This ended up being far more complicated than anticipated -- the utils
now make use of filterExceptAttributes() in order to properly preserve
the index type when passing a MeshData through. Before an
implementation-specific index type was lost in the process due to
passing just the view to MeshIndexData constructor and not also the
type, and because keeping the metadata is a bit involved process, this
had to wait until there's an alternative to reference() that allows me
to drop all attribute information. And that's what
filterExceptAttributes() does.
Then, interleave() implicitly doesn't preserve strided index buffers
because GPUs don't like them, and since implementation-specific index
types always result in strided index buffers, this was then immediately
dying inside interleave() as it's not possible to tightly-pack an index
buffer of an unknown type. Thus, the function now has to expose
InterleaveFlags and propagate them to interleave() in order to allow the
user to preserve the index buffer without repacking it.
Apmong other things where it's useful for end users as a more convenient
alternative to recreating the MeshData by hand, I need to use this
inside the transform() utilities to preserve all index buffer properties
without copypasting nasty code everywhere.
As implementation-specific index types cause the index buffer to be
strided. which is not allowed by default, these work only if
InterleaveFlag::PreserveStridedIndices is set. Originally I thought
about straight-out disallowing implementation-specific types here,
unfortunately that broke some valid use cases so I backtracked on that
decision.
In the future I might change the default to preserve
implementation-specific indices with sane strides (i.e., positive powers
of 2), but that will only be done once similar treatment is finished for
attributes (currently zero and negative strides are just disallowed
completely).
Otherwise it makes a tightly-packed copy. This also means that arbitrary
padding before/after the index buffer is preserved only in case it would
mean the data can be transferred without a copy -- otherwise it's faster
to just drop the padding and copy only the used part.
The reference() and mutableReference() utils preserve that implicitly as
MeshData::indices() now returns a strided array view containing all
needed information.
I tried, I really did, but the downsides eventually just outweighed the
potential of this feature and I gave up. May try to tackle this again in
the future, but not now.
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.
Leads to an assertion inside StridedArrayView constructor, because I
just pass through the original attribute offsets/strides even though
interleavedMutableData() gives me a much smaller stride.
Moreover, because I just perform the deduplication on the original
vertex data *including* all random padding, the duplicate removal won't
always be able to find all duplicates. So test that case here as well --
turns out we just have to repack the data completely, after all.
This took me quite a while to realize -- not always it's desirable to
have the original layout unconditionally preserved, especially if for
example filtering a MeshData to just a subset of attributes.
It was already done in removeDuplicatesFuzzy(), so just make it
consistent with the rest (shorter message, listing also the offending
attribute index).
Haha, I even had a TODO here. An exception in this case is when the mesh
is already interleaved -- then the layout is kept intact and thus it's
not needed to know what vertex format sizes to repack.
Vladimír Vondruš