triton-ascend-case-reduction-amin-medium

What problem does it solve?

Large-scale 2D amin reductions suffer from inefficiencies when the reduce-axis is very large. This skill introduces a tiling strategy that allocates bigger blocks on the reduce axis to maximize the compute unit utilization, reducing the number of iterations while balancing the workload of a single iteration.

Core Features & Use Cases

• Large-reduce-axis tiling strategy to maximize device utilization and minimize total iterations.
• Autotune configurations exploring BLOCK_SIZE_M and BLOCK_SIZE_N for trade-offs between parallelism and per-iteration load.
• Suitable for scenes with mid-sized non-reduce axes and very large reduce axes (e.g., hundreds of thousands to millions elements).

Quick Start

Run the Triton Ascend case to evaluate and apply large-block tiling for large reduce axes.