triton-ascend-case-matmul-swizzle2d

What problem does it solve?

Large-scale matrix multiplication on Ascend suffers from poor cache locality and load imbalance when launching many blocks. This Skill applies a Swizzle2D-based grouping and a fixed-core startup to improve data reuse and balance across cores for large A[M,K] x B[K,N] matrices.

Core Features & Use Cases

• Fixed-core startup: grid=(num_cores,), each core processes multiple blocks.
• Swizzle2D block reordering: group-level data locality, improved cache hits.
• Adaptive grouping direction: choose row- or column-first grouping based on M vs N.
• Block-size tuning: select BLOCK_M, BLOCK_K, BLOCK_N for the data type and cache size.

Quick Start

Run the Swizzle2D matmul optimization on Ascend for a large A[M,K] x B[K,N] workload to evaluate performance improvements.