attenuating waves based on ocean depth around shorelines.
To control the ocean shape, we introduce an intuitive and fun shape authoring interface - an equalizer style editor which makes it fast and easy to achieve surface shape. Art direction such as small choppy waves with longer waves rolling in from a storm at the horizon is simple to achieve in this framework. We also support empirical ocean spectra from the literature (Phillips, JONSWAP, etc) which can be used directly or as a comparison.
For interactivity we add a dynamic wave simulation on top of the ocean waves. The simulation is computed on a heightfield and then converted into displacements, which trigger foam generation and enable boat wakes to be generated. This is demonstrated in the boat and threeboats example scenes.
The final shape is asynchronously read back to the CPU for gameplay/physics use. This gives access to the full, rich shape without requiring expensive CPU calculations or pipeline stalls.
We implement a 100% pop-free meshing solution, which follows the same unified multi-scale structure/layout as the shape data. The vertex densities and locations match the shape texels 1:1. This ensures that the shape is never over-sampled or under-sampled, giving the same guarantees as described above.
Our meshing approach requires only simple shader instructions in a vertex shader, and does not rely on tessellation or compute shaders or any other advanced shader model features. The geometry is composed of tiles which have strictly no overlap, and support frustum culling. These tiles are generated quickly on startup.
The multi-resolution representation (shape textures and geometry) is scaled horizontally when the camera changes altitude to ensure appropriate level of detail and good visual range for all viewpoints. To further extend the surface to the horizon we also add a strip of triangles at the mesh boundary.
Normal maps are elegantly incorporated into our multi-scale framework. Normal map textures are treated as a slightly different form of shape that is too detailed to be efficiently sampled by geometry, and are sampled at a scale just below the shape textures. This combats typical normal map sampling issues such as lack of detail in the foreground, or a glassy flat appearance in the background.
The ocean colour comes from a subsurface scattering approximation based on view and normal vectors, and primary light direction.
Foam is shaded in two layers. Underwater bubbles have parallax and refraction offsets to give an impression of depth. White foam on top of the water is shaded with a simple 3D lighting model using procedurally generated normals.
Transparency and refraction are also supported, and Schlick's fresnel approximation selects between the refracted colour and a reflected colour. There is an option on the material to boost specular highlights by adding directional lighting if needed.
All shading features are on static switches and can be disabled if not required.
As an area of future work, the branch fx_test explores dynamically generating spray particle effects by randomly sampling points on the surface to detect wave peaks.
On startup, the OceanBuilder script creates the ocean geometry as a LODs, each composed of geometry tiles and a shape camera to render the displacement texture for that LOD.
At run-time, the viewpoint is moved first, and then the Ocean object is placed at sea level under the viewer. A horizontal scale is computed for the ocean based on the viewer height, as well as a _viewerAltitudeLevelAlpha that captures where the camera is between the current scale and the next scale (x2), and allows a smooth transition between scales to be achieved using the two mechanisms described in the SIGGRAPH course.
Once the ocean has been placed, the ocean surface shape is generated by rendering Gerstner wave components into the shape LODs. These are visualised on screen if the Show shape data debug option is enabled. Each wave component is rendered into the shape LOD that is appropriate for the wavelength, to prevent over- or under- sampling and maximize efficiency. A final pass combines the results down the shape LODs (from largest to most-detailed), disable the Shape combine pass debug option to see the shape contents before this pass.
The ocean geometry is rendered with the Ocean shader. The vertex shader snaps the verts to grid positions to make them stable. It then computes a lodAlpha which starts at 0 for the inside of the LOD and becomes 1 at the outer edge. It is computed from taxicab distance as noted in the course. This value is used to drive the vertex layout transition, to enable a seemless match between the two. The vertex shader then samples the current LOD shape texture and the next shape texture and uses lodAlpha to interpolate them for a smooth transition across displacement textures. A foam value is also computed using the determinant of the Jacobian of the displacement texture. Finally, it passes the LOD geometry scale and lodAlpha to the pixel shader.
The ocean pixel shader samples normal maps at 2 different scales, both proportional to the current and next LOD scales, and then interpolates the result using lodAlpha for a smooth transition. Two layers of foam are added based on different thresholds of the foam value, with black point fading used to blend them.