There are multiple solutions to optimize extractions of tris.

• D11290: T88352: Use threaded ibo.tris extraction for single material meshes. adds an optimized version for single material meshes.
• Do bucket filtering when looping and joining buckets at the end of a loop and/or inside the finish function. When using memcpy we utilize AVX instructions. to construct the main buffer. this would require a loop start and loop end callback to init the buffer for each thread and sync with the master bucket.
• Only sort when subbuffers are requested. Needs some reinitialization code in DRW_mesh_batch_cache_create_requested
• Make the subbuffers normal buffers. More data transfer needed.

This patch feels to be an evolution of the current solution.

I would strive to make the init function constant, enable threading. Looking at the above solutions the bucket filtering only matches this, but adds quite some complexity.
A different patch we should make should ensure that tris are only recreated when needed (cache invalidatation).

LGTM.

Eventually it would be possible to do the indexing in a threaded non blocking way. TLS contains an array per mat during reduction these arrays are appended to the master array and its indexes are updated to the master array. The reduction could be vectorized but would add more complexity to the code.

So I am fine with the current solution.