25 lines (19 loc) · 793 Bytes

Experiment 01: Dispatch Basics

1. Focus

Establish the minimum viable Vulkan compute benchmark and validation path.

2. Question

Can the project execute upload, dispatch, and readback with correct outputs and stable GPU timing?

3. Variants

contiguous_write
noop
problem-size and dispatch-count sweep

4. Method

Use a deterministic 1D write kernel and a no-op control path.
Sweep problem size and dispatch count while keeping timing and synchronization policy fixed.

5. Outputs

Median GPU dispatch time.
End-to-end time.
Throughput and correctness pass rate.

6. Interpretation

This is a harness-validation experiment first and a performance result second.
Later experiments should only be trusted if they remain consistent with this baseline.