|
02:53
MasterDuke joined
|
|||
| MasterDuke | timo, lizmat, et al.: that leibniz benchmark is interesting. i made some modifications (github.com/niklas-heer/speed-compa...bniz.raku) that sped it up, but across some of the different languages i get surprising results on x86 vs apple silicon | 02:57 | |
| on my laptop (apple silicon), raku takes twice as long as on my desktop (amd x86). but the python version reversed, it takes twice as long on the laptop | 03:00 | ||
| and just changing `$pi = $pi + $x / (2 * $i - 1).Num;` to `$pi = $pi + nqp::div_n($x, (2 * $i - 1));` speeds it up by at least 5x on both laptop and desktop | 03:03 | ||
| and a bunch of the savings is that now there are 0 GCs, where before there were ~8k (doing only 100000000 rounds instead of the default 1000000000, just to get results quicker) | 03:06 | ||
| it'd be great it spesh could figure out that it doesn't need to allocate a new Num each iteration, but can just reuse one | 03:08 | ||
| (oops, typoed the bit about python. it's twice as long on the *desktop*) | 03:09 | ||
| off to bed & | 03:11 | ||
|
03:16
MasterDuke left
05:48
melezhik joined
|
|||
| melezhik | I am going to post summary of fundings. But right now diff head vs 2025.11 does not show much | 05:48 | |
| brw.sparrowhub.io/report/brw-reporter/2957 | 05:49 | ||
| ab5tract | MasterDuke: so wuold spesh essentially install a state variable in the scope and reference that? That also sounds like the kind of thing I would expect a RakuAST based optimizer to be able to do as well | 07:32 | |
|
07:46
Voldenet left
07:47
Voldenet joined
|
|||