/soft-float-starter-pack

Software implementation of floating point numbers and operations

Primary LanguageC#MIT LicenseMIT

Soft float starter pack

Software implementation of floating point numbers and operations.
Soft floats can be used for deterministic calculations, e.g. for physics simulation.
They will give the same results every time, on every platform, on every processor.

This repository uses the work of:

  • SoftFloat, which implements basic soft float functionality
  • libm, which implements various operations for floating point numbers, including square root, trigonometric functions, transcendental functions, etc. (ported to C#)

How to use

The sfloat type is the main type that you'll need to use for soft float calculations.

The sfloat type can be constructed in three ways:

  • Explicit cast from float:
sfloat a = (sfloat)1.0f;
sfloat b = (sfloat)(-123.456f);
sfloat c = (sfloat)float.PositiveInfinity;
sfloat d = (sfloat)float.NaN;

This cast is basically free, since the internal representations are identical.

  • Explicit cast from int:
sfloat a = (sfloat)1;
sfloat b = (sfloat)(-123);
sfloat c = (sfloat)int.MaxValue;
  • Create from raw byte representation
sfloat a = sfloat.FromRaw(0x00000000); // == 0
sfloat b = sfloat.FromRaw(0x3f800000); // == 1
sfloat c = sfloat.FromRaw(0xc2f6e979); // == -123.456
sfloat d = sfloat.FromRaw(0x7f800000); // == Infinity

This cast is also basically free, it's just the byte representation of the value.

The rest of the operations work just like with floats (addition, multiplication, etc.).
Note that you should always use a float literal (or a variable that was assigned a float literal before) for explicit casts from floats, since any operation done on floats can be non-deterministic.

// OK
float a = 1.0f;
sfloat b = (sfloat)a + (sfloat)123.456f;


// NOT OK
float a = 1.0f;
sfloat b = (sfloat)(a + 123.456f); // <-- float addition here, which may be non-deterministic

Using libm

You can use libm just like a regular mathematics library:

sfloat x = (sfloat)2.0f;
sfloat squareRoot = libm.sqrtf(x);

sfloat c = libm.cosf((sfloat)3.1415f);

sfloat e = libm.expf((sfloat)1.0f);

All functions have the f suffix, which comes from the rust libm implementation. You can rename them if you want to.