D958 - Shader costs are measured from kept translations

Status: decided Date: 2026-10-24 known_by: measured

What a title's shader programs cost is measured from its kept translations, not by running the title. orbistoun-cli shader-census makes each kept translation again from its bytes and context with this build's translator, and records module words against guest instructions and the model chosen. With --compile it has the driver compile each distinct module on up to four threads, with no pipeline cache and outside the compile watch, and times a compute module's whole pipeline through the layout a dispatch uses; a draw's module is timed as a shader module alone. Rows are written as TSV, so two builds compare row by row.

Why: a translation that grew with the wave's width, and a driver compile of minutes, surfaced one program at a time inside forty-minute runs of Grand Theft Auto V. The store already holds every program the title ran with its context, so all of them are measured in minutes - its 1,084 programs in 2 s, their compute pipelines compiled in 3.5 min - and a translator change is confirmed across every program at once.

Rejected:

• Timing compiles inside a run: one program a run, serialised behind the guest.
• Reading sizes from the kept modules: they are the build that made them, not this one.
• Compiling draws' full pipelines: the store does not keep the render state a draw pipeline needs, and inventing it would time a pipeline no run builds.

Amended 2026-10-26: the census also tries each kept compute program as a run would on the device-addressed route (D964), its slots' formats and flags renumbered as that route binds them, and gives every refusal a class: its category, the decisions it cites, its offset and instruction, ranked by how many programs each reaches.