triton-ascend-case-reduction-sum-large

What problem does it solve?

Large-scale 2D reduction workloads in Triton-Ascend kernels suffer from excessive reduction calls on small axes; this Skill reorganizes computation to maximize UB occupancy and minimize reduction steps, delivering faster performance for very large non-reduce axes.

Core Features & Use Cases

• Autotune configurations explore different tile sizes (e.g., BLOCK_SIZE_M and BLOCK_SIZE_N) to balance compute and memory.
• Compute-tiling optimizations that reorganize reduction to reduce the number of iterations.
• Applicable to 2D reductions with extremely large non-reduce axes and moderate reduce axes, achieving significant latency improvements.

Quick Start

Run autotuning on the Triton Ascend kernel to discover optimal BLOCK_SIZE_M and BLOCK_SIZE_N settings for large non-reduce axes.