Skip to content

Latest commit

 

History

History
24 lines (18 loc) · 843 Bytes

File metadata and controls

24 lines (18 loc) · 843 Bytes

Experiment 04: Sequential Indexing

1. Focus

  • Establish the good-path baseline for fully contiguous indexing without indirection or stride.

2. Question

  • How fast is a simple sequential read-modify-write kernel before access-pattern penalties are introduced?

3. Variants

  • sequential_indexing
  • problem-size sweep

4. Method

  • Map one invocation to one contiguous element in both source and destination buffers.
  • Keep arithmetic and memory footprint fixed while scaling only the logical workload size.

5. Outputs

  • Median GPU time.
  • Throughput and effective GB/s.
  • Baseline reference for later access-pattern experiments.

6. Interpretation

  • This is the comparison point for gather, scatter, stride, and reuse studies.
  • Later slowdowns should be explained relative to this contiguous baseline, not in isolation.