21 Apr
2005
21 Apr
'05
8:52 p.m.
Tim Blechmann wrote:
oh i have forgotten to mention that zexy's [matrix~] and [multiline~] were written for exactly this purpose... (although not SIMD-optimized)
and if i remember correctly they are not even loop unrolled / vectorized ... haven't checked the code, though ...
neither is sickle version of matrix~, nor any other sickle clone. When I was experimenting back then with gcc-2.95, the consistent pattern was that -funroll-loops performed slightly better than unrolling by hand.
Wonder about similar measurements with -ftree-vectorize once it matures...
Krzysztof