Constrained Randomization and Reproducibility

Generate varied protocol-correct transactions and reproduce a failure exactly.

Random, but valid

Constrained randomization generates combinations within a defined space. It does not replace the verification plan. It explores more interactions than a short list of directed tests.

class dma_request;
  rand bit [31:0] address;
  rand int unsigned length;
  rand bit write;
 
  constraint legal_length {
    length inside {[1:256]};
  }
 
  constraint aligned_address {
    address[1:0] == 2'b00;
  }
 
  constraint boundary {
    address[11:0] + length <= 4096;
  }
endclass

These constraints define a legal command: bounded length, aligned address, and a transfer that does not cross a 4 KiB page.

Always check the result

randomize() returns 1 when it finds a solution and 0 when constraints are inconsistent.

dma_request req = new();
 
assert (req.randomize() with {
  write == 1'b1;
  length inside {16, 32, 64};
}) else $fatal(1, "Request constraints have no solution");

Ignoring the return value can send an old or default value to the driver. A randomization failure is a testbench error that needs diagnosis.

Distribution and interesting cases

dist gives extra weight to selected values without making them mandatory.

constraint length_distribution {
  length dist {
    1       := 5,
    [2:63]  := 1,
    64      := 5,
    256     := 5
  };
}

Boundary values have extra weight here. Distribution should follow the verification plan and observed coverage rather than an arbitrary choice.

randc cycles through a small domain before repeating values. It is not suitable for wide vectors whose complete cycle would be unrealistic.

Reproduce a bug

A pseudo-random generator is deterministic for a given seed as long as environment, call order, and tool remain comparable. Test logs should record:

• the global or local seed;
• test name and configuration;
• DUT parameters;
• failing transaction and sequence number;
• code and simulator version.

Rerun with the same seed first. Then reduce the scenario to a few transactions to speed up diagnosis.

Key points

• Constraints define the valid or deliberately invalid space to explore.
• Always check the return value of randomize().
• Distributions target important cases and coverage holes.
• randc fits small domains.
• Record seed, configuration, and the failing transaction.

📝 Test your knowledge - Chapter quiz