Survey of Floating Point Implementations for Maximum

float_max_survey.md

Survey of Floating Point Implementations for Maximum

Following some IEEE Standard

IEEE Std 754-2008 maxNum (NaN Absorb, 0 implementation defined)

• C17 fmax
• Rust f64::max
• LLVM llvm.maxnum
• ARM FMAXNM (either mode)

IEEE Std 754-2019 maximum (NaN Prop, -0 < +0)

• EcmaScript Math.max
• WebAssembly fmax
• ARM FMAX (canonical mode)
• Java Math.max
• Go math.Max (doesn't have -0 literals?!)
• LLVM llvm.maximum
• Rust f64::maximum

IEEE Std 754-2019 maximumNumber (NaN Absorb, -0 < +0)

• ARM FMAXNM (canonical mode)

Following a fast implementation

if a < b { b } else { a } (prefer left hand side)

• C++20 std::max
• Python math.max
• WebAssembly SIMD pmax

if b < a { a } else { b } (prefer right hand side)

• Intel x86 MAXPS

Others

NaN Prop, -0 == +0, prefer right hand side

• C# Math.Max
• ARM FMAX (ARMv8.7-A FEAT_AFP + AH bit = 1)

Ignore NaN, prefer left hand side, except for -0 and +0, then prefer right hand side

• Swift max