Make timeout limit for fifo sim configurable

[LeahLS]

Prerequisites

None

Details

New behavior

It might be useful to make the timeout limit for the set-fifo-depth simulation configurable, for example using an env var like for the rtlsim liveness_treshold.

Motivation

While building the finn-examples for the AUP-ZU3 board, the issue of an fifo sim timeout occurred. Currently, this issue can only be resolved by manually modifying the code.

Parts of FINN affected

Transformation, FPGA Dataflow, Set FIFO depths

[SwordTechCorp]

Agree. Got similar issue during the "setting FIFO depth" step. A simple workaround would be setting "max_iter" to None, in case there is no deadlock? Or we can come up with a looser upper bound.

[merkelmarrow]

Bumping this, had a similar issue. Was able to workaround by monkeypatching xsi_fifosim. In general the bound seems too tight.

Hacky fix without modifying the source:

import finn.transformation.fpgadataflow.set_fifo_depths as fifodepth
_xsi_sim_orig = fifodepth.xsi_fifosim
def _xsi_fifosim_patched(model, n_inferences, max_iters, throttle_cycles=0):
    max_iters = 10000000
    return _xsi_sim_orig(model, n_inferences, max_iters, throttle_cycles)
fifodepth.xsi_fifosim = _xsi_fifosim_patched

[mumallaeng]

@auphelia would it be okay for me to work on this?

Looking at the code, xsi_fifosim already accepts max_iters=None as the default in its signature, but the hardcoded value at the call site acts as an empirical safety cap. Without it, a misconfigured FIFO could cause the simulation to hang indefinitely, especially in CI.

The AUP-ZU3 timeout may be happening because a smaller device requires heavier folding and deeper pipelines, causing the network to take more clock cycles to reach steady state than the fixed cap allows.

Rather than going fully unbounded, which risks hanging indefinitely, or simply increasing the constant, which would only postpone the issue, a middle ground similar to the existing get_watchdog_timeout_cycles pattern might work well: derive a rough default from model properties (e.g., node count × a per-node cycle estimate), then allow FINN_FIFOSIM_MAX_ITERS (or a similar environment variable) to override it.

That way, the safety cap scales with the network instead of using a one-size-fits-all constant, while still protecting against infinite hangs.