plan.md

Experiment 32: Subgroup Stream Compaction Variants

1. Focus

• Compare per-workgroup compaction built on shared atomics versus subgroup ballot ranking.

2. Question

• How much does subgroup ballot reduce the overhead of local compaction across sparsity regimes?

3. Variants

• shared_atomic_block
• subgroup_ballot
• valid ratios 5%, 25%, 50%, 75%, 95%

4. Method

• Compact each workgroup into its own fixed output segment.
• Validate exact counts for both variants and stable ordering only for the subgroup-ballot path.

5. Outputs

• Median GPU time by valid ratio.
• Speedup of subgroup ballot vs shared atomic append.
• Effective payload GB/s using actual valid-count writes.

6. Interpretation

• This isolates subgroup compaction mechanics without the added cost of a global scan pipeline.
• Ordering guarantees should be interpreted separately from raw throughput.