triton-ascend-case-reduction-amin-atomic

What problem does it solve?

Optimizes atomic reductions for tensors where the non-reduce axis is small and the reduce axis is large, enabling mapping the reduce axis across multiple cores to improve throughput on Ascend devices.

Core Features & Use Cases

• Atomic-based reduction in Triton kernels with two complementary strategies: per-block atomic updates and post-aggregation, enabling trade-offs between memory usage and contention.
• Axis-aware tiling and fine-grained partitioning to respect hardware caches and maximize throughput for extreme shapes (e.g., M << N).
• Use case coverage includes large matrix reductions in ML workloads and operations requiring synchronized minima or sums across threads.

Quick Start

Run the amin-atomic Triton Ascend example to benchmark the two atomic strategies and observe performance improvements on your hardware.