plan.md

Experiment 22: Prefix Sum or Scan

1. Focus

• Measure block-scan performance and the effect of batching multiple items per thread.

2. Question

• What is the best items_per_thread tradeoff for the shared-memory scan baseline?

3. Variants

• items_per_thread_1
• items_per_thread_4
• items_per_thread_8

4. Method

• Run the same inclusive-scan kernel shape while changing only how many items each thread handles.
• Keep output semantics, correctness rules, and timing policy fixed across the sweep.

5. Outputs

• Median GPU time by items_per_thread.
• Throughput by configuration.
• Practical default for later scan and compaction work.

6. Interpretation

• More work per thread is not automatically better if it reduces occupancy or inflates local state.
• This is the shared-memory scan baseline that later subgroup variants should be compared against.