tune-incore

First complete swimlane analysis and graph-level tuning, then re-collect fresh profiling data with debug options. Analyze the slow task's instructions and apply targeted fixes: special shape preprocessing, L2 cache policies, redundant computation to break dependencies, and pipeline overlap via submit_before_loop.

Use Vector operations to preprocess irregular shapes into standard ones, for example concatenating small right matrices into a larger diagonal block matrix before matmul. One documented case reduced a (884736,16)x(16,16) matmul from 500us to 40us.

Apply NONE_CACHEABLE to large weight matrices that are read only once, so they bypass L2 cache and free capacity for frequently accessed data. In fused operators, set all weights together; avoid it for small inputs and output tensors, and always verify with measurement.

Setting a single weight in a fused operator can unbalance L2 cache contention and slow other weight accesses. The documented fix is to set NONE_CACHEABLE on all large weights simultaneously, which released L2 capacity and achieved a 19.1% improvement.

Swimlane analysis and graph-level tuning must be complete, precision validation must pass, and a slow task must be identified. You must also re-run tests with debug options to collect fresh swimlane data, since reusing old data leads to wrong conclusions.

Use the valid_shape parameter in pypto.view to mark the actual data range of the last tile when it is smaller than BLOCK_SIZE. Only tail blocks need valid_shape; if the tail axis itself is too small, first enlarge it with concat, transpose, or reshape.