ca60419b3a
Use a thread local buffer to gather info about the custom normal space
for each corners group in threading to concatenate them later. This
avoids need to lock a mutex to write into a buffer shared betwen threads
in the corner groups traversal hot loop.
The performance improvement depends on the mesh size and the number of
CPU threads. In some files there might be no change, in other files an
improvement of over 2x was observed, mostly because we can now remove
the compromise from c8a4026984.
Pull Request: https://projects.blender.org/blender/blender/pulls/144660