Skip to content

Lecture 05 — Test the checker with deliberate RTL faults

Course guide · Previous · Next: Lecture 06

Before trusting a checker, make it reject errors you understand. The course's RTL protocol lab includes three deterministic fault injections. They modify a temporary copy of the design, run the actual simulator, and require the expected assertion category. They do not edit the checked-in reference or accept a failed run as a successful design.

Run three negative cases

From the repository root:

python3 'Phase 4 - Track A - Xilinx FPGA/1. Xilinx FPGA Development/labs/run_stream_lab.py' --inject-fault overwrite-stall
python3 'Phase 4 - Track A - Xilinx FPGA/1. Xilinx FPGA Development/labs/run_stream_lab.py' --inject-fault wrap-overflow
python3 'Phase 4 - Track A - Xilinx FPGA/1. Xilinx FPGA Development/labs/run_stream_lab.py' --inject-fault retain-on-reset

Each command should report fault_detected with the associated assertion. The wrapper's successful exit means the negative-test experiment caught the expected bug. The candidate simulation itself exits nonzero. This distinction prevents a reporting mistake: a successful fault-detection test is not a passing RTL candidate.

Fault Violated requirement Expected category
Overwrite while stalled Undelivered data must remain stable STALL
Wrap 255+1 to zero Arithmetic saturates at 255 DATA
Retain valid state through reset Reset flushes buffered validity VALID

These mutations were executed with Icarus 12.0 during the course review. The lab also runs multiple deterministic pseudorandom backpressure seeds; see its validation record for the actual runs.

Read a counterexample as a timeline

For overwrite-stall, reconstruct the first failing sequence: which value was accepted, whether the output was valid, whether the consumer was ready, and which next input changed the stored value. The root cause is an incorrect state-update condition, not merely “the scoreboard expected another byte.”

A useful debug note contains the contract clause, first divergence cycle, relevant state and handshakes, proposed cause, and a testable repair. Avoid changing several independent behaviors before rerunning; otherwise you cannot tell which modification fixed the error or introduced another one.

Strengthen the test deliberately

Add a directed reset while full, repeated saturating inputs, prolonged downstream stall, or simultaneous consume/replace. For each added case, predict the reference queue and expected counts before running. If reset flushes an item, the conservation rule is accepted = delivered + flushed + remaining, using a consistent measurement interval.

Formal properties can complement these traces. State assumptions about the input producer, reset, and output consumer. A liveness property needs an environment progress assumption: if ready is allowed to stay low forever, eventual delivery is impossible. A safety property such as output stability during a stall does not require eventual readiness.

Exercise: the overwrite mutation passes when downstream ready is always high. Is the fault injection broken? Answer: no; that stimulus never exercises the faulty transition. Add the relevant stall rather than increasing random runtime without checking coverage.

Your artifact is a fault-to-property table and one annotated failing trace. Continue to Lecture 06 to use these results as structured feedback in a bounded repair loop.