> >> We won't get zero from exponential function, so expecting zero result > >> is flawed anyway. > >> > >> If we would like to introduce universal epsilon comparisons into the > >> testsuite, then please read [1]. Being overly pedantic, the definition > >> should be "|(v[i] - u.a[i]) / v[i]|", as stated in [2]. > >> > >> [1] > >> http://randomascii.wordpress.com/2012/02/25/comparing-floating-point-numbers-2012-edition/ > >> [2] http://en.wikipedia.org/wiki/Relative_error > >> > > > > We get zero from testing zero-masking. Currently we produce 0/0 = NaN. > > Comparison with NaN is always false, so tests pass. But I think that > > this should be fixed to avoid division by zero. As for being more > > pedantic about comparison, I doubt that its useful, when we use > > 0.0001 as eps. > > In this case, please add simple check for zero, with the above > comment. We don't test exp function, but masking. >
Something like this? --- gcc/testsuite/gcc.target/i386/m512-check.h | 10 ++++++++++ 1 file changed, 10 insertions(+) diff --git a/gcc/testsuite/gcc.target/i386/m512-check.h b/gcc/testsuite/gcc.target/i386/m512-check.h index 3209039..a96a103 100644 --- a/gcc/testsuite/gcc.target/i386/m512-check.h +++ b/gcc/testsuite/gcc.target/i386/m512-check.h @@ -58,6 +58,16 @@ check_rough_##UINON_TYPE (UINON_TYPE u, const VALUE_TYPE *v, \ \ for (i = 0; i < ARRAY_SIZE (u.a); i++) \ { \ + /* We will always have v[i] == 0 == u.a[i] for some i, \ + when we test zero-masking. */ \ + if (v[i] == 0.0 && u.a[i] == 0.0) \ + continue; \ + if (v[i] == 0.0 && u.a[i] != 0.0) \ + { \ + err++; \ + PRINTF ("%i: " FMT " != " FMT "\n", \ + i, v[i], u.a[i]); \ + } \ VALUE_TYPE rel_err = (u.a[i] - v[i]) / v[i]; \ if (((rel_err < 0) ? -rel_err : rel_err) > eps) \ { \ -- 1.8.3.1