Profiling Skill

Use this skill when the user needs to investigate performance issues, generate flamegraphs, or analyze kernel execution times. IMPORTANT: All profiling runs that produce useful data MUST be recorded in the project's benchmark log.

Tools

•.agent/skills/profiling/scripts/auto_profile.sh: Primary Tool. Automates the entire data collection and reporting pipeline.

Workflows

1. Automated Profiling (Recommended)

Use the auto_profile.sh script to Run -> Visualize -> Update Log in one step.

Command:

bash

./.agent/skills/profiling/scripts/auto_profile.sh \
    --cmd "<full_android_command>" \
    --output-name "<identifier>"

Example:

bash

./.agent/skills/profiling/scripts/auto_profile.sh \
    --cmd "/data/local/tmp/generate --model-path /data/local/tmp/llm_rs2/models/llama3.2-1b --prompt-file /data/local/tmp/llm_rs2/eval/short_len.txt --num-tokens 128 -b cpu" \
    --output-name "cpu_short_run"

2. Manual Profiling (Legacy)

If you need granular control, refer to benchmarks/GUIDE.md for individual script usage (android_profile.py, visualize_profile.py, etc.).

Best Practices

•Build First: Ensure the binary is built and pushed (use testing skill for this).
•Clean State: Ensure no other heavy tasks are running on the device.