Ascend/msagent / skills/operator/ops-performance-tuning
ops-performance-tuning
VerifiedCompile, profile, diagnose, optimize, and compare Ascend NPU operators across CANN versions and A2/A3/A5 for Ascend C, CATLASS, Triton-Ascend, TileLang-Ascend, PyPTO, and SHMEM/MC2. Use when an operator already has a runnable implementation and the user asks for msOpProf collection, bottleneck analysis, source-level tuning, or a reproducible before/after performance report. Do not use as the primary workflow for operator creation, migration, or unresolved correctness failures.
references/cases/triton/pr_triton_lgamma_simt.md
View on GitHub# PR #693 "[libdevice] fix: Optimize the performance of lgamma under the SIMT mode"【PR】 后被回退
## 基本信息
- 算子类别:vector-elementwise
- DSL/框架:triton
- 类型:PR
- 来源可信度:一手 PR 原文
## 来源链接
- PR/出处链接:<https://github.com/triton-lang/triton-ascend/pull/693>(2026-06-23 合入)( 已验证可达,验证日期 2026-08)
- 优化代码查看:<https://github.com/triton-lang/triton-ascend/pull/693/files>
- 回退/重启用 PR:<https://github.com/triton-lang/triton-ascend/pull/815>("[libdevice] revert: Revert high-performance optimization scheme",2026-06-26 合入)、<https://github.com/triton-lang/triton-ascend/pull/833>(重新 Enable,Draft 未合入)( 已验证可达,验证日期 2026-08)
## 问题与瓶颈
SIMT 模式下 lgamma 性能低(原文性能分仅 0.12,L20 平台)。
## 优化方法(理论手段)
1. 在 SIMT 模式下优化 libdevice lgamma 实现(高性能方案)。原文未详述具体机制。
## 性能对比
(PR 正文原文):"The original lgamma implementation achieved a performance score of only 0.12 on the L20, which has been optimized to over 0.5 after improvement."
| 指标 | 优化前 | 优化后 | 变化 |
|---|---|---|---|
| lgamma 性能分(L20) | 0.12 | 0.5+ | 约 4 倍 |
## 适用范围与警示
- 风险说明:后续 PR #815("[libdevice] revert: Revert high-performance optimization scheme",2026-06-26 合入)因 libdevice.tanh 测试失败回退了 libdevice 高性能方案的默认启用及 lgamma 优化;PR #833 尝试重新 Enable(Draft 未合入)。引用时说明其反复状态。
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| planning-and-task-breakdownaddyosmani/agent-skillsBreaks work into ordered tasks. Use when you have a spec or clear requirements and need to break work into implementable tasks. Use when a task feels too large to start, when you need to estimate scope, or when parallel work is possible. | Stars: 119k | Updated | Claude CodeCodex | Verified | Install |
| screenpipethedotmack/claude-memSearch screen and audio history recorded by Screenpipe and optionally save selected context into claude-mem. Use when the user asks to recall desktop activity, a meeting, or something seen in another app through Screenpipe. | Stars: 118k | Updated | Claude CodeCodex | Verified | Install |
| contact-widgetnexu-io/open-designSelf-contained floating chat widget with welcome screen, social links, meeting button, and message input. Single HTML file, zero dependencies. | Stars: 108k | Updated | Claude CodeCodex | Verified | Install |
| meeting-notesnexu-io/open-designMeeting notes page — title bar with attendees, agenda checklist, decisions block, action items table with owners + dates, and a "next meeting" footer. Use when the brief mentions "meeting notes", "minutes", "1:1 notes", "all-hands recap", or "会议纪要". | Stars: 108k | Updated | Claude CodeCodex | Verified | Install |