ascendc

AscendC transformer operator development is guided by structured templates for op_host definitions, op_kernel implementations, tiling, and infershape logic. These patterns apply directly to FFN, GMM, and MoE style operators within the CANN framework.

Scaffolding a new AscendC operator uses the genop tool in ops-transformer to generate initial structures. This creates end-to-end workflows from YAML frontmatter discovery to production-ready host and kernel code aligned with existing references.

AscendC operator development strictly requires consistent DataType and Format mappings across host and kernel interfaces. Ensuring this consistency, along with proper YAML frontmatter presence, prevents errors in tiling and shape inference logic.

Implementing tiling and infershape logic for CANN examples is fully supported by this guide. It provides specific patterns and scaffolding to align your operator definitions and kernels with existing CANN reference implementations.

AscendC op_host definitions fail to align with op_kernel implementations when tiling patterns and DataType mappings are inconsistent. This guide enforces structured, pattern-based alignment across host, kernel, and CANN examples to resolve such errors.