#asahi · 2026-09-12 22:44
linkmauveIt is marked #[inline] though.#asahi · 2026-09-12 22:43
linkmauveOh uh, _mm_hadd_ps() doesn’t get inlined? That seems like an omission in the core crate.#asahi · 2026-09-12 22:38
linkmauveIt’d be more about whether the computed values are correct than about performances atm.#asahi · 2026-09-12 22:34
chadmed(can you tell i dropped engineering after a year of undergrad lol)#asahi · 2026-09-12 22:34
chadmedim not entirely sure how youd go about writing a synthetic test for the actual algorithm's performance#asahi · 2026-09-12 22:33
chadmedif the benchmark completes, then it (most likely) works#asahi · 2026-09-12 22:33
linkmauveWith some tests I’d be much more confident that things are working, but otherwise it’s pretty much the exact same algorithm as the NEON one.#asahi · 2026-09-12 22:32
chadmedidk anything about x86 simd intrinsics so im just gonna say lgtm #asahi · 2026-09-12 22:31
linkmauveNew version, with both SSE and NEON versions: https://linkmauve.fr/files/hilbert_iir.rs#asahi · 2026-09-12 21:46
chadmedif this would be useful to folks on not-apple silicon macs then i wouldnt say no to optimisations to make their lives better#asahi · 2026-09-12 21:46
linkmauveInteresting.#asahi · 2026-09-12 21:46
chaos_princess(that is kinda why muvm has x86 support to some extent)#asahi · 2026-09-12 21:46
chadmedlinkmauve: the great thing about FLOSS is that you can just do that #asahi · 2026-09-12 21:45
chaos_princessCI. github actions is x86.#asahi · 2026-09-12 21:45
linkmauveYes. :3#asahi · 2026-09-12 21:45
chaos_princessdo you _want_ to do it and are looking for an excuse? (ready to provide one)#asahi · 2026-09-12 21:44
linkmauveWould it also be useful to write a SSE2 or AVX version btw, for x86 computers?#asahi · 2026-09-12 21:43
linkmauveIdeally we’d have tests though. ^^'#asahi · 2026-09-12 21:42
chadmedlinkmauve: neat! by all means please file a PR, this looks like a good change :)#asahi · 2026-09-12 17:44
chaos_princesssince i am stupid and can't resist - the thing above is not malware, just some kind of a weird scam? Don't bother anyway#asahi · 2026-09-12 17:32
L00pers630 url sqli H00pS https://bin.disroot.org/?72ded627b80bb968#FLRdPsUQz7E6HC1Bqs3iX3tCQbvjSJvs5X4Tb5zQHRUv#asahi · 2026-09-12 15:32
linkmauveBut I don’t have access to any Apple hardware newer than G4, so…#asahi · 2026-09-12 15:32
linkmauveI could do a PR later if you’re still interested, if that improves things in any way on your target hardware.#asahi · 2026-09-12 15:31
linkmauveBut this would only help slightly with cache locality, and probably not at all with actual throughput.#asahi · 2026-09-12 15:29
linkmauveld1/2/3/4 are some pretty fun ARM instructions, they deinterleave arrays, and st1/2/3/4 are the instructions to interleave them back.#asahi · 2026-09-12 15:28
linkmauveAnother slight optimization could be to use an array of Complex, and use ld2 instead of ld1 to deinterleave into two registers at once.#asahi · 2026-09-12 15:27
linkmauveAlso, the loops get fully unrolled, it might not be the best depending on the icache and prefetch characteristics of the core.#asahi · 2026-09-12 15:26
linkmauveWe could do away with the asserts by changing from an array of slices to an array of arrays, but those shouldn’t be that bad probably.#asahi · 2026-09-12 15:21
linkmauveAnyway, here is my experiment, maybe the difference will be faster on Apple CPUs: https://linkmauve.fr/files/hilbert_iir.rs#asahi · 2026-09-12 15:20
linkmauveAlso you don’t have any test, so I don’t know if I got it right or not.#asahi · 2026-09-12 15:16
linkmauveThis function was completely scalar before.#asahi · 2026-09-12 15:15
linkmauveOn this phone it seems bottlenecked by memory bandwidth, not by the CPU (at least not much).#asahi · 2026-09-12 15:15
linkmauveJamesCalligeros[m], chadmed, autovectorization was indeed not working at all in the hottest function, Filter::process(). I just wrote some NEON intrinsics to speed things up and the example went from ~830ms on my Cortex-A53 phone down to ~775ms.#asahi · 2026-09-12 08:05
monthebiffmight be just skyrim things.#asahi · 2026-09-12 08:05
monthebiffoh one thing i did notice with games/ skyrim... trying to render at 1080p was weird, you could do a 1080p window and then full screen it, but the UI was still locked a the 1080 dimensions... and the visual quality didn't seem to be any different#asahi · 2026-09-12 07:58
monthebifflooks like 1200 bucks AUD for an m3 air... mileage varies on ram / storage configs.#asahi · 2026-09-12 07:55
monthebiffSD card slot maybe, especially when i start playing with my cameras again#asahi · 2026-09-12 07:54
monthebiffkind of a blocker for modding stuff to not have that off the bat#asahi · 2026-09-12 07:54
monthebiffi stopped trying when launching skyrim script extender failed.#asahi · 2026-09-12 07:54
monthebifftrying to think what i'd miss off the pro vs the air... the display?#asahi · 2026-09-12 07:54
JamesCalligeros[m]Can you even do that for mods on nexus#asahi · 2026-09-12 07:53
JamesCalligeros[m]They might work if you install them manually#asahi · 2026-09-12 07:52
monthebiffno mods though, didn't seem to get whatever that required working#asahi · 2026-09-12 07:52
monthebiffwhenever ive booted skyrim on this m1pro its seemed happy#asahi · 2026-09-12 07:51
JamesCalligeros[m]Bugthesda games barely run properly on actual pcs :p#asahi · 2026-09-12 07:51
monthebiffI will never let go of that dream#asahi · 2026-09-12 07:50
monthebiffyay!#asahi · 2026-09-12 07:50
monthebiffand when it works, world of warcraft#asahi · 2026-09-12 07:50
monthebiff... skyrim#asahi · 2026-09-12 07:50
chadmedi find it noticeably better for daily driver stuff that isnt heavily threaded but yeah emerge -auvDN @world is.... not quick