plan.md

Experiment 21: Parallel Reduction

1. Focus

• Compare baseline and hierarchical reduction strategies for a classic GPU primitive.

2. Question

• When does shared-memory tree reduction beat a simpler global-atomic approach?

3. Variants

• global_atomic
• shared_tree

4. Method

• Reduce the same deterministic input values with the same final operator in both variants.
• Sweep problem size across the available scratch budget and dispatch limit.

5. Outputs

• Median GPU time by variant and size.
• Speedup of shared_tree vs global_atomic.
• Effective input-read throughput.

6. Interpretation

• Hierarchical reduction usually pays off only after the workload is large enough to amortize setup overhead.
• This establishes the pre-subgroup baseline for later reduction work.