Skip to content

Latest commit

 

History

History
26 lines (20 loc) · 872 Bytes

File metadata and controls

26 lines (20 loc) · 872 Bytes

Experiment 12: Gather Access Pattern

1. Focus

  • Measure the penalty of indirect indexed reads relative to sequential access.

2. Question

  • How much does gather performance depend on index coherence and locality?

3. Variants

  • sequential
  • block_coherent
  • random
  • clustered_random

4. Method

  • Read from a source buffer through a host-generated index buffer with controlled distributions.
  • Keep output writes and arithmetic fixed so the result is dominated by read-side access behavior.

5. Outputs

  • Median GPU time by index distribution.
  • Relative slowdown vs the sequential baseline.
  • Throughput and useful-payload GB/s.

6. Interpretation

  • Gather cost should track how much locality survives the indirection pattern.
  • This is the baseline for later systems that depend on neighbor lookups, sparse reads, or visibility indirection.