Library Flocq.Core.Fcore_FLT

This file is part of the Flocq formalization of floating-point arithmetic in Coq: http://flocq.gforge.inria.fr/
Copyright (C) 2010-2013 Sylvie Boldo
Copyright (C) 2010-2013 Guillaume Melquiond
This library is free software; you can redistribute it and/or modify it under the terms of the GNU Lesser General Public License as published by the Free Software Foundation; either version 3 of the License, or (at your option) any later version.
This library is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the COPYING file for more details.

Floating-point format with gradual underflow

Require Import Fcore_Raux.
Require Import Fcore_defs.
Require Import Fcore_rnd.
Require Import Fcore_generic_fmt.
Require Import Fcore_float_prop.
Require Import Fcore_FLX.
Require Import Fcore_FIX.
Require Import Fcore_ulp.
Require Import Fcore_rnd_ne.

Section RND_FLT.

Variable beta : radix.

Notation bpow e := (bpow beta e).

Variable emin prec : Z.

Context { prec_gt_0_ : Prec_gt_0 prec }.

Definition FLT_format (x : R) :=
   f : float beta,
  x = F2R f (Zabs (Fnum f) < Zpower beta prec)%Z (emin Fexp f)%Z.

Definition FLT_exp e := Zmax (e - prec) emin.

Properties of the FLT format
Links between FLT and FLX
Links between FLT and FIX (underflow)
Theorem canonic_exp_FLT_FIX :
   x, x 0%R
  (Rabs x < bpow (emin + prec))%R
  canonic_exp beta FLT_exp x = canonic_exp beta (FIX_exp emin) x.

Theorem generic_format_FIX_FLT :
   x : R,
  generic_format beta FLT_exp x
  generic_format beta (FIX_exp emin) x.

Theorem generic_format_FLT_FIX :
   x : R,
  (Rabs x bpow (emin + prec))%R
  generic_format beta (FIX_exp emin) x
  generic_format beta FLT_exp x.

Theorem ulp_FLT_small: x, (Rabs x < bpow (emin+prec))%R
    ulp beta FLT_exp x = bpow emin.

Theorem ulp_FLT_le :
   x, (bpow (emin + prec - 1) Rabs x)%R
  (ulp beta FLT_exp x Rabs x × bpow (1 - prec))%R.

Theorem ulp_FLT_gt :
   x, (Rabs x × bpow (-prec) < ulp beta FLT_exp x)%R.

FLT is a nice format: it has a monotone exponent...
and it allows a rounding to nearest, ties to even.
Hypothesis NE_prop : Zeven beta = false (1 < prec)%Z.

Global Instance exists_NE_FLT : Exists_NE beta FLT_exp.

End RND_FLT.