plan.md

Experiment 16: Shared or Workgroup Memory Tiling

1. Focus

• Compare direct global reads against explicit workgroup-memory tiling for a stencil-like workload.

2. Question

• When does local staging outperform direct global access once halo loads and barriers are included?

3. Variants

• direct_global
• shared_tiled
• reuse-radius sweep

4. Method

• Use the same 1D sliding-window sum for both implementations.
• Sweep neighborhood radius while keeping output count, arithmetic, and validation rules fixed.

5. Outputs

• Median GPU time by variant and radius.
• Speedup of shared_tiled vs direct_global.
• Shared-memory footprint and estimated global-traffic reduction.

6. Interpretation

• Shared memory only wins when reuse is large enough to amortize halo loading and barrier cost.
• This is the baseline for all later shared-memory tuning experiments.