27 lines (21 loc) · 897 Bytes

Experiment 19: Branch Divergence

1. Focus

Measure how different predicate distributions affect warp or wave efficiency.

2. Question

How much slowdown comes from divergent control flow when useful work stays otherwise similar?

3. Variants

uniform_true
alternating
random_p25
random_p50
random_p75

4. Method

Run the same kernel body with controlled branch masks and deterministic predicate generation.
Keep writes and arithmetic comparable so the dominant change is branch coherence.

5. Outputs

Median GPU time by branch pattern.
Slowdown relative to the uniform baseline.
Divergence sensitivity across several predicate mixes.

6. Interpretation

Divergence penalties depend on pattern shape, not just branch probability.
This experiment is strongest when read together with memory-behavior experiments that also vary access coherence.