Repository navigation
AGaMEMnon v0.5.0
AGaMEMnon turns ordinary Verilog into an AG32 L48 fabric image with open tools only.
This release finds and removes the routing-table errors that made some default builds
produce silently wrong images, bounds every place-and-route attempt, fences one BRAM
mode that failed on silicon, and adds opt-in router-chosen LUT input pins. It is a
bounded L48 release, not vendor parity. See release validation
for the evidence behind each claim.
Tested on hardware
Results are from the release candidate on one AG32VF303CCT6 (L48) board, loaded into
SRAM, with a known-good reference image between designs.
| Check | Result |
|---|---|
| Maintained corpus and holdouts, 18 fresh default builds | 18/18 pass; 19/19 references |
| Stronger fast and slow RAM holdouts, 2 loads x 32 resets | 64/64 and 64/64 (0/32 before this release) |
| Stronger FIFO and LFSR holdouts, 2 loads x 32 resets | 61/64 and 61/64; see Known limitations |
| Designs that built wrong images during release preparation, rebuilt | pass |
| Router-chosen LUT inputs on, 38 designs | 38/38 pass (38/38 with it off) |
Software: the full Python suite and the native-toolchain complement pass with no
failures; installed wheels pass on Linux, Windows and macOS; the SDK archives build.
What changed
- Wrong routing-table rows are refused. Five default images that read back wrong
on silicon were each traced on the board to one routing connection: rerouting only
the net that used it made the image pass, and none was a timing effect (the first
failed identically at 4, 10 and 40 MHz). All five take their selector codeword from
a table that generalises codewords seen at one row or column to others, and the
vendor's own bitgen has none of them. A vendor-bitgen check of every unconfirmed
entry of that table found nine more, and six further connections select a different
source than modelled. The flow refuses these 20 and retries around them; the device
graph is unchanged. Evidence. - Never-proven route shapes are refused. Routes through a source/destination pair
and tile offset that no board-passing design has used are refused and retried (2,416
pips, about 1% of fabric routing), as are 207 pips isolated by pass/fail board
comparisons. Evidence. - Every place-and-route attempt is bounded at 300 s. nextpnr's router2 could
oscillate indefinitely; a timed-out attempt moves on to the next seed. - x1 single-port BRAM writes need
PORTA_OUTREG=1. The unregistered form failed on
silicon while the registered form and the vendor image passed. - A completed build is kept when an optional alternative reset mapping fails its
bounded search. - Opt-in router-chosen LUT inputs (
AGAMEMNON_LUT_PIN_SWAP=1). The router may land
a signal on any eligible input of a LUT and the truth table is permuted to match, as
the vendor router does, with the vendor's freeze rules for feedback, register-bypass
and carry inputs. Off, output is byte-identical to before. - Earlier 0.5 changes (memory lowering, asynchronous clear, route retries, the PLL ratio
model, BRAM source profiles, SDK packaging) are listed in the changelog.
Known limitations
- Synchronize external resets. A design that feeds an asynchronous input pin
straight into a synchronous reset across many registers can start some registers a
clock later than others. The stronger FIFO and LFSR holdouts do this and miss about
3 in 64 reset releases; the same RTL with a two-flop reset synchronizer passed 96/96
and 128/128. The vendor's images of the unsynchronized RTL pass, most likely because
its reset routing has less skew; skew-aware reset routing is future work. - Some dense designs refuse to build, for example with a forced
--seed, when every
legal route needs a refused connection. The build stops with a message instead of
emitting an image. - Router-chosen LUT inputs are opt-in. Their board A/B found no defect, but they
change most images, including every byte-exact qualified fixture, so enabling them by
default needs its own requalification. - Build success is not silicon evidence. Test a new design on hardware.
- BRAM: x18 and x2 dual-port writes, x4 dual-port writes and read-only modes at X13Y4
are the supported set. Timing is conservative and does not model clock skew. Other
packages, the MCU AHB master and DMA, analog, and most hard-peripheral modes remain
outside the tested scope. See supported scope and the
roadmap.
Installation artifacts
The release workflow builds one wheel and embeds that exact wheel in the Windows
and Linux SDK archives. Each artifact has a SHA-256 sidecar:
agamemnon_ag32-0.5.0-py3-none-any.whl
agamemnon-sdk-linux-x64.tar.gz
agamemnon-sdk-windows-x64.zip
Verify the sidecars, activate the matching SDK environment, and check that
agamemnon --version reports agamemnon 0.5.0. A wheel alone does not install
all native FPGA/MCU tools. Follow installation and
agamemnon doctor --no-hardware for the required capabilities. The OpenOCD
transport remains a separate installation. Downloads are on the
release page.