ngspice / Bugs / #622 Incorrect simulation results on aarch64

KiCad device models:
https://gitlab.com/kicad/code/kicad/-/tree/master/qa/data/eeschema

Failing circuit netlist:

i13162 KiCad schematic

.include "./TL072-dual.lib"
.include "./VDMOS_models.lib"
.save all
.probe alli
.ic v(Tj1)={envtemp} v(Tj2)={envtemp}
.temp {envtemp}
.param envtemp=25
.tran 200u 500m
.option RELTOL=.01 ABSTOL=1N VNTOL=10u
.control
set controlswait
if $?sharedmode
rusage
else
run
rusage
settype temperature  tj1 tj2 tcase1 tcase2
plot tj1 tj2 tcase1 tcase2
plot in out xlimit 5.2 5.3
end
.endc
Rload1 out GND 8
C7 Net-_C7-Pad1_ GND 300m
Vamb1 Net-_R11-Pad1_ GND {envtemp}
R14 Net-_R11-Pad1_ Net-_C7-Pad1_ 3
C6 Net-_C6-Pad1_ GND 300m
R13 Net-_C7-Pad1_ Tcase2 200m
R12 GND Tj2 1G
R9 GND Tj1 1G
R10 Net-_C6-Pad1_ Tcase1 200m
R11 Net-_R11-Pad1_ Net-_C6-Pad1_ 3
R4 Net-_U1A--_ GND 1k
C3 Net-_U1B-+_ GND 1u
R2 Net-_U1B-+_ GND 10k
XU1 Net-_R16-Pad2_ Net-_U1A--_ Net-_U1A-+_ GND Net-_U1B-+_ Net-_U1B--_ Net-_R17-Pad2_ VCC TL072c
R5 Net-_M2-D_ Net-_U1A--_ 19.5k
R17 Net-_M2-G_ Net-_R17-Pad2_ 100
R6 Net-_M2-S_ Net-_U1B--_ 100k
Vin1 in GND dc 0 ac 1 sin(0 0.5 100 20m)
C1 VCC GND 1u
R16 Net-_M1-G_ Net-_R16-Pad2_ 100
C2 Net-_U1A-+_ in 330n
R3 Net-_U1A-+_ Net-_U1B--_ 100k
R1 VCC Net-_U1B-+_ 390k
R7 Net-_M1-S_ Net-_M2-D_ 100m
C4 Net-_M2-D_ out 10m
R15 out GND 1k
C5 out Net-_M2-D_ 1u
M2 Net-_M2-D_ Net-_M2-G_ Net-_M2-S_ Tj2 Tcase2 IRFP240 thermal
M1 VCC Net-_M1-G_ Net-_M1-S_ Tj1 Tcase1 IRFP240 thermal
V1 VCC GND 36
R8 Net-_M2-S_ GND 800m
.end

Stefan Brüns - 2023-02-17

Error occurs with both ngspice-38 and ngspice-39.

If you would like to refer to this comment somewhere else in this project, copy and paste the following link:

Holger Vogt - 2023-02-18

The circuit is delicate about finding the operatig point. You might see this by watching the line
.option RELTOL=.01 ABSTOL=1N VNTOL=10u
where some base parameters have been changed to achieve op convergence. The reason is the inclusion of self heating in the power devices, which may lead to instabilities.

I have run the circuit on Linux, Windows, or macOS, but I do not have aarch64. When I remove the cited line, I will get the buggy result (tested on macOS).

So you may play with the parameters of the cited .option line. When I remove the line, but change the simulation command to
.tran 200u 10 uic
and thus skipping the op calculation, the result is o.k. again, even without the .option line.

It is questionable why KiCad has chosen such a circuit for their qa procedures.

Last edit: Holger Vogt 2023-02-18

If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
- Stefan Brüns - 2023-02-18
  
  So, the op determination fails on aarch64 due to problems with numerical stability? Is this a problem of the circuit (no error margin when there should be some), or of the calculation (e.g. catastrophic error propagation)?
  
  I have seen differences between x87 FP, x86_64, PPC and AArch64 math over the years, so this is not totally unexpected, though most of the times adjusting error margins was sufficient.
  
  I am not specifically interested in making this circuit work on all architectures, more on deterministic behavior across architectures.
  
  I agree the complexity of the used circuit is out of scope for a regression test in KiCad. Probably a significantly simpler circuit (single voltage source and resistor?) would have sufficed to trigger the error. Unfortunately I don't have an affected KiCad nightly at hand, so can't really provide a better (simpler) test case.
  
  If you would like to refer to this comment somewhere else in this project, copy and paste the following link:

Stefan Brüns - 2023-02-18

Btw, creating a fully emulated aarch64 machine on a x86_64 machine is fairly trivial with virt-manager. It shouldn't take more than 5 minutes to set one up (putting aside any download and installation time running in the background).
Performance on a current x86_64 host is fairly decent, similar to an RPi3, so sufficient to build ngspice and simulate moderately sized circuits.

If you would like to refer to this comment somewhere else in this project, copy and paste the following link:

This now (ngspice-42) apparently also affects x86_64.

Output from kicad-7.0.10, on aarch64, ngspice-40:

[ 1917s] + /usr/bin/ctest --output-on-failure --force-new-ctest-process -j8 --tests-regex qa_eeschema
[ 1917s] Test project /home/abuild/rpmbuild/BUILD/kicad-7.0.10/build
[ 1917s]     Start 4: qa_eeschema
[ 1923s] 1/1 Test #4: qa_eeschema ......................***Failed    6.32 sec
[ 1923s] Running 104 test cases...
[ 1923s] /home/abuild/rpmbuild/BUILD/kicad-7.0.10/qa/unittests/eeschema/./test_netlist_exporter_spice.h(242): [1;31;49merror: in "DualNMOSAmp": check abs( yVector[i] - refValue ) <= maxError has failed [10.510899464823588 > 0.011140000000000001]
[ 1923s] Failure occurred in a following context:
[ 1923s]     X vector name: time, X value: 0.029999999999999999
[ 1923s]     Y vector name: V(out), Ref value: 0.55700000000000005, Actual value: 11.067899464823588[0;39;49m
[ 1923s] /home/abuild/rpmbuild/BUILD/kicad-7.0.10/qa/unittests/eeschema/./test_netlist_exporter_spice.h(242): [1;31;49merror: in "DualNMOSAmp": check abs( yVector[i] - refValue ) <= maxError has failed [11.06536531291361 > 0.028740000000000002]
[ 1923s] Failure occurred in a following context:
[ 1923s]     X vector name: time, X value: 0.035000000000000003
[ 1923s]     Y vector name: V(out), Ref value: -1.4370000000000001, Actual value: 9.6283653129136102[0;39;49m
[ 1923s] /home/abuild/rpmbuild/BUILD/kicad-7.0.10/qa/unittests/eeschema/./test_netlist_exporter_spice.h(84): [1;31;49merror: in "DualNMOSAmp":

nspice-42, x86_64:

[ 2960s] [0;39;49m
[ 2960s] /home/abuild/rpmbuild/BUILD/kicad-7.0.10/qa/unittests/eeschema/./test_netlist_exporter_spice.h(242): [1;31;49merror: in "DualNMOSAmp": check abs( yVector[i] - refValue ) <= maxError has failed [0.35642603373163484 > 0.011140000000000001]
[ 2960s] Failure occurred in a following context:
[ 2960s]     X vector name: time, X value: 0.029999999999999999
[ 2960s]     Y vector name: V(out), Ref value: 0.55700000000000005, Actual value: 0.20057396626836521[0;39;49m
[ 2960s] /home/abuild/rpmbuild/BUILD/kicad-7.0.10/qa/unittests/eeschema/./test_netlist_exporter_spice.h(242): [1;31;49merror: in "DualNMOSAmp": check abs( yVector[i] - refValue ) <= maxError has failed [0.32961448774402657 > 0.028740000000000002]
[ 2960s] Failure occurred in a following context:
[ 2960s]     X vector name: time, X value: 0.035000000000000003
[ 2960s]     Y vector name: V(out), Ref value: -1.4370000000000001, Actual value: -1.1073855122559735[0;39;49m
[ 2960s] /home/abuild/rpmbuild/BUILD/kicad-7.0.10/qa/unittests/eeschema/./test_netlist_exporter_spice.h(84): [1;31;49merror: in "DualNMOSAmp":

ngspice-42, aarch64:

[ 3018s] /home/abuild/rpmbuild/BUILD/kicad-7.0.10/qa/unittests/eeschema/./test_netlist_exporter_spice.h(242): [1;31;49merror: in "DualNMOSAmp": check abs( yVector[i] - refValue ) <= maxError has failed [10.448457439819462 > 0.011140000000000001]
[ 3018s] Failure occurred in a following context:
[ 3018s]     X vector name: time, X value: 0.029999999999999999
[ 3018s]     Y vector name: V(out), Ref value: 0.55700000000000005, Actual value: 11.005457439819462[0;39;49m
[ 3018s] /home/abuild/rpmbuild/BUILD/kicad-7.0.10/qa/unittests/eeschema/./test_netlist_exporter_spice.h(242): [1;31;49merror: in "DualNMOSAmp": check abs( yVector[i] - refValue ) <= maxError has failed [11.135914448610931 > 0.028740000000000002]
[ 3018s] Failure occurred in a following context:
[ 3018s]     X vector name: time, X value: 0.035000000000000003
[ 3018s]     Y vector name: V(out), Ref value: -1.4370000000000001, Actual value: 9.6989144486109318[0;39;49m
[ 3018s] /home/abuild/rpmbuild/BUILD/kicad-7.0.10/qa/unittests/eeschema/./test_netlist_exporter_spice.h(84): [1;31;49merror: in "DualNMOSAmp":

Incorrect simulation results on aarch64

Group

Searches

Help

#622 Incorrect simulation results on aarch64

Discussion