What problem does it solve? Migrating AscendC device kernels from ops-nn or ops-transformer repositories to the ops-tensor Blaze/Tensor API is error-prone: teams risk silently changing host tiling, ABI, workspace, support domains, or mathematical semantics while lacking a rigorous equivalence proof. This Skill enforces a serial, evidence-bound migration workflow that guarantees byte-identical functional results and comparable performance between the original and migrated kernels. ## Core Features & Use Cases - Serial Gate Workflow (G0-G6): Enforces sequential gates from environment preparation, contract freezing, Blaze development, validation asset generation, original baseline verification, Blaze acceptance, to final code delivery, with strict ownership and evidence rules. - Byte-Exact Equivalence Validation: Requires original and Blaze outputs, dynamic metadata, and inplace after-state to match byte-for-byte via a shared ACLNN runner; tolerance-based comparisons are diagnostic only. - CMCT/CGMCT Elimination and Anti-Pattern Scanning: Scans the active dependency closure and changed files to ensure zero CMCT/CGMCT hits and no low-level AscendC APIs leaking into Kernel, Block, or Scheduler orchestration layers. - Use Case: Migrate a GroupMatmul kernel on Ascend 950 (DAV_3510) to Blaze: the Skill freezes the migration contract in G1, develops the Blaze implementation in G2, builds both OPP packages in G3, establishes the original baseline in G4, and accepts the migration in G5 only after byte-exact and msprof performance comparisons pass. ## Quick Start Ask the assistant to use the ascendc-blaze-migration skill to migrate a specific AscendC kernel from ops-nn or ops-transformer to the Blaze Tensor API on Ascend 950, starting with gate G0 environment and repository preparation.
