aprof-performance-workflow

Plugin aprof-performance-workflow from marketplace aprof in jieran-zhang/Aprof. Commands read from the repo's manifest — 21 days ago.

Skills (21)

• ascendc-remote-kernel-deploy

  通过 SSH 将本地 Ascend C 算子工程同步到远程 CANN 主机，在远端完成 kernel 编译、msprof 采集（simulator / 上板 msprof / msprof op），并把 trace.json、instr_exe.csv、op_summary CSV 等报告拉回本地。触发：远程部署算子、远程 msprof 仿真或上板 profiling、SSH/SFTP 拉取报告时使用。深度指标解读见 ops-profiling skill。

• ascendc-api-best-practices

  Check Ascend C copy, queue, synchronization, reduction and matrix API constraints against the installed SDK before an AProf execution edit.

• ascendc-aprof-diagnosis

  Build source, workload, and profile evidence for an Ascend C performance task. Use to activate auditable evidence predicates, produce non-exclusive anchor facets and provisional mechanism hypotheses, request missing metrics, or interpret existing profiling data before SkillGraph routing. Treat every

• ascendc-aprof-optimization

  Instantiate a selected AProf transformation contract into an isolated Ascend C candidate. Use after evidence collection and SkillGraph routing to propose one bounded optimization intent, preserve action parameters and provenance, and submit the draft to the machine candidate gate. This skill cannot

• ascendc-aprof-parallel-planning

  Optimize the algorithm and task decomposition of an Ascend C operator for a verified architecture and workload groups before tiling or pipeline execution tuning. Produce hash-bound parallel and dispatch plans while preserving semantics and precision.

• ascendc-aprof-profiling

  Measure complete kernel tasks and collect Ascend C mechanism evidence with source replay. Use msprof Task Duration for qualified 910C/A3 paired timing, separate PMU captures for mechanisms, and npu-compute on supported 950 systems.

• ascendc-aprof-workflow

  Run a bounded Ascend C optimization session with reusable preparation, one hypothesis and one candidate, exploratory device timing, and separate formal acceptance. Use for operator optimization or plan-only diagnosis.

• ascendc-code-review

  Review an isolated AProf candidate for plan consistency, dispatch coverage, memory bounds, precision and synchronization before its build gate.

• ascendc-crash-debug

  Triage an Ascend C timeout, illegal access, AIC exception or synchronization hang from preserved AProf gate evidence.

• ascendc-docs-search

  Resolve an Ascend C API uncertainty from installed CANN headers and official versioned documentation without an external search skill or service.

• ascendc-env-check

  Inspect the local CANN toolchain, device visibility, profiling tools, and target identity before an AProf build or measurement.

• ascendc-kernel-direct-invoke

  Build a runnable AscendC kernel direct-invoke scaffold from a single kernel file or kernel folder. Use when an agent needs to turn a kernel source into a host-run direct-launch project, generate CMake/run/data/simulator glue, validate the launch ABI, or optionally collect hardware msprof or simulato

• ascendc-perf-optimize

  Separate algorithm parallel planning from execution tuning in an AProf operator candidate, using workload evidence and the published SkillGraph route.

• ascendc-performance-best-practices

  Interpret AProf execution bottlenecks and choose evidence-backed copy, pipeline, buffer, vector, or cube changes within a frozen parallel plan.

• ascendc-precision-debug

  Diagnose a failed AProf correctness gate while preserving the reference contract and recording the first divergent stage.

• ascendc-runtime-debug

  Triage an AProf candidate runtime failure from exact command, binary identity, runtime logs and SDK/device compatibility.

• ascendc-tiling-design

  Design capacity-safe workload dispatch and per-core tiling after an AProf parallel plan, including direct-launch tiling-key plumbing.

• npu-arch

  Build an evidence-backed AProf architecture profile and distinguish 910B/A2, 910C/A3 and 950/3510 capabilities before parallel planning or execution tuning.

• ops-profiling

  Route AProf hardware collection to npu-compute on supported Ascend 950 systems or msprof on 910C/A3, and preserve original reports for runtime source replay.

• ops-simulator

  Inspect Ascend simulator traces and instruction/pipeline gaps with bundled tools, preserving proxy status and target-specific capability limits.

• tool-npu-compute

  Collect and interpret npu-compute operator reports for AProf on supported Ascend 950 systems, or inspect existing .npu-rep, Section CSV, summary.jsonl, HardwareInfo.jsonl and PipeTrace.json evidence. Use after architecture-aware profiling selection or for an explicit npu-compute report request.

Subagents (4)

• aprof-diagnosis-wrapper

  AProf workflow 内部诊断 wrapper。负责把 workflow 输入规范化后交给 aprof-diagnosis-agent。

• aprof-optimization-wrapper

  AProf workflow 内部优化 wrapper。负责规范化 evidence 与 graph route，调用 optimization skill 生成隔离 candidate draft，并把 draft/evidence 交给机器 gate；不自行裁决候选。

• aprof-parallel-planning-wrapper

  Construct architecture-aware AProf algorithm/task plans and workload dispatch before concrete execution optimization.

• aprof-profiling-wrapper

  AProf workflow 内部 profiling wrapper。负责把 diagnosis metric 转换为 profiling_plan.json，并按架构采集、解析原始 report、提交可回放的 metric 来源。

Similar plugins