On 12/25/10 3:07 PM, bearophile wrote:
Andrei:
Third, it looks like larger unrolling limits is better - I only got a
plateau at 128!
But this is true for a microbenchmark. In a real program the code half part of
the CPU L1 cache is quite limited, so the more code you have to push through
that little cache (code of different functions), the more cache misses you
have, and this slows down the code. This is why too much unrolling or too much
inlining is bad, and this is why I have unrolled my sum() only once.
Yah, that's what I think unroll should be a generic function leaving it
to the user to choose the parameters. Before that I'd like to generalize
the function a bit more - right now it can only do reduce-type workloads
on associative functions.
Andrei
