plan.md

Experiment 25: Spatial Binning or Clustered Culling Capstone

1. Focus

• Combine earlier memory and primitive lessons into a rendering-adjacent compute pipeline.

2. Question

• Which list-building strategy performs best across sparse, dense, and clustered scene distributions?

3. Variants

• naive_append
• local_staging
• compacted_lists
• coherent and incoherent scene ordering

4. Method

• Build bin or cluster membership lists from the same deterministic scene generator.
• Compare pipeline variants under several scene distributions while keeping object counts and correctness rules fixed.

5. Outputs

• Total GPU time by pipeline variant.
• Stage breakdown for list construction.
• Practical winner by scene distribution.

6. Interpretation

• This is an end-to-end engineering experiment, not a pure microbenchmark.
• A good result is one that explains which earlier primitive or access-pattern choice mattered in the final pipeline.