catlass-operator-performance-optim

To optimize Catlass operator performance, you iteratively adjust tiling parameters like TileShape, DispatchPolicy, and Swizzle, then recompile and compare against a baseline to find the best configuration.

The best way to benchmark Catlass operators is by collecting baseline profiler data, applying single-variable iterative changes to tiling settings, and generating side-by-side performance comparison reports.

Iterative profiling improves NPU operator performance by enforcing single-variable variation when adjusting tiling parameters, allowing you to isolate the impact of each change and validate results with PRE/POST reports.

Yes, following the official Catlass optimization guide is required to identify tunable parameters, establish baseline profiler data, and systematically apply disciplined iterative changes for performance tuning.

You should document every configuration change during performance tuning to ensure single-variable variation, track the impact of each tiling adjustment, and deliver a final validated performance comparison.

Yes, the workflow generates reproducible performance comparison reports for CO-based systems by documenting iterative tiling adjustments and delivering clear PRE/POST results against a defined baseline.