The thought behind Array2D was that it would be more efficient from a
memory allocation perspective to have one fixed buffer that grows than
it would be to have a vector of vectors. In reality, the runtime of
inserting into the middle of one large buffer, and shifting all the
resulting elements, ends up far outweighing any memory allocation
overhead.
In the mesh optimizer, things are added to the end of each sub-vector
one by one. It isn't known up front how many elements each sub-vector
will have. With vertex welding enabled, it is far more likely that a
given vertex will be influenced by a large number of joints. When using
the Array2D to store the influences, and a vertex with a low index has
more than 4 influences, those influences have to be inserted into close
to the front of the big vector, and all the other elements shifted.
Array2D was doing this with a linear shift, shifting the elements by 1
at a time, and not providing any exponential growth on the amount of
elements reserved by each sub-vector. The result is a linear insertion
time.
By contrast, using vector<vector<Influence>> instead gives us back the
amortized constant insertion time. Since the skin influences are just
added to the end of each sub-vector, no shifting of the elements is
necessary. And since the amount of original vertex indices is known up
front, the number of sub-vectors can still be pre-created, so the
sub-vectors themselves never need to shift.
Signed-off-by: Chris Burel <burelc@amazon.com>
* Final update copyright headers to reference license files at the repo root
Signed-off-by: spham <spham@amazon.com>
* Fix copyright validator unit tests to support the stale O3DE header scenario
Signed-off-by: spham <spham@amazon.com>
Chris Burel
Steve Pham
alexpete