triton-ascend-example-attention

What problem does it solve? Writing Flash Attention kernels for Ascend A5 (Ascend950) hardware requires coordinating Cube and Vector cores with explicit synchronization, which is difficult to get right. This Skill provides a complete, working Triton-Ascend reference implementation showing the serial Cube/Vector cooperation pattern. ## Core Features & Use Cases - Serial Cube/Vector Cooperation: Cube cores run two matmuls (logits = Q·K^T and partial = P·V) with fixpipe, while Vector cores run online softmax and flash accumulation. - Explicit Synchronization Pattern: Demonstrates three sync_block events (ids 0/1/2) pairing PIPE_FIX/V/MTE to coordinate Cube and Vector execution. - Online Softmax with NZ Layout: Shows row-max rescaling, denominator accumulation, and writing probability tiles to L1 in NZ 16x16 fractal format. - Use Case: When generating an A5 serial attention-style operator, reference this kernel structure to implement the Cube/Vector split, event pairing, and compile options (disable_auto_inject_block_sync, vf_merge_level=1). ## Quick Start Ask the agent to generate an A5 serial Flash Attention kernel following the Cube/Vector cooperation structure in this Triton-Ascend example.