Underwater Caustics & Lighting

Underwater lighting reads as fake the moment the caustics stop agreeing with the water above them. The usual shortcut — a looping pre-baked caustic texture — breaks as soon as the surface moves, the light angle changes, or the camera goes under.

I wanted caustics that are actually derived from the live water surface, running fast enough to be usable in a real frame budget, and I wanted to be able to prove the cost rather than guess at it.

The surface is driven by a Gerstner wave simulation that runs entirely in a compute shader, producing a normal map for the current frame. Keeping it on the compute queue means the CPU is not doing per-vertex wave math every frame.

That normal map feeds the caustics pass, which projects light through the surface via a light matrix and samples it in the fragment shader on submerged geometry. Explicit memory barriers synchronise the compute output with the graphics passes that consume it.

// Colour attachment for the caustics resolve pass
VkAttachmentDescription colorAttachment{};
colorAttachment.format         = swapChainImageFormat;
colorAttachment.samples        = VK_SAMPLE_COUNT_1_BIT;
colorAttachment.loadOp         = VK_ATTACHMENT_LOAD_OP_CLEAR;
colorAttachment.storeOp        = VK_ATTACHMENT_STORE_OP_STORE;
colorAttachment.stencilLoadOp  = VK_ATTACHMENT_LOAD_OP_DONT_CARE;
colorAttachment.stencilStoreOp = VK_ATTACHMENT_STORE_OP_DONT_CARE;

Rather than eyeball the framerate, I built a testing harness into the renderer: selectable water modes, scripted test suites, and automatic CSV export. Each run reports mean and median FPS, 1% lows, mean frame time, standard deviation, and how many frames were rejected as outliers.

That made the quality-versus-cost trade-off an actual measurement. Two configurations, both over 95 valid frames with zero outliers:

The two runs use different test suites, so they describe two configurations rather than a single before/after. The point of the harness is that the numbers are repeatable and exportable — I can re-run a suite after a change and compare like for like.

Everything is live-tunable through an ImGui layer: sun, secondary and ambient lighting, rim light, water surface and underwater parameters, particles, effects, and a wireframe toggle. Being able to isolate one term while the scene runs is what made the lighting tractable to debug.