On 01/28/2013 05:29 PM, Tuomo Keskitalo wrote:
On 01/27/2013 05:47 PM, Andreas Schwab wrote:
The ode-initval2 test no longer fails, but others remain.
Hmm. I removed part of code which divided and multiplied a set of
values by the same constant (doing stuff in between). Rhys said earlier:
> FWIW, I am aware of FMA gives the wrong answer for comparisons like
> (a*x + b*y) == 0 when x == y and b == -a. The fix is never check
> strict equality for floats and instead check if |a*x + b*y| < tol.
> The problem in those cases is not the use of FMAs but rather the code
> being incorrect in the presence of floating point arithmetic.
So, does this mean that whenever an algorithm relies on a=a+a-a or
a=a*a/a, the algorithm is going to fail when high precision is
required? I don't know details about sf, poly or cfd functions, but
there might be a few problematic places.
I wonder if you should post separate bug reports against those, this
one was for ode-initval2 only.
Sorry, meant a=a+b-b and a=a*b/b.
No. You need both addition and multiplication to involve
FusedMultiplyAdd. No FMA will take place in a==a*b/b. It might fail or
not, that depends on specific numbers and rounding mode.
Unless I missed some new floating point instructions or new quirks of
compiler optimization.
