lazyuvm 0.1.0

Local-first RTL verification helper for UVM scaffolds, protocol checks, logs, and reports.

LazyUVM

LazyUVM is a local-first RTL verification helper that turns simple Verilog/SystemVerilog modules into starter testbenches, protocol guesses, UVM skeletons, and simulation-log triage reports.

It is intentionally small: one Python file, no required Python packages, and optional local AI through Ollama.

The LazyUVM checked badge means the repository's smoke checks pass on GitHub Actions. It is a quick project-health signal, not silicon sign-off.

What It Does Today

• Scans RTL files and reports modules, parameters, ports, widths, and directions.
• Generates starter SystemVerilog testbenches.
• Infers common bus protocols from port names:
  • APB
  • AHB
  • AXI-Lite
  • AXI-Stream
• Prints a simple multi-file hierarchy tree from module instances.
• Reads simulator-style filelists with source files, include dirs, defines, and nested -f entries.
• Emits copy/paste Verilator, VCS, and Questa command lines from a project filelist.
• Resolves simple parameterized port widths such as [WIDTH-1:0] and [(DATA_WIDTH/8)-1:0].
• Runs lightweight static RTL checks for unresolved instances, unconnected ports, reset oddities, and obvious width mismatches.
• Cross-checks simple text spec rules against simulator log timelines.
• Generates UVM skeleton files:
  • interface
  • sequence item
  • driver
  • monitor
  • sequencer
  • sequence
  • agent
  • env
  • smoke test
  • top module
  • filelist
• Emits APB-aware driver and monitor handshake timing.
• Emits AXI-Lite-aware driver and monitor handshake timing across write address, write data, write response, read address, and read data channels.
• Emits protocol-aware SVA and coverage starter files:
  • APB setup/access/wait-state assertions
  • AXI-Lite valid/ready hold assertions
  • APB and AXI-Lite covergroup starter models
• Emits SystemVerilog bind files that attach generated SVA and coverage sentinels to the DUT.
• Runs a local auto-run loop that generates artifacts, lints with Verilator, compiles with iverilog, runs vvp, triages logs, and writes a Markdown report.
• Runs a seed-based local regression loop over selected modules and writes a PASS/FAIL case matrix.
• Runs a small regress.yaml test plan so projects can keep repeatable smoke tests in source control.
• Triages text coverage reports and suggests concrete next stimulus targets for low or missing coverage.
• Writes starter SystemVerilog directed-stimulus skeletons from coverage holes.
• Summarizes simulator logs with built-in rules, a time-ordered event timeline, and optional local Ollama AI.
• Stores and reuses explicit local-only log and coverage issue patterns when --learn-cache and --use-cache are used.

Why This Exists

Research frameworks such as UVLLM and UVMarvel explore ambitious LLM-driven RTL verification flows. LazyUVM is deliberately narrower: a lightweight CLI that a student, junior engineer, or small team can clone and run locally in a few minutes.

Area	Research Frameworks	LazyUVM
Goal	Broad automated verification research	Fast local UVM scaffold generation
Setup	Larger framework and experiment flow	Single Python script
AI	Often benchmarked with large LLM workflows	Optional local Ollama model
First use	Read paper, configure stack, run experiments	Run one command on an RTL file
Current strength	Full research pipeline	Protocol-aware starter files

LazyUVM is not a replacement for a signoff verification flow. It is a bootstrap tool for getting from "I have RTL" to "I have a first UVM environment shape" quickly.

Installation

From a local checkout:

python3 -m pip install -e .
lazyuvm --version
lazyuvm --help

With pipx from GitHub:

pipx install git+https://github.com/jeonghunx/LazyUVM.git

After a PyPI release, the target install flow is:

pipx install lazyuvm

Contributors can still run python3 rtl_scout.py ... directly from the repository, but lazyuvm ... is the preferred CLI.

Release Status

Current package version: 0.1.0

• See CHANGELOG.md for the release history.
• See docs/releases/v0.1.0.md for the first release note draft.
• The GitHub Actions badge should be green before cutting a release tag.
• PyPI publishing is configured through .github/workflows/release.yml and PyPI Trusted Publishing.

Quick Start

cd "/mnt/c/Users/jeonghun-dev/Documents/New project"

lazyuvm --version
lazyuvm --filelist examples/files.f --hierarchy
lazyuvm --filelist examples/files.f --checks
lazyuvm --filelist examples/files.f --auto-run --no-ai
lazyuvm --regress-file examples/regress.yaml --no-ai
lazyuvm --uvm-dir generated_uvm examples/apb_timer.sv

For focused examples, see the sections below for filelists, tool commands, regression, coverage triage, hierarchy, auto-run, log analysis, and reviewed pattern packs.

Generated files appear under:

generated_uvm/apb_timer/

The UVM generator now also writes assertion and coverage sentinels:

generated_uvm/apb_timer/lazyuvm_apb_timer_sva.sv
generated_uvm/apb_timer/lazyuvm_apb_timer_coverage.sv
generated_uvm/apb_timer/lazyuvm_apb_timer_bind.sv
generated_uvm/axi_lite_gpio/lazyuvm_axi_lite_gpio_sva.sv
generated_uvm/axi_lite_gpio/lazyuvm_axi_lite_gpio_coverage.sv
generated_uvm/axi_lite_gpio/lazyuvm_axi_lite_gpio_bind.sv

The bind files use SystemVerilog bind plus implicit port connections, so the generated sentinels can be attached to the DUT without editing the original RTL.

Generated outputs are ignored by Git because they can be recreated at any time.

Filelist Project Linker

Real RTL projects often compile from a files.f file instead of a single source path. LazyUVM can now use that same project entry point:

lazyuvm --filelist examples/files.f --hierarchy
lazyuvm --filelist examples/files.f --checks
lazyuvm --filelist examples/files.f --auto-run --no-ai

Supported filelist syntax:

+incdir+include
+define+SIMULATION
+define+WIDTH=32
-f nested_files.f
rtl/top.sv
rtl/block.sv

File paths and nested filelists are resolved relative to the filelist that mentions them. --auto-run passes include dirs and defines through to Verilator and Icarus Verilog, then records the exact reproduce commands in the Markdown report.

Tool Command Generator

LazyUVM can turn a project filelist into copy/paste commands for common RTL tools without launching licensed software:

lazyuvm --filelist examples/files.f --emit-tool verilator
lazyuvm --filelist examples/files.f --emit-tool vcs
lazyuvm --filelist examples/files.f --emit-tool questa

Example output:

verilator --lint-only --sv -f examples/files.f
vcs -sverilog -full64 -f examples/files.f
vlog -sv -f examples/files.f

When you pass --module, LazyUVM treats it as the top/module hint:

lazyuvm --filelist examples/files.f --emit-tool verilator --module chip_top
lazyuvm --filelist examples/files.f --emit-tool vcs --module chip_top
lazyuvm --filelist examples/files.f --emit-tool questa --module chip_top

Regression Runner

For a small local regression:

lazyuvm --filelist examples/files.f --regress --seeds 1,2,3 --no-ai

This writes:

LazyUVM_Regression.md
.lazyuvm_regress/

The regression runner:

• generates a smoke testbench for each selected module
• passes +LAZYUVM_SEED=<seed> into vvp
• repeats each module across the requested seed list
• keeps each case in its own workspace folder
• records lint, compile, simulation, and overall PASS/FAIL status in a Markdown matrix

Seed ranges are supported:

lazyuvm --filelist examples/files.f --regress --seeds 1-10 --module chip_top --no-ai

This is still an early screening loop, not signoff. Its job is to make local smoke regressions cheap enough to run constantly.

For a project-owned regression plan:

lazyuvm --regress-file examples/regress.yaml --no-ai

Example regress.yaml:

tests:
  - name: apb_smoke
    files: files.f
    module: apb_timer
    seeds: 1,2

  - name: axi_smoke
    files: files.f
    module: axi_lite_gpio
    seeds: 1,2

Paths in the YAML file are resolved relative to that YAML file, so a test plan under examples/ can simply use files: files.f.

Supported keys:

• name: test name shown in reports
• files or filelist: one filelist path or a short list
• module or top: module to run
• seeds: comma list or range such as 1,2,3 or 1-10

Coverage Triage

For a text coverage report:

lazyuvm --analyze-coverage examples/coverage_report.txt --no-ai

LazyUVM extracts:

• coverage percentages such as line, branch, toggle, and functional coverage
• common vendor-style table rows with covered/total/percent columns
• ratio-only rows such as 0/1 when the row looks like coverage data
• low coverage metrics below a threshold
• explicit miss lines such as 0 hits, Hits: 0, Count = 0, uncovered, not covered, and never hit
• rule-based next stimulus ideas for APB, AXI-Lite, reset, FSM, and toggle holes

To change the low-coverage threshold:

lazyuvm --analyze-coverage examples/coverage_report.txt --coverage-threshold 80 --no-ai

To write the analysis to a file:

lazyuvm --analyze-coverage examples/coverage_report.txt --coverage-md LazyUVM_Coverage.md --no-ai

With Ollama available, omit --no-ai to add a local summary on top of the rule-based findings.

To generate starter SystemVerilog stimulus tasks from the same coverage holes:

lazyuvm --analyze-coverage examples/coverage_report.txt --coverage-stimulus-dir generated_stimulus --module apb_timer --no-ai
lazyuvm --analyze-coverage examples/vendor_coverage_report.txt --coverage-stimulus-dir generated_stimulus --module apb_timer --no-ai

This writes a skeleton such as:

generated_stimulus/lazyuvm_apb_timer_coverage_stimulus.sv

The generated file contains task stubs for holes such as APB wait states, AXI-Lite backpressure, reset edges, toggles, and generic directed bins. It also records parsed targets such as source locations, coverpoints, bins, branch conditions, and likely protocol signals. It intentionally leaves BFM/UVM calls as TODO comments because real projects name their sequence APIs differently.

Hierarchy Scan

For a folder with multiple RTL files:

lazyuvm --hierarchy examples
lazyuvm --hierarchy --module chip_top examples

This prints top-module candidates and a simple instance tree, for example:

chip_top [AXI-Lite 95%, APB 80%]
|-- apb_timer u_timer [APB 90%]
`-- axi_lite_gpio u_gpio [AXI-Lite 100%]

Width Solver

For parameterized designs:

lazyuvm --widths examples/axi_lite_gpio.sv

This resolves simple integer expressions in packed ranges:

axi_lite_gpio:
  parameters:
    ADDR_WIDTH = 12 => 12
    DATA_WIDTH = 32 => 32
  ports:
    input awaddr: [ADDR_WIDTH-1:0] => 12 bits
    input wstrb: [(DATA_WIDTH/8)-1:0] => 4 bits

Static Checks

For a quick local RTL review:

lazyuvm --checks examples

LazyUVM checks what it can prove from the local source:

• unresolved module instances
• unconnected named child ports
• obvious parent/child port width mismatches
• clock-like modules without reset-like inputs
• reset-like ports with suspicious directions

These checks are intentionally conservative: if LazyUVM cannot infer a width confidently, it avoids guessing a mismatch.

Auto-Run Report

For a one-command local triage loop:

lazyuvm --auto-run examples/apb_timer.sv

This writes:

LazyUVM_Report.md
.lazyuvm/

The auto-run loop:

• scans the RTL
• runs lightweight static RTL checks
• generates a starter testbench
• generates UVM skeleton files
• lints the RTL with Verilator when available
• compiles with iverilog
• runs with vvp
• triages lint/compile/simulation logs
• optionally asks local Ollama for an explanation
• writes patch suggestions under .lazyuvm/patches/

If Verilator is not installed or you only want the simulator path:

lazyuvm --auto-run --no-lint examples/apb_timer.sv

Safety rule: LazyUVM does not modify your RTL source files during auto-run. Patch suggestions are isolated for manual review.

Local AI Log Analysis

AI is optional. The rule-based analyzer works without Ollama:

lazyuvm --analyze-log examples/sim_error.log --no-ai
lazyuvm --analyze-log examples/timing_trace.log --no-ai

The log analyzer extracts suspicious lines and a compact timeline before local AI is used:

Timeline:
  100ns L3 [event] signals: psel=1, penable=0, pwrite=1
  120ns L5 [error] signals: penable=1, penable=0, psel=1, pready=0
  130ns L7 [fatal] signals: psel=1, penable=0, pready=0

With Ollama:

curl -fsSL https://ollama.com/install.sh | sh
ollama pull llama3.2:1b

lazyuvm --analyze-log examples/sim_error.log --ollama-model llama3.2:1b

Spec-Aware Log Cross-Check

For a small spec/log comparison:

lazyuvm --analyze-log examples/timing_trace.log --spec examples/spec.txt --no-ai

LazyUVM extracts simple requirement lines containing words such as must, shall, within N cycles, after reset, and remain high until, then compares the supported rules against the log timeline.

Example requirement:

PENABLE must assert one cycle after PSEL is asserted.
PSEL must remain high until PREADY.

The checker is intentionally conservative. It reports direct evidence from the timeline, for example:

Spec expects `penable` after `psel`, but timeline never shows `penable=1` before the failure evidence.

This is a fact cross-check, not a full natural-language spec verifier.

Local Pattern Cache

LazyUVM does not upload project data. By default, it only prints analysis results. If you explicitly add --learn-cache, it stores a small local pattern entry under .lazyuvm_cache/patterns.json.

lazyuvm --analyze-log examples/timing_trace.log --spec examples/spec.txt --learn-cache --no-ai
lazyuvm --analyze-coverage examples/coverage_report.txt --learn-cache --no-ai
lazyuvm --cache-report

After patterns are learned, use --use-cache to compare the current log or coverage report with the local project memory:

lazyuvm --analyze-log examples/timing_trace.log --spec examples/spec.txt --use-cache --no-ai
lazyuvm --analyze-coverage examples/coverage_report.txt --use-cache --no-ai

The cache keeps normalized fingerprints, categories, repeated-hit counts, signal names found in timelines, and short rule-based hints. Cache matching uses local fingerprints plus category, signal, and hint-token overlap. It is meant to become a private project memory, not hidden telemetry.

Use a custom cache folder when you want to keep experiments separate:

lazyuvm --analyze-log examples/timing_trace.log --learn-cache --cache-dir /tmp/lazyuvm_cache --no-ai
lazyuvm --analyze-log examples/timing_trace.log --use-cache --cache-dir /tmp/lazyuvm_cache --no-ai

For team sharing without a central server, export and import the cache as a plain JSON file:

lazyuvm --cache-export lazyuvm_patterns.json
lazyuvm --cache-import lazyuvm_patterns.json
lazyuvm --cache-report

Import merges by fingerprint. Re-importing the same file updates matching entries instead of duplicating them, so teams can pass around a reviewed pattern pack in source control, chat, or an internal artifact store.

For a reviewed team pattern pack, mark only trusted entries before exporting:

lazyuvm --cache-report
lazyuvm --cache-review b5064cf --review-note "APB setup/access timeout seen in smoke regressions"
lazyuvm --cache-export-reviewed lazyuvm_reviewed_patterns.json
lazyuvm --cache-import lazyuvm_reviewed_patterns.json

--cache-review and --cache-unreview accept unique fingerprint prefixes, so you do not need to type the full hash.

For a self-hosted team registry, publish reviewed packs into a shared folder:

lazyuvm --registry-dir /shared/lazyuvm_registry --registry-publish apb_smoke
lazyuvm --registry-dir /shared/lazyuvm_registry --registry-list
lazyuvm --registry-dir /shared/lazyuvm_registry --registry-pull apb_smoke

The registry is just files:

lazyuvm_registry.json
packs/apb_smoke.json

That makes it usable with an internal Git repo, shared drive, artifact store, or any company-controlled file server. No external service is required.

As the cache grows, inspect and prune it:

lazyuvm --cache-stats
lazyuvm --cache-prune 100
lazyuvm --cache-stats

Pruning keeps the highest-count and most recently seen entries first. Use export before pruning if you want a backup:

lazyuvm --cache-export lazyuvm_patterns_backup.json
lazyuvm --cache-prune 100

License

LazyUVM is released under GPLv3. Commercial licensing can be considered separately for companies that need different terms.

This is not legal advice; talk to a lawyer before using any open-source license strategy commercially.