i stopped hyperoptimizing a long time ago, it just isn't worth the effort 90% of the time if it work

i stopped hyperoptimizing a long time ago, it just isn't worth the effort 90% of the time
if it works and its not stupid slow then its good

TechPizza•10/20/23, 10:56 PM

it does not get vectorized per say. modern software doesn't use x87

AAqua i stopped hyperoptimizing a long time ago, it just isn't worth the effort 90% of...

Maxine•10/20/23, 10:56 PM

i don't optimize anything unless it's for shit like this

Lyris the Kitori•10/20/23, 10:57 PM

math stuff like this you want optimized

Maxine•10/20/23, 10:57 PM

where like yeah, you want this to have as little overhead as regular Vector3

MMaxine are you saying the non-specialized version gets vectorized?

Perksey•10/20/23, 10:57 PM

which you is this

Maxine•10/20/23, 10:57 PM

you perks

Maxine•10/20/23, 10:57 PM

forgor to reply

Perksey•10/20/23, 10:57 PM

i was not saying that

Perksey•10/20/23, 10:57 PM

i gave you a specialised version

Maxine•10/20/23, 10:57 PM

yeah i'm using that one

Perksey•10/20/23, 10:57 PM

JIT cannot detect SIMD patterns today

Maxine•10/20/23, 10:58 PM

sadge

Maxine•10/20/23, 10:58 PM

then again neither can LLVM half the time

MMaxine i don't optimize anything unless it's for shit like this

AquaOP•10/20/23, 10:58 PM

yeah i get it, for math you want things as fast as you can

AquaOP•10/20/23, 10:58 PM

write the math library in C and call into it with p/invoke when

Perksey•10/20/23, 10:59 PM

i must say though with the xplat helpers SIMD is so much easier

Perksey•10/20/23, 10:59 PM

like have any of you seen the

feature/math-simd

feature/math-simd

branch

PPerksey like have any of you seen the `feature/math-simd` branch

TechPizza•10/20/23, 11:00 PM

i'm scared. what did tanner invent this time

Perksey•10/20/23, 11:00 PM

nono this was the WIP work to vectorize silk.net.maths

Perksey•10/20/23, 11:00 PM

https://github.com/dotnet/Silk.NET/tree/feature/math-simd/src/Maths/Silk.NET.Maths/Simd

Perksey•10/20/23, 11:00 PM

so much work whereas now we don't have to care about the individual ISA

TechPizza•10/20/23, 11:01 PM

oh yeah, basically exactly what net8 does

TechPizza•10/20/23, 11:01 PM

no wait, did we get crossplat in net7?

TechPizza•10/20/23, 11:01 PM

i'm too tired for code rn

Perksey•10/20/23, 11:01 PM

net7

Perksey•10/20/23, 11:02 PM

but yeah we were essentially writing our own

Vector128

Vector128

xplat helper until the real thing came along

TechPizza•10/20/23, 11:03 PM

"my" software rasterizer got quite a nice boost in net8, although it may be the avx512 doing... something, rather than JIT improvements

TechPizza•10/20/23, 11:03 PM

but that code is completely destroyed by inlining limits sadge

Perksey•10/20/23, 11:05 PM

when aggressive isn't aggressive enough

Maxine•10/20/23, 11:05 PM

you can't do extension operators, can you

Maxine•10/20/23, 11:05 PM

shame

PPerksey when aggressive isn't aggressive enough

TechPizza•10/20/23, 11:07 PM

JIT literally gives up in the middle of my hot methods waaaaaaaaaahhhhhh

Maxine•10/20/23, 11:07 PM

incredibly disgusting

Maxine•10/20/23, 11:08 PM

i'll make it an instance method at least

khitiara•10/20/23, 11:16 PM

why do you need conversion and division to be together?

khitiara•10/20/23, 11:17 PM

like an as method should be good enough and then implement division using the same type

Kkhitiara why do you need conversion and division to be together?

Maxine•10/20/23, 11:33 PM

hm.

Maxine•10/20/23, 11:34 PM

my thought is

Vector3 / float

Vector3 / float

is a very common pattern and useless with most integer values

Maxine•10/20/23, 11:35 PM

alright, yeeted

Maxine•10/20/23, 11:41 PM

design RFC: should static methods be provided

on Vector3<T> (when doing generics, it'd be Vector3<T>.Lerp<TFloat> for instance)
on static Vector3 (Vector3.Lerp<T, TFloat>, but for non-generic methods it wouldn't require specifying the <T>)
on both (pick whichever you want)
on static Vector3 only if the method is generic, otherwise on Vector3<T> (what i'm currently doing, but may be confusing to some)

Maxine•10/20/23, 11:42 PM

oh no, i've invoked tanner

Perksey•10/20/23, 11:46 PM

i don't know what's happening

Perksey•10/20/23, 11:46 PM

is this for silk

Maxine•10/20/23, 11:46 PM

someone was talking about generic Vector3, so i made a PoC

Maxine•10/20/23, 11:46 PM

i think it's for silk?

Maxine•10/20/23, 11:47 PM

i'm just entertained by things like this, and it's late on a friday

MMaxine the implementation of sin(x) for `byte` should just be a cast to float, get the ...

tannergooding•10/20/23, 11:47 PM

the problem here isn't how to do the computation
yes, it is very much intended to be done as floating-point and then cast to integer
(compute the infinitely precise result and round to the nearest representable result)

the problem comes in from how to implement that support generically and correctly
especially over time using where DIM implementations need to be provided

there also lies a slightly more general issue in that not everything can be represented equivalently
for example,

Sqrt(-1.0f)

Sqrt(-1.0f)

produces

NaN

NaN

, but it would have to throw for integer

likewise you end up not being able to represent infinites for trig functions
you end up with many other computations resulting in 0 or 1, regardless of input, due to it requiring fractional results

and so many operations functionally end up "pointless" for integrals

Maxine•10/20/23, 11:47 PM

good point

Maxine•10/20/23, 11:48 PM

i ended up not implementing cast to integer, though

Maxine•10/20/23, 11:48 PM

a function like Vector3.Distance, you obviously want a float

i stopped hyperoptimizing a long time ago, it just isn't worth the effort 90% of the time if it work

Similar Threads

Similar Threads

Similar Threads