ascendc-performance-best-practices

Query the MatMul family guide, which covers optimizations like pingpong, SWAT, StreamK, full-load, scale coalescing, MTE2 preload, and constant folding. Variants such as matmul_mxfp4, batch_matmul, and matmul_a16w16 all map to the same matmul family documentation.

Follow the lookup chain from SKILL.md to references/reduce/guide.md, then select a template in the templates directory based on your R dimension versus UB capacity. Convert .template files to .h per usage_guide.md, copy them into your op_kernel directory, and never rewrite the MicroAPI code from scratch.

Use the common tail-block handling guide at references/common/tail_block_design.md. It covers norm/tail kernel separation, GatherMask vectorized tail processing, and 32B/64B alignment with DataCopyPad to avoid out-of-bounds access and load imbalance.

Yes. Fused operators are decomposed into base operator families, and each family's optimization guide is consulted separately. For example, AddRelu splits into Broadcast (Add) plus Elementwise (Relu), then the practices are adapted to the fusion scenario.

Broadcast performance depends on selecting the right implementation among OneDim fast path, NDDMA, DataCopyPad, and UB Broadcast. Choosing UB Broadcast when no expandable dimension exists in the tile causes empty broadcast and nan; the family decision tree explains the selection criteria.

Convolution, NN, and Random operator families are planned but not yet covered, and queries for unlisted families return a not-yet-collected response. Most guides target DAV_3510, and template code requires minimal adaptation such as namespace aliases to compile against your CANN version.