DECODE_OP

const DECODE_OP:

import { DECODE_OP } from '@driftengine/texture';

In depth

This is the file where the expensive mistake would be made, so the reasoning is here. The decode program is uploaded as a uniform buffer and walked by one shader. There is no code generation, no #define, no shader per material and no variant — because ARCHITECTURE.md has already measured what the other way costs, with the number: a fifth permutation flag took the generated WGSL corpus from 914 KB to 1,906 KB and cost 196,910 gzipped bytes on every consumer, including those who never enabled it, because deflate's window is 32 KB and near-identical permutations do not deduplicate. A decode graph as permutations would be that mistake with a far larger exponent — per material rather than per feature.

As data it costs bytes, linearly, and a new operation is one more case in one switch.

A node is four words: op, a, b, out. Which of a and b name registers rather than immediates is a property of the operation, held in OP_ARGS so that the validator and both interpreters read one table rather than three copies of a convention.