25 lines (19 loc) · 876 Bytes

Experiment 05: Global ID Mapping Variants

1. Focus

Compare common indexing formulas used to map global IDs onto logical work.

2. Question

What overhead is introduced by offset and grid-stride indexing relative to direct mapping?

3. Variants

direct
fixed_offset
grid_stride

4. Method

Keep arithmetic, memory traffic, and output semantics constant.
Change only the mapping from invocation ID to logical element and measure the result across larger workloads.

5. Outputs

Median GPU time by mapping strategy.
Throughput by mapping strategy.
Relative overhead of flexible indexing vs direct mapping.

6. Interpretation

Grid-stride loops buy scalability and launch flexibility, but the control-flow cost still needs to be measured.
Direct mapping remains the simplest reference path for kernels that can launch exact coverage.