Agent Skills › uw-syfi/vibesys › kernel-rocprof-analysis

kernel-rocprof-analysis

GitHub

分析AMD rocprof性能数据，诊断ROCm内核瓶颈。通过追踪和计数器识别调度、占用率及内存带宽问题，提供优化假设与验证建议。

resources/skills/kernel-rocprof-analysis/SKILL.md uw-syfi/vibesys

Trigger Scenarios

需要诊断ROCm内核性能瓶颈拥有rocprof捕获数据分析GPU计算性能

Install

npx skills add uw-syfi/vibesys --skill kernel-rocprof-analysis -g -y

More Options

Non-standard path

npx skills add https://github.com/uw-syfi/vibesys/tree/main/resources/skills/kernel-rocprof-analysis -g -y

Use without installing

npx skills use uw-syfi/vibesys@kernel-rocprof-analysis

指定 Agent (Claude Code)

npx skills add uw-syfi/vibesys --skill kernel-rocprof-analysis -a claude-code -g -y

安装 repo 全部 skill

npx skills add uw-syfi/vibesys --all -g -y

预览 repo 内 skill

npx skills add uw-syfi/vibesys --list

SKILL.md

Frontmatter

{
    "name": "kernel-rocprof-analysis",
    "description": "Interpret AMD rocprofv3 and rocprof-compute evidence for a ROCm kernel task. Use when a rocprof capture is available and a measured bottleneck needs diagnosis."
}

Kernel rocprof analysis

Identify the candidate's dispatches from a system trace before collecting counters. Library and JIT-compiled kernel names rarely match a friendly operator name, so take the exact name from the trace instead of guessing a filter. Match each dispatch to the benchmarked candidate, input shape, dtype, device, and warmup state; exclude first-call JIT compilation and tuning from the captured window.

Each PMC pass reruns the profiled workload, and counter and thread-trace collection perturb timing. Use the task benchmark for before and after performance decisions.

Form one hypothesis from the relevant counters. Occupancy on CDNA follows wavefronts of 64 lanes and is bounded by VGPR, AGPR, LDS, and scratch use; low occupancy alone does not prove it limits throughput. HBM bandwidth and L2 hit rates need byte counts and access patterns to be meaningful. For matrix-heavy kernels, compare MFMA issue rate with the architecture's peak before blaming memory. LDS bank conflicts matter only when LDS traffic is on the critical path. Reach for instruction-level thread trace only after counters point at a specific stall class.

Report the capture identity, dominant dispatches, counter values, uncertainty or unavailable counters, and the next experiment. Confirm any optimization with the accuracy checker and scored benchmark.