triton-ascend-reduce

What problem does it solve?

Optimize reduce-type and composite operators that involve reductions, including multi-axis reductions and normalization scenarios, to improve performance on Ascend using Triton.

Core Features & Use Cases

• Supports non-final-dimension reductions with correct multi-dimensional indexing to avoid costly reshapes.
• Describes two-stage reduction workflows for complex operators like normalization (layernorm, rmsnorm, groupnorm, batchnorm) and statistical computations (variance, std).
• Provides practical guidance for implementing efficient Triton kernels on Ascend hardware, including axis-aware tiling and atomic reductions.

Quick Start

Start by identifying a non-final axis reduction and implement a two-stage Triton kernel on Ascend to validate performance gains.