Synthesis-script search, then place and route to GDSII, on open-source EDA -- with a verifier that can veto at every stage.
You write Verilog. Something has to turn it into gates. That something is a synthesiser with tens of knobs and a script language, and the choice of script changes the resulting circuit by 8-30% in area and up to 61% in depth. Nobody knows the right script in advance, so you try many.
Below the search, agentic/physflow.py carries a verified netlist all the way to
GDSII: OpenROAD for floorplan/place/CTS/route, magic for stream-out and signoff
DRC, open_pdks sky130A as the PDK. 11/11 benchmarks route, 11/11 have signoff
violations this OpenROAD cannot repair, and the README says which number is which.
This is that search, plus the part most such attempts skip:
A smaller circuit that computes the wrong function is not progress.
Every candidate is proven equivalent to the original before its number is allowed to mean anything. A candidate that fails equivalence is recorded as rejected, never as a small one.
RTL (.v)
| yosys -- elaborator only: synth -top X -noabc -> BLIF
| abc -- the search subject: <script>; print_stats; write_blif
| verifier -- chosen by shape, not by habit
v
manifest.json (hash-chained receipt) + mapped BLIF
combinational -> abc "cec <reference.blif>" counterexample if it fails
sequential -> yosys equiv_make / equiv_simple -seq 5
/ equiv_induct -seq 5 / equiv_status (0 unproven required)
13 hand-written strategies x 11 benchmarks, all generated locally so the run
needs no GitHub access. All 110 verified candidates are in bench/table.json;
every one passed its verifier and its manifest check, and the verifier's own
self-test refused the deliberately corrupted netlist in 11/11 cases.
Improvement over the naive balance; rewrite baseline:
benchmark area depth
adder32 -30.1% -61.2%
adder16 -27.8% -42.9%
arbiter16 -25.9% -18.8%
prio32 -16.3% -50.0%
mult4 -11.8% -6.2%
mult8 -8.3% -22.2%
mult12 -8.0% -34.0%
mult16 -7.5% -41.2%
adder8 -26.0% -22.2%
barrel16 0.0% 0.0%
lfsr16 0.0% 0.0%
barrel16 and lfsr16 are 0.0% on purpose. They are already shallow and
regular, the strategy set saturates immediately, and no script beats the
baseline. That is a real result, not a bug, and it is left in rather than
tuned away.
agentic/evolve_search.py mutates and selects scripts -- parameters, step
order, insertion, deletion -- under the same verifier. A candidate owns its own
strash position, because fx, mfs and renode act on the logic network
and ABC rejects them outright once the network is strashed.
Against the hand-written set on 5 benchmarks:
prio32 area 77 -> 68 (-11.7%) fx; mfs; strash; rewrite -l; resub -K 8; dc2
adder32 depth 26 -> 25
mult8 / barrel16 / lfsr16 parity, or 0.2-0.3% worse
Honest reading: the search wins on one shape. The winner appeared in generation zero, in the seed set, so it was not found by mutation. Attributing a win to "the search works" needs repeated experiments across seeds with a controlled budget, which has not been done here. Three of five benchmarks show no gain.
These are kept because they are the actual cost of the work, and because each one is the kind of failure that reports success.
1. cec with no argument is a rubber stamp. Bare cec compares nothing
and prints "equivalent" for an AND network against an OR network. The first
equivalence check passed 7/7 that way. The correct form is cec <reference>,
which returns a=1 b=0 as a counterexample. The fix is accompanied by a
mandatory self-test: corrupt one 11 1 to 10 1 in a real product and the
verifier must refuse it. If it does not, the run aborts rather than reporting.
2. A different metric is not a failure. After if, ABC reports a mapped
network as aig = 1157 lev = 5 with no and field at all, and an area regex
finds nothing. Separately, and = 473 (choice = 22) lev = 35 carries an
optional (choice = N) infix that a \s+lev pattern never matches. Ten
legal candidates were scored as failures because of this, which had produced
the false conclusion that the search loses to the hand-written set. Statistics
are now parsed field by field, and an absent metric is reported as unmeasured
rather than failed.
3. A timestamp is not non-determinism. ABC writes
# Benchmark "x" written by ABC on <date> as line 1. Two identical runs
hashed differently. Normalising that line before hashing shows the products are
byte-identical across independent runs.
Also worth writing down, from the toolchain: 2>&1 | tail -4 swallows the
statistics line, because WSL injects localhost proxy warnings on stderr;
concurrent wsl.exe calls intermittently return empty stdout, which in a
threaded harness is indistinguishable from a candidate producing nothing.
Requires WSL with yosys and berkeley-abc (apt install yosys berkeley-abc), Python 3.10+.
python agentic/run_bench.py # 11 benchmarks through the hand-written set
python evolve_run.py # search vs the hand-written set
python determinism_check.py # two independent runs, compare hashes
python selftest_spine.py # manifest integrity: tamper and truncate
python agentic/physflow.py # 11 benchmarks netlist -> GDSII + signoff DRC
The physical flow additionally needs, in WSL:
micromamba install -p ~/eda-env -c litex-hub -c conda-forge open_pdks.sky130a magic
apt install yosys openroad # openroad also ships on litex-hub
and the PDK is 2.7 GB unpacked under ~/eda-env/share/pdk/sky130A (the 1.26 GB
tarball lands in micromamba's own cache; if share/pdk is empty afterwards the
package downloaded and unpacked but was not linked into the environment, and a
symlink is the fix -- not another download).
agentic/spine.py typed stages, hash-chained manifest, verify_manifest
agentic/synth_search.py elaboration, search, strategy set, parse_stats, gate self-test
agentic/evolve_search.py mutation and selection over scripts
agentic/bench_gen.py 11 benchmark generators (no external corpus needed)
agentic/run_bench.py the 11-benchmark sweep
agentic/llm_agent.py the model-driven searcher (key read from outside the repo)
agentic/physflow.py netlist -> GDSII: OpenROAD, magic, sky130 signoff DRC
bench/table.json the synthesis result
evolve/results.json the search result, every candidate row included
phys/results.json the physical result: per-design hashes, both violation counts
The synthesis search above stops at a verified netlist. agentic/physflow.py
carries that netlist the rest of the way to silicon layout on the same open
toolchain, and it does it as a driver rather than a recipe:
verified netlist
| OpenROAD floorplan -> tracks -> pin place -> global place
| -> detailed place -> CTS -> filler -> route -> DEF
| magic stream out with the PDK's own GDSII cell library -> GDS
| magic signoff DRC, the PDK's rule deck, by name -> violations
v
results.json (per-design artifact hashes, both violation counts)
11 benchmarks, every stage green on 11/11, 430 s for the sweep:
design sky130 cells area um2 util router signoff GDS
adder8 3 229 34% 0 27 50 KB
adder16 18 512 31% 0 63 184 KB
adder32 30 1085 32% 0 177 363 KB
arbiter16 14 214 33% 0 37 122 KB
barrel16 20 596 31% 0 74 205 KB
lfsr16 6 502 33% 0 39 107 KB (sequential, CTS ran)
mult4 30 382 33% 0 54 238 KB
mult8 37 1912 30% 0 293 476 KB
mult12 50 4792 30% 0 640 938 KB
mult16 57 8647 30% 0 1221 1509 KB
prio32 20 425 32% 0 44 195 KB
Two violation counts, because they are two different claims. The router
reports 0 on all 11 designs: TritonRoute is satisfied with what it built. The
signoff deck disagrees on all 11: 27-1221 boxes. The rules are minimum area,
met1.6 (0.083 um^2), met2.6 (0.0676 um^2), met3.6 (0.24 um^2), and once a met1
spacing. They are in the technology LEF -- AREA 0.083 sits right there under
LAYER met1 -- but this OpenROAD does not repair them: there is no min-area
repair pass, no add_met_fill, and detailed_route -help is not even
supported. A flow that printed only the router's number would call all 11
layouts clean.
These three are the same kind of bug as the harness lies above: they report success. The order of two commands is the whole difference between them:
- The technology LEF is
techlef/sky130_fd_sc_hd__nom.tlef. The corner suffix is part of the name. The LEF underlef/carries MACROs and zeroLAYERstatements, so reading it alone yields 0 technology layers and a design that reads, links, floorplans -- and then cannot route. make_tracksmust come afterread_liberty(it takes microns; without a liberty it returns "command units uninitialized" and creates nothing) and afterinitialize_floorplan(tracks are laid over the current die). Wrong order creates zero tracks, and the failure surfaces two stages later inplace_pinsas "no horizontal tracks for met3".
And one that looks like a layout bug and is not: arbiter16 has a 1'b1
constant, yosys turns it into a net, OpenROAD's Verilog reader labels that net
GROUND by name, and TritonRoute refuses it -- "[DRT-0305] Net zero_ of signal
type GROUND is not routable". Nothing is wrong with the layout; the net's type
is. The driver demotes GROUND/POWER back to SIGNAL on every net that is not one
of the PDK's own power nets, before the floorplan. Designs without a constant
never hit it.
No timing closure, no analogue layout, no tapeout. The GDS files are real, stream-out, DRC-checked layout geometry -- and every one of them has signoff violations that this OpenROAD version cannot repair, so they are evidence the flow works, not evidence of manufacturability. A newer OpenROAD (litex-hub is at 2.0_9990, this is 2.0_3175) may close the min-area gap; that has not been measured here.
The analogue and in-memory-compute directions still need what they needed before: a device model, an analogue benchmark corpus. The PDK part of that gap is now closed -- open_pdks sky130A is installed and in use.
Two more flows, same shape as the ones above -- propose, run, veto, feed back -- and they close the two loops that were open.
agentic/packflow.py starts from the die physflow.py actually produced and
answers the question the die cannot answer for itself. The pad ring does not fit:
adder8 is 229 um^2 of logic, a 15 um square, and 23 pads at 60 um pitch need a
390 um edge. The die has to grow, and the flow computes by how much instead of
discovering it at assembly.
design pins die grew to substrate part chosen layers Tj
adder8 23 390 um 4560 um Texas_DSBGA-12_1.36x1.86mm 2 26.1 C
adder16 43 690 um 5700 um BGA-81_4.496x4.377mm_P0.4mm 1 25.8 C
mult4 20 330 um 4460 um Texas_S-PDSO-N10_EP1.2x2mm 2 26.1 C
mult8 40 630 um 6100 um Texas_VQFN-RHL-20_ThermalVias 3 25.8 C
mult16 60 930 um 5700 um BGA-81_4.496x4.377mm_P0.4mm 2 25.8 C
prio32 40 630 um 6100 um Texas_VQFN-RHL-20_ThermalVias 3 25.8 C
The parts are real: 11,994 footprints from the KiCad library, indexed and read,
their pad coordinates taken as given. Substrates are 2-4 layer BT or alumina --
that choice is a search, and each iteration changes exactly one thing the
previous veto named, with the reason recorded in pack/results.json.
pack/mult8/mult8_substrate_top.gbr is exported RS-274X, one copper layer, and
the writer refuses to emit a coordinate outside the outline it declares.
Package selection is the weakest step and says so: the shortlist is ranked by outline area and then it walks to a coarser-pin part, which is not the same as an area-and-cost trade study, and no assembly yield or coplanarity is modelled. The 0.4 mm pitch of a BGA is the binding constraint on how many pins can escape one substrate layer, and that number is measured (14.0-69.3 um between an escape and a foreign pad) rather than vetoed, because a constant-width straight line cannot satisfy a constant clearance rule across a pad field whose own separation varies 20x. That took three wrong versions to admit.
agentic/lessons.py is the other closed loop, and it is the one this repo was
missing: what was learned now lives in records instead of in this file. A lesson
has four required prose fields -- problem, decision, applicability,
prevention_action -- and evidence with the measured numbers, bound to the hash of
the artefact it came from. lessons.db holds 14, every one of them a failure this
repo actually hit: the BLIF net defined in the last token not the second, the
signoff deck's 1221 against the router's 0, make_tracks before read_liberty,
three separate factor-of-1000 unit errors, Stoney curvature refusing ceramic
packages because tf/ts was 0.37 and the model needs 0.1.
The discipline is borrowed whole from Milkdomeda/ai-mechanical-3dcad-design-agent
(Apache-2.0): knowledge is an accelerator, never a precondition. search returns
[] on a missing or empty store and never raises; every run records which lessons
it consulted, so "ran without experience" and "ran against experience" are
different facts afterwards. Lessons are evaluated after the artefact is confirmed
and nothing publishes itself, and the scope split is theirs -- a lesson that is
true only here is project_only and is excluded from publication, so it cannot be
read next month as general.
The signature of the mechanical-agent project, and it is the right one: a passing report proves the checks that ran against one exact revision. It is not FEA, not a manufacturing release, not a safety certification, not a legal standards certification. Final engineering responsibility stays with the user or an authorised engineer.
For this repo, concretely:
- The GDS files are stream-out geometry. Every one carries signoff DRC violations this OpenROAD cannot repair (27-1221 boxes, min-area). Not manufacturable, and not claimed to be.
- The thermal number is a lumped theta-JA = 7.7/sqrt(area_cm2). No board copper, no thermal vias, no neighbours, no FEM.
- The warpage number is Stoney curvature, valid only for tf/ts < 0.1, and at 0.37 it says NOT EVALUATED rather than guessing.
- The substrate "DRC" is a clearance and containment check on geometry this file wrote. It is not a fab's deck, and there is no via, impedance, EM, solder-mask or drill check at all.
- The ATPG coverage numbers come from a fault simulator cross-checked against
iverilogand integer arithmetic on 1470 random probes, not against a commercial tool and not against silicon.
agentic/testflow.py is the third flow and it is the simplest to state: take the
netlist the synthesis search produced, ask how much of it a stuck-at test could
observe, and be exact about which of the two questions a number answers.
design gates faults coverage ceiling provably redundant enumerated vectors closed
adder8 ? 316 0.9324 - - - 400 no
adder16 ? 302 0.9669 - - - 400 no
mult4 ? 316 0.4747 0.4747 166 256 400 yes
mult8 ? 1422 0.5120 - - - 400 no
mult12 ? 3307 0.5205 - - - 400 no
prio32 ? 424 0.5307 - - - 499 no
arbiter16 ? 296 0.7736 - - - 422 no
barrel16 ? 221 1.0000 - - - 400 yes
lfsr16 ? 74 0.0135 0.0135 73 4 403 yes
Two thirds of the faults in a 4-bit multiplier are redundant: no input vector
distinguishes them from the fault-free circuit. That is not a coverage shortfall,
it is the ceiling, and it is a proof rather than a gap because 2**8 vectors is the
whole input space and all 256 were simulated. The mult4 row is the flow working.
lfsr16 at 0.0135 is the same statement about a shift register, under the
assumption the driver states out loud: latches are treated as pseudo-inputs and
pseudo-outputs, full-scan style. Under that assumption 73 of 74 faults are
redundant. Without it the number means something else entirely, which is why the
assumption is written down instead of left in the tool.
Six designs are not closed and the report says why in each case: the input
space was too large to enumerate, so not_detected is the honest word and
undetectable is not available. repair is not attempted and the wall-clock
budget of the guided stage is a first-class parameter, because the operator's
point is that cost is the difficulty, not the algorithm.
The numbers come from a fault simulator cross-checked against iverilog and
against integer arithmetic on 1470 random probes -- not against a commercial ATPG
tool, and not against silicon.
MIT.
The same claim -- a number is meaningless until it is shown to survive its own verification -- is made and measured in other domains:
- autoforge -- a tool's fitness, until an oracle outside the tool agrees
- debt-verify -- a debt clause decision, until it survives the published revision record
- tool-market -- a tool's liveness, until the hash chain says which revision is live
- agent-safety-bench -- a model's safety compliance, measured rather than assumed