Render

Layer 2. GPU-driven, bindless, deferred-shading renderer. The engine records the entire pipeline each frame; user code mostly chooses material types, light kinds, and a post-process effect stack — the pipeline itself is fixed.

Why deferred + light volumes

Forward shading scales as pixels × lights. The deferred pass writes a G-buffer (position, normal, albedo, metallic-roughness, emissive, depth) once per visible fragment, then runs one fullscreen ambient pass (IBL + BRDF LUT) plus a per-light volume pass for each direct light. A light’s volume (sphere for point, cone for spot, fullscreen triangle for directional) is rasterized with reversed depth test so only fragments actually reached by that light get shaded. Avoids the “shade every pixel against every light” loop without needing a tile or cluster pass.

Bindless + GPU-driven culling

All textures live in one descriptor array; all geometry lives in a handful of giant vertex / index / skinning buffers. A draw is just “indirect dispatch + bindless indices” — there is no per-mesh descriptor binding. Visibility culling runs as a compute pass per camera (frustum + depth-pyramid occlusion) and writes the indirect draw commands the next graphics pass will consume.

Shadows

Two strategies, picked per light type:

• 2D shadow maps for directional + spot lights. One compute pass per light clips draws against the light frustum; one depth-only graphics pass writes a 512² R32_SFLOAT slice.
• Cubemap shadow maps for point lights. One compute pass clips against the light’s sphere; one geometry-shader-amplified draw writes all six faces. Requires -define:REQUIRE_GEOMETRY_SHADER=true at build time.

Shadow buffers are allocated lazily on the first cast and released when the light despawns.

Post-process

Each camera has an effect stack pushed via render.post_process.add_*. Effects run in insertion order inside the POST_PROCESS pass: tonemap, bloom, blur, fog, outline, DoF, crosshatch, grayscale. Built as small fullscreen passes against the camera’s FINAL_IMAGE attachment.

Camera attachments

Each camera owns per-frame Texture2DHandles for every G-buffer slot. mjolnir.get_camera_attachment(engine, cam, .FINAL_IMAGE) returns a bindless handle — feed it to a UI quad or a material to composite that camera’s output anywhere (minimaps, mirrors, render-to-texture surfaces).