M8 crowds and animation LOD #862

Activity

1. paul-hammant

   closed this as completedin #891on Nov 1, 2026
2. nicolas-maman

   reopened this on Nov 1, 2026
3. nicolas-maman commented on Nov 1, 2026

   

   nicolas-maman

   on Nov 1, 2026

   CollaboratorAuthor

   More actions

   Reopened: #891's merge (6fda12b) closed this before its list was met. Not started. The list above stands, and these closed issues are part of it:

   • A crowd's instances take a position and a yaw but no scale of their own: every person one height #766: a crowd's instances take a position and a yaw but no scale of their own, so every person in a crowd is one height.
   • Real human figures and walks for the horde: an importer for Mixamo/CC0 characters into the pipeline #342's horde half: real CC0 figures in the horde, baked from the same clips as the graph. (Its importer half is on main; see M1 skeleton, pose, clips, retargeting and root motion (L0, L1) #855.)

   Done when the list above is met, with a cost per figure per frame at each LOD against the 2,000-figure budget.
4. nicolas-maman

   closed this as completedin #905on Nov 2, 2026
5. nicolas-maman

   reopened this on Nov 2, 2026
6. nicolas-maman commented on Nov 2, 2026

   

   nicolas-maman

   on Nov 2, 2026

   CollaboratorAuthor

   More actions

   Reopened: #905's merge (68934a3) closed this, but it carried only the hand-off between writers (animator.pause, resume, share_take, safe behaviour removal). The remaining work is as written above.
7. nicolas-maman commented on Nov 2, 2026

   

   nicolas-maman

   on Nov 2, 2026

   CollaboratorAuthor

   More actions

   Added from the joint plan with Siltreach (2026-11-02):

   • A budget allocator by significance (distance and visibility into an update rate, interpolated between updates, as UE's Animation Budget Allocator does) over the LOD tiers.
   • Animators evaluated in parallel on ae3d.jobs: the pose phase is independent per figure (shared takes are read-only), but the feet's ground rays query physics, which must be safe to read from workers. Today the phase runs on one thread.
   • Measured costs first. Only take sampling is measured today (0.35 µs for 53 joints). A whole figure's frame (graph, feet with their ground rays, aim, hands) is not, and the targets get set from those numbers.

   Siltreach proposed about 50 µs per near figure, and 2 ms for 100 near figures plus 2,000 in the crowd. At 50 µs a figure, 100 near figures are 5 ms on one thread, so that budget needs either about 20 µs a figure or the parallel evaluation above.