fla-ascend-performance

name

fla-ascend-performance

description

Guidelines for Ascend NPU kernel / Triton-Ascend backend performance work in the FLA repo. Covers profiling with torch_npu, PipeUtilization/MemoryUB CSV analysis, Cube/Vector/MTE/UB bottleneck diagnosis, and kernel optimization (UB tiling, grid splits, fusion/split, varlen, G_T_CONTIG gate loading, constexpr DMA-path split / TAIL_MODE, extract_slice, MTE OOB, int32 address overflow, tl.cast vs constexpr .to, make_block_ptr int32 offsets, correctness gates, tl.dot left-operand clobber). NPU kernels must not use num_warps/num_stages. Per-kernel catalog: references/cases.md (incl. causal_conv1d core-grid). Use when working on NPU profiling, kernel_details/op_statistic, aic_metrics, fla triton_ascend backends (ops or modules), g transpose stride-1, UB overflow, dual-path DCE, grid limits, int64 pointer math, tl.dot reuse, or Ascend performance.

FLA Ascend NPU: Profiling → Bottlenecks → Optimization

Use this skill for Ascend operator performance work on all files under any triton_ascend directory (**/triton_ascend/**).

Multi-round iteration discipline (frozen tests, task contract, when to stop): fla-optimization-loop. MR packaging: fla-mr-readiness.

Collection must use this skill's generic scripts — do not copy torch_npu.profiler boilerplate per op.

Environment: Use the Python/NPU environment already active in the current terminal (including any activated conda/venv). Run collection, analysis, and benchmarks in the same shell; do not spawn a new shell or switch environments mid-workflow. If the terminal has no NPU stack loaded yet, activate the project's Ascend environment first, then continue in that same session. Metrics, failure modes, code index: reference.md. Past kernel notes: cases.md.

Make the target backend semantically correct before optimizing; never hide missing capability or kernel bugs behind a Torch fallback. What generalizes: UB modeling, grid splits, layout/precision, and verification. Values like , K slabs of 64, or specific are starting points only — do not copy them as rules.