Skip to main content

Regressions

A regression runs the test suites listed in one manifest and combines their results.

Configure a regression​

rtl-buddy-filetype: reg_config

test-configs:
  - design/block_a/verif/tests.yaml
  - design/block_b/verif/tests.yaml

Paths resolve from the directory containing regression.yaml. Each suite keeps its own artefacts and detailed log; the manifest directory receives the regression log and merged outputs.

See YAML Formats: regression.yaml for the schema.

Resolve the manifest​

An explicit config wins:

rb regression --reg-config path/to/regression.yaml

Without it, RTL Buddy checks:

  1. ./regression.yaml in the invocation directory
  2. cfg-rtl-reg.reg-cfg-path in root_config.yaml

Other flow regressions use the same order: explicit -c, ./<flow>_regression.yaml, then the matching cfg-rtl-reg.<flow>-reg-cfg-path. Declare non-root flow manifests in cfg-rtl-reg so graph discovery can find them.

Filter by regression level​

Tests with reglvl in the selected inclusive range run; others report SKIP:

rb regression --reg-level 2000
rb regression --start-level 1000 --reg-level 3000

The default upper level is 0, so an unqualified regression runs must-run tests with reglvl: 0. A test may define one level or builder-specific levels. See Tests.

Reuse compilation​

When tests share compile inputs, reuse a compiled build:

rb regression --share-build

Verilator, VCS, and Icarus support cross-test sharing. Reuse is reported once per build directory per process on the console, and every test's compile.log (and the log file) records its own reuse; add --rebuild to compile even when the stamp says the build is current. See Sharing compiled builds for invalidation and backend limitations.

Run in parallel​

The default --dispatch local runs tests sequentially in the current process. For parallel execution:

rb regression --dispatch local-parallel -j 4
rb regression --dispatch slurm

Dispatch implies shared builds. RTL Buddy expands each suite, creates one build job per unique compile key, then runs dependent simulation jobs and combines their normal results.

local-parallel uses subprocesses on the current host and needs no scheduler. It cannot enforce resources: reservations or collect usage telemetry.

Slurm dispatch requires a Linux submit host, Slurm client commands, and a filesystem shared with compute nodes. See Parallel Dispatch for cluster configuration, resources, failure recovery, and job accounting.