log1pexp

src/basicfuns.jl

@@ -153,8 +153,8 @@ Return `log(1+exp(x))` evaluated carefully for largish `x`.
 This is also called the ["softplus"](https://en.wikipedia.org/wiki/Rectifier_(neural_networks))
 transformation, being a smooth approximation to `max(0,x)`. Its inverse is [`logexpm1`](@ref).
 """
-log1pexp(x::Real) = x < 18.0 ? log1p(exp(x)) : x < 33.3 ? x + exp(-x) : oftype(exp(-x), x)
-log1pexp(x::Float32) = x < 9.0f0 ? log1p(exp(x)) : x < 16.0f0 ? x + exp(-x) : oftype(exp(-x), x)
+log1pexp(x::Real) = x ≤ -37 ? exp(x) : x ≤ 18 ? log1p(exp(x)) : x ≤ 33.3 ? x + exp(-x) : float(x)
+log1pexp(x::Float32) = x < 9f0 ? log1p(exp(x)) : x < 16f0 ? x + exp(-x) : oftype(exp(-x), x)
 
 """
 $(SIGNATURES)

test/basicfuns.jl

@@ -110,15 +110,17 @@ end
 # log1pexp, log1mexp, log2mexp & logexpm1
 
 @testset "log1pexp" begin
-    @test log1pexp(2.0)    ≈ log(1.0 + exp(2.0))
-    @test log1pexp(-2.0)   ≈ log(1.0 + exp(-2.0))
-    @test log1pexp(10000)  ≈ 10000.0
-    @test log1pexp(-10000) ≈ 0.0
-
-    @test log1pexp(2f0)      ≈ log(1f0 + exp(2f0))
-    @test log1pexp(-2f0)     ≈ log(1f0 + exp(-2f0))
-    @test log1pexp(10000f0)  ≈ 10000f0
-    @test log1pexp(-10000f0) ≈ 0f0
+    # test every branch
+    for x in (0, 1, 2, 10, 20, 40), T in (Int, Float32, Float64)
+        @test (@inferred log1pexp(-T(x))) ≈ log1p(exp(big(-x)))
+        @test (@inferred log1pexp(+T(x))) ≈ log1p(exp(big(+x)))
+    end
+
+    # large arguments
+    @test (@inferred log1pexp(1e4)) ≈ 1e4
+    @test (@inferred log1pexp(1f4)) ≈ 1f4
+    @test iszero(@inferred log1pexp(-1e4))
+    @test iszero(@inferred log1pexp(-1f4))
 end
 
 @testset "log1mexp" begin