We have a start for testing input, output accuracy for ufuncs in the numpy/core/tests/test_umath_accuracy.py tests (thanks @r-devulap!). It is patterned loosely after the CPython cmath tests, which use a data file of values. The test is a good start, here are a few ways it could be improved:

• the data files (for instance here are binary only. Maybe the float value could be in a comment at the end of the line?
• in the CPython cmath tests, every test has a name. The test runs to completion and the failing names are printed
• the test should run as a pytest.mark.parametrize test for each ufunc tested
• currently the tests are AVX only since various glibc implementations of math functions vary. Those variations should either be documented or worked around
• more functions should be tested, and more dtypes as well. Various other tests should be consilidated into this one: like the ones in test_half
• the README about the data files should more clearly document how to generate correct values.
• ufuncs should have a way to report the loop chosen so we don't have to guess which SIMD option is active