bench: crowd load test — N-avatar ceiling (~22 @ 60fps) gating #199 - #350
Draft
arkavo-com wants to merge 2 commits into
Draft
bench: crowd load test — N-avatar ceiling (~22 @ 60fps) gating #199#350arkavo-com wants to merge 2 commits into
arkavo-com wants to merge 2 commits into
Conversation
) VRMBenchmark --mode crowd [--count N] [--stack]: N renderers share one loaded model, each drawn once per frame into a shared color+depth target (depth stored so avatars occlude), positioned by baking a grid offset into the view matrix with a shared projection. Sweeps N upward and reports the avatar count that fits the 60/120fps budget. --stack overlaps avatars (heavy occlusion) to measure how much occlusion culling could recover. Findings on this M4: ~22 avatars @ 60fps, ~11 @ 120fps, ~linear (~0.62ms/avatar). Stacked (63 of 64 occluded) costs ~the same as spread (only ~8% cheaper from GPU early-Z), confirming the renderer does NO occlusion culling — so #199 would help dense/occluded crowds (recover hidden avatars' full cost) but not spread ones (which need instancing/LOD). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Library-level test for the pattern the crowd load test (and future #199 work) relies on: N renderers share one loaded model and draw once each into a shared color+depth target, positioned via a per-renderer view offset under a shared projection. Asserts it renders, doesn't deadlock on the per-renderer in-flight semaphore (5 avatars > 3 slots), and places distinct avatars (5 cover more than 1). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
A
VRMBenchmark --mode crowdload test to empirically find the N-avatar ceiling of the current single-avatar-per-renderer architecture, plus a library-level test of the crowd render path. This gates the decision on #199 (GPU occlusion culling): build it only if the target crowd size/density warrants it.Draft — opened to share the measurements; not intended to merge as-is.
How it works
N renderers share one loaded model (1× geometry memory), each draws once per frame into a shared color+depth target (depth stored so avatars occlude), positioned by baking a grid offset into the view matrix under a shared projection (so depth stays consistent and avatars occlude correctly). Static pose to isolate render throughput.
--count Nsets the sweep ceiling;--stackoverlaps avatars to measure the occlusion opportunity.Results (M4 mini, 1024², MSAA 1×, static)
Ceiling: ~22 avatars @ 60 fps, ~11 @ 120 fps. Cost is linear, ~0.62 ms/avatar — no batching/LOD/occlusion culling.
The occlusion measurement (gates #199)
--stack(63 of 64 avatars hidden behind the front one) costs 39.7 ms vs 43.3 ms spread — only ~8% cheaper. The renderer pays ~full price for invisible avatars; the 8% is just GPU early-Z skipping some occluded fragments. The dominant per-avatar cost (CPU encode, skinning, draw-call overhead) is paid regardless of visibility.Conclusion on #199
spriteCacheSystem/DetailLevel).Tests
CrowdRenderTests— validates the shared-model, shared-depth multi-renderer path renders, doesn't deadlock (5 avatars > 3 in-flight slots), and places distinct avatars.🤖 Generated with Claude Code