To be clear -- it would work even without, as it'd fall back to
calculating the range of the weights attribute and choosing an epsilon
based on that, but now it takes a shortcut as Weights are known to be in
a [0, 1] range.
I realized those are too annoying when writing a glTF exporter which
contains a lot of switches over enums. And as further shown by the diff,
those were only inflicting additional pain in *all* switch statements,
nothing else, no other added value. And everywhere else the helpers are
the designated way to deal with those, so there's no point in having an
explicit enum value denoting start of a "custom range".
It wasn't even any convenient to have it in the enum, as the extra
effort needed for casting actually made it *exactly* the same length as
if I'd just use a separately-defined constant.
This is not something the classic GPU vertex pipeline can handle
(except maybe Vulkan, which can handle zero strides for instanced
attributes?), but useful for other scenarios. This means existing code
needs to be aware of and handle the new corner case.
Since the main speed advantage of the function is that it hashes a
*whole* vertex together instead of going through the attribute arrays
one by one, it can't really operate on whatever funny interleaved layout
it was given as there may be padding bytes with random content, breaking
the deduplication.
Since checking that a layout is really padding-less is rather complex,
the repacking is now performed always. This also means the && overload
makes no sense anymore and thus it was dropped.
This makes the test added in the previous commit not assert anymore, and
behave the same as the padding-less case.
It was already done in removeDuplicatesFuzzy(), so just make it
consistent with the rest (shorter message, listing also the offending
attribute index).
This is a breaking change that changes the signature, sorry -- if you
were using concatenate() for mesh concatenation before, enjoy the new
less strange signature, if you were using it for making the mesh owned
before (which was a strange and not very well thought out use case),
please use the recently added owned() instead. I thought about adding an
overload for backwards compatibility, but it would need to allocate to
work. This way with the breakage it's ensured you actually change to the
right API.
This also cleans up a lot of ugly code in the internals and resolves one
XFAIL in removeDuplicates().
The array size is always last, defaulting to 0. This makes it consistent
with the offset-only constructor and removes two unnecessary overloads.
It's a breaking change, but I don't think array attributes have many
users yet -- and better to do this now than later. In any case, sorry
about breaking your code.
I also tried storing only a 32-bit index and have the base pointer stored
in ArrayEqual/ArrayHash also, but that didn't really improve amything
much (probably because the allocated items are 8-byte aligned anyway) and
only made the code a lot less clear.
Basically using the same idea as with the discrete version -- having the
second dimension dynamic, together with restricting the implementation to
just Float and Double.
According to the SubdivideRemoveDuplicatesBenchmark, this makes the
implementation slightly slower. I presume this is due to how minmax and
offsets are calculated which is quite cache-inefficient as it goes over
the same memory block multiple times. Added a TODO for later.
Vladimír Vondruš