rtl-p5s-perf-policy

name: rtl-p5s-perf-policy description: "Policy rules, performance monitor naming conventions, BFM comparison thresholds (10% deviation), metric definitions, and verification checklists. Pure reference — no orchestration." user-invocable: false

Performance Verification Policy

Performance Monitor Naming Conventions

Performance monitor instrumentation and testbenches MUST follow project conventions (CLAUDE.md):

• Signal references: i_ prefix for inputs, o_ prefix for outputs (e.g., i_valid, o_stall)
• Clock: clk (single domain) or {domain}_clk (multiple domains, e.g., sys_clk) — NOT clk_i
• Reset: rst_n (single domain) or {domain}_rst_n (multiple domains, e.g., sys_rst_n) — NOT rst_ni
• Performance counter instances: u_ prefix (e.g., u_perf_counter)
• Use logic for all signal declarations (NOT reg/wire)

Escalation & Stop Conditions

• Performance deficit >20% vs BFM → escalate to rtl-architect for pipeline analysis
• BFM baseline file missing → run bfm-develop first, halt rtl-p5s-perf-verify
• Performance monitor uses wrong signal names → perf-verifier must fix before re-run

Final Checklist

• Performance monitors use correct signal names (i_/o_ prefix, sys_clk)
• Deterministic performance vectors used
• sim/{module}/*_perf.json written for all modules
• RTL vs BFM comparison done per metric
• All deviations >10% flagged with root cause analysis

Stress Vector Patterns and Cycle-Accurate Measurement

Performance vectors should stress maximum throughput (back-to-back i_valid high, no gaps). Also run a stall-stress vector (frequent o_ready deassertion) to expose stall handling bugs. Use $time with sys_clk edges for cycle-accurate measurement in SV testbenches.