kali20gakki/msAgent / skills/operator/msprof-op-simulator-pipeline-analysis
msprof-op-simulator-pipeline-analysis
Verified使用 msprof op simulator 采集并分析昇腾 SIMD、SIMT 及混合算子的流水空泡、同步等待、忙碌阶段耗时和核间尾部,结合 trace.json、仿真 dump、热点与源码,生成 Markdown 分析报告和优化建议;经用户授权后修改代码、复测并生成 Markdown 优化对比报告。适用于指令流水诊断与有证据的优化,不用于仅有上板指标的性能归因。
| File | Kind | Size |
|---|---|---|
| SKILL.md | skill md | 13.8 KB |
| references/pipeline-analysis.md | markdown | 8.7 KB |
| references/simd-optimization-map.md | markdown | 10.2 KB |
| references/simt-optimization-map.md | markdown | 10.0 KB |
references/simt-optimization-map.md
View on GitHub# SIMT 与混合编程:从流水位置到线程级原因
## 先判断数据走哪条路径
从源码与实际启动接口确认 SIMT Kernel 或 `__simt_vf__`,并记录线程块数、每块线程数、编译期最大线程数及索引公式。外层 SIMD 的核数、SIMT VF 的线程数和独立 SIMT Kernel 的 grid/block 不是可互换的参数。
- **直接访问 GM**:SIMT 线程的 GM load/store 在 VF 内发生,MTE2/MTE3 可以没有有效搬运。先查线程地址合并、缓存与指令效率,不能将“没有 MTE 重叠”判为单缓冲问题。
- **UB 中转**:外层 MTE2 → SIMT VF(UB) → MTE3,分别检查外层数据块依赖和内部 SIMT 访存;前者适用[共同参考](pipeline-analysis.md),后者按本表分析。
- **SIMD/SIMT 混合**:按 VF 片段分别归因;出现 PUSHQ、VECTOR、RVEC 不能独自证明是 SIMT。AIC/AIV MIX 的矩阵同步是另一类协作,不直接对应 Warp 发散。
## SIMT 技巧与流水证据
| 现象 / 位置 | 定因前的补证 | 官方技巧与适用改动 | 最小对照应验证什么 |
| --- | --- | --- | --- |
| 少数核持续工作,或大量短任务反复调度 | 全核覆盖、线程块数、物理核数、每线程工作量;仅一个核或采样核不能定因 | [线程块配置][grid]:结合规模调整 gridDim,必要时用 grid-stride 循环摊薄调度 | 对相同逻辑元素比较核间尾部、调度次数与完整耗时;不能将“块越多越快”或固定64核作为通用结论 |
| VF 内出现大量栈 load/store,调线程数后耗时反常 | 目标版本编译器资源报告、最大线程数、实际线程数、反汇编栈访问;区分栈用途 | [线程数与寄存器溢出][register]:调整最大线程数、寄存器需求和局部变量活跃范围 | 栈流量/指令与相关 stall 减少,同时观察并行度、每线程循环代价;实际线程数不得超过编译期上限,不照搬其他架构寄存器档位 |
| GM 访问段长,算术少;MTE 可以不活跃 | 展开同一 Warp 同一指令上各 lane 的地址、stride、有效字节与目标 Cache Line;有计数再查实际流量 | [访存合并][coalescing]:重新设计线程到连续数据的映射,必要时 UB 中转转置 | 更少的访存事务或无效字节、较短的访存阶段;逻辑数据量不变。MTE 的512B建议不是 Warp 合并访问的替代条件 |
| GM 已连续,但 UB 访问或转置阶段仍拖长 | 对照目标型号的 bank/group/subbank 低位映射,计算 Warp 的并发地址;相应证据支持后再归因 | [避免 SIMT UB subbank 冲突][bank]:padding、布局或线程映射 | 冲突相关代价和 UB 阶段减少;记录 padding 占用,不将 SIMD Bank 指标直接等同于 SIMT subbank 冲突 |
| 重复访问少量表数据时 VF 访存成本高 | 热点/流式数据的复用范围、工作集、当前 DCache 路径与支持的缓存策略;只有时序则是候选 | [DCache 策略][dcache]、哈希探测场景的 [SoA 布局][soa] | 热点有效复用增加、无用字段或重载流量减少;不能由 VF 时长计算 Cache 命中率,也不能强制所有数据驻留/旁路 |
| 大量窄 load/store 或标量元素访问支撑很少有效数据 | 指令位宽、结构体/指针对齐、lane 地址、奇数尾块和现有运算形式 | [短向量类型][short]、[类型对齐][align]:语义允许时用 half2 等或调整类型对齐 | 访存/发射指令减少、有效字节比例改善;保持数值与越界约束,宽类型并不自动保证访存合并或总吞吐提高 |
| 控制流热点,VF 忙但有效 lane 少;Warp 结束时间分散 | 同一 Warp 内真实输入对应的分支条件、控制流/掩码、反汇编或可用采样;Warp 寿命差不是发散证明 | [减少 Warp 内分支分歧][branch]、简单分支的 [predicate 优化][predicate] | 分支切换、串行分支或无效 lane 成本减少;比较实际指令,编译器可能已谓词化。不能将有副作用/越界分支改为无条件执行 |
| 整数除法、取模或循环寻址是主要指令热点 | 除数是否固定、类型/符号、输入范围、源码与真实指令;排除只是等待访存 | [快除预计算][division]、提取循环不变量及简化可证明冗余寻址 | 目标指令或依赖链缩短、有效搬运工作可保持不变;验证除零、符号和舍入边界,不能把无符号定除数算法直接套到一般除法 |
| 共享地址原子更新集中,线程进度被更新操作拖住 | 实际 atomic 指令、目标地址分布、返回值是否被使用、数据范围与必要的跨线程语义 | [原子优化][atomic]:寄存器/UB 局部聚合后 GM 汇总,合法时减少未用返回值或选择足够位宽 | GM 原子次数/竞争与相应执行代价下降;记录局部聚合和同步成本,验证溢出、浮点累加顺序和结果,不凭图删除必要原子 |
没有支持的线程级事件或补充计数时,仍可报告外层 VF 时序和源码推导的候选,但不能伪造 Warp stall、lane 效率或 occupancy。资源报告中的 Stack size 与动态 spill 代价也不是同一个量。
PCSampling、Warp Stall、全核 Occupancy 等数据可作为补证,其获取方式和适用范围先从 msopprof 仓库读取;采样占比不是精确周期分解,独立 Warp 重放不能与仿真对齐成同一条时间轴。
## SIMD/SIMT 混合的衔接
| 外层与内部现象 | 有条件的改动 | 观察目标与限制 |
| --- | --- | --- |
| SIMT 直接 GM 的代价高,而连续块可装入 UB | [UB 中转][ub],再评估可独立复用的 Ping/Pong 与输入/输出同步 | 新出现 MTE2/MTE3 与 VF 的跨块重叠,内部 GM 访问转为 UB;同时核算 DCache、系统保留及 VF 栈,不能直接复用纯 SIMD 的全部可用 UB 数值 |
| 外层频繁调用极短 SIMT VF,启动/调度成本压过实际计算 | [线程块与 VF 调用配置][hybrid-grid],在容量和依赖允许时合并工作粒度、减少调用 | VF 调用/外层准备次数减少,内部指令与每线程工作增加可能抵消收益;不通过增大外层核数解决所有问题 |
| SIMD 标量串行处理大量分支判断,成为下一阶段生产者 | [SIMT 并行分支判断][hybrid-branch],仅将合适片段交给 SIMT | 外层 SCALAR 代价减少,但新增 VF 调用、访存和汇合同样计入;保持同一 Warp 的分支分析及程序语义 |
引入 UB 或重分块会改变物理搬运、调用与调度数量。比较相同逻辑输入和精度,分别验证“外层重叠”和“内部执行效率”;外层气泡消失并不证明内部访存或分支已经最优。
## 来源与版本
依据官方 asc-devkit 提交 `fba3e6925684e20cae884aa86300814e9cc64d05` 的 [SIMT 优化概述][overview]和[混合编程概述][hybrid]及下列专题整理。流水映射与验证顺序是本 skill 的分析方法,具体 API、Cache Line、Bank 与资源上限按目标产品文档确认。
[overview]: https://gitcode.com/cann/asc-devkit/blob/fba3e6925684e20cae884aa86300814e9cc64d05/docs/zh/guide/operator_practice/simt_operator_optimization/overview.md
[hybrid]: https://gitcode.com/cann/asc-devkit/blob/fba3e6925684e20cae884aa86300814e9cc64d05/docs/zh/guide/operator_practice/simd_simt_hybrid_optimization/overview.md
[grid]: https://gitcode.com/cann/asc-devkit/blob/fba3e6925684e20cae884aa86300814e9cc64d05/docs/zh/guide/operator_practice/simt_operator_optimization/execution_config/thread_block_config.md
[register]: https://gitcode.com/cann/asc-devkit/blob/fba3e6925684e20cae884aa86300814e9cc64d05/docs/zh/guide/operator_practice/simt_operator_optimization/execution_config/thread_count_register_overflow.md
[coalescing]: https://gitcode.com/cann/asc-devkit/blob/fba3e6925684e20cae884aa86300814e9cc64d05/docs/zh/guide/operator_practice/simt_operator_optimization/memory_access/memory_access_merge.md
[bank]: https://gitcode.com/cann/asc-devkit/blob/fba3e6925684e20cae884aa86300814e9cc64d05/docs/zh/guide/operator_practice/simt_operator_optimization/memory_access/avoid_ub_bank_conflict_simt.md
[dcache]: https://gitcode.com/cann/asc-devkit/blob/fba3e6925684e20cae884aa86300814e9cc64d05/docs/zh/guide/operator_practice/simt_operator_optimization/memory_access/dcache_access.md
[soa]: https://gitcode.com/cann/asc-devkit/blob/fba3e6925684e20cae884aa86300814e9cc64d05/docs/zh/guide/operator_practice/simt_operator_optimization/memory_access/soa_layout_efficiency.md
[short]: https://gitcode.com/cann/asc-devkit/blob/fba3e6925684e20cae884aa86300814e9cc64d05/docs/zh/guide/operator_practice/simt_operator_optimization/memory_access/short_vector_efficiency.md
[align]: https://gitcode.com/cann/asc-devkit/blob/fba3e6925684e20cae884aa86300814e9cc64d05/docs/zh/guide/operator_practice/simt_operator_optimization/memory_access/type_alignment_transfer.md
[branch]: https://gitcode.com/cann/asc-devkit/blob/fba3e6925684e20cae884aa86300814e9cc64d05/docs/zh/guide/operator_practice/simt_operator_optimization/control_flow/avoid_warp_branch.md
[predicate]: https://gitcode.com/cann/asc-devkit/blob/fba3e6925684e20cae884aa86300814e9cc64d05/docs/zh/guide/operator_practice/simt_operator_optimization/control_flow/branch_handing.md
[division]: https://gitcode.com/cann/asc-devkit/blob/fba3e6925684e20cae884aa86300814e9cc64d05/docs/zh/guide/operator_practice/simd_simt_hybrid_optimization/instruction_optimization/fast_integer_division.md
[atomic]: https://gitcode.com/cann/asc-devkit/blob/fba3e6925684e20cae884aa86300814e9cc64d05/docs/zh/guide/operator_practice/simt_operator_optimization/instruction_optimization/atomic_instruction_optimization.md
[ub]: https://gitcode.com/cann/asc-devkit/blob/fba3e6925684e20cae884aa86300814e9cc64d05/docs/zh/guide/operator_practice/simd_simt_hybrid_optimization/memory_access/ub_memory_access.md
[hybrid-grid]: https://gitcode.com/cann/asc-devkit/blob/fba3e6925684e20cae884aa86300814e9cc64d05/docs/zh/guide/operator_practice/simd_simt_hybrid_optimization/execution_control/thread_block_config.md
[hybrid-branch]: https://gitcode.com/cann/asc-devkit/blob/fba3e6925684e20cae884aa86300814e9cc64d05/docs/zh/guide/operator_practice/simd_simt_hybrid_optimization/compute_optimization/simt_branch_judgment.md
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| writing-skillsobra/superpowersUse when creating new skills, editing existing skills, or verifying skills work before deployment | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |
| xlsxanthropics/skillsUse this skill any time a spreadsheet file is the primary input or output. This means any task where the user wants to: open, read, edit, or fix an existing .xlsx, .xlsm, .xltx, .csv, or .tsv file (e.g., adding columns, computing formulas, formatting, charting, cleaning messy data); create a new spreadsheet from scratch or from other data sources; or convert between tabular file formats. Trigger especially when the user references a spreadsheet file by name or path — even casually (like "the xlsx in my downloads") — and wants something done to it or produced from it. Also trigger for cleaning or restructuring messy tabular data files (malformed rows, misplaced headers, junk data) into proper spreadsheets. The deliverable must be a spreadsheet file. Do NOT trigger when the primary deliverable is a Word document, HTML report, standalone Python script, database pipeline, or Google Sheets API integration, even if tabular data is involved. | Stars: 198k | Updated | Claude CodeCodex | Verified | Install |
| shipping-and-launchaddyosmani/agent-skillsPrepares production launches. Use when preparing to deploy to production, or when asking what needs to be in place before shipping. Use when you need a pre-launch checklist, when setting up monitoring, when planning a staged rollout, or when you need a rollback strategy. | Stars: 119k | Updated | Claude CodeCodex | Verified | Install |
| childhoodthedotmack/claude-memBabysit a shipped release after merge, publication, or deployment. Verify published packages, production behavior, installed plugins, and late CI or user reports until the observation window is clean; use babysit for PR readiness before release. | Stars: 118k | Updated | Claude CodeCodex | Verified | Install |