커밋 d894fc0의 AUTHORING.md를 그대로 옮긴 것으로, 본문은 아직 영어입니다. 패키지는 0.1.0에서 쓸 typeshade라는 이름으로 가져옵니다.

GLSL 부동소수점 정밀도

The GLSL ES 3.00 backend emits precision highp float;. Mobile GPUs pay real bandwidth and power for highp where mediump suffices, so emitGlslModule / emitGlslStages take a build-time knob:

const fs = emitGlslModule(m, 'fragment', { floatPrecision: 'mediump' })

'highp' is the default and is byte-neutral — omit the option and you get the exact bytes the backend has always emitted.

Caveats — read before reaching for it.

• It is a whole-stage default, so it covers positions and coordinates too. mediump is ~fp16: roughly 3 decimal digits over a ±65504 range. A projected map coordinate does not survive that; f32 already collapses at deep zoom, which is the entire reason the df64 emulation exists (§7 above, and examples/fp64-deep-zoom.ts for the picture). Use this only for fragment-colour-class shaders whose output is a bounded, low-dynamic-range colour — never for a stage computing a position, a tile/world coordinate, or a df64 lane.
• It touches the float line only. precision highp int; stays highp — the storage→data-texture emulation’s index math and the bitcast lanes need the full int range — and so does the precision highp sampler2DArray; line (§4.5.4, a separate requirement).
• Build-time, not a runtime device probe. Emitted GLSL is cached under a shaderRequestKey with no precision component, so a runtime-varying precision would hand a mediump program to a highp request.
• CI cannot judge the numeric effect, and does not pretend to. The census in playground/e2e/_glsl-compile-gate.spec.ts measured the CI rasterizer (ANGLE/SwiftShader, Vulkan 1.3): getShaderPrecisionFormat advertises MEDIUM_FLOAT as {rangeMin: 15, rangeMax: 15, precision: 10} against HIGH_FLOAT’s {127, 127, 23} — i.e. it claims fp16 — yet a shader compiled under precision mediump float; there behaves as f32: the probe ((1.0 + 2⁻¹²) − 1.0) × 4096.0 reads 255 (survived) on the mediump arm, identical to the highp control. (Either the stack computes mediump at ≥f32, or its compiler reassociates the (1+ε)−1 form — GLSL ES 3.00 has no precise qualifier to forbid that — and the two are indistinguishable from outside, which layer of ANGLE/SwiftShader is responsible included.) A precision format is a declared minimum, not a promise. Either way, no CI pixel gate can distinguish a highp emit from a mediump one; the gates here cover header shape (unit) and compile + link validity (Playwright), and real-device mediump behavior — the actual bandwidth win, the banding it can cause, the range clipping — is an explicit skip, verifiable only on real mobile hardware.

이 페이지 편집