The ClockGen module has an off-by-one error when deriving a clock using .derive or .calculate. The output clock period is one input clock cycle shorter than necessary. This happens only when the ratio between the input and output frequencies is >= 3.
Amaranth playground link.
Here, the input clock is 4Hz and the output clock is 1Hz. Since the input is evenly divisible by the output, we should get an output clock that is exactly 4 input clock cycles long. And, since the ratio is divisible by 2, we should get a 2 clock high pulse and a 2 clock low pulse on the output. Instead, we get 3 cycles - 1 high, 2 low.
The issue is that calculate computes cyc = round(input_hz // output_hz) - 1, and then the implementation uses this value for the counter if it's >= 2. If cyc == 0 or cyc == 1, then we get into one of the two special cases (same frequency or half frequency), and there everything works fine.
I think the fix should be something like:
---
+++
@@ -41,33 +41,33 @@
self.stb_r = Signal()
self.stb_f = Signal()
def elaborate(self, platform):
m = Module()
- if self.cyc == 0:
+ if self.cyc == 1:
# Special case: output frequency equal to input frequency.
# Implementation: wire.
m.d.comb += [
self.clk.eq(ClockSignal()),
self.stb_r.eq(1),
self.stb_f.eq(1),
]
- if self.cyc == 1:
+ if self.cyc == 2:
# Special case: output frequency half of input frequency.
# Implementation: flip-flop.
m.d.sync += [
self.clk.eq(~self.clk),
]
m.d.comb += [
self.stb_r.eq(~self.clk),
self.stb_f.eq(self.clk),
]
- if self.cyc >= 2:
+ if self.cyc >= 3:
# General case.
# Implementation: counter.
counter = Signal(range(self.cyc))
clk_r = Signal()
m.d.sync += counter.eq(counter - 1)
@@ -110,14 +110,14 @@
.format(output_hz / 1000, input_hz / 1000))
if min_cyc is not None and output_hz * min_cyc > input_hz:
raise ValueError("output frequency {:.3f} kHz requires a period smaller than {:d} "
"cycles at input frequency {:.3f} kHz"
.format(output_hz / 1000, min_cyc, input_hz / 1000))
- cyc = round(input_hz // output_hz) - 1
- actual_output_hz = input_hz / (cyc + 1)
+ cyc = round(input_hz // output_hz)
+ actual_output_hz = input_hz / cyc
deviation_ppm = round(1000000 * (actual_output_hz - output_hz) // output_hz)
if max_deviation_ppm is not None and deviation_ppm > max_deviation_ppm:
raise ValueError("output frequency {:.3f} kHz deviates from requested frequency "
"{:.3f} kHz by {:d} ppm, which is higher than {:d} ppm"
.format(actual_output_hz / 1000, output_hz / 1000,
@@ -139,15 +139,15 @@
if logger is not None:
if clock_name is None:
clock = "clock"
else:
clock = f"clock {clock_name}"
- if cyc in (0, 1):
+ if cyc in (1, 2):
duty = 50
else:
duty = (cyc // 2) / cyc * 100
logger.debug("%s in=%.3f req=%.3f out=%.3f [kHz] error=%d [ppm] duty=%.1f%%",
clock, input_hz / 1000, output_hz / 1000, actual_output_hz / 1000,
deviation_ppm, duty)
return cyc
Simulation with the fix.
On a related note, the special case where the output frequency is half the input frequency has a strobe behaviour different from the normal case. In the special case stb_r and stb_f rise one cycle before their respective clock edges. In the normal case they rise on the same cycle. Not sure which behaviour is the correct one, but I think they should at least be consistent.
In terms of actual impact the whole thing may not be a big deal. Assuming most applets derive clocks that are much slower than the system clock, the off-by-one deviation should be negligible.
The
ClockGenmodule has an off-by-one error when deriving a clock using.deriveor.calculate. The output clock period is one input clock cycle shorter than necessary. This happens only when the ratio between the input and output frequencies is >= 3.Amaranth playground link.
Here, the input clock is 4Hz and the output clock is 1Hz. Since the input is evenly divisible by the output, we should get an output clock that is exactly 4 input clock cycles long. And, since the ratio is divisible by 2, we should get a 2 clock high pulse and a 2 clock low pulse on the output. Instead, we get 3 cycles - 1 high, 2 low.
The issue is that
calculatecomputescyc = round(input_hz // output_hz) - 1, and then the implementation uses this value for the counter if it's >= 2. Ifcyc == 0orcyc == 1, then we get into one of the two special cases (same frequency or half frequency), and there everything works fine.I think the fix should be something like:
Simulation with the fix.
On a related note, the special case where the output frequency is half the input frequency has a strobe behaviour different from the normal case. In the special case
stb_randstb_frise one cycle before their respective clock edges. In the normal case they rise on the same cycle. Not sure which behaviour is the correct one, but I think they should at least be consistent.In terms of actual impact the whole thing may not be a big deal. Assuming most applets derive clocks that are much slower than the system clock, the off-by-one deviation should be negligible.