import Base import ./format.bend as Fmt # num.bend: the contract a float implementation meets. # # import ./num.bend as Num # # An implementation is a carrier type T, the exact value each T stands # for, and its operations. It is correctly rounded for its format when # each operation's result stands for the spec result on the operands' # values (AddOk, MulOk). Hardware F32 and the software SF32 are both # implementations over the carrier F32 (see sf32.bend); anyone can add # another carrier, format or algorithm, and code written against Impl # works with all of them. type Impl<-T: Data> is Type: Impl{fmt: Fmt.Format, value: T -> Fmt.Val, add: T -> T -> T, mul: T -> T -> T} def Impl.fmt(-T: Data, i: Impl) -> Fmt.Format: match i: case Impl{fmt, value, add, mul}: fmt def Impl.value(-T: Data, i: Impl, x: T) -> Fmt.Val: match i: case Impl{fmt, value, add, mul}: value(x) def Impl.add(-T: Data, i: Impl, a: T, b: T) -> T: match i: case Impl{fmt, value, add, mul}: add(a, b) def Impl.mul(-T: Data, i: Impl, a: T, b: T) -> T: match i: case Impl{fmt, value, add, mul}: mul(a, b) # add is correctly rounded at a and b def AddOk(~T: Data, ~i: Impl, a: T, b: T) -> Type: {Fmt.canon(Impl.value(T, i, Impl.add(T, i, a, b))) == Fmt.canon(Fmt.spec_add(Impl.fmt(T, i), Impl.value(T, i, a), Impl.value(T, i, b))) : Fmt.Val} # mul is correctly rounded at a and b def MulOk(~T: Data, ~i: Impl, a: T, b: T) -> Type: {Fmt.canon(Impl.value(T, i, Impl.mul(T, i, a, b))) == Fmt.canon(Fmt.spec_mul(Impl.fmt(T, i), Impl.value(T, i, a), Impl.value(T, i, b))) : Fmt.Val} # an implementation that is correctly rounded everywhere type Correct<-T: Data, -i: Impl> is Type: Correct{add: @a: T -> @b: T -> AddOk(~T, ~i, a, b), mul: @a: T -> @b: T -> MulOk(~T, ~i, a, b)}