What problem does it solve? Writing Cube/Vector affinity kernels on Ascend A5 (Ascend950) hardware is difficult because developers must coordinate cube-core GEMM computation, fixpipe data movement to UB, and vector-core post-processing with explicit synchronization. This Skill provides complete, runnable Triton-Ascend reference implementations that show exactly how to structure these fused kernels. ## Core Features & Use Cases - MatMul + ReLU fused kernel: A single-kernel CV fusion example using UB buffers, ROW_SPLIT fixpipe mode, sub_vec_id for dual sub-vector utilization, and explicit cube/vector sync_block primitives. - Vision MLP + GELU backward mixed-form example: Demonstrates how to split 7 backward operations across one fused-cv kernel, three plain Triton matmul kernels, and two pure-vector reduce kernels, eliminating an intermediate (S, I) tensor's GM write/read round trip. - Fusion selection guidance: Explains when affinity (Cube/Vector fusion) pays off versus when native Triton GEMM is the correct choice, including host-side requirements like disable_auto_inject_block_sync=True. - Use Case: When porting a transformer MLP backward pass to Ascend A5, follow the K1/K2/K3 kernel decomposition pattern to fuse only the GEMM-plus-epilogue boundary that benefits, keeping other GEMMs as plain Triton. ## Quick Start Ask the agent to implement a MatMul with ReLU epilogue as a fused Cube/Vector Triton-Ascend kernel for Ascend A5 following this example's structure.
