PCIe vs kernel timing breakdown

[SolidRegardless]

Summary

Break down benchmark timings to separately report PCIe transfer time (host↔device) versus kernel execution time, giving clear visibility into where time is actually spent.

Motivation

The current benchmark reports total wall-clock time per iteration, but this conflates two very different costs: data transfer over the PCIe bus and actual GPU kernel execution. For bandwidth-bound kernels (e.g. VectorAdd), transfer time may dominate. For compute-bound kernels (e.g. SHA-256), kernel time dominates. Without this breakdown, it's impossible to identify the real bottleneck or measure the benefit of optimisations like double-buffering (#6).

Acceptance Criteria

• Instrument the benchmark harness to separately time:
  • Host → Device transfer
  • Kernel execution
  • Device → Host transfer
• Report these as separate columns/rows in the benchmark output alongside existing metrics
• Maintain backward compatibility — existing total time and throughput metrics should still be reported
• Timing should use GPU-side events/synchronisation where possible for accuracy (not just host-side stopwatch)
• Include transfer vs compute ratio/percentage in the summary

Technical Notes

• ILGPU provides stream synchronisation primitives that can be used to isolate transfer and compute phases
• BenchmarkResult will need new fields for transfer and kernel timings
• BenchmarkRunner will need to wrap each phase with timing instrumentation
• Consider adding a --verbose or --breakdown flag to show the detailed timing (keeping default output clean)