Skip to main content

How luma.gl fits together

The luma.gl layers solve different parts of one GPU application. They compose, but they are not mandatory steps: a small application can use Core directly, an ordinary rendered scene often starts with Engine, and GPU scheduling becomes useful only when the GPU workflow itself needs scheduling.

Connect data and shaders through Model.

Geometry describes attributes, ShaderInputs owns module values, and Model manages pipelines, bindings, draw state, and redraw needs.

One rendered application, four views​

Consider a colored triangle whose color behavior is reusable:

  1. Shadertools defines the color module, its typed props, dependencies, and WGSL/GLSL source.
  2. Engine gives Model the geometry and ShaderInputs, creates compatible bindings, and tracks whether another draw is needed.
  3. Core owns the buffer, shader and pipeline resources, render pass, command encoding, submission, presentation, and destruction.
  4. GPU scheduling, if the application later adds culling, compaction, indirect drawing, and analysis, schedules those operations and their intermediate resources.

Move up or down deliberately​

If you need…Start with…Move when…
Reusable hooks, injections, or shader dependenciesShadertoolsEngine must bind inputs and draw
Geometry, models, redraw, picking, animation, or passesEngineCore-level resource control is necessary
Buffers, textures, pipelines, passes, and submissionCoreSeveral GPU operations need shared scheduling
Hazards, transient aliasing, indirect work, or frame budgetsGPU schedulingA single pass or model is simpler without a graph

The WebGPU ideas that unlock graphs​

GPU scheduling is not mysterious machinery. A few WebGPU capabilities make the design practical:

  • Storage buffers let compute stages share large structured results without CPU copies.
  • Indirect draw and dispatch let GPU-produced counts control later work.
  • Explicit command encoding makes work order and resource access visible enough to validate and schedule.
  • Compute shaders make scans, compaction, sorting, indexing, and aggregation first-class GPU work.

GPU scheduling names and composes those capabilities. WebGL applications can still use Core, Engine, and Shadertools without adopting the graph layer.

Where to continue​