tune-swimlane

Run analyze_swimlane.py on the output directory containing merged_swimlane.json, passing a manually computed --outer-loops value. The script reports per-leafHash timing statistics, hashOrder merge keys, subGraphCount, and a Merge Tuning Guide with suggested granularities.

Stitch tuning controls how many root functions are dispatched together via the stitch_function_max_num runtime option in pypto.frontend.jit. Larger values compact the swimlane and reduce scheduling overhead, but excessively large values increase scheduling latency and workspace usage.

cube_l1_reuse_setting merges subgraphs that repeatedly load the same L1 data, while cube_nbuffer_setting merges structurally identical AIC subgraphs to cut scheduling overhead. Tune L1 reuse first, avoid setting both too large, and only apply them after cores are fully utilized.

Merging before cores are fully utilized reduces parallelism by packing few tasks onto even fewer cores. Overly large merge granularity can also overflow L1/UB memory. Always run core-usage analysis first and validate each change with freshly collected swimlane data.

Follow the decision tree based on M, N, K sizes: large shapes target compute-bound with tiles like [128,128],[64,256],[256,256]; small M or N favors keeping that dimension unsplit with A resident in L1; small M and N with huge K uses split-K via enable_split_k or manual K loops.

No. The skill mandates recollecting swimlane data with runtime_debug_mode enabled after every code change, because performance characteristics shift with each modification and stale data leads to incorrect tuning decisions.