Profile performances of arbitrary attributes

[eliemichel]

We should profile the amount of new memory allocation when manipulating arbitrary attributes, and organize the API in a way that hints the user into using the most efficient approach. For instance:

v1.attributes["uv"] = new FloatAttributeValue(0.0f, 0.0f);

is less efficient than

var uv = v1.attributes["uv"].asFloat();
uv.data[0] = 0.0f;
uv.data[1] = 0.0f;

but the latter is a bit annoying to write.

[MechWarrior99]

So I was going to add some QoL to the attribute API to encourage better (and easier) usage, but I realized that it is kind of limited and clunky as is. For example doesn't support other types like bool, or string.
So I was thinking refactoring it, but I saw in a comment in AttributeType you mention using data buffers. Are you able to elaborate on what you had in mind?

[eliemichel]

Storing vertex attribute values on a per-vertex basis fragments the memory and does a lot of repeated allocations when adding a new attribute. It would be much more efficient to allocate an array with all values for an attribute written consecutively (making the vertex list a struct of arrays rather than an array of structs), but this makes the addition/deletion of vertices trickier so I am not decided yet about the best design. (ofc it goes the same for edge, loop and face attributes).