#FLOATING

#Start Here

• New user: Getting Started
• Choose a package: Package Guide
• Copy a minimal example: Quick Start
• Understand algorithms and boundaries: Architecture
• Check numerical claims: Verification
• See 0.8.0 changes: CHANGELOG
• Read the optimization evidence: 0.7.1 audit
• Other languages: 简体中文 · 日本語

#Install

moon add Luna-Flow/floating@0.8.0

import {
  "Luna-Flow/floating/bin_float"
  "Luna-Flow/floating/decimal"
  "Luna-Flow/floating/decimal_gda"
  "Luna-Flow/floating/ball_float"
}

#Quick Start

///|
test "floating 0.8.0 quick start" {
  let binary = @bin_float.BinFloat::make(
    @bin_float.BinCoeff::from_uint64(3UL),
    -1,
    53,
  )
  inspect(binary.to_double(), content="1.5")

  let context = @decimal.DecimalContext::decimal64()
  let (decimal, flags) = @decimal.Decimal::from_string_ctx("12.3400", context)
  inspect(decimal.quantum(), content="-4")
  inspect(flags.has_error(), content="false")

  let interval = @ball_float.BallFloat::from_bounds(
    @bin_float.BinFloat::from_int(1, precision=53),
    @bin_float.BinFloat::from_int(2, precision=53),
  )
  inspect(interval.contains(binary), content="true")
}

#Package Guide

#0.8.0 At A Glance

• BinFloat, Decimal, and BallFloat expose certified elementary-function paths with bounded refinement and structured certification failure.
• Binary and decimal coefficient kernels use target-specific, exact-fallback dispatch across schoolbook, Karatsuba, Toom-3, NTT, block division, and reciprocal algorithms.
• decimal and decimal_gda are independent state models: IEEE per-operation flags are not GDA sticky status/traps.
• BinFloat implements the contextual arithmetic traits, so binary, IEEE decimal, and GDA decimal all compose through the same ArithmeticContext.
• Converting an ArithmeticContext into a binary or decimal context now carries e_min, e_max, and clamp; contextual operations honour the caller's exponent range instead of running unbounded.
• ball_float covers the declared strict IEEE 1788 phases with bare/decorated intervals, critical-point/pole handling, and conservative total fallbacks.
• Benchmarks moved into the unified bench/* Maremark hierarchy with explicit crossover and regression analysis.
• The 0.7.1 optimization audit records exact-kernel, directed-rounding, and interval-monotonicity proofs for the optimized paths.
• The native benchmark artifact covers all four core suites; non-monotonic auto-tune observations remain evidence only until independently replicated.

• Binary conformance · performance
• IEEE decimal conformance · performance
• GDA decimal conformance · performance
• Interval conformance · performance
• Elementary capability matrix

#Development

just pr 8

just fmt
just docs
just gate binary 8
just gate decimal 8
just gate decimal_gda 8
just gate interval 8
just bench bin-float --target native
just bench auto-tune --target native

just conformance smoke binary
just conformance run decimal --run-target native --run-target wasm
just conformance run interval --phase trigonometric --strict-supported

just ci 8

#License