MCP‑EDA Example Flow

An end‑to‑end digital implementation reference that combines Synopsys Design Compiler for RTL‑to‑gate synthesis and Cadence Innovus for physical implementation, wrapped in a set of lightweight Python/FastAPI micro‑services plus automation scripts.

The goal is to let you:

• spin up every step (setup → compile → floor‑plan → place → CTS → power‑plan → route → save) with a single command
• or drive / debug individual steps via simple REST calls
• keep all tool settings under version control (CSV + TCL) for fully repeatable runs

---

1. Repository Layout

.
├── config/                  # CSV knobs for each stage
│   ├── synthesis.csv
│   ├── imp_global.csv
│   ├── placement.csv
│   └── cts.csv
├── designs/
│   └── des/                 # example AES‑DES design
│       ├── rtl/             # Verilog sources
│       └── implementation/  # auto‑generated results
├── libraries/               # FreePDK45 timing/lib/LEF
├── scripts/
│   └── FreePDK45/
│       ├── tech.tcl         # common tech setup
│       ├── frontend/        # Design‑Compiler TCL
│       └── backend/         # Innovus TCL (1–8)
├── server/                  # FastAPI micro‑services
│   ├── synth_setup_server.py
│   ├── synth_compile_server.py
│   ├── floorplan_server.py
│   ├── placement_server.py
│   ├── cts_server.py
│   ├── powerplan_server.py
│   ├── route_server.py
│   └── save_server.py
└── run_pipeline.sh          # one‑shot driver

---

2. Prerequisites

Tool	Version (known good)	Notes
Synopsys Design Compiler	V‑2023.12‑SP2	dc_shell in $PATH
Cadence Innovus	21.1 s075	innovus in $PATH
Python	3.9 +	tested with venv
FreePDK45	Nangate OpenCell v1.3	timing/LEF already included

Licenses: You need valid DC & Innovus licenses on the host.

---

3. Quick Start – one line

./run_pipeline.sh       # runs ALL 8 stages, logs to ./logs/

Behind the scenes the script:

1. Spins up (or re‑uses) the 8 REST services on localhost :3333…3340.
2. Calls them in sequence, passing the parameter sets from config/*.csv.
3. Streams JSON back to the console for each stage; artefacts go under
   designs/des/FreePDK45/{synthesis,implementation}/….

If you see ✅ at the end you have a routed & saved GDS/V files ready.

---

4. Running individual stages

Every stage is a tiny FastAPI app exposing POST /<stage>/run.

Example – placement:

curl -X POST http://localhost:3337/place/run      -H 'Content-Type: application/json'      -d '{
           "design": "des",
           "tech":   "FreePDK45",
           "impl_ver":"cpV1_clkP1_drcV1__g0_p0",
           "g_idx":0, "p_idx":0,
           "force": true
         }' | jq .

All servers honour two rules:

• idempotent – rerun with force=true to wipe previous logs/reports.
• stateless – all context (version tags, CSV indices) comes from the JSON.

Port	Service	Endpoint	Depends on
3333	synth_setup_server	/setup/run	source RTL
3334	synth_compile_server	/compile/run	setup artefacts
3335	floorplan_server	/floorplan/run	.mapped.v/.sdc
3337	placement_server	/place/run	floorplan.enc.dat
3338	cts_server	/cts/run	placement
3336	powerplan_server	/power/run	floorplan (striped)
3339	route_server	/route/run	CTS
3340	save_server	/save/run	route

See docs/api_examples.http for ready‑to‑paste requests in VS Code.

---

5. Configuration – CSV → environment → TCL

Instead of duplicating parameters in TCL, every tunable knob lives in CSV:

• imp_global.csv – global flow effort, target util, etc.
• placement.csv – setPlaceMode switches, density caps …
• cts.csv – CCOpt properties.
• synthesis.csv – compile options, clock period …

At runtime the Python server:

1. Picks a row by index (g_idx, p_idx, c_idx …).
2. Exports the key/values as environment variables.
3. The stage‑specific *.tcl scripts reference them, e.g.:

setDesignMode -flowEffort $env(design_flow_effort)
floorPlan -r $env(ASPECT_RATIO) $env(target_util)

Change a cell in CSV → commit → rerun only the affected stage.

---

6. Logs, Reports & Artifacts

Path	What
logs/<stage>/	Full tool std‑out per run (timestamped)
designs/.../synthesis/<syn_ver>/results	.mapped.v, .sdc, QOR
designs/.../implementation/<impl_ver>/pnr_reports/	floorplan_summary.rpt, place_timing.rpt.gz, …
pnr_save/	floorplan.enc.dat, final.enc.dat – for restore
pnr_out/	DEF, SPEF, net‑delays, etc.

Reports that matter (WNS/TNS, DRC counts, routing congestion) are collected and returned in JSON by each service – so your CI can grep them directly.

---

7. How to Re‑use for your Design

1. Place your RTL under designs/<name>/rtl/*.v.
2. Add a minimal config.tcl under designs/<name>/ containing at least

   set TOP_NAME "<top_module>"
   set RTL_LIST  [glob $::env(BASE_DIR)/designs/<name>/rtl/*.v]
   

3. Keep using the default CSV rows or add new ones.
4. Update the REST calls / run_pipeline.sh with design="<name>".

No other path hard‑coding required.

---

8. Troubleshooting

Symptom	Likely Cause	Fix
“*.mapped.sdc not found” during floorplan	RTL top name ≠ file prefix	either rename file or use the new auto‑detect logic in backend/1_setup.tcl
Placement fails with row not legalised	target util too high	tweak placement.csv → place_global_max_density
Route DRC over 1 M	missing power stripes on M4/M5	revisit 3_powerplan.tcl stripe pitch
REST call returns error: … enc.dat not found	previous stage crashed	look at logs/<stage>/…log and fix first

---

9. Roadmap

• 🚧 add OpenROAD support for fully open‑source flow
• parameter sweep driver (sweep.py) to explore CSV grids
• Prometheus metrics endpoint on each server

---

10. License

See LICENSE. Free to use for academic & evaluation purposes; no guarantees.

---

Made with ❤️ by the MCP team – feel free to open issues / PRs!