That is a bit surprising. Not sure but I have used the combined integer division maybe once. (ignori

That is a bit surprising. Not sure but I have used the combined integer division maybe once. (ignoring things in algorithms I didn't come up with)
like finding the average of integers I think is super common to just do integer division cause the precision isn't that important

KKai not sure I get this - if I have a separate point type what does - return?

tannergooding•10/21/23, 10:13 PM

Point2<T>.DistancePoint2<T>.Distance returns a Point2<T>Point2<T>, which represents the per dimension stepping
which then allows you to do the correct and efficient thing for integers

and for floating-point, you can convert to Vector2<T>Vector2<T> to get it in terms of length (at the same total cost)
or to do per dimension stepping, which is incredibly useful for interpolation (especially with physics)

khitiara•10/21/23, 10:14 PM

where we have a loosely constrained type with few members and a bunch of extensions as we have been considering the whole time

Aqua•10/21/23, 10:14 PM

that's how my math library works

Aqua•10/21/23, 10:14 PM

and it works well for me

Aqua•10/21/23, 10:14 PM

i can't say on a global scale but

khitiara•10/21/23, 10:14 PM

yeah and that was the thing being worked on hours ago

Ttannergooding `Point2<T>.Distance` returns a `Point2<T>`, which represents the per dimension s...

KaiOP•10/21/23, 10:14 PM

Right, but if Point2<T>.Distance returns Point2 again, then we can't really say Point is the absolute type, since that returned point is meaningless without the source it came from

KaiOP•10/21/23, 10:15 PM

it'll find us some point on the grid that just doesn't have any meaning at all

khitiara•10/21/23, 10:15 PM

before the argument started on whether vector3<int> should be permissable out of almost nowhere

Aqua•10/21/23, 10:15 PM

have we considered that we may be overthinking this

khitiara•10/21/23, 10:16 PM

yeah im with just continuing as is, allowing vector3<int> is easy within the framework as proposed and anything more involved is probably overthinking imo

KKai That is a bit surprising. Not sure but I have used the combined integer division...

tannergooding•10/21/23, 10:16 PM

division isn't often used for things like computing averages, at least not for integers
you typically use floating/fixed point (user facing apps, they expect fractional results)

in practice it gets used for things like bucketing and finding the relevant offset in an array
which includes using the remainder to get the actual final offset

KKai Right, but if Point2<T>.Distance returns Point2 again, then we can't really say ...

tannergooding•10/21/23, 10:17 PM

there ends up having to be some allowances in the type system, in a way that makes sense
nothing is going to be perfect, it's more a question of can you separate the two core needs enough
so that you aren't ending up with a type that can do 10% of the functionality if you have an integer and 100% with a float

tannergooding•10/21/23, 10:17 PM

because having a type where the user doesn't have access to 90% of the functionality, isn't good design

tannergooding•10/21/23, 10:17 PM

especially if that is expected/common functionality

khitiara•10/21/23, 10:18 PM

i feel like you might be overestimating how much gamedevs expect a vector-ish type to have length

khitiara•10/21/23, 10:18 PM

is that because gamedev has bad conventions? yes

khitiara•10/21/23, 10:18 PM

but still

khitiara•10/21/23, 10:18 PM

idk for sure

tannergooding•10/21/23, 10:19 PM

maybe I'm contrasting AAA vs indie development too much

but, length is a core need for using vectors correctly and efficiently

Kkhitiara where we have a loosely constrained type with few members and a bunch of extensi...

Perksey•10/21/23, 10:19 PM

this is the closing remark on which my brain has shut off here

Perksey•10/21/23, 10:19 PM

but these are all important discussion points, as it's bad to have a preconceived conlusion without considering all relevant factors

tannergooding•10/21/23, 10:19 PM

as is normalization, if you aren't normalizing your vectors and local coordinate space
then you're likely losing precision and introducing physics and rendering bugs/quirks

tannergooding•10/21/23, 10:20 PM

(and lots of rubber banding/jitter)

Ttannergooding because having a type where the user doesn't have access to 90% of the functiona...

KaiOP•10/21/23, 10:20 PM

Yeah I mean this I fully agree with. I'm just not sure separating into multiple types is the way I want out of this.
I like the 2.X idea of allowing anything on anything, but obviously "try it and see what happens" isn't great.
The current 3.0 design IMO goes too much into having too many things unavailable.
Having more types is of course a solution, but then we have to somehow get the user to choose the correct types.

tannergooding•10/21/23, 10:21 PM

Another alternative is to force users to use floating-point (it's better/more appropriate anyways)
and provide some specialized functions for creating a Vector2/3/4<T>Vector2/3/4<T> from integer types

khitiara•10/21/23, 10:21 PM

separate types would actually make my reflection-based vertex attribute stuff easyer tbh, it just feels awkward and the integer one doesnt have a natural name imo since point overlooks a bunch of uses for integer tuples supporting some basic vector ops

Ttannergooding maybe I'm contrasting AAA vs indie development too much but, length is a core n...

KaiOP•10/21/23, 10:21 PM

This is true IMO, but I like the rational for HalfHalf a lot. If you choose the type you likely know of the trade offs and putting further barriers in place just to confirm you actually know what you're doing is unecessary. Similar for VectorN<byte> etc.

KaiOP•10/21/23, 10:22 PM

(also - ignoring any arguments around naming here, idk if I would feel good about Point either, given we also have line and other geometry but that's irrelevant right now)

Perksey•10/21/23, 10:23 PM

unless we explicitly decide on multiple types the name is irrelevant really

Ttannergooding Another alternative is to force users to use floating-point (it's better/more ap...

tannergooding•10/21/23, 10:23 PM

the other benefit of this, is that it also better meshes with GPU compute
doing integers on the GPU is generally a bad time, and not always well supported

Perksey•10/21/23, 10:23 PM

it's worth noting there is some precedent to multiple types in silk maths e.g. see Box2DBox2D vs RectangleRectangle

Ttannergooding the other benefit of this, is that it also better meshes with GPU compute doing ...

tannergooding•10/21/23, 10:25 PM

(not well supported meaning, GPUs don't often have many integer compute units; so you're often getting 25% of the throughput by using integers)

KaiOP•10/21/23, 10:27 PM

That's the other thing, and I'd need to do have a look around there. How important is it to have Vector3<T>Vector3<T> including integers to avoid extra overloads in bindings?
@Perksey do you have a clue? I'm fairly confident that there's a couple of cases in Vk that take Vector3<T> and you get to put whatever type

khitiara•10/21/23, 10:27 PM

oh hmm

khitiara•10/21/23, 10:27 PM

im not remembering anything

KaiOP•10/21/23, 10:28 PM

Cause Silk isn't just about extending the BCL maths to vectors here, we also (and maybe primarily) want this to work with bindings so CPUGPU is fairly painless

tannergooding•10/21/23, 10:28 PM

extending the BCL maths to vectors here

(this is happening in .NET 9, it's already been partially started in an OOB in 8)

KaiOP•10/21/23, 10:30 PM

given your preferred design I'm assuming float only

tannergooding•10/21/23, 10:30 PM

for the .NET 8 OOB feature it's float
but for .NET 9 it will be all floating-point types

KKai That's the other thing, and I'd need to do have a look around there. How importa...

Perksey•10/21/23, 10:31 PM

we haven't discussed this in terms of SilkX yet. for OpenGL i know there are a lot of manual overloads today, i'll check BT to see if there's any automated ones

tannergooding•10/21/23, 10:32 PM

the OOB feature is TensorPrimitivesTensorPrimitives, which is math APIs over ROSpan<T>ROSpan<T> (so TensorPrimitives.Log(ROSpan<float> x, Span<float> destination)TensorPrimitives.Log(ROSpan<float> x, Span<float> destination))
in .NET 9, that will expand to where T : IFloatingPointIeee754<T>where T : IFloatingPointIeee754<T>
and also to Vector128.Log(Vector128<float> x)Vector128.Log(Vector128<float> x) and friends (+ double overloads)

KaiOP•10/21/23, 10:32 PM

floating-point types

yeah that's sort of what I ment

Ttannergooding the OOB feature is `TensorPrimitives`, which is math APIs over `ROSpan<T>` (so `...

KaiOP•10/21/23, 10:32 PM

ooooh that is really cool

tannergooding•10/21/23, 10:32 PM

ah, well then yes
because Cos(int)Cos(int) doesn't make sense

Perksey•10/21/23, 10:32 PM

doesn't appear to be any automated ones

Perksey•10/21/23, 10:33 PM

so there is technically no precedent here

PPerksey doesn't appear to be any automated ones

KaiOP•10/21/23, 10:33 PM

not because we will generate 10 overloads for each possible T? Or because the APIs don't have any

Perksey•10/21/23, 10:33 PM

BuildTools does not generate any Silk.NET.Maths or System.Numerics overloads today

tannergooding•10/21/23, 10:33 PM

we'll likely have some APIs where it can minimally make sense for integers though
like Log2(int)Log2(int) (and already have these for scalars, of course)

That is a bit surprising. Not sure but I have used the combined integer division maybe once. (ignori

Similar Threads