amaranth-lang
Pysim performance dependent on module structure
I've stumbled upon a performance problem where the simulation performance depends heavily on which module a given signal is assigned in. Some minor changes triggered a 2x slowdown, half of which I was able to reclaim just by changing the place of assignment of some signals. See https://github.com/kuznia-rdzeni/coreblocks/pull/834.
The following code reproduces the problem. On my computer, running with HOW_DEEP=200 gives ~2.2s runtime; HOW_DEEP=1 gives 0.2s. Changing the HOW_DEEP variable only influences the nesting depth of modules, the number of signals and their connections remain the same.
from amaranth import *
from amaranth.lib import wiring
from amaranth.lib.wiring import In, Out
from amaranth.sim import Simulator
NUM_SIGNALS = 200
HOW_DEEP = 100
class Deep(Elaboratable):
def __init__(self, k, o):
self.k = k
if k:
self.d = Deep(k-1, o)
self.s = self.d.s
else:
self.o = o
self.s = Signal(32)
def elaborate(self, platform):
m = Module()
if self.k:
m.submodules.d = self.d
else:
m.d.comb += self.o.eq(self.s + 1)
return m
class Test(wiring.Component):
i: In(32)
o: Out(32)
def elaborate(self, platform):
m = Module()
for i in range(NUM_SIGNALS):
d = Deep(HOW_DEEP, self.o)
m.submodules[f"d{i}"] = d
m.d.comb += d.s.eq(self.i + 1)
return m
dut = Test()
async def testbench(ctx):
for x in range(100000):
ctx.set(dut.i, x)
await ctx.tick()
sim = Simulator(dut)
sim.run()
Yes, this is not particularly surprising. The solution for this problem would be to use the NIR netlist to generate code instead of the current approach. I don't expect this will happen before 2026.
