Commit graph

1 commit

Author SHA1 Message Date
Fredrik Johansson
0244fe0e46 Fix evil bug in floating-point arithmetic and strengthen test code.
When rounding up, the error bound sometimes came out as 1/2 ulp instead of 1 ulp due to the mantissa rounding up to the next power of two, without a corresponding exponent correction. Fortunately, this should not affect ball arithmetic (which always rounds down).
2014-04-18 16:19:53 +02:00