minor tweak to pow accuracy and tests (JuliaLang#48233)

* minor tweak to pow accuracy and tests
marcom · Jan 16, 2023 · 687433b · 687433b
1 parent f6a5de5
commit 687433b
Show file tree

Hide file tree

Showing 2 changed files with 15 additions and 8 deletions.
diff --git a/base/special/exp.jl b/base/special/exp.jl
@@ -238,8 +238,10 @@ end
     r = muladd(N_float, LogBo256L(base, T), r)
     k = N >> 8
     jU, jL = table_unpack(N)
-    very_small = muladd(jU, expm1b_kernel(base, r), jL)
-    small_part =  muladd(jU,xlo,very_small) + jU
+    kern = expm1b_kernel(base, r)
+    very_small = muladd(kern, jU*xlo, jL)
+    hi, lo = Base.canonicalize2(1.0, kern)
+    small_part = fma(jU, hi, muladd(jU, (lo+xlo), very_small))
     if !(abs(x) <= SUBNORM_EXP(base, T))
         x >= MAX_EXP(base, T) && return Inf
         x <= MIN_EXP(base, T) && return 0.0

diff --git a/test/math.jl b/test/math.jl
@@ -1352,9 +1352,9 @@ end
 
 @testset "pow" begin
     # tolerance by type for regular powers
-    POW_TOLS = Dict(Float16=>[.51, .51, 2.0, 1.5],
-                    Float32=>[.51, .51, 2.0, 1.5],
-                    Float64=>[1.0, 1.5, 2.0, 1.5])
+    POW_TOLS = Dict(Float16=>[.51, .51, .51, 2.0, 1.5],
+                    Float32=>[.51, .51, .51, 2.0, 1.5],
+                    Float64=>[.55, 0.8, 1.5, 2.0, 1.5])
     for T in (Float16, Float32, Float64)
         for x in (0.0, -0.0, 1.0, 10.0, 2.0, Inf, NaN, -Inf, -NaN)
             for y in (0.0, -0.0, 1.0, -3.0,-10.0 , Inf, NaN, -Inf, -NaN)
@@ -1372,9 +1372,11 @@ end
             got, expected = x^y, widen(x)^y
             if isfinite(eps(T(expected)))
                 if y == T(-2) # unfortunately x^-2 is less accurate for performance reasons.
-                    @test abs(expected-got) <= POW_TOLS[T][3]*eps(T(expected)) || (x,y)
-                elseif y == T(3) # unfortunately x^3 is less accurate for performance reasons.
                     @test abs(expected-got) <= POW_TOLS[T][4]*eps(T(expected)) || (x,y)
+                elseif y == T(3) # unfortunately x^3 is less accurate for performance reasons.
+                    @test abs(expected-got) <= POW_TOLS[T][5]*eps(T(expected)) || (x,y)
+                elseif issubnormal(got)
+                    @test abs(expected-got) <= POW_TOLS[T][2]*eps(T(expected)) || (x,y)
                 else
                     @test abs(expected-got) <= POW_TOLS[T][1]*eps(T(expected)) || (x,y)
                 end
@@ -1385,7 +1387,7 @@ end
             x=rand(T)*floatmin(T); y=rand(T)*3-T(1.2)
             got, expected = x^y, widen(x)^y
             if isfinite(eps(T(expected)))
-                @test abs(expected-got) <= POW_TOLS[T][2]*eps(T(expected)) || (x,y)
+                @test abs(expected-got) <= POW_TOLS[T][3]*eps(T(expected)) || (x,y)
             end
         end
         # test (-x)^y for y larger than typemax(Int)
@@ -1396,6 +1398,9 @@ end
     # test for large negative exponent where error compensation matters
     @test 0.9999999955206014^-1.0e8 == 1.565084574870928
     @test 3e18^20 == Inf
+    # two cases where we have observed > 1 ULP in the past
+    @test 0.0013653274095082324^-97.60372292227069 == 4.088393948750035e279
+    @test 8.758520413376658e-5^70.55863059215994 == 5.052076767078296e-287
 end
 
 # Test that sqrt behaves correctly and doesn't exhibit fp80 double rounding.