Skip to content

plan.md

Latest commit

 

History

History
26 lines (20 loc) · 837 Bytes

plan.md

File metadata and controls

26 lines (20 loc) · 837 Bytes

Experiment 33: 2D Locality and Transpose Study

1. Focus

  • Move beyond 1D buffers into matrix-shaped access and transpose behavior.

2. Question

  • How much does tiling recover from the poor locality of naive transpose?

3. Variants

  • row_major_copy
  • naive_transpose
  • tiled_transpose
  • tiled_transpose_padded

4. Method

  • Use a square matrix sized from the scratch budget.
  • Keep input data and output semantics fixed while comparing naive and tiled access patterns.

5. Outputs

  • Median GPU time by variant.
  • Effective GB/s by variant.
  • Speedup of tiled transpose over naive transpose.

6. Interpretation

  • This experiment connects the project's memory story to image-style and rendering-adjacent workloads.
  • The padded tiled variant shows whether bank-conflict mitigation matters in the transpose case.