SAIFIINDUSTRIES's picture
Add Repo: KohakuBlueleaf_KohakuTPU
dc3de35 verified
|
Raw
History Blame
2.63 kB

sim/verilator — a simulated card

The goal is one thing:

An external AXI master → station bus → system node → mesh, running a real program, driven from the software stack. A card you can test a driver against without touching silicon.

Not a faster bench runner. The benches under tests/ are a way to validate the model; they are not the product. The product is a C++ object the Python driver talks to exactly as it talks to JTAG or XDMA.

There is no wall

Every "Verilator can't do X" in this tree turned out to be a shim or a harness question. Recorded so nobody re-derives them:

Claimed blocker Actual status
XPM macros Four cells, four shims, cross-checked against the real Xilinx cells. shims.md
deassign in Vivado's xpm_memory.sv Real, and confined to xpm_memory_base. Routed around, not fought.
Double-pumped matmul needs BUFGCE_DIV Never a blocker. The repo already guards it with `ifdef SYNTHESIS and simulates a behavioural divider. cluster_node_pump passes: 7524 checks.
MX_MODEL=0 / DSP48E2 Not needed for a card. MX_MODEL=1 is the default everywhere.
X-propagation loss Irrelevant to a card model — the driver supplies stimulus, not X. divergences.md
disable, $random in benches Testbench-only. A C++ harness has no testbench.

The documents

docs/card-backend.md The point of all this. Verilator as a real backend for the software stack, and how it plugs into driver/kohakuaccel with no driver changes.
docs/status.md Work log. What runs today, what is open, what was measured. Updated as work lands.
docs/setup.md Install, and scripts/py/vlt.py.
docs/shims.md The primitive replacements, why each exists, and the cross-check method that validates them.
docs/divergences.md Where Verilator and xsim disagree, and which disagreements matter.

Layout

sim/verilator/
  README.md            this file
  docs/                the five documents above
  shims/               simulation-only primitive replacements
  examples/            cross-check benches: same bench, both simulators
scripts/py/vlt.py      runs a bench under Verilator, reusing xsim.py's BENCHES

Nothing here modifies src/, tests/, or scripts/py/xsim.py. The xsim flow is untouched and stays the flow of record.

One command

python scripts/py/vlt.py cluster_node
python scripts/py/vlt.py ctrlpe_mesh --lint-only