Am 20.05.2017 um 17:54 schrieb William F Pokorny:

> On my I3-4130 two core 4 thread at 3.4GZ, Ubuntu 16.06 and the
> --benchmark, I am seeing:
> 
> Threads==cores (+wt2) : -6.59%
> 
> Threads==threads (+wt4) : -12.40%
> 
> In a bit of a surprise to me the new noise code especially helps where
> threads > cores.

<shrugs>
That's the magic of hyperthreading optimizations.


> However, I did find one case running a couple isosurfaces using the
> f_hetero_mf function where I am seeing a 3.8% slowdown for +wt4. I've
> done no digging. Any ideas as to why this might be so?

`f_hetero_mf` is a noise-based function, so /some/ effect is to be expected.

Maybe on your machine the hand-optimized implementation of the `Noise`
function (which is used by `f_hetero_mf`) happens to be slower than the
compiler-optimized one, while the hand-optimized implementation of the
`DNoise` function (which is used by turbulence warp and similar effects)
is faster than the compiler-optimized one.

Post a reply to this message