Skip to content

test(EDyadicFloat): conformance vs native Float32 and Float (FP64) #792

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
bollu wants to merge 1 commit into
base: edyadicfloat-ops
Choose a base branch
from
Open

test(EDyadicFloat): conformance vs native Float32 and Float (FP64) #792

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
1 change: 1 addition & 0 deletions UnitTest/FP.lean
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -5,3 +5,4 @@ public import UnitTest.FP.EDyadic
public import UnitTest.FP.FastFloat
public import UnitTest.FP.FastFloatSmoke
public import UnitTest.FP.EDyadicFloat
public import UnitTest.FP.EDyadicFloat.ConformanceNative
88 changes: 88 additions & 0 deletions UnitTest/FP/EDyadicFloat/ConformanceNative.lean
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,88 @@
module

import Veir.Data.FP.EDyadicFloat
import Veir.Data.FP.PackedFloat.OfFloat

meta import Veir.Data.FP.EDyadicFloat
meta import Veir.Data.FP.PackedFloat.OfFloat

namespace UnitTest.Fp.EDyadicFloat.ConformanceNative

open Veir.Data.FP

/-! # `EDyadicFloat` vs native IEEE-754

`EDyadicFloat e s .RNE` computes each operation on the **exact** dyadic value and
rounds once to `(e, s)` under round-to-nearest-ties-to-even — the IEEE-754
correctly-rounded result. We check it bit-for-bit against Lean's native `Float32`
(single, `(8, 23)`) and `Float` (double, `(11, 52)`), which are themselves
correctly rounded. Unlike `FastFloat`, there is no internal double-rounding,
so the agreement is exact.

Output is routed back through the packed encoding: `EDyadicFloat → PackedFloat → native`. -/

/-! ## Single precision: `EDyadicFloat 8 23 .RNE` vs `Float32` -/

private def mk32 (f : Float32) : EDyadicFloat 8 23 .RNE :=
EDyadicFloat.ofPackedFloat (PackedFloat.ofFloat32 f)
private def out32 (x : EDyadicFloat 8 23 .RNE) : Float32 :=
PackedFloat.toFloat32 (EDyadicFloat.toPackedFloat x)
private def bits32 (a b : Float32) : Bool := Float32.toBits a == Float32.toBits b

-- addition
#guard bits32 (out32 (mk32 1.0 + mk32 2.0)) (1.0 + 2.0)
#guard bits32 (out32 (mk32 1.5 + mk32 2.5)) (1.5 + 2.5)
#guard bits32 (out32 (mk32 0.1 + mk32 0.2)) (0.1 + 0.2)
#guard bits32 (out32 (mk32 (-1.0) + mk32 1.0)) ((-1.0) + 1.0)
#guard bits32 (out32 (mk32 1.0e10 + mk32 1.0)) (1.0e10 + 1.0)
#guard bits32 (out32 (mk32 1.0e-30 + mk32 1.0e-30)) (1.0e-30 + 1.0e-30)
-- subtraction
#guard bits32 (out32 (mk32 3.0 - mk32 1.0)) (3.0 - 1.0)
#guard bits32 (out32 (mk32 1.0 - mk32 2.0)) (1.0 - 2.0)
#guard bits32 (out32 (mk32 1.0e10 - mk32 1.0)) (1.0e10 - 1.0)
-- multiplication
#guard bits32 (out32 (mk32 1.5 * mk32 2.0)) (1.5 * 2.0)
#guard bits32 (out32 (mk32 0.1 * mk32 0.1)) (0.1 * 0.1)
#guard bits32 (out32 (mk32 1.0e15 * mk32 1.0e15)) (1.0e15 * 1.0e15)
#guard bits32 (out32 (mk32 (-2.0) * mk32 3.0)) ((-2.0) * 3.0)
-- division
#guard bits32 (out32 (mk32 1.0 / mk32 3.0)) (1.0 / 3.0)
#guard bits32 (out32 (mk32 22.0 / mk32 7.0)) (22.0 / 7.0)
#guard bits32 (out32 (mk32 (-1.0) / mk32 4.0)) ((-1.0) / 4.0)
-- specials
#guard bits32 (out32 (mk32 (1.0 / 0.0) + mk32 1.0)) ((1.0 / 0.0) + 1.0)
#guard bits32 (out32 (mk32 0.0 * mk32 0.0)) (0.0 * 0.0)
#guard bits32 (out32 (mk32 (-0.0) + mk32 0.0)) ((-0.0) + 0.0)

/-! ## Double precision: `EDyadicFloat 11 52 .RNE` vs `Float` -/

private def mk64 (f : Float) : EDyadicFloat 11 52 .RNE :=
EDyadicFloat.ofPackedFloat (PackedFloat.ofFloat f)
private def out64 (x : EDyadicFloat 11 52 .RNE) : Float :=
PackedFloat.toFloat (EDyadicFloat.toPackedFloat x)
private def bits64 (a b : Float) : Bool := Float.toBits a == Float.toBits b

-- addition
#guard bits64 (out64 (mk64 1.0 + mk64 2.0)) (1.0 + 2.0)
#guard bits64 (out64 (mk64 0.1 + mk64 0.2)) (0.1 + 0.2)
#guard bits64 (out64 (mk64 (-1.0) + mk64 1.0)) ((-1.0) + 1.0)
#guard bits64 (out64 (mk64 1.0e100 + mk64 1.0)) (1.0e100 + 1.0)
#guard bits64 (out64 (mk64 1.0e-300 + mk64 1.0e-300)) (1.0e-300 + 1.0e-300)
-- subtraction
#guard bits64 (out64 (mk64 3.0 - mk64 1.0)) (3.0 - 1.0)
#guard bits64 (out64 (mk64 1.0e16 - mk64 1.0)) (1.0e16 - 1.0)
-- multiplication
#guard bits64 (out64 (mk64 1.5 * mk64 2.0)) (1.5 * 2.0)
#guard bits64 (out64 (mk64 0.1 * mk64 0.1)) (0.1 * 0.1)
#guard bits64 (out64 (mk64 3.141592653589793 * mk64 2.718281828459045))
(3.141592653589793 * 2.718281828459045)
-- division
#guard bits64 (out64 (mk64 1.0 / mk64 3.0)) (1.0 / 3.0)
#guard bits64 (out64 (mk64 22.0 / mk64 7.0)) (22.0 / 7.0)
#guard bits64 (out64 (mk64 2.0 / mk64 3.0)) (2.0 / 3.0)
-- specials
#guard bits64 (out64 (mk64 (1.0 / 0.0) + mk64 1.0)) ((1.0 / 0.0) + 1.0)
#guard bits64 (out64 (mk64 0.0 * mk64 0.0)) (0.0 * 0.0)
#guard bits64 (out64 (mk64 (-0.0) + mk64 0.0)) ((-0.0) + 0.0)

end UnitTest.Fp.EDyadicFloat.ConformanceNative
Loading