https://gcc.gnu.org/bugzilla/show_bug.cgi?id=93009
Matthias Hochsteger <matthias.hochsteger at tuwien dot ac.at> changed:
What |Removed |Added
----------------------------------------------------------------------------
Status|RESOLVED |UNCONFIRMED
Resolution|INVALID |---
--- Comment #5 from Matthias Hochsteger <matthias.hochsteger at tuwien dot
ac.at> ---
Thanks for the fast replies. Anyway, I think there was a misunderstanding. The
issue is not about accuracy of fma vs. mult+add.
The attached code should clarify the issue (I still couldn't simplify it much
though). It basically boils down to a single call of multiplyAndAdd:
template <typename T>
T multiplyAndAdd(T a, T b, T c)
{
return a*b+c;
}
template <class S>
__attribute__ ((__always_inline__)) inline S P1(S x) const
{
cout << "a = " << S(coefsal[1][0]) << endl;
cout << "b = " << S(x) << endl;
cout << "c = " << S(coefsal[1][1]) << endl;
auto res = multiplyAndAdd (S(coefsal[1][0]),S(x),S(coefsal[1][1]));
cout << "res: " << res << endl;
return res;
}
The data type is "AutoDiffRec<3, SIMD<double, 2>>", which basically contains 4
_m128d values.
>$ g++ -std=c++17 -march=skylake-avx512 -O1 test_fma.ii && ./a.out
> a = 1 1, D = 0 0 0 0 0 0
> b = 3 4, D = 0 0 0 0 0 0
> c = 2 2, D = 0 0 0 0 0 0
> res: 5 6, D = 0 0 0 0 0 0
>$ g++ -std=c++17 -march=skylake-avx512 -O1 -fexpensive-optimizations
>test_fma.ii && ./a.out
> a = 1 1, D = 0 0 0 0 0 0
> b = 3 4, D = 0 0 0 0 0 0
> c = 2 2, D = 0 0 0 0 0 0
> res: 3 3, D = 0 0 0 0 0 0
