Ascendc

AscendC Operator Development (All-in-One)

This skill drives a new AscendC custom operator from a spec to a production-ready, benchmarked operator inside an ascend-kernel project (a PyTorch custom-op project that exposes operators as torch.ops.npu.<op> via csrc/ops/, csrc/register.cpp, and build.sh). It is self-contained: every phase, template, and reference lives under this skill directory.

Scope note: this skill targets the ascend-kernel / csrc/ops PyTorch custom-op workflow (vector / row / index / sort / pool operators, FP16/BF16 up-cast, two-level tiling). It is not for the ops-transformer aclnn/genop flow.

How to use this skill

1. Read this SKILL.md fully first (lifecycle, gates, anti-patterns).
2. For each phase, MUST open the matching references/NN-*.md before acting.
3. Reuse the bundled templates/, examples/, and scripts/; do not invent project structure or APIs.
4. Honor the stage gates: never start a phase until the previous phase's checklist passes. Never skip the in-chat result display rules.

When to use

• "Develop / implement / create a new AscendC operator <name>" (e.g. acosh, rms_norm).
• "Continue operator development" → detect the current phase from artifacts and resume.
• Any single phase on an existing operator: design only, code-gen only, precision eval only, performance benchmark only, optimization only, or code review only.

Lifecycle overview

Phase 0  Environment + requirements
Phase 1  Project init        -> references/01-project-init.md
Phase 2  Design (design.md)  -> references/02-design.md
Phase 3  Test cases          -> references/03-testcase-gen.md
Phase 4  Code generation     -> references/04-code-gen.md (+ 04a-kernel-api.md)
Phase 5  Compile / debug     -> references/05-compile-debug.md
Phase 6  Interface docs      -> references/06-doc-gen.md
Phase 7  Precision eval      -> references/07-precision-eval.md (fail -> 07b-precision-debug.md)
Phase 8  Performance eval    -> references/08-performance-eval.md
Phase 9  Performance optim   -> references/09-performance-optim.md
Phase 10 Code review         -> references/10-code-review.md
(opt)    Memory check        -> references/11-mssanitizer.md

Input: operator name (snake_case) + functional/math spec. Output: built & installed operator, design.md, unified test-case doc, PyTorch-style README, precision report, performance report (and optimization/review reports if run).

Phase 0 — Environment and requirements

Confirm the build/run environment before any development action.

• CANN: echo $ASCEND_HOME_PATH. If set, use it as CANN_PATH. If unset, MUST ask the user for the CANN install path. Activate per shell with source ${CANN_PATH}/*/set_env.sh.
• Conda: echo $CONDA_DEFAULT_ENV. If non-empty and not base, use it. Otherwise MUST ask the user for the conda env name; activate with conda activate <env>.
• Requirements: operator name (snake_case, required), functional spec / math formula (required), supported dtypes (optional, default float16, float32, may add bfloat16), SoC (optional, default ascend910b, obtained via platform API at runtime).

Details and the decision tree: references/00-environment.md.

Gate: CANN path resolved and activatable; conda env resolved and activatable; operator name and functional spec confirmed.

Phase 1 — Project init

Locate or create the ascend-kernel project, then scaffold csrc/ops/<op>/.

• Detect with scripts/detect_ascend_kernel_project.sh. If none, copy the bundled template templates/ascend-kernel/ and chmod +x build.sh.
• Create csrc/ops/<op>/{op_host/<op>.cpp, op_kernel/<op>.cpp, CMakeLists.txt, design.md} (placeholders).
• Flag the three registration update points for later phases: csrc/ops.h, csrc/register.cpp, csrc/CMakeLists.txt.

Read references/01-project-init.md.

Gate: project exists (build.sh, CMakeLists.txt, csrc/); csrc/ops/<op>/ skeleton created with the four files.

Phase 2 — Design (design.md)

Produce a complete design document; it is the direct input for code generation.

• Pick an implementation path (default AscendC Kernel; CATLASS only for matmul/cube).
• Map the math to an AscendC API call sequence; design two-level tiling (block-level inter-core + UB-level intra-core); fill the UB allocation table and derive bufferCoefficient per dtype; describe the FP16/BF16 → FP32 up-cast path.
• Fill templates/design-template.md → write to csrc/ops/<op>/design.md.

Read references/02-design.md (tiling-by-op-type, UB allocation, API map, hardware constraints).

Gate: design.md has function signature, supported dtypes, API pseudocode, UB allocation table with bufferCoefficient per dtype, tiling struct, and up-cast path.

Phase 3 — Test cases

Generate one unified test-case document reused by precision and performance later.

• Read design.md; produce SUPPORTED_DTYPES, TEST_SHAPES, GENERAL_SHAPES, BOUNDARY_VALUES, and the operator baseline (CPU reference + NPU call).
• Total cases (TEST_SHAPES + GENERAL_SHAPES) x SUPPORTED_DTYPES >= 30; keep single-shape element count reasonable (<= ~200K for regular cases).
• Fill templates/test-cases-template.md → write csrc/ops/<op>/test/<op>-test-cases.md.

Read references/03-testcase-gen.md.

Gate: test-case doc exists with dtypes, shapes, boundary values, and baseline; values respect design.md constraints.

Phase 4 — Code generation + framework adaptation

Generate op_host and op_kernel, then wire them into the framework.

• Select a template pair from templates/code-gen/ by operator type (elementwise / row / index / index-per-elem / sort / pool); copy into csrc/ops/<op>/ and adapt.
• op_host: signature, input checks, platform API for coreNum/ubSize (never hardcode), bufferCoefficient, left-value EXEC_KERNEL_CMD args.
• op_kernel: BUFFER_NUM=2, Init core offsets, InitBuffer sizes, Compute logic, tail-tile alignment, FP16/BF16 up-cast to FP32, DataCopyPad for GM↔UB, backup before Reduce.
• Framework: add declaration to csrc/ops.h, m.def+m.impl to csrc/register.cpp, host+kernel sources to csrc/CMakeLists.txt.

Read references/04-code-gen.md and the API essentials in references/04a-kernel-api.md.

Gate: both sources generated; three registration points updated; checklist in the reference satisfied.

Phase 5 — Compile, install, test, debug

Build the project, install the wheel, generate a basic test, run it, and debug.

• chmod +x build.sh && bash build.sh; confirm output/ascend_kernel*.whl.
• pip install output/ascend_kernel*.whl --force-reinstall --no-deps.
• Generate tests/test_<op>.py; run functional test (python ...) then precision test (pytest -v). Source the environment before every shell command.
• On failure, run the debug loop (max 3 attempts) using the error decision trees.

Read references/05-compile-debug.md.

Gate: wheel built and installed; functional test exits 0; precision pytest green.

Phase 6 — Interface docs

Extract interface facts from source and emit a PyTorch-style README.

• Pull schema from register.cpp (m.def), C++ signature from ops.h, algorithm/dtype/ constraints from design.md, TORCH_CHECK from op_host, example from the test file.
• Assemble the fixed sections and write csrc/ops/<op>/README.md. Default language is English; switch to Chinese on request.
• MUST display the full README content in chat, not just the path.

Read references/06-doc-gen.md.

Gate: README has signature, params, dtypes, shape, constraints, example, returns; matches register.cpp schema; displayed in chat.

Phase 7 — Precision evaluation

Run a comprehensive precision suite and produce a report.

• Load the Phase 3 test-case doc; adapt (shapes + boundary) x dtypes >= 30.
• Generate test_<op>_precision.py and run_<op>_precision_report.py from templates/precision/; run pytest then the report generator.
• Pass rule (MERE/MARE, ecosystem open-source standard): MERE < Threshold and MARE < 10 x Threshold. Standards table in references/07a-precision-standards.md.
• On precision failure that thresholds cannot fix, run root-cause debugging per references/07b-precision-debug.md.
• MUST display in chat: overview (totals + pass rate), any failures, and >=3 key findings — then the report path.

Read references/07-precision-eval.md.

Gate: pytest green; JSON + Markdown report written; results displayed in chat.

Phase 8 — Performance evaluation

Benchmark the custom operator against a baseline with torch_npu.profiler.

• Build a JSONL case file (>= 8 cases) from the test-case doc + design.md; always run a dual-path comparison (custom vs baseline; baseline must run on NPU — use a small-op composition when no equivalent API exists).
• Fixed schedule warmup=5, active=5; aggregate Total Time(us) from ASCEND_PROFILER_OUTPUT/op_statistic.csv. Copy examples/layer_norm_profiler_reference/ as the starting point.
• MUST display in chat: the unified comparison table (with DType column), the summary, and >=3 short conclusions — then the report path.

Read references/08-performance-eval.md, references/08a-profiler-and-metrics.md, and references/08b-perf-case-jsonl.md.

Gate: dual-path report written; displayed in chat with table + summary + conclusions.

Phase 9 — Performance optimization (optional, closed loop)

Investigate, modify, and verify — at most 3 rounds.

• Investigate across 5 dimensions (tiling, data copy, API usage, memory, pipeline) and emit a ranked report; snapshot a baseline.
• Apply changes obeying the anti-pattern list; re-run precision (must pass) then the same performance cases; compare to baseline; iterate.

Read references/09-performance-optim.md.

Gate: precision still passes; performance compared to baseline; results displayed in chat.

Phase 10 — Code review (optional)

Hypothesis-testing security review against the coding red lines (numeric, memory/pointer, resource, input validation, concurrency, operator interface, ABI compatibility).

Read references/10-code-review.md.

Optional — Memory check

Run mssanitizer to detect illegal access / leaks / UB out-of-bounds. Read references/11-mssanitizer.md.

Unified anti-patterns (NEVER)

• NEVER skip design and write code directly; code-gen consumes design.md.
• NEVER let FP16/BF16 go through complex math directly — up-cast to FP32 first.
• NEVER use DataCopy for GM↔UB — use DataCopyPad.
• NEVER pass r-values (temporaries/literals/expressions) into EXEC_KERNEL_CMD.
• NEVER hardcode core count or UB size — query the platform API.
• NEVER modify files under cmake/ or csrc/utils/.
• NEVER reuse a source tensor right after ReduceSum/ReduceMax (reduction may modify it).
• NEVER use std::min/max/abs/sqrt/exp etc. inside a kernel.
• NEVER pass repeatTime > 255 to high-dim split APIs (silent uint8 truncation).
• NEVER use a non-profiler timing method as a performance conclusion.
• NEVER report only a file path for precision/performance/optimization — show the tables and conclusions in chat.

Resume from interruption

Status tracker

Dependencies

This skill is self-contained (templates, references, examples, and scripts are bundled). It only needs a working CANN toolkit and a PyTorch + torch_npu conda environment on a host with Ascend NPUs to compile and run the operator.