plan.md

Experiment 20: Barrier and Synchronization Cost

1. Focus

• Isolate the runtime cost of workgroup barriers and synchronization placement.

2. Question

• How much overhead comes from barriers themselves, and does that cost depend on how work is tiled?

3. Variants

• flat_loop_no_barrier
• tiled_regions_no_barrier
• flat_loop_with_barrier
• tiled_regions_with_barrier

4. Method

• Use the same logical output and arithmetic while changing whether the kernel runs as a flat loop or staged tiled regions.
• Add or remove barriers without changing the final output contract.

5. Outputs

• Median GPU time by synchronization strategy.
• Barrier overhead relative to the no-barrier forms.
• Placement sensitivity between flat and tiled execution shapes.

6. Interpretation

• A barrier cost is only meaningful relative to the work it protects.
• This experiment explains why some shared-memory kernels fail even when their memory traffic looks favorable on paper.