precision deviations) than the vector contents, the graph will
not be plotted in this area.
Therefore extend the limits a bit (0.5%).
Gnuplot: Use snprintf instead of sprintf. Reduce limits extension to 0.005.
Enhancement-241 fixed an amplitude normalization that divided by the ZERO-PADDED
transform size instead of the number of input samples -- in the `fft` COMMAND
(frontend/com_fft.c). The identical mistake survived in maths/cmaths/cmath4.c, the
vector-expression function reached by `let F = fft(v)`: a separate implementation of
the same computation.
Found by continuing the oracle campaign that produced E-241, and located with E-241's
own discriminator, the DC bin. One signal, 4001 samples padded to 4096, DC offset 2.0:
fft s ; the COMMAND -> mag(s)[0] = 2.000000 correct
let F = fft(s) ; the FUNCTION -> mag(F)[0] = 1.953613 = 2.0 * 4001/4096
X[0] is the sum of the samples, D*length for a DC offset D, so dividing by the padded N
reads back D*length/N. As E-241 put it: a DC value cannot depend on how many samples
were taken.
This is a contradiction, not a matter of convention. cx_fft holds TWO complete
implementations -- one for complex input, one for real -- and each has an FFTW branch
and a Green's radix-2 branch. In BOTH, the FFTW branch already used the input length
while Green's used the padded size:
real branch FFTW: scale = ((double)length)/2.0 Green: ((double)N)/2 <- wrong
complex branch FFTW: scale = (double) fpts Green: (double) N <- wrong
the correct version sitting a few lines from the wrong one inside the same function --
the same shape as `avg` disagreeing with `integ` in E-302. HAVE_LIBFFTW3 is undefined in
this build, so Green's is the live path and the defect was reachable.
oracle before after closed form
real-input, DC bin 1.953613 2.000000 2.0
real-input, ifft(fft(x)) 2.3e-02 1.1e-16 0
complex-input, bin 0 0.9766780 0.9998683 0.9998683
The round trip is an INDEPENDENT confirmation: nothing here touches ifft, so its going
from 2.3% error to machine precision is evidence from a direction the fix did not aim
at -- the pair only inverts when the forward normalization is right. cx_ifft was
audited and deliberately left alone for that reason.
Every caller of the Green's radix-2 kernel was audited, not only the one that failed:
com_fft.c's fft (x2) and spec/PSD (x2) are correct from E-241; cx_ifft is correct;
trannoise/1-f-code.c cannot pad at all, since n_pts is grown to 2^n_exp by construction;
and fft/ifft are the only transform functions in the expression table, so there is no
spec twin to miss.
Verification: examples/fftexpr_examples/verify_fftexpr.py -- 6 checks under both
solvers, all against closed form (DC bin on both paths, round trip padded and unpadded,
complex-input bin 0 against the analytic mean of an RC response). It scores 3/6 on the
pre-fix binary, so it is a real regression guard. E-241's own suite (fftnorm_examples)
and ifftreal_examples pass unchanged. Full sweep 241/241 OK.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Found while sweeping the file-parser code models (d_state/d_source/xfer/
file_source); confirmed with AddressSanitizer. Unlike the earlier code-model
finds (E-246/247/250, reads or bounded errors), these are out-of-bounds WRITES --
silent heap corruption.
1. xfer (analog/xfer/cfunc.mod, read_file): reads a transfer-function file (a
Touchstone-style `#` option line then data). It sscanf's up to 9 values per
line and a state machine stores every value (freq/real/imag triples), so a
line with more than one record stores more than 3. The allocation check
reserved only 3 (`if (i + 3 > size)`, ALLOC=1024), so a multi-record line
wrote past the buffer at the 1024-double boundary (ASan: heap-buffer-overflow
WRITE, cm_xfer). Fix: reserve the sscanf maximum of 9 (`if (i + 9 > size)`).
2. file_source (analog/file_source/cfunc.mod): stores one record per line -- a
timepoint plus `size` channel values = stepsize (size+1) doubles -- but
reserved only `size` (`vecallocated - size`), one short. At the reallocation
boundary the final channel wrote one double past the end (ASan:
heap-buffer-overflow WRITE, cm_filesource). Fix: reserve a full record
(`- stepsize`).
Both are heap OOB writes reachable from a valid-syntax netlist with a crafted
data file; the release build corrupts adjacent heap silently rather than always
crashing (UB either way). The sibling parsers were checked: d_source validates
its per-line token count against the declared width, and d_state's fixed line
buffer is fgets-bounded -- no analogous overrun.
Code-model-only change: analog.cm regenerated via cmpp and redeployed under
bin/*/codemodels/; the ngspice binary is unchanged. Verify
(examples/filefix_examples, 4 checks x2 solvers): valid transfer-function and
file_source files simulate; an xfer multi-record file and a boundary-crossing
file_source file run without overrunning. Both reproduced under ASan and shown
fixed. Full regression 208/208.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
First numerical-correctness (wrong-number, not crash) find of a correctness-audit
campaign that checked ngspice analyses against analytic and numpy oracles.
com_fft.c has two paths: FFTW3 (exact length-point transform, scale=length/2,
correct) and a Green's radix-2 FFT (no FFTW3) that zero-pads the `length` input
samples up to the next power of two N. The non-FFTW path normalized the
single-sided amplitude by the PADDED size (scale=N/2) instead of the true sample
count (length/2), so every FFT whose sample count is not a power of two reported
amplitudes too small by length/N -- up to 2x. A .tran essentially never yields
exactly 2^k points, so this bit the common case silently. The DC bin is the clean
discriminator (always exactly bin 0, no scalloping): a 2.0 V DC offset read back
as 2.0*length/N, e.g. 1.0009 for a 1025-sample record padded to 2048. spec had
the same error in its power normalization (intres = N*N instead of length*length;
the FFTW path already used length*length).
Fix: scale = length/2 and intres = length*length in the non-FFTW path, matching
FFTW. The frequency axis was already padding-aware, so only the magnitude scale
needed changing; power-of-2 records (length == N) and FFTW-linked builds are
unchanged. Validated against numpy: after the fix ngspice's fft matches
numpy.fft.rfft of the same zero-padded record (normalized by length) bin-by-bin
to ~1e-9. fft/spec amplitudes now change for non-power-of-2 records (they become
correct). Regression 199/199.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
similar to 'gnuplot' (Meisam Bahadori and Claude 4.8).
Commit contains Enhancements 94, 95, 98, 99, 182, 183.
Add a plotting flag 'boxesplot', in addition to 'combplot'
This fixes a bug in gnuplot and pyplot,
now offering both distinct combplot and boxesplot. (H. Vogt).
matplotlib -- a Python counterpart to ngspice's 'gnuplot' command. Same
syntax (pyplot <file> <plot expressions...>): the first word is the output file name only when it is NOT a plot expression, the rest are ordinary plot expressions handled by plotit().
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
the start of a file written by a Windows editor. A BOM is never
legal SPICE syntax; left in place it glues to the first token
(fatal ".subckt/.ends mismatch" class errors in included libs).
coefficient list) one element at a time, calling askInstanceQuest for
every element. Two bugs made that catastrophic on a large chip:
* IFvalue val was uninitialised, so a query that did not set
v.numValue left `n' as stack garbage -- the element loop then ran
essentially forever.
* askInstanceQuest TMALLOCs the coefficient vector (vsrcask.c) on
every call and printvals()/printvals_old() never freed it, so the
listing leaked O(n^2) allocations -- tens of GB on a deck with many
stimulus sources.
Fix: memset val before the query and clamp n (kills the runaway loop),
and free the returned vector after use via a new free_if_vec() helper.
Verified: pulse coefficients still print correctly; show on MOS/BSIM4/
OSDI/vsource is unchanged and does not crash.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Add CKTtopologyReduce() (cktsetup.c) to remove degree-1 dangling
capacitors/resistors -- e.g. dead-end opamp compensation caps -- that
otherwise cause spurious 'Timestep too small' aborts. Node degree is
counted generically over all device types; the floating node is pinned
with a unit diagonal. Disable with 'set no_topo_reduce'.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
running in parallel on a multi-core machine (12 min execution time
on a i9 9900, ngspice compiled with gcc, debug mode enabled.
Linux with Valgrind and Parallel are required.
Holger Vogt
Meisam Bahadori
dwarning
Brian Taylor
Peer Hofstra
alto555
Giles Atkinson
Brian Taylor