26 lines (20 loc) · 924 Bytes

Experiment 18: Register Pressure Proxy Study

1. Focus

Approximate the cost of rising register pressure without changing the overall memory pattern.

2. Question

How much performance is lost as the kernel carries more live temporaries?

3. Variants

temp_4
temp_8
temp_16
temp_32

4. Method

Keep the same memory traffic and output contract while increasing the amount of per-thread temporary state.
Validate every variant against the same deterministic CPU reference.

5. Outputs

Median GPU time by temporary-count variant.
Slowdown relative to the lightest register-pressure path.
Evidence of occupancy or scheduling cliffs.

6. Interpretation

This is a proxy experiment, so it should be read as pressure sensitivity rather than a literal register count measurement.
Large slowdowns indicate that arithmetic-only optimizations can still fail if they inflate live state.