dse-loop logo

dse-loop

dse loop

SKILL.md

Full skill instructions

DSE Loop: Autonomous Design Space Exploration

🔁 Do not wrap this skill in /loop / CronCreate. It already loops internally until its objective is met or it times out. Unlike the verdict-bearing review/audit skills, its stop gate is an objective machine-checkable metric (Type-A), so its self-termination is safe same-model — the reason not to wrap it is scheduler duplication, not the verdict fence. See shared-references/external-cadence.md.

Autonomously explore a design space: run → analyze → pick next parameters → repeat, until the objective is met or timeout is reached. Designed for computer architecture and EDA problems.

Context: $ARGUMENTS

Safety Rules — READ FIRST

NEVER do any of the following:

  • sudo anything
  • rm -rf, rm -r, or any recursive deletion
  • rm any file you did not create in this session
  • Overwrite existing source files without reading them first
  • git push, git reset --hard, or any destructive git operation
  • Kill processes you did not start

If a step requires any of the above, STOP and report to the user.

Constants (override via $ARGUMENTS)

ConstantDefaultDescription
TIMEOUT2hTotal wall-clock budget. Stop exploring after this.
MAX_ITERATIONS50Hard cap on number of design points evaluated.
PATIENCE10Stop early if no improvement for this many consecutive iterations.
OBJECTIVEminimizeminimize or maximize the target metric.

Override inline: /dse-loop "task desc — timeout: 4h, max_iterations: 100, patience: 15"

Typical Use Cases

ProblemProgramParametersObjective
Microarch DSEgem5 simulationcache size, assoc, pipeline width, ROB size, branch predictormaximize IPC or minimize area×delay
Synthesis tuningyosys/DC scriptoptimization passes, target freq, effort levelminimize area at timing closure
RTL parameterizationverilator simdata width, FIFO depth, pipeline stages, buffer sizesmeet throughput target at min area
Compiler flagsgcc/llvm build + benchmark-O levels, unroll factor, vectorization, schedulingminimize runtime or code size
Placement/routingopenroad/innovusutilization, aspect ratio, layer configminimize wirelength / timing
Formal verificationabc/sbybound depth, engine, timeout per propertymaximize coverage in time budget
Memory subsystemcacti / ramulatorbank count, row buffer policy, schedulingoptimize bandwidth/energy

Workflow

Phase 0: Parse Task & Setup

  1. Parse $ARGUMENTS to extract:

    • Program: what to run (command, script, or Makefile target)
    • Parameter space: which knobs to tune and their ranges/options (may be incomplete — see step 2)
    • Objective metric: what to optimize (and how to extract it from output)
    • Constraints: hard limits that must not be violated (e.g., timing must close)
    • Timeout: wall-clock budget
    • Success criteria: when is the result "good enough" to stop early?
  2. Infer missing parameter ranges — If the user provides parameter names but NOT ranges/options, you MUST infer them before exploring:

    a. Read the source code — search for the parameter names in the codebase:

    • Look for argparse/click definitions, config files, Makefile variables, module parameters, #define, parameter (SystemVerilog), localparam, etc.
    • Extract defaults, types, and any comments hinting at valid values

    b. Apply domain knowledge to set reasonable ranges:

    Parameter typeInference strategy
    Cache/memory sizesPowers of 2, typically 1KB–16MB
    AssociativityPowers of 2: 1, 2, 4, 8, 16
    Pipeline width / issue widthSmall integers: 1, 2, 4, 8
    Buffer/queue/FIFO depthPowers of 2: 4, 8, 16, 32, 64
    Clock period / frequencyBased on technology node; try ±50% from default
    Bound depth (BMC/formal)Geometric: 5, 10, 20, 50, 100
    Timeout valuesGeometric: 10s, 30s, 60s, 120s, 300s
    Boolean/enum flagsEnumerate all options found in source
    Continuous (learning rate, threshold)Log-scale sweep: 5 points spanning 2 orders of magnitude around default
    Integer counts (threads, cores)Linear: from 1 to hardware max

    c. Start conservative — begin with 3-5 values per parameter. Expand range later if the best result is at a boundary.

    d. Log inferred ranges — write the inferred parameter space to dse_results/inferred_params.md so the user can review:

    # Inferred Parameter Space
    
    | Parameter | Source | Default | Inferred Range | Reasoning |
    |-----------|--------|---------|---------------|-----------|
    | CACHE_SIZE | config.py:42 | 32768 | [8192, 16384, 32768, 65536, 131072] | powers of 2, ±2x from default |
    | ASSOC | config.py:43 | 4 | [1, 2, 4, 8] | standard associativities |
    | BMC_DEPTH | run_bmc.py:15 | 10 | [5, 10, 20, 50] | geometric, common BMC depths |
    

    e. Boundary expansion — during the search, if the best result is at the min or max of a range, automatically extend that range by one step in that direction (but log the extension).

  3. Read the project to understand:

    • How to run the program
    • Where results are produced (stdout, log files, reports)
    • How to parse the objective metric from output
    • Current/baseline configuration (if any)
  4. Create working directory: dse_results/ in project root

    • dse_results/dse_log.csv — one row per design point
    • dse_results/DSE_REPORT.md — final report
    • dse_results/DSE_STATE.json — state for recovery
    • dse_results/inferred_params.md — inferred parameter space (if ranges were not provided)
    • dse_results/configs/ — config files for each run
    • dse_results/outputs/ — raw output for each run
  5. Write a parameter extraction script (dse_results/parse_result.py or similar) that takes a run's output and returns the objective metric as a number. Test it on a baseline run first.

  6. Run baseline (iteration 0): run the program with default/current parameters. Record the baseline metric. This is the point to beat.

Phase 1: Initial Exploration

Goal: Quickly survey the space to understand which parameters matter most.

Strategy: Latin Hypercube Sampling or structured sweep of key parameters.

  1. Pick 5-10 diverse design points that span the parameter ranges
  2. Run them (in parallel if independent, via background processes or sequential)
  3. Record all results in dse_log.csv:
    iteration,param1,param2,...,metric,constraint_met,timestamp,notes
    0,default,default,...,baseline_val,yes,2026-03-13T10:00:00,baseline
    1,val1a,val2a,...,result1,yes,2026-03-13T10:05:00,initial sweep
    ...
    
  4. Analyze: which parameters have the most impact on the objective?
  5. Narrow the search to the most sensitive parameters

Phase 2: Directed Search

Goal: Converge toward the optimum by making informed choices.

Strategy: Adaptive — pick the approach that fits the problem:

  • Few parameters (≤3): Fine-grained grid search around the best region from Phase 1
  • Many parameters (>3): Coordinate descent — optimize one parameter at a time, holding others at current best
  • Binary/categorical params: Enumerate promising combinations
  • Continuous params: Binary search or golden section between best neighbors
  • Multi-objective: Track Pareto frontier, explore along the front

For each iteration:

  1. Select next design point based on results so far:

    • Look at the trend: which direction improves the metric?
    • Avoid re-running configurations already evaluated
    • Balance exploration (untested regions) vs exploitation (near current best)
  2. Modify parameters: edit config file, command-line args, or source constants

  3. Run the program: execute and capture output

  4. Parse results: extract the objective metric and check constraints

  5. Log to dse_log.csv: append the new row

  6. Check stopping conditions:

    • Timeout reached? → stop
    • Max iterations reached? → stop
    • Patience exhausted (no improvement in N iterations)? → stop
    • Success criteria met (metric is "good enough")? → stop
    • Constraint violation pattern detected? → adjust search bounds
  7. Update DSE_STATE.json:

    {
      "iteration": 15,
      "status": "in_progress",
      "best_metric": 1.23,
      "best_params": {"cache_size": 32768, "assoc": 4, "pipeline_width": 2},
      "total_iterations": 15,
      "start_time": "2026-03-13T10:00:00",
      "timeout": "2h",
      "patience_counter": 3
    }
    
  8. Decide next step → back to step 1

Phase 3: Refinement (if time allows)

If the search converged and there's still time budget:

  1. Local perturbation: try ±1 step on each parameter from the best point
  2. Sensitivity analysis: which parameters can be relaxed without hurting the metric?
  3. Constraint boundary: if a constraint is nearly binding, explore near-feasible points

Phase 4: Report

Write dse_results/DSE_REPORT.md:

# Design Space Exploration Report

**Task**: [description]
**Date**: [start] → [end]
**Total iterations**: N
**Wall-clock time**: X hours Y minutes

## Objective
- **Metric**: [what was optimized]
- **Direction**: minimize / maximize
- **Baseline**: [value]
- **Best found**: [value] ([improvement]% better than baseline)

## Best Configuration
| Parameter | Baseline | Best |
|-----------|----------|------|
| param1    | default  | best_val |
| param2    | default  | best_val |
| ...       | ...      | ... |

## Search Trajectory
| Iteration | param1 | param2 | ... | Metric | Notes |
|-----------|--------|--------|-----|--------|-------|
| 0 (baseline) | ... | ... | ... | ... | baseline |
| 1 | ... | ... | ... | ... | initial sweep |
| ... | ... | ... | ... | ... | ... |
| N (best) | ... | ... | ... | ... | ★ best |

## Parameter Sensitivity
- **param1**: [high/medium/low impact] — [brief explanation]
- **param2**: [high/medium/low impact] — [brief explanation]

## Pareto Frontier (if multi-objective)
[Table or description of non-dominated points]

## Stopping Reason
[timeout / max_iterations / patience / success_criteria_met]

## Recommendations
- [actionable insights from the exploration]
- [which parameters matter most]
- [suggested follow-up explorations]

Also generate a summary plot if matplotlib is available:

  • Convergence curve (metric vs iteration)
  • Parameter sensitivity bar chart
  • Pareto frontier scatter (if multi-objective)

State Recovery

If the context window compacts mid-run, the loop recovers from DSE_STATE.json + dse_log.csv:

  1. Read DSE_STATE.json for current iteration, best params, patience counter
  2. Read dse_log.csv for full history
  3. Resume from next iteration

Key Rules

  • Work AUTONOMOUSLY — do not ask the user for permission at each iteration
  • Every run must be logged — even failed runs, constraint violations, errors. The log is the ground truth.
  • Never re-run an identical configuration — check dse_log.csv before each run
  • Respect the timeout — check elapsed time before starting a new iteration. If the next run is likely to exceed the timeout, stop and report.
  • Parse metrics programmatically — write a parsing script, don't eyeball logs
  • Keep raw outputs — save each run's full output in dse_results/outputs/iter_N/
  • Constraint violations are not improvements — a design point that violates constraints is never "best", regardless of the metric
  • If a run crashes, log the error, skip that point, and continue with the next
  • If the same crash repeats 3 times with different configs, the harness code itself is the suspect — discard and reimplement the run/parse script cleanly from the spec (a peer move to another patch; delete only the script, never dse_log.csv / dse_results/; see shared-references/external-cadence.md § Let a broken attempt restart, not just patch). Before resuming the sweep, re-validate metric comparability: re-parse one COMPLETED iteration's raw output from dse_results/outputs/iter_N/ with the new parser and confirm it reproduces that row of dse_log.csv; on mismatch, either fix the parser or re-parse and flag all affected rows — never mix two parsing semantics in one log. If a clean reimplement crashes the same way, stop and report — the spec or the environment is then in question, which is what needs the human

Example Invocations

# Minimal — just name the parameters, let the agent figure out ranges
/dse-loop "Run gem5 mcf benchmark. Tune: L1D_SIZE, L2_SIZE, ROB_ENTRIES. Objective: maximize IPC. Timeout: 3h"

# Partial — some ranges given, some not
/dse-loop "Run make synth. Tune: CLOCK_PERIOD [5ns, 4ns, 3ns, 2ns], FLATTEN, ABC_SCRIPT. Objective: minimize area at timing closure. Timeout: 1h"

# Fully specified — explicit ranges for everything
/dse-loop "Simulate processor with FIFO_DEPTH [4,8,16,32], ISSUE_WIDTH [1,2,4], PREFETCH [on,off]. Run: make sim. Objective: max throughput/area. Timeout: 2h"

# Real-world: PDAG-SFA formal verification tuning
/dse-loop "Run python run_bmc.py. Tune: BMC_DEPTH, ENGINE, TIMEOUT_PER_PROP. Objective: maximize properties proved. Timeout: 2h"