Professional CPU performance optimization skill covering 10 performance dimensions, Top-Down Microarchitecture Analysis (TMA) workflow, and multi-agent optimization.
Perf-Tune is a professional performance optimization skill for AI coding agents. It provides systematic CPU performance analysis and source-level optimization.
- 5 Languages: C, C++, Rust, Go, Python (native extensions)
- 10 Performance Dimensions: P1-P10 coverage matrix (Memory Access through I/O)
- TMA Workflow: Top-Down Microarchitecture Analysis from L1-L3
- Multi-Agent: Parallel optimization agents for memory, computation, branch prediction, code layout, multithreading
- 100+ Optimization Techniques: Cache-friendly data structures, vectorization, loop transforms, branchless code, PGO/BOLT, false sharing elimination, huge pages, system tuning
- Tool Integration: Linux perf, Intel VTune, AMD uProf, Apple Instruments, cargo bench, go test -bench
cp -r perf-tune ~/.claude/skills/
cp -r perf-tune ~/.codex/skills/
cp -r perf-tune ~/.config/opencode/skills/"Optimize this code", "Profile this program", "Find the bottleneck", "/perf", "/tune", "/optimize"
| Mode | Use Case | Scope |
|---|---|---|
| Quick | Hotspot identification | PMU counters, top bottlenecks, quick wins |
| Standard | Module-level optimization | Full TMA, multi-dimension, 1-2 rounds |
| Deep | Whole-program optimization | All 10 dims, iterative, multi-agent, 2-3 rounds |
User: /tune deep ./my_project → [MODE] deep → [RECON] C++/CMake, x86-64 → [MEASURE] Backend:Memory 45%, IPC 0.7 → [PLAN] 3 Agents... → [REPORT] IPC +35%, cache-miss -60%
| Track | Dimensions | Method | What It Optimizes |
|---|---|---|---|
| Memory-driven | P1, P7 | Measure miss rates → locate access → apply cache/prefetch/layout | Present memory bottlenecks |
| Compute-driven | P2, P6 | Measure IPC/ports → check vectorize/loop/inline → apply | Present computation inefficiencies |
| Layout-driven | P4, P9 | Measure Frontend Bound → check layout/PGO/ITLB → apply | Code layout issues |
| Concurrency-driven | P5 | Measure scaling → check false sharing/locks/scheduling → apply | Multithreading issues |
| Config-driven | P8, P10 | Check system config → compare baseline → adjust | Misconfigurations |
| # | Dimension | Coverage |
|---|---|---|
| P1 | Memory Access | Cache-friendly data structures, DTLB, prefetching |
| P2 | Computation | Vectorization, loop optimization, inlining |
| P3 | Branch Prediction | Branchless, lookup tables, predication |
| P4 | Front-End | Code layout, BB placement, PGO/BOLT, ITLB |
| P5 | Multithreading | Scaling, false sharing, task scheduling |
| P6 | Data Dependencies | Critical path analysis, chain breaking, ILP |
| P7 | Memory Bandwidth | Compression, quantization, streaming stores |
| P8 | System Tuning | NUMA, huge pages, kernel params, C-states |
| P9 | Compiler | Flags, pragmas, function attributes, LTO |
| P10 | I/O & External | System calls, external latency, async I/O |
Phase 1: Recon + Measurement
Phase 2: TMA Diagnosis → Agent allocation
Phase 3: Parallel Agents (R1)
Phase 4: Coverage Evaluation → R2 if gaps
Phase 5: Report with before/after metrics
perf-tune/
├── SKILL.md
├── agent.md
├── README.md / README_CN.md
└── references/
├── core/ (6) # Measurement, microarchitecture, terminology, analysis, TMA, CPU features
├── optimization/ (15) # Memory(4) + Computation(4) + Branch(1) + CodeLayout(3) + Other(2) + Multithreading(1)
├── checklists/ (5) # Coverage matrix + C++, Rust, Go, Python
├── languages/ (3) # C++, Rust, Go deep-dive patterns
└── tools/ (3) # perf, VTune, benchmarking
Languages: C, C++, Rust, Go, Python | Architectures: x86-64, ARM64, RISC-V | Tools: Linux perf, Intel VTune, AMD uProf, Apple Instruments, cargo bench, go test -bench
This skill provides optimization recommendations. Always verify correctness by running tests after any code change. Some optimizations may be architecture-specific and not portable.