Add LRC microbenchmarks for L2 request coalescing analysis

[William-An]

Summary

• Add lrc_max_merged microbenchmark to measure the maximum number of L2 sector requests that can be merged per LRC (L2 Request Coalescer) entry on NVIDIA GPUs
• Add lrc_queue_size stub with documented TODO for future work on discovering LRC queue depth per L2 sub-partition
• Update .gitignore to exclude IDE/AI tool directories

Details

lrc_max_merged

Measures how many concurrent sector requests from multiple warps/threadblocks can be coalesced into a single L2 lookup by the LRC. Supports three launch modes:

• Normal: Standard kernel launch
• Cluster: Uses threadblock clusters with mbarrier-based synchronization to co-locate blocks in the same GPC
• Cooperative: Uses cooperative kernel launch for grid-wide synchronization

Designed to be profiled with ncu — includes a run script (run_lrc_merged.sh) with relevant L2 sector metrics.

lrc_queue_size (stub)

Documents the open challenges for measuring LRC queue depth, including the need to reverse-engineer L2 address-to-sub-partition mapping.

Corresponds to Accel-Sim config parameter: -gpgpu_lrc_max_entries

Test plan

• Build lrc_max_merged with make on SM_90+ target
• Run run_lrc_merged.sh with ncu and verify L2 sector counts
• Verify lrc_queue_size builds and prints stub message

🤖 Generated with Claude Code