Randomized Config from Test: Controlled Variation, Stable Triage

Randomization strategy

Randomize cfg at test level to explore mode space, then push one coherent policy object into the env.

[TEST][UVM][ENV] random cfg flow

cfg = create()
cfg.apply_defaults()
if !cfg.randomize() with scenario constraints -> fatal
emit cfg summary
set cfg to env*
create env

seed + cfg summary must reproduce behavior

virtual function void build_phase(uvm_phase phase);
  super.build_phase(phase);
  cfg = env_cfg::type_id::create("cfg");
  cfg.apply_defaults();
  if (!cfg.randomize() with {
      num_agents inside {[1:4]};
      timeout_ns inside {[100_000:2_000_000]};
      !(enable_low_power && enable_max_perf);
    }) begin
    `uvm_fatal("CFG_RAND", "env_cfg randomize failed")
  end
  log_cfg(cfg);
  uvm_config_db#(env_cfg)::set(this, "env*", "cfg", cfg);
  env = my_env::type_id::create("env", this);
endfunction

Key takeaways

• Randomize cfg once, then treat it as immutable test policy.

• Constrain for legal architecture combinations.

• Always log solved cfg fields with the seed.

Common pitfalls

• Randomizing cfg in multiple components independently.

• Leaving legality checks only to late runtime assertions.

• No compact cfg summary in failure logs.

Legality and reproducibility

Separate constraint solving, legality validation, and scenario tagging to make random cfg runs debuggable.

Two-step quality gate

function bit is_legal_cfg(env_cfg c);
  if (c.num_agents == 0) return 0;
  if (c.enable_cov && !c.enable_scoreboard) return 0;
  if (c.max_outstanding > 64 && c.mode == "LOW_POWER") return 0;
  return 1;
endfunction

function void finalize_cfg(env_cfg c);
  if (!is_legal_cfg(c))
    `uvm_fatal("CFG_LEGAL", "illegal cfg combination")
endfunction

• Do not rely on constraints alone for cross-field policy checks.

• Tag cfg with scenario labels for dashboard grouping.

• Promote high-value randomized cfgs into directed regression tests.

Debug packet format

[TEST] cfg debug packet

seed=12345
test=stress_backpressure_mix_test
cfg:
  num_agents=3
  timeout_ns=900000
  mode=PERF
  max_outstanding=48
  enable_cov=1
  enable_scoreboard=1

[UVM] triage rule

first reproduce with same:
  - test class
  - seed
  - cfg packet
then mutate one cfg dimension at a time

Common pitfalls

• Changing constraints and expecting old seeds to behave identically.

• Huge cfg dumps with no prioritized key fields.

• Combining random cfg and random sequence chaos in first triage pass.