Shader variants and prewarming: reduce first-use stalls deliberately

Separate three decisions

A shader can build successfully and still cause a pause when first used. Unity documents a driver step that creates a GPU-specific variant when rendering first needs it. [1] Treat selecting variants, keeping required variants in the build and preparing runtime pipelines as separate decisions. A small executable is not evidence that every required material state survived stripping.

Count the combinations

Consider a deliberately simplified shader pass with four independent binary feature switches. The combination count is 2 × 2 × 2 × 2 = 16 before additional passes, stages or target configurations. If the product only allows three documented combinations, the other combinations are candidates for removal, provided no runtime transition needs them.

Make that allowed-state list explicit. Include quality changes, equipment swaps and dynamically loaded materials. Unity's stripping guidance warns that removing a required variant can make the engine select a similar available variant instead. [2] Visual correctness therefore belongs in the stripping check.

Understand pipeline preparation

A pipeline state object, or PSO, includes more than a shader name. Unreal's precaching documentation includes vertex information and rendering state, and distinguishes missed requests from requests that finish too late. Its compilation runs asynchronously and consumes CPU and memory resources. [3] Preparing one material in one preview scene does not establish coverage for every draw configuration.

Design a practical warmup experiment

Imagine a game with a hub, an ice arena and a vehicle garage. Build a representative first-use sequence for each area. Exercise the allowed material states, relevant mesh layouts, shadow participation and target output configuration. Record which pipeline preparations happen before entry and which still occur during play.

Compare a run with a cold driver cache against a separate warmed run. Keep graphics API, driver, build and scene route fixed. Measure loading time, memory pressure and slow frames together. A warmup that removes an arena hitch but excessively delays entry or competes with asset loading may require a smaller staged set. These are proposed experiments, not performance measurements.

Keep failure visible

Preserve a diagnostic build that exposes missing variants and pipeline misses. Check both the first encounter and later transitions, including a quality change after loading. Do not accept a fallback material merely because the game stopped pausing.

Cache coverage changes with content, renderer configuration and target hardware. Revalidate the first-use sequence after those changes. Prewarming moves or avoids some preparation work; it cannot guarantee that streaming, synchronization or unrelated gameplay work will never produce a long frame.

Sources

[1] Shader loading — Unity Technologies.

[2] Strip shader variants — Unity Technologies.

[3] PSO Precaching — Epic Games.