FunctionAccuracyTests.jl

ULP testing for Floating Point special functions.

This package provides 2 useful things for testing the accuracy of functions. A FloatIterator type, and a function test_acc.

• FloatIterator{T}(min, max)

Construct an indexable iterator over all floating point values between min and max. This is useful for testing Float16 and Float32 implimentations of functions. For relatively cheap functions, testing all Float16 values takes about .1 second, and testing all Float32 values takes around 90 seconds. Do not use this for testing Float64 inputs.

test_acc - Function

• test_acc(Dict(test_fun=>reference_fun), xs; tol=1.5)

This tests whether test_fun and reference_fun produce the same result on an iterator of x values. tol is the tolerance specified in ULPs (Units in Last Place). The reference_fun should be the accurate "reference" function you're comparing your new "test" function with.

• test_acc(f::Function, xx; tol = 1.5)

This is the same, except it tests the BigFloat method of a function against the method for eltype(xx).

• test_acc(f::Function, min, max; tol = 1.5)

This tests whether all floating point values between min and max match.

• test_acc(f::Function, T::Type; tol = 1.5)

This tests whether all values of a Floating Point type match.