Add 2026 spring T3-1-1 skill submission for junkai-kay

[junkai-kay]

🏷 Skill Information

Skill name	ninetoothed-operator-skill
Competition	2026 Spring AI Contest
Track	T3-1-1 — NineToothed .skill Innovation Challenge
Participant	Zhao Junkai
GitHub ID	@junkai-kay
Branch	2026-spring-junkai-kay-t3-1-1
Skill path	skills/competition/ninetoothed-operator-skill/

---

📖 Summary

本 skill 指导 AI 编码智能体完成 NineToothed GPU 算子开发全流程——从需求理解、DSL 实现、正确性测试、性能 benchmark，到失败诊断和 PR 集成。提供一套可执行的标准操作流程，涵盖 5 种 arrangement/application 模式，使 AI 面对陌生算子任务时自主完成语义提取、模式匹配、实现选型、测试覆盖和故障闭环。

所有 API 模式均基于 NineToothed 0.26 官方源码逐行对照验证，T1/T2/T3 共 37 个测试用例全部通过。

🎯 Scope

Category	Capability	Evidence
Elementwise / Broadcast	Pattern B: 2D tile + expand；mask 条件、dtype、broadcast 边界	T1 — 12/12 PASSED
Reduction / Block	Pattern D: (1, BLOCK_SIZE) tile + 无参归约；max-subtraction 数值稳定	T2 — 13/13 PASSED
Layout-Sensitive	4 种布局变体 (contiguous/transposed/sliced/offset)	T3 — 12/12 PASSED
Performance / Diagnosis	17 配置 benchmark + 6 步诊断闭环 + Bug 复现	T4 — 完成

❌ Out of Scope

NVIDIA GPU only (CUDA 11.8+), float32 primary, static shapes, single GPU. CPU/macOS/Windows/dynamic shapes/float64/multi-GPU not tested. 详见 SKILL.md.

📦 Files Added

All under skills/competition/ninetoothed-operator-skill/ — SKILL.md, README.md, HONOR_CODE.md, REFERENCE.md, references/index.md, scripts/, tests/, logs/, reports/, examples/ (4 tasks).

pytest output: ```text
Environment: Tesla T4, CUDA 12.8, PyTorch 2.11.0, NineToothed 0.26.0
T1 — Masked Add with Broadcast: 12 passed in 122.26s
T2 — Tiled Softmax: 13 passed in 8.04s
T3 — Layout-Sensitive GELU: 12 passed in 7.69s

## 📊 Benchmark

Environment: Tesla T4, CUDA 12.8, PyTorch 2.11.0

**T1 — Masked Add (5 configs):** NineToothed overtakes PyTorch at large sizes (0.44x at 4096²)

**T2 — Softmax (6 configs):** Competitive or faster at most shapes. Single-row (1,4096): 3.53x slower due to launch overhead.

**T3 — GELU (4096×4096):** contiguous 1.00x, transposed **4.35x slower**, sliced 1.01x — directly proving layout sensitivity matters.

**T4 — 8-shape benchmark:** Completed on Tesla T4.

Raw logs: `logs/` directory.

## 🔬 Skill vs No-Skill

| | Without Skill | With Skill |
|---|---|---|
| API selection | Used `block_size()`, `ntl.max(dim=,keepdims=)`, `ntl.tanh()` — all unavailable | Reads official `tests/test_softmax.py` → correct `Symbol` + `@ninetoothed.jit` |
| T2 result | 13/13 CompilationError | 13/13 PASSED |
| T3 result | 12/12 CompilationError | 12/12 PASSED |
| Test coverage | 1 case | 10+ parameterized (boundaries, extremes, non-contiguous, dtype) |

## 🔧 Failure Diagnosis — Real Bug

Bug reproduced on Tesla T4 (2026-07-12): `ntl.max(input_row, dim=1, keepdims=True)` → `TypeError: max() got an unexpected keyword argument 'dim'`. Fix: `ntl.max(input_row)` without arguments, tile as `(1, BLOCK_SIZE)`. Full record in `examples/task4_benchmark_debug/diagnosis_record.md`.

## 🔗 Documents

- HONOR_CODE.md, REFERENCE.md in skill root
- Proposal, Mid-term Report, Final Report PDFs in `reports/`
- PR附件（三个PDF）已拖入下方评论区

## ✅ Compliance

- [x] Only files under `skills/competition/ninetoothed-operator-skill/`
- [x] No API keys, credentials, or secrets
- [x] No hidden evaluation answers
- [x] No deleted or bypassed tests
- [x] No fabricated results — all data traceable to `logs/`
- [x] ruff check: All checks passed
- [x] Branch: `2026-spring-junkai-kay-t3-1-1` (kebab-case, 32 chars)
- [x] AI tool usage disclosed in REFERENCE.md

[junkai-kay]

ZhaoJunkai_九齿skill创新挑战_proposal.pdf
ZhaoJunkai_九齿skill创新挑战_T3-1-1_赛题报告.pdf
ZhaoJunkai_九齿skill创新挑战_中期报告.pdf