Evidence · FSL

topup and applytopup

topup estimates the off-resonance field that displaces echo-planar images along their phase encoding, from images acquired with opposite phase encodings, and applytopup corrects images with it; fMRIPrep runs topup for every reversed-phase-encoding (PEPOLAR) field map. The replica (larmorx-nc) gives FSL 6.0.7.23’s files, messages and exit status byte for byte in all 107 of its compared cases, out of 177, fMRIPrep’s own calls among them. The clean room (larmorx) is within its record’s thresholds in all 106 of its compared cases, out of 187, and 14 more are expected divergences, each with its reason. On fMRIPrep’s two calls it is as close to FSL as FSL at another thread count, but for the field’s correlation on ds005040 (0.999958, against FSL’s own 0.999966), and it runs 10 to 12 times as fast as FSL on one thread, about 40 times on 14.

Ports
  • replica larmorx-nc FSL Licence: non-commercial use only: translated from FSL 6.0.7.23's source (fsl-topup 2203.5, with fsl-basisfield, fsl-warpfns and fsl-miscmaths' optimisers and sparse matrices); it runs as a separate program
  • clean room larmorx Apache-2.0: crate larmorx-fsl, written from specs/topup.md, the topup papers, FSL's user guides and runs of FSL 6.0.7.23, without FSL's source (2026-11-03; brought up to the specification's answers on 2026-11-05)
Status
replica experimental; clean room experimental
Records
the replica's: standard comparison (FSL 6.0.7.23) · the clean room's: standard comparison (FSL 6.0.7.23) · benchmark fsl-topup
Reference
topup, applytopup · API page fsl-topup

What is compared

An echo-planar image is displaced along its phase-encoding axis wherever the magnetic field departs from its nominal value: by the off-resonance (in Hz) times the read-out time, in voxels. Its intensity is compressed or stretched by the displacement’s Jacobian. Two images with opposite phase encodings are displaced in opposite directions, so the field that makes them agree can be estimated.

topup models the field as a cubic (or quadratic) B-spline. It resamples each image through the displacement the field gives it (and, with --estmov, through its own rigid-body movement), multiplies by the Jacobian, and minimises the squared differences between the corrected images plus a regularisation of the field’s bending energy (or membrane energy), weighted by --lambda. It works level by level, each level with its knot spacing (--warpres), subsampling, smoothing and number of iterations, by Levenberg-Marquardt or scaled conjugate gradients.

It writes the field’s coefficients and the movements (--out). On request it also writes the corrected images (--iout), the field in Hz (--fout), each image’s displacement field (--dfout), rigid-body matrix (--rbmout) and Jacobian (--jacout), and a log. applytopup corrects other images with topup’s result: by Jacobian modulation of each image alone (--method=jac), by the least-squares restoration of pairs with opposite phase encodings into one image (lsr, the default), or by one of three undocumented variational Bayes restorations (vb2D, vb3D, vb4D).

fMRIPrep calls topup for every reversed-phase-encoding field map, in sdcflows’ PEPOLAR workflow (fmap_preproc_wf/<estimator>/topup). It gives topup the average of each phase encoding’s images, merged into one series, with sdcflows’ configuration b02b0.cnf (seven levels from 14 mm to 4 mm knots, without movements): topup --config=b02b0.cnf --datain=<encoding file> --imain=<merged pair> --out --iout --fout --jacout=jac --logout --rbmout=xfm --dfout=warpfield. sdcflows turns the coefficients into the field map fMRIPrep uses. fMRIPrep does not call applytopup: its own resampler applies the field.

The two ports run beside FSL 6.0.7.23’s topup and applytopup on the same files with the same arguments. fMRIPrep’s own outputs are the files it wrote in its run, with topup in fMRIPrep 25.2.6’s image.

bash
larmorx fsl topup --imain=b0_pair --datain=acqparams.txt --config=b02b0.cnf --out=tu --fout=field --iout=corrected
larmorx fsl applytopup --imain=dwi_up,dwi_dn --inindex=1,2 --datain=acqparams.txt --topup=tu --out=hifi

The first estimates the field from a pair and writes it with the corrected pair. The second corrects two images of opposite phase encodings with that result.

python
import larmorx as lx

acq = [[0, 1, 0, 0.0138], [0, -1, 0, 0.0138]]   # phase-encoding vector and read-out time (s)
r = lx.fsl.topup("b0_pair.nii.gz", acq, config="b02b0.cnf", fsldir="/opt/fsl", corrected=True)
field_hz = r.field                     # lx.Image on the input's grid
hifi = lx.fsl.applytopup(["dwi_up.nii.gz", "dwi_dn.nii.gz"], acq, [1, 2], r).corrected

The calls above leave implementation at "auto". The API page describes every option. How to read what follows: how to read the evidence.

Its two ports

FSL comes under the FSL Licence, which allows non-commercial use only. So LARMORX has two ports of its FSL programs. The replica is translated from FSL’s source and lives in its own package, larmorx-nc (pip install larmorx-nc). The clean room is in larmorx itself. It is written without FSL’s code, from topup’s behavioural specification (specs/topup.md), the papers (Andersson, Skare and Ashburner 2003; Smith et al. 2004), FSL’s user guides and runs of FSL’s programs as a black box.

implementation="auto" | "replica" | "clean-room" chooses between them ("original" in 0.1.0rc1). On the command line, larmorx --implementation clean-room fsl topup ... does the same from the next release; in 0.1.0rc1 it is the environment variable LARMORX_IMPLEMENTATION. auto, and larmorx fsl topup, take the replica when larmorx-nc is installed, else the clean room.

The replica follows FSL 6.0.7.23. It runs none of FSL’s programs and needs no FSL installation; FSL’s configurations by name need its etc/flirtsch folder, or fsldir. The FSL libraries topup needs are translated once into larmorx-nc’s shared core, beside those of mcflirt, BET, FLIRT and FAST: the B-spline fields with their bending- and membrane-energy regularisers, the sparse matrices, the non-linear optimisers, the warps, the spline interpolation with its derivatives, and LAPACK’s and OpenBLAS’s arithmetic as FSL’s build calls it.

The clean room has every option, message, exit status and failure case of both programs: the configuration files, both minimisers per level, both regularisers, both spline orders, the precisions and interpolations, the movements, every output, --debug, --trace and --profile, and applytopup’s jac, lsr, vb2D, vb3D and vb4D. It computes sin, cos and exp correctly rounded on every platform. It is built for speed, and its result is the same bits at any thread count (speed):

  • the B-spline field’s products run along whole lines and planes;
  • the Levenberg-Marquardt steps are solved by preconditioned conjugate gradients on the Hessian stored by its diagonals;
  • the arrays are recycled, and every parallel loop runs in fixed blocks summed in a fixed order.

Where the specification leaves a convention open, the clean room follows FSL’s as black-box runs measured it:

  • an oblique phase encoding wraps along no axis;
  • the scaled conjugate gradients’ second gradient keeps the current point’s regulariser weight;
  • --debug’s images are written in FSL’s order.

The C library’s last digits (since 2026-11-04). topup smooths the images at each level of its schedule (--fwhm) with a Gaussian whose weights FSL computes with glibc’s exp. glibc is the C library of Linux, and its results are now and then one unit in the last place away from the exact value rounded once. LARMORX computes exp correctly rounded, the same on every platform. Since 2026-11-04 the clean room’s reproduce computes it exactly as glibc 2.35 does, and its default, fix, correctly rounded. That is GLIBC-07 in the register, and topup names it nowhere: whether a fit moved with such last bits cannot be told without running it twice (a decision of the project of 2026-11-04, recorded with GLIBC-07). On topup it changed no value. The weights are kept in single precision, no number moved on any case of the record, and the two modes still agree on fMRIPrep’s calls.

The clean-room catch-up (2026-11-05). The specification’s answers of 2026-11-03 (specs/topup.md §13) are now in the clean room, and every example below was run again with it.

  • Its reproduce takes FSL’s forward differences of single-precision rotation matrices for the movements’ derivatives (FSL-201): the first gradients of FSL’s --debug files to 2 × 10−5 and 10−6. Its default keeps the exact derivatives and names FSL-201 on every run that estimates the movements.
  • Its reproduce lays FSL-145’s coefficient count exactly, and stops with FSL’s error where a level refits a single coefficient. It follows FSL-133’s index on lines whose length does not divide 232, and lays FSL-134’s field of another shape as FSL does.
  • In fix and in reproduce, the regulariser samples the quadratic B-splines’ bending energy as FSL does, the spline prefilter starts from FSL’s truncated sums on short axes, and applytopup’s least-squares restoration builds its matrices row by row as FSL walks them.
  • The record’s cases that FSL-133, FSL-134 and FSL-145 kept apart now pass.
  • Random testing after the catch-up found one more defect of FSL, registered and corrected by default: FSL-242, a level of scaled conjugate gradients whose last step is rejected ends at that worse point. It also found that FSL counts every such iteration toward --miter, which the clean room now does too (below).

The verdicts at a glance

The replica is a translation of FSL’s source, so the question for it is whether it gives FSL’s bytes: every file topup and applytopup write (each image’s header and data, the coefficients, the movements, the matrices, the log), their standard output and error, and their exit status.

Agreement, replica against the original: FSL topup and applytopup 6.0.7.23 (standard comparison, 177 cases of the record), by category and verdict
fMRIPrep's calls2
--debug, --trace and --profile5
options37
configuration files16
errors and messages39
thread counts3
applytopup75

Select a segment to list its cases below.

The replica’s record runs 177 cases against FSL 6.0.7.23 (the record, made again on 2026-11-06). 107 give FSL’s files byte for byte. In 67 FSL refuses the input or stops, and the replica does the same, with the same message and exit status (rejected by both). 3 are refused by the replica in fix and in reproduce (expected divergence): a negative --warpres, where FSL needs more memory than any machine holds (FSL-124), and fields on another grid, which FSL reads past its lines (FSL-134).

Some of the cases came later:

  • three with the specification’s answers to the clean room’s questions (2026-11-03): quadratic splines on an axis shorter than two knot spacings, and on one shorter than one spacing at a level that refits it, where FSL stops with an index error and the replica’s default finishes (FSL-145); and four volumes encoded along y and x (FSL-146);
  • one with an answer written for PRELUDE’s specification (2026-11-06): --method=jac with a coefficient field whose voxel sizes differ from the images’ (at/jac-coef-aniso; with a field image, at/jac-field-aniso). FSL scales the Jacobian by the voxel sizes’ ratio along each encoded axis, and the replica’s default takes the Jacobian of the displacement applied (FSL-273).

The cases cover:

  • fMRIPrep’s call on its two fMRIPrep runs, whose files are fMRIPrep’s own too;
  • every option of topup: double and single precision, spline and linear interpolation, both regularisers, quadratic and cubic splines, the scaling, the regridding, the minimisers per level, the movements, FSL’s shipped configurations and configurations of their own, --acqp, x, y and oblique phase encodings, three and four volumes, --debug, --trace, --profile;
  • its thread counts;
  • every option and method of applytopup, on lines of 30, 32 and 96 voxels and in every data type;
  • every message, exit status and failure case of both command lines.

Each case also runs the replica’s default, fix, where LARMORX corrects FSL’s deterministic defects and names them (below).

The clean room is written without FSL’s source, so its record holds it to the thresholds of specs/topup.md §12 instead:

  • FSL’s exit status, files and messages, the numbers in them to 4 significant digits, the headers and the log identical;
  • inside the brain, the field within 0.5 Hz at the 99th percentile and 0.1 Hz on average, with a correlation of at least 0.999;
  • the corrected images within 10−3 relative;
  • the movements within 0.05 mm and 0.001 rad.

It also shows how close the clean room comes. Each case it passes is run again and classified from every file against FSL’s, and against FSL’s own thread band where the record measures it: FSL at four threads against FSL at one (FSL-116), on fMRIPrep’s two calls and the thread cases.

Agreement, clean room against the original: FSL topup and applytopup 6.0.7.23 (standard comparison, 187 cases of the record), by category and verdict
fMRIPrep's calls2
--debug, --trace and --profile5
options37
configuration files16
errors and messages39
thread counts3
applytopup75
random testing10

Select a segment to list its cases below.

The clean room’s record runs 187 cases against FSL 6.0.7.23 (the record, made again on 2026-11-07). 106 are within the thresholds, 67 are rejected by both alike (the same message and exit status), and 14 are expected divergences, each with its reason in the record. None is outside.

  • Bit-identical, 9 of the 106. They give FSL’s values bit for bit.
  • Within FSL’s own band, 2. fMRIPrep’s call on ds001600 and the case at --nthr=4 lie within FSL’s own thread band in every metric of the record.
  • Open, 95. They are within the thresholds, but not bit for bit: the clean room computes in its own order of operations. On 92 of them FSL’s own band is not measured; that band is measured in the final validation phase. On the other three the clean room lies beyond the band FSL’s thread settings give: fMRIPrep’s call on ds005040 in two metrics of 24 (the field’s correlation with FSL’s, for --fout and for the coefficients’ field: 0.999958 against FSL’s own 0.999966; the field’s 99th percentile, 0.61 Hz, is inside FSL’s 0.64), and the two thread cases whose settings leave FSL’s own result all but unchanged.
  • Expected divergences, 14. In 6 the movements are estimated and have not converged. There reproduce takes FSL’s forward differences (FSL-201), but FSL’s own path is not determined beyond its inputs’ last bits, so the record holds them to FSL’s own spread with the read-out times changed by 1 to 256 last bits (below). 4 are applytopup’s variational restorations: FSL-143’s diverging iterations, reproduced, which amplify the last bits, and vb2D’s transients in a few slices. One is FSL-126’s negative weight, with which both minimisations diverge. And 3 inputs are refused in fix and in reproduce: a negative --warpres (FSL-124: FSL needs more memory than a machine holds) and fields on another grid (FSL-134: FSL reads past its lines).

The five cases that FSL-133’s lines and FSL-134’s field of another shape kept apart pass since the catch-up, and opt/fsl145-one-coefficient is rejected by both alike. Two cases came on 2026-11-07:

  • at/jac-coef-aniso, with the specification’s answer on applytopup’s Jacobian (FSL-273, specs/topup.md §7.3): --method=jac with a coefficient field of other voxel sizes. FSL divides the field’s derivatives by the field’s own voxel sizes while it lays the displacement in the images’, so its Jacobian is scaled along each encoded axis by the ratio of the two. The clean room takes the Jacobian of the displacement it applies by default (scale 1, naming FSL-273), and FSL’s scales in reproduce (findings).
  • random testing’s permanent case difftest/time-step-zero-series: fMRIPrep’s call on a series whose time step is 0, its --iout compared (the time step as FSL’s image layer reads it).

In fMRIPrep: the field map from a reversed-phase-encoding pair

fMRIPrep’s own call on ds005040’s pair, given to FSL 6.0.7.23 at one thread (as fMRIPrep runs it) and at four, to the replica in reproduce and in fix (and at four threads), and to the clean room in reproduce and in fix on 14 threads (and on one). The two input panels are the pair before the correction. Choose an output above the panels to see each image corrected, the field, the Jacobians and the displacements.

The replica’s differences from FSL are zero in every one. FSL’s own result at four threads differs from its result at one (FSL-116). That comparison, FSL topup, --nthr=4 − --nthr=1, is the yardstick of how much FSL itself moves, and the replica gives FSL’s one-thread result at any thread count.

The clean room’s differences from FSL are of about the yardstick’s size, on the same scale beside it. This call’s scaled conjugate gradients at full resolution (60 iterations with almost no regularisation at a knot spacing of one voxel) carry any last-bit difference forward, the clean room’s as FSL’s own at another thread count. Below the viewer: every file topup writes, with fMRIPrep’s own outputs (topup in fMRIPrep’s image), and the record’s metrics of the clean room beside FSL’s own thread band.

Worked example

fMRIPrep's field map from a reversed-phase-encoding pair

What this shows: On fMRIPrep’s own call, the replica gives FSL’s files, every value, at one thread and at four, and the same values as fMRIPrep’s own. FSL itself moves at four threads (FSL-116: its field by up to 26.1 Hz), the yardstick. The clean room, the same at 1 and 14 threads, lies within that band in every metric of its record but the field’s correlation with FSL’s (0.999958, where FSL’s own at four threads is 0.999966), so the comparison stays open. No correction by default changes this call, in either port.

The topup node fMRIPrep 25.2.6 ran on ds005040-sub-002’s reversed-phase-encoding pair (sdcflows’ PEPOLAR estimation, its configuration b02b0.cnf: seven levels from 14 mm to 4 mm knots, the last four by scaled conjugate gradients), given to FSL 6.0.7.23 at one thread (as fMRIPrep runs it) and at four, to the replica (larmorx-nc) in reproduce and in fix and at four threads, and to LARMORX’s clean room in reproduce and in fix at 14 threads and at one. The two input panels are the pair before the correction, distorted in opposite directions along the phase encoding; choose an output above the panels: each volume corrected, the field in Hz, each volume’s Jacobian and its displacement along the phase encoding. Below the viewer, every file topup writes against FSL’s, with fMRIPrep’s own, and the record’s metrics of the clean room beside FSL’s own thread band.

Data
OpenNeuro ds005040, sub-002 (CC0-1.0): the field-map images fMRIPrep merged (each run averaged); shown inside sdcflows’ brain mask of the corrected reference, dilated by 2 voxels (1,397 ml), so no face shows
Node
fMRIPrep 25.2.6, work/fmriprep_25_2_wf/sub_002_wf/fmap_preproc_wf/wf_auto_00000/topup
Command
topup --config=<sdcflows' b02b0.cnf> --datain=<nipype's encoding file> --imain=<fMRIPrep's merged pair> --out=sub-002_dir-A_epi_average_merged_sliced_volreg_base --iout=sub-002_dir-A_epi_average_merged_sliced_volreg_corrected.nii.gz --fout=sub-002_dir-A_epi_average_merged_sliced_volreg_field.nii.gz --jacout=jac --logout=sub-002_dir-A_epi_average_merged_sliced_volreg_topup.log --rbmout=xfm --dfout=warpfield
Acquisition
phase encodings +y, −y; read-out time 0.01379510 s (nipype’s encoding file)
Programs
FSL 6.0.7.23 (fsl-topup 2203.5; ~/fsl-6.0.7.23, never on the PATH) (its run served from the cache of the original’s runs, as the record ran it; at --nthr=4 run through it); larmorx-nc 0.0.1 (the replica), from LARMORX at commit b2a352b8b3ba; LARMORX 0.0.1 (the clean room, crate larmorx-fsl, Apache-2.0), from LARMORX at commit 80cc208fa4a3
Displacements
warpfield_01.nii.gz, warpfield_02.nii.gz: three components each, the x and z ones 0 everywhere (phase encoding along y), so the y component is shown
Grid
64 × 64 × 40 × 2 voxels of 3 × 3 × 3.3 mm; shown: 52 × 62 × 40 around the brain
Inputds005040-sub-002 pair, volume 1 (+y)topup's input in fMRIPrep: the reversed-phase-encoding images, each the average of its run, merged (float32)
Inputds005040-sub-002 pair, volume 2 (−y)topup's input in fMRIPrep: the reversed-phase-encoding images, each the average of its run, merged (float32)
OriginalFSL topup 6.0.7.23the conda installation, as built, one thread (the record's run, from the cache of the original's runs)
Replicareplicalarmorx-nc, FSL Licence · reproduce · against FSL topup 6.0.7.23
Clean roomclean roomlarmorx, Apache-2.0 · reproduce, 14 threads · against FSL topup 6.0.7.23
the original against itself: 1 and 4 threadsFSL topup, --nthr=4 − --nthr=1scale ±7.13e3

Bit-identical163,840 of 163,840 voxels equal

the replica against the originalreplica − FSL topup 6.0.7.23
the clean room against the originalclean room − FSL topup 6.0.7.23scale ±7.13e3

Axial slice 20 of 40; the crosshair where the clean room differs most from FSL at one thread in this slice, beside FSL at four threads against one, its own band, on the same scale (the replica's output equals FSL's everywhere). The interactive viewer needs JavaScript; the metrics below are complete without it.

Every comparison, computed in float64 over every voxel of the outputs

Comparison (A − B), verdict and reasonOutputmax |Δ|Values that differRMS of Δmax |Δ| ÷ range1 − NCC
The original against itself: the yardstick
FSL topup, --nthr=4 − --nthr=1referenceFSL’s own thread band (FSL-116): above one thread FSL sums the sparse matrix-vector products of its Hessian solves in per-thread pieces, so its result changes with --nthr: here the field by up to 26.1 Hz and the corrected images by up to 6.26e+03 over the whole field of view (inside the brain: the thread-band table below the viewer). The replica gives FSL’s one-thread result at any --nthr; this row is the yardstick. FSL-116Corrected: volume 1 (+y)6.06e3150,463 of 163,840 (91.8 %)1150.04228.16e−6
Corrected: volume 2 (−y)6.26e3150,494 of 163,840 (91.9 %)1060.04846.85e−6
Corrected: mean of the volumes2.57e3162,732 of 163,840 (99.3 %)23.10.02033.27e−7
Field (Hz)26.1163,796 of 163,840 (100 %)0.1970.037.96e−6
Jacobian: volume 1 (+y)0.204163,179 of 163,840 (99.6 %)0.001950.04649.37e−5
Jacobian: volume 2 (−y)0.204163,241 of 163,840 (99.6 %)0.001950.04649.37e−5
Displacement along y: volume 1 (+y)1.08163,782 of 163,840 (100 %)0.008170.037.96e−6
Displacement along y: volume 2 (−y)1.08163,782 of 163,840 (100 %)0.008170.037.96e−6
The replica against the original
replica − FSL topup 6.0.7.23Bit-identicalThe same values, every one, in every image topup writes (the record compares every file, the messages and the exit status too).every outputidentical: 0 of 163,840 values differ
replica − fMRIPrep's outputBit-identicalfMRIPrep’s own files (topup in fMRIPrep 25.2.6’s image, an older build) hold the same values.every outputidentical: 0 of 163,840 values differ
The replica against itself: thread counts, fix and reproduce
replica, --nthr=4 − --nthr=1Bit-identicalThe replica gives FSL’s one-thread result at any --nthr (FSL-116: its sums are taken in one fixed order).every outputidentical: 0 of 163,840 values differ
replica, fix − reproduceBit-identicalNo correction of the register applies to this call: fix and reproduce give the same values.every outputidentical: 0 of 163,840 values differ
The clean room against the original
clean room − FSL topup 6.0.7.23Open differenceWithin the record’s thresholds (specs/topup.md §12), but further from FSL than FSL’s own thread band in 2 metric(s): _fieldcoef (its field, Hz): correlation inside the brain 0.999958 (§12: 0.999; FSL’s own 0.999966); --fout (Hz): correlation inside the brain 0.999958 (§12: 0.999; FSL’s own 0.999966). findingsCorrected: volume 1 (+y)7.13e3150,547 of 163,840 (91.9 %)1230.04979.43e−6
Corrected: volume 2 (−y)7.13e3150,549 of 163,840 (91.9 %)1130.05527.85e−6
Corrected: mean of the volumes3.27e3163,207 of 163,840 (99.6 %)250.02583.86e−7
Field (Hz)31.3163,780 of 163,840 (100 %)0.2260.03591.05e−5
Jacobian: volume 1 (+y)0.275163,063 of 163,840 (99.5 %)0.002210.06260.00012
Jacobian: volume 2 (−y)0.275163,107 of 163,840 (99.6 %)0.002210.06260.00012
Displacement along y: volume 1 (+y)1.29163,781 of 163,840 (100 %)0.009370.03591.05e−5
Displacement along y: volume 2 (−y)1.29163,781 of 163,840 (100 %)0.009370.03591.05e−5
clean room − fMRIPrep's outputOpen differencefMRIPrep’s own files are FSL’s (the replica’s row): within the record’s thresholds (specs/topup.md §12), but further from FSL than FSL’s own thread band in 2 metric(s): _fieldcoef (its field, Hz): correlation inside the brain 0.999958 (§12: 0.999; FSL’s own 0.999966); --fout (Hz): correlation inside the brain 0.999958 (§12: 0.999; FSL’s own 0.999966).Corrected: volume 1 (+y)7.13e3150,547 of 163,840 (91.9 %)1230.04979.43e−6
Corrected: volume 2 (−y)7.13e3150,549 of 163,840 (91.9 %)1130.05527.85e−6
Field (Hz)31.3163,780 of 163,840 (100 %)0.2260.03591.05e−5
Jacobian: volume 1 (+y)0.275163,063 of 163,840 (99.5 %)0.002210.06260.00012
Jacobian: volume 2 (−y)0.275163,107 of 163,840 (99.6 %)0.002210.06260.00012
Displacement along y: volume 1 (+y)1.29163,781 of 163,840 (100 %)0.009370.03591.05e−5
Displacement along y: volume 2 (−y)1.29163,781 of 163,840 (100 %)0.009370.03591.05e−5
The clean room against itself: thread counts, fix and reproduce
clean room, 1 thread − 14 threadsBit-identicalThe thread count changes nothing: the clean room’s parallel loops (fixed blocks of lines, planes and coefficients, sums in a fixed order) give its one-thread result bit for bit at 14 threads.every outputidentical: 0 of 163,840 values differ
clean room, fix − reproduceBit-identicalNo correction of the register applies to this call: fix and reproduce give the same values (the smoothing’s exp, glibc’s in reproduce and correctly rounded by default, GLIBC-07, gives the same result here).every outputidentical: 0 of 163,840 values differ

Δ is A − B, computed in float64 over all 163,840 voxels of the outputs, not only those shown. The replica’s record’s threshold: every file FSL writes bit-identical (images decompressed, descrip left out), with FSL’s exit status and messages. The clean room’s record (fsl-topup) holds it to specs/topup.md §12’s thresholds instead: FSL’s exit status, files and messages (numbers to 4 significant digits), the headers and the log identical; inside the record’s brain the field within 0.5 Hz at the 99th percentile and 0.1 Hz on average (correlation at least 0.999), the corrected images within 1e-3 relative at the 99th percentile and 1e-3 of the mean intensity on average, the movements within 0.05 mm and 0.001 rad, the Jacobians and displacements within the field’s bounds times the read-out time; a metric beyond its threshold passes within FSL’s own band where the record measures it.

Every file topup writes, both ports against the original: FSL topup 6.0.7.23 on one thread

FileTypeSizereplica, reproducereplica, reproduce, --nthr=4replica, fixFSL topup 6.0.7.23, --nthr=4fMRIPrep's outputclean room, reproduceclean room, fix
jac_01.nii.gzfloat3264 × 64 × 40identicalidenticalidenticaldiffersidentical163,063 of 163,840 values differ163,063 of 163,840 values differ
jac_02.nii.gzfloat3264 × 64 × 40identicalidenticalidenticaldiffersidentical163,107 of 163,840 values differ163,107 of 163,840 values differ
sub-002_dir-A_epi_average_merged_sliced_volreg_base_fieldcoef.nii.gzfloat3264 × 64 × 40identicalidenticalidenticaldiffersidentical163,782 of 163,840 values differ163,782 of 163,840 values differ
sub-002_dir-A_epi_average_merged_sliced_volreg_base_movpar.txttext2 linesidenticalidenticalidenticalidentical–identicalidentical
sub-002_dir-A_epi_average_merged_sliced_volreg_corrected.nii.gzfloat3264 × 64 × 40 × 2identicalidenticalidenticaldiffersidentical301,096 of 327,680 values differ301,096 of 327,680 values differ
sub-002_dir-A_epi_average_merged_sliced_volreg_corrected.nii.gz_mean.nii.gzfloat3264 × 64 × 40identicalidenticalidenticaldiffers–163,207 of 163,840 values differ163,207 of 163,840 values differ
sub-002_dir-A_epi_average_merged_sliced_volreg_field.nii.gzfloat3264 × 64 × 40identicalidenticalidenticaldiffersidentical163,780 of 163,840 values differ163,780 of 163,840 values differ
sub-002_dir-A_epi_average_merged_sliced_volreg_topup.logtext42 linesidenticaldiffersidenticaldiffers–identicalidentical
warpfield_01.nii.gzfloat3264 × 64 × 40 × 3identicalidenticalidenticaldiffersidentical163,781 of 491,520 values differ163,781 of 491,520 values differ
warpfield_02.nii.gzfloat3264 × 64 × 40 × 3identicalidenticalidenticaldiffersidentical163,781 of 491,520 values differ163,781 of 491,520 values differ
xfm_01.mattext4 linesidenticalidenticalidenticalidenticalidenticalidenticalidentical
xfm_02.mattext4 linesidenticalidenticalidenticalidenticalidenticalidenticalidentical

Each image decompressed and compared byte for byte with FSL’s: identical, the header (its descrip, FSL’s build string, aside) and every value; data identical, every value but not the header; differs, values that differ (see the comparisons above); text files byte for byte; – where the run keeps no such file. fMRIPrep’s own files are the node’s outputs in the run’s working directory (topup in fMRIPrep 25.2.6’s image); it keeps no log or movement parameters. The log records the options as given, so the --nthr=4 runs’ logs differ by that option. For the clean room, where the values are not the same, how many of the file’s values differ from FSL’s (its record holds it to specs/topup.md §12’s thresholds; the comparisons above give the sizes).

FSL’s own thread band, the original against itself: FSL topup 6.0.7.23 at --nthr=4 and at one thread

OutputValues that differLargest |Δ|Largest |Δ|, brain99th percentile |Δ|, brainMedian |Δ|, brainUnit
Corrected: volume 1 (+y)150,4636.06e36.06e357513.2–
Corrected: volume 2 (−y)150,4946.26e36.26e356113.6–
Corrected: mean of the volumes162,7322.57e328544.30.977–
Field (Hz)163,79626.13.190.4920.0491Hz
Jacobian: volume 1 (+y)163,1790.2040.05090.007090.000246–
Jacobian: volume 2 (−y)163,2410.2040.05090.007090.000246–
Displacement along y: volume 1 (+y)163,7821.080.1320.02040.00203mm
Displacement along y: volume 2 (−y)163,7821.080.1320.02040.00203mm

Computed in float64 over every voxel of FSL’s two runs’ files; brain: inside sdcflows’ brain mask of the corrected reference (not dilated). The replica at --nthr=4 gives FSL’s one-thread values (the comparison replica, --nthr=4 − --nthr=1 above: identical); FSL-116 in the register.

The record’s metrics, the clean room against the original: FSL topup 6.0.7.23, beside FSL’s own thread band

Metric (specs/topup.md §12)clean room − FSL topup 6.0.7.23FSL topup, --nthr=4 − --nthr=1§12's threshold
jac_01: |d| inside the brain, 99th percentile0.008220.00840.0069within FSL’s band
jac_01: |d| inside the brain, mean0.0007950.0008290.00138within FSL’s band
jac_02: |d| inside the brain, 99th percentile0.008220.00840.0069within FSL’s band
jac_02: |d| inside the brain, mean0.0007950.0008290.00138within FSL’s band
_fieldcoef (its field, Hz): |d| inside the brain, 99th percentile0.6070.640.5within FSL’s band
_fieldcoef (its field, Hz): |d| inside the brain, mean0.08270.08850.1within FSL’s band
_fieldcoef (its field, Hz): correlation inside the brain0.9999580.9999660.999within the threshold, beyond FSL’s band
sub-002_dir-A_epi_average_merged_sliced_volreg_base_movpar.txt: translations000.05within FSL’s band
sub-002_dir-A_epi_average_merged_sliced_volreg_base_movpar.txt: rotations000.001within FSL’s band
sub-002_dir-A_epi_average_merged_sliced_volreg_corrected: relative difference inside the brain, 99th percentile0.01010.01020.001within FSL’s band
sub-002_dir-A_epi_average_merged_sliced_volreg_corrected: mean |d| / mean intensity inside the brain0.0008540.0008830.001within FSL’s band
sub-002_dir-A_epi_average_merged_sliced_volreg_corrected.nii.gz_mean: relative difference inside the brain, 99th percentile0.002570.002790.001within FSL’s band
sub-002_dir-A_epi_average_merged_sliced_volreg_corrected.nii.gz_mean: mean |d| / mean intensity inside the brain9.45e−59.82e−50.001within FSL’s band
–fout (Hz): |d| inside the brain, 99th percentile0.6070.640.5within FSL’s band
–fout (Hz): |d| inside the brain, mean0.08270.08850.1within FSL’s band
–fout (Hz): correlation inside the brain0.9999580.9999660.999within the threshold, beyond FSL’s band
warpfield_01 (mm): |d| inside the brain, 99th percentile0.02510.02650.0207within FSL’s band
warpfield_01 (mm): |d| inside the brain, mean0.003420.003660.00414within FSL’s band
warpfield_02 (mm): |d| inside the brain, 99th percentile0.02510.02650.0207within FSL’s band
warpfield_02 (mm): |d| inside the brain, mean0.003420.003660.00414within FSL’s band
xfm_01.mat: translations000.201within FSL’s band
xfm_01.mat: rotations000.001within FSL’s band
xfm_02.mat: translations000.201within FSL’s band
xfm_02.mat: rotations000.001within FSL’s band

Each metric as the record computes it (measures of the comparison suite parity.fsl_topup), over the record’s brain (FSL’s corrected mean above 10 % of its 99th percentile): the clean room (reproduce, 4 threads; its result is the same at any thread count) against FSL 6.0.7.23 at one thread, and FSL at --nthr=4 against FSL at one (its own thread band, FSL-116; the run the replica’s evidence made, served by the cache of the original’s runs); a correlation is better when higher, every other metric when lower.

Provenance of this example
  • Data: OpenNeuro ds005040, sub-002, CC0-1.0
  • Inputs: merged reversed-phase-encoding pair (sha256 69ae4ec0347584a3…); b02b0.cnf (sha256 034f7c42027e500d…); brain mask (sha256 6f66568a84cafa3c…)
  • Runs: figures cache entry fsl-topup/fmriprep-ds005040-sub-002-noreconall, computed 2026-11-03T06:22:09+00:00; 1 of its 5 runs made again since, last on 2026-11-06T19:55:18+00:00, with larmorx b2a352b8b3ba; FSL 6.0.7.23 (fsl-topup 2203.5; ~/fsl-6.0.7.23, never on the PATH); larmorx-nc 0.0.1; every command, its environment, threads and exit status are in its RUN_INFO.json
  • Runs: figures cache entry fsl-topup/original-fmriprep-ds005040-sub-002-noreconall, computed 2026-11-03T14:29:36+00:00; its runs made again since, last on 2026-11-07T02:06:39+00:00, with larmorx 80cc208fa4a3; FSL 6.0.7.23 (fsl-topup 2203.5; ~/fsl-6.0.7.23, never on the PATH); “larmorx 0.0.1 (the clean-room original, crate larmorx-fsl, Apache-2.0)”, as the cache entry records it; every command, its environment, threads and exit status are in its RUN_INFO.json

Every number above is in fsl-topup.json, exported by python -m larmorx_validation evidence fsl-topup.

A second run: ds001600

fMRIPrep’s call on the data panel’s ds001600, a Siemens Prisma pair whose two phase encodings have different read-out times and were acquired on different grids (sdcflows resampled one onto the other first). The same runs and comparisons as above; the viewer opens on the field. Here the clean room lies within FSL’s own thread band in every metric of the record.

Worked example

A second run: ds001600's pair

What this shows: On fMRIPrep’s call on ds001600’s pair, the replica gives FSL’s files, every value, at one thread and at four, and the same values as fMRIPrep’s own. The clean room, the same at 1 and 14 threads, is within FSL’s own thread band in every metric of its record. No correction by default changes this call, in either port.

The topup node fMRIPrep 25.2.6 ran on ds001600-sub-1’s reversed-phase-encoding pair (sdcflows’ PEPOLAR estimation, its configuration b02b0.cnf: seven levels from 14 mm to 4 mm knots, the last four by scaled conjugate gradients), given to FSL 6.0.7.23 at one thread (as fMRIPrep runs it) and at four, to the replica (larmorx-nc) in reproduce and in fix and at four threads, and to LARMORX’s clean room in reproduce and in fix at 14 threads and at one. The two input panels are the pair before the correction, distorted in opposite directions along the phase encoding; choose an output above the panels: each volume corrected, the field in Hz, each volume’s Jacobian and its displacement along the phase encoding. Below the viewer, every file topup writes against FSL’s, with fMRIPrep’s own, and the record’s metrics of the clean room beside FSL’s own thread band.

Data
OpenNeuro ds001600, sub-1 (CC-BY-SA): the field-map images fMRIPrep merged (each run averaged); shown inside sdcflows’ brain mask of the corrected reference, dilated by 2 voxels (1,360 ml), so no face shows
Node
fMRIPrep 25.2.6, work/fmriprep_25_2_wf/sub_1_wf/fmap_preproc_wf/wf_auto_00003/topup
Command
topup --config=<sdcflows' b02b0.cnf> --datain=<nipype's encoding file> --imain=<fMRIPrep's merged pair> --out=sub-1_dir-PA_epi_average_regrid001_merged_sliced_volreg_base --iout=sub-1_dir-PA_epi_average_regrid001_merged_sliced_volreg_corrected.nii.gz --fout=sub-1_dir-PA_epi_average_regrid001_merged_sliced_volreg_field.nii.gz --jacout=jac --logout=sub-1_dir-PA_epi_average_regrid001_merged_sliced_volreg_topup.log --rbmout=xfm --dfout=warpfield
Acquisition
phase encodings +y, −y; read-out time 0.03465010, 0.04266030 s (nipype’s encoding file)
Programs
FSL 6.0.7.23 (fsl-topup 2203.5; ~/fsl-6.0.7.23, never on the PATH) (its run served from the cache of the original’s runs, as the record ran it; at --nthr=4 run through it); larmorx-nc 0.0.1 (the replica), from LARMORX at commit b2a352b8b3ba; LARMORX 0.0.1 (the clean room, crate larmorx-fsl, Apache-2.0), from LARMORX at commit 80cc208fa4a3
Displacements
warpfield_01.nii.gz, warpfield_02.nii.gz: three components each, the x and z ones 0 everywhere (phase encoding along y), so the y component is shown
Grid
80 × 80 × 60 × 2 voxels of 3 × 3 × 3 mm; shown: 53 × 64 × 50 around the brain

Data shown

  • Example fieldmaps (OpenNeuro ds001600, version 1.0.0), Creative Commons Attribution-ShareAlike (CC BY-SA; the dataset gives no version). Attribution: Cieslak M., Elliott M., Satterthwaite T. Example Fieldmaps. OpenNeuro ds001600, version 1.0.0 (Siemens Prisma, University of Pennsylvania). The images shown are adapted from it: reoriented to RAS, cropped to the region shown, masked outside it, and shown as the outputs of the programs compared. Share-alike: the images adapted from this dataset are shared under the same licence, CC BY-SA.
Inputds001600-sub-1 pair, volume 1 (+y)topup's input in fMRIPrep: the reversed-phase-encoding images, each the average of its run, merged (float32)
Inputds001600-sub-1 pair, volume 2 (−y)topup's input in fMRIPrep: the reversed-phase-encoding images, each the average of its run, merged (float32)
OriginalFSL topup 6.0.7.23the conda installation, as built, one thread (the record's run, from the cache of the original's runs)
Replicareplicalarmorx-nc, FSL Licence · reproduce · against FSL topup 6.0.7.23
Clean roomclean roomlarmorx, Apache-2.0 · reproduce, 14 threads · against FSL topup 6.0.7.23
the original against itself: 1 and 4 threadsFSL topup, --nthr=4 − --nthr=1scale ±0.0872

Bit-identical384,000 of 384,000 voxels equal

the replica against the originalreplica − FSL topup 6.0.7.23
the clean room against the originalclean room − FSL topup 6.0.7.23scale ±0.0872

Axial slice 25 of 50; the crosshair where the clean room differs most from FSL at one thread in this slice, beside FSL at four threads against one, its own band, on the same scale (the replica's output equals FSL's everywhere). The interactive viewer needs JavaScript; the metrics below are complete without it.

Every comparison, computed in float64 over every voxel of the outputs

Comparison (A − B), verdict and reasonOutputmax |Δ|Values that differRMS of Δmax |Δ| ÷ range1 − NCC
The original against itself: the yardstick
FSL topup, --nthr=4 − --nthr=1referenceFSL’s own thread band (FSL-116): above one thread FSL sums the sparse matrix-vector products of its Hessian solves in per-thread pieces, so its result changes with --nthr: here the field by up to 5.83 Hz and the corrected images by up to 29.9 over the whole field of view (inside the brain: the thread-band table below the viewer). The replica gives FSL’s one-thread result at any --nthr; this row is the yardstick. FSL-116Corrected: volume 1 (+y)2.81359,002 of 384,000 (93.5 %)0.05360.0023.79e−8
Corrected: volume 2 (−y)29.9360,369 of 384,000 (93.8 %)0.1220.01852.22e−7
Corrected: mean of the volumes15.4377,750 of 384,000 (98.4 %)0.05670.01094.51e−8
Field (Hz)5.83383,680 of 384,000 (99.9 %)0.02330.00929.06e−8
Jacobian: volume 1 (+y)0.129380,479 of 384,000 (99.1 %)0.0005560.01741.82e−6
Jacobian: volume 2 (−y)0.159382,322 of 384,000 (99.6 %)0.0006840.01741.82e−6
Displacement along y: volume 1 (+y)0.606383,613 of 384,000 (99.9 %)0.002430.00929.06e−8
Displacement along y: volume 2 (−y)0.746383,675 of 384,000 (99.9 %)0.002990.00929.06e−8
The replica against the original
replica − FSL topup 6.0.7.23Bit-identicalThe same values, every one, in every image topup writes (the record compares every file, the messages and the exit status too).every outputidentical: 0 of 384,000 values differ
replica − fMRIPrep's outputBit-identicalfMRIPrep’s own files (topup in fMRIPrep 25.2.6’s image, an older build) hold the same values.every outputidentical: 0 of 384,000 values differ
The replica against itself: thread counts, fix and reproduce
replica, --nthr=4 − --nthr=1Bit-identicalThe replica gives FSL’s one-thread result at any --nthr (FSL-116: its sums are taken in one fixed order).every outputidentical: 0 of 384,000 values differ
replica, fix − reproduceBit-identicalNo correction of the register applies to this call: fix and reproduce give the same values.every outputidentical: 0 of 384,000 values differ
The clean room against the original
clean room − FSL topup 6.0.7.23Within the original's own variabilityIn every metric of the record (specs/topup.md §12, inside the record’s brain) the clean room is no further from FSL at one thread than FSL at another thread count is (its own thread band, FSL-116): sub-1_dir-PA_epi_average_regrid001_merged_sliced_volreg_corrected: relative difference inside the brain, 99th percentile 0.00108 (§12: 0.001; FSL’s own 0.00155); sub-1_dir-PA_epi_average_regrid001_merged_sliced_volreg_corrected.nii.gz_mean: relative difference inside the brain, 99th percentile 0.000608 (§12: 0.001; FSL’s own 0.000903); sub-1_dir-PA_epi_average_regrid001_merged_sliced_volreg_corrected: mean |d| / mean intensity inside the brain 0.000126 (§12: 0.001; FSL’s own 0.000175). findings FSL-116Corrected: volume 1 (+y)3.01357,784 of 384,000 (93.2 %)0.04750.002142.97e−8
Corrected: volume 2 (−y)19.1359,172 of 384,000 (93.5 %)0.09050.01181.21e−7
Corrected: mean of the volumes9.3376,810 of 384,000 (98.1 %)0.03960.006562.2e−8
Field (Hz)4.89383,532 of 384,000 (99.9 %)0.0190.007725.97e−8
Jacobian: volume 1 (+y)0.109376,612 of 384,000 (98.1 %)0.0004580.01471.24e−6
Jacobian: volume 2 (−y)0.134380,038 of 384,000 (99 %)0.0005640.01471.24e−6
Displacement along y: volume 1 (+y)0.508383,535 of 384,000 (99.9 %)0.001970.007725.97e−8
Displacement along y: volume 2 (−y)0.626383,531 of 384,000 (99.9 %)0.002430.007725.97e−8
clean room − fMRIPrep's outputWithin the original's own variabilityfMRIPrep’s own files are FSL’s (the replica’s row): in every metric of the record (specs/topup.md §12, inside the record’s brain) the clean room is no further from FSL at one thread than FSL at another thread count is (its own thread band, FSL-116): sub-1_dir-PA_epi_average_regrid001_merged_sliced_volreg_corrected: relative difference inside the brain, 99th percentile 0.00108 (§12: 0.001; FSL’s own 0.00155); sub-1_dir-PA_epi_average_regrid001_merged_sliced_volreg_corrected.nii.gz_mean: relative difference inside the brain, 99th percentile 0.000608 (§12: 0.001; FSL’s own 0.000903); sub-1_dir-PA_epi_average_regrid001_merged_sliced_volreg_corrected: mean |d| / mean intensity inside the brain 0.000126 (§12: 0.001; FSL’s own 0.000175).Corrected: volume 1 (+y)3.01357,784 of 384,000 (93.2 %)0.04750.002142.97e−8
Corrected: volume 2 (−y)19.1359,172 of 384,000 (93.5 %)0.09050.01181.21e−7
Field (Hz)4.89383,532 of 384,000 (99.9 %)0.0190.007725.97e−8
Jacobian: volume 1 (+y)0.109376,612 of 384,000 (98.1 %)0.0004580.01471.24e−6
Jacobian: volume 2 (−y)0.134380,038 of 384,000 (99 %)0.0005640.01471.24e−6
Displacement along y: volume 1 (+y)0.508383,535 of 384,000 (99.9 %)0.001970.007725.97e−8
Displacement along y: volume 2 (−y)0.626383,531 of 384,000 (99.9 %)0.002430.007725.97e−8
The clean room against itself: thread counts, fix and reproduce
clean room, 1 thread − 14 threadsBit-identicalThe thread count changes nothing: the clean room’s parallel loops (fixed blocks of lines, planes and coefficients, sums in a fixed order) give its one-thread result bit for bit at 14 threads.every outputidentical: 0 of 384,000 values differ
clean room, fix − reproduceBit-identicalNo correction of the register applies to this call: fix and reproduce give the same values (the smoothing’s exp, glibc’s in reproduce and correctly rounded by default, GLIBC-07, gives the same result here).every outputidentical: 0 of 384,000 values differ

Δ is A − B, computed in float64 over all 384,000 voxels of the outputs, not only those shown. The replica’s record’s threshold: every file FSL writes bit-identical (images decompressed, descrip left out), with FSL’s exit status and messages. The clean room’s record (fsl-topup) holds it to specs/topup.md §12’s thresholds instead: FSL’s exit status, files and messages (numbers to 4 significant digits), the headers and the log identical; inside the record’s brain the field within 0.5 Hz at the 99th percentile and 0.1 Hz on average (correlation at least 0.999), the corrected images within 1e-3 relative at the 99th percentile and 1e-3 of the mean intensity on average, the movements within 0.05 mm and 0.001 rad, the Jacobians and displacements within the field’s bounds times the read-out time; a metric beyond its threshold passes within FSL’s own band where the record measures it.

Every file topup writes, both ports against the original: FSL topup 6.0.7.23 on one thread

FileTypeSizereplica, reproducereplica, reproduce, --nthr=4replica, fixFSL topup 6.0.7.23, --nthr=4fMRIPrep's outputclean room, reproduceclean room, fix
jac_01.nii.gzfloat3280 × 80 × 60identicalidenticalidenticaldiffersidentical376,612 of 384,000 values differ376,612 of 384,000 values differ
jac_02.nii.gzfloat3280 × 80 × 60identicalidenticalidenticaldiffersidentical380,038 of 384,000 values differ380,038 of 384,000 values differ
sub-1_dir-PA_epi_average_regrid001_merged_sliced_volreg_base_fieldcoef.nii.gzfloat3280 × 80 × 60identicalidenticalidenticaldiffersidentical383,587 of 384,000 values differ383,587 of 384,000 values differ
sub-1_dir-PA_epi_average_regrid001_merged_sliced_volreg_base_movpar.txttext2 linesidenticalidenticalidenticalidentical–identicalidentical
sub-1_dir-PA_epi_average_regrid001_merged_sliced_volreg_corrected.nii.gzfloat3280 × 80 × 60 × 2identicalidenticalidenticaldiffersidentical716,956 of 768,000 values differ716,956 of 768,000 values differ
sub-1_dir-PA_epi_average_regrid001_merged_sliced_volreg_corrected.nii.gz_mean.nii.gzfloat3280 × 80 × 60identicalidenticalidenticaldiffers–376,810 of 384,000 values differ376,810 of 384,000 values differ
sub-1_dir-PA_epi_average_regrid001_merged_sliced_volreg_field.nii.gzfloat3280 × 80 × 60identicalidenticalidenticaldiffersidentical383,532 of 384,000 values differ383,532 of 384,000 values differ
sub-1_dir-PA_epi_average_regrid001_merged_sliced_volreg_topup.logtext42 linesidenticaldiffersidenticaldiffers–identicalidentical
warpfield_01.nii.gzfloat3280 × 80 × 60 × 3identicalidenticalidenticaldiffersidentical383,535 of 1,152,000 values differ383,535 of 1,152,000 values differ
warpfield_02.nii.gzfloat3280 × 80 × 60 × 3identicalidenticalidenticaldiffersidentical383,531 of 1,152,000 values differ383,531 of 1,152,000 values differ
xfm_01.mattext4 linesidenticalidenticalidenticalidenticalidenticalidenticalidentical
xfm_02.mattext4 linesidenticalidenticalidenticalidenticalidenticalidenticalidentical

Each image decompressed and compared byte for byte with FSL’s: identical, the header (its descrip, FSL’s build string, aside) and every value; data identical, every value but not the header; differs, values that differ (see the comparisons above); text files byte for byte; – where the run keeps no such file. fMRIPrep’s own files are the node’s outputs in the run’s working directory (topup in fMRIPrep 25.2.6’s image); it keeps no log or movement parameters. The log records the options as given, so the --nthr=4 runs’ logs differ by that option. For the clean room, where the values are not the same, how many of the file’s values differ from FSL’s (its record holds it to specs/topup.md §12’s thresholds; the comparisons above give the sizes).

FSL’s own thread band, the original against itself: FSL topup 6.0.7.23 at --nthr=4 and at one thread

OutputValues that differLargest |Δ|Largest |Δ|, brain99th percentile |Δ|, brainMedian |Δ|, brainUnit
Corrected: volume 1 (+y)359,0022.812.810.4120.0424–
Corrected: volume 2 (−y)360,36929.92.710.4510.0481–
Corrected: mean of the volumes377,75015.40.6360.08950.00537–
Field (Hz)383,6805.830.05040.02180.00439Hz
Jacobian: volume 1 (+y)380,4790.1290.00210.0005679.28e−5–
Jacobian: volume 2 (−y)382,3220.1590.002580.0006980.000114–
Displacement along y: volume 1 (+y)383,6130.6060.005240.002260.000456mm
Displacement along y: volume 2 (−y)383,6750.7460.006450.002790.000562mm

Computed in float64 over every voxel of FSL’s two runs’ files; brain: inside sdcflows’ brain mask of the corrected reference (not dilated). The replica at --nthr=4 gives FSL’s one-thread values (the comparison replica, --nthr=4 − --nthr=1 above: identical); FSL-116 in the register.

The record’s metrics, the clean room against the original: FSL topup 6.0.7.23, beside FSL’s own thread band

Metric (specs/topup.md §12)clean room − FSL topup 6.0.7.23FSL topup, --nthr=4 − --nthr=1§12's threshold
jac_01: |d| inside the brain, 99th percentile0.0006510.0009180.0173within FSL’s band
jac_01: |d| inside the brain, mean0.0001120.0001710.00347within FSL’s band
jac_02: |d| inside the brain, 99th percentile0.0008020.001130.0213within FSL’s band
jac_02: |d| inside the brain, mean0.0001380.0002110.00427within FSL’s band
_fieldcoef (its field, Hz): |d| inside the brain, 99th percentile0.02310.02960.5within FSL’s band
_fieldcoef (its field, Hz): |d| inside the brain, mean0.004290.006750.1within FSL’s band
_fieldcoef (its field, Hz): correlation inside the brain110.999within FSL’s band
sub-1_dir-PA_epi_average_regrid001_merged_sliced_volreg_base_movpar.txt: translations000.05within FSL’s band
sub-1_dir-PA_epi_average_regrid001_merged_sliced_volreg_base_movpar.txt: rotations000.001within FSL’s band
sub-1_dir-PA_epi_average_regrid001_merged_sliced_volreg_corrected: relative difference inside the brain, 99th percentile0.001080.001550.001within FSL’s band
sub-1_dir-PA_epi_average_regrid001_merged_sliced_volreg_corrected: mean |d| / mean intensity inside the brain0.0001260.0001750.001within FSL’s band
sub-1_dir-PA_epi_average_regrid001_merged_sliced_volreg_corrected.nii.gz_mean: relative difference inside the brain, 99th percentile0.0006080.0009030.001within FSL’s band
sub-1_dir-PA_epi_average_regrid001_merged_sliced_volreg_corrected.nii.gz_mean: mean |d| / mean intensity inside the brain2.91e−54.34e−50.001within FSL’s band
–fout (Hz): |d| inside the brain, 99th percentile0.02310.02960.5within FSL’s band
–fout (Hz): |d| inside the brain, mean0.004290.006750.1within FSL’s band
–fout (Hz): correlation inside the brain110.999within FSL’s band
warpfield_01 (mm): |d| inside the brain, 99th percentile0.00240.003080.052within FSL’s band
warpfield_01 (mm): |d| inside the brain, mean0.0004460.0007020.0104within FSL’s band
warpfield_02 (mm): |d| inside the brain, 99th percentile0.002960.003790.064within FSL’s band
warpfield_02 (mm): |d| inside the brain, mean0.0005490.0008640.0128within FSL’s band
xfm_01.mat: translations000.242within FSL’s band
xfm_01.mat: rotations000.001within FSL’s band
xfm_02.mat: translations000.242within FSL’s band
xfm_02.mat: rotations000.001within FSL’s band

Each metric as the record computes it (measures of the comparison suite parity.fsl_topup), over the record’s brain (FSL’s corrected mean above 10 % of its 99th percentile): the clean room (reproduce, 4 threads; its result is the same at any thread count) against FSL 6.0.7.23 at one thread, and FSL at --nthr=4 against FSL at one (its own thread band, FSL-116; the run the replica’s evidence made, served by the cache of the original’s runs); a correlation is better when higher, every other metric when lower.

Provenance of this example
  • Data: OpenNeuro ds001600, sub-1, CC-BY-SA
  • Inputs: merged reversed-phase-encoding pair (sha256 f573a275b924eebd…); b02b0.cnf (sha256 034f7c42027e500d…); brain mask (sha256 aa9f5a88e712a0ff…)
  • Runs: figures cache entry fsl-topup/fmriprep-ds001600-sub-1-noreconall, computed 2026-11-03T06:24:18+00:00; 1 of its 5 runs made again since, last on 2026-11-06T19:56:31+00:00, with larmorx b2a352b8b3ba; FSL 6.0.7.23 (fsl-topup 2203.5; ~/fsl-6.0.7.23, never on the PATH); larmorx-nc 0.0.1; every command, its environment, threads and exit status are in its RUN_INFO.json
  • Runs: figures cache entry fsl-topup/original-fmriprep-ds001600-sub-1-noreconall, computed 2026-11-03T14:29:51+00:00; its runs made again since, last on 2026-11-07T02:06:49+00:00, with larmorx 80cc208fa4a3; FSL 6.0.7.23 (fsl-topup 2203.5; ~/fsl-6.0.7.23, never on the PATH); “larmorx 0.0.1 (the clean-room original, crate larmorx-fsl, Apache-2.0)”, as the cache entry records it; every command, its environment, threads and exit status are in its RUN_INFO.json

Every number above is in fsl-topup.json, exported by python -m larmorx_validation evidence fsl-topup.

Membrane energy, and a correction by default

With --regmod=membrane_energy and Levenberg-Marquardt (the record’s case opt/membrane-lm, on the half-resolution pair), FSL takes the regulariser’s Hessian in other units than its cost and gradient: each first derivative divided by the voxel size in millimetres, where the cost counts in voxels. The replica’s default takes the Hessian in the cost’s units (FSL-117), so its steps follow the cost they minimise; its reproduce gives FSL’s bytes.

This case estimates the movements too (topup’s default), so since 2026-11-03 the default also takes their exact rotation derivatives (FSL-201, below). The comparison replica, fix − reproduce shows both corrections together, in the field. The clean room’s default makes the same two corrections.

Worked example

Membrane energy with Levenberg-Marquardt, and a correction by default

What this shows: With the membrane energy and Levenberg-Marquardt, the replica gives FSL’s files, every value, and the clean room is within its record’s thresholds, its field at most 3.2 × 10⁻⁶ Hz from FSL’s. In fix, their default, both correct FSL-117 (the energy’s Hessian in other units than its cost) and FSL-201 (the movements’ derivatives: this call estimates them), which moves the replica’s field by up to 0.019 Hz.

The record’s case opt/membrane-lm: topup with the membrane energy as its regulariser (--regmod=membrane_energy; fMRIPrep’s configuration uses the bending energy) and Levenberg-Marquardt, one short level on the half-resolution pair of ds005040. FSL 6.0.7.23 and the replica’s reproduce give the same bytes; the replica’s default corrects FSL-117 (the energy’s Hessian in other units than its cost and gradient), and so takes other steps: the difference replica, fix − reproduce shows what that changes in the field. LARMORX’s clean room runs the same call in reproduce (at 14 threads and at one) and in fix, its default, which corrects FSL-117 and FSL-201 as the replica’s does.

Data
OpenNeuro ds005040, sub-002 (CC0-1.0): the record’s inputs derived from fMRIPrep’s merged pair (every second voxel; oracles/inputs/fsl-topup-replica/); shown inside sdcflows’ brain mask of the corrected reference at the same resolution, dilated by one voxel
Case
opt/membrane-lm of the records (fsl-topup-replica, fsl-topup)
Command
topup --imain=<half.nii.gz: derived from ds005040's fMRIPrep pair> --datain=<acq.txt: acquisition parameters written by the suite> --out=o --warpres=20 --fwhm=8 --miter=3 --lambda=0.005 --regmod=membrane_energy --fout=f
Programs
FSL 6.0.7.23 (fsl-topup 2203.5; ~/fsl-6.0.7.23, never on the PATH) (its run served from the cache of the original’s runs, as the record ran it); larmorx-nc 0.0.1 (the replica), from LARMORX at commit b2a352b8b3ba; LARMORX 0.0.1 (the clean room, crate larmorx-fsl, Apache-2.0), from LARMORX at commit 80cc208fa4a3
Grid
32 × 32 × 20 × 2 voxels of 6 × 6 × 6.6 mm; shown: 28 × 32 × 20 around the brain
Inputds005040 pair, half resolution, volume 1phase encoding +y; derived from fMRIPrep's merged pair (every second voxel: 6 × 6 × 6.6 mm)
Inputds005040 pair, half resolution, volume 2phase encoding −y; derived from fMRIPrep's merged pair (every second voxel: 6 × 6 × 6.6 mm)
OriginalFSL topup 6.0.7.23the conda installation, as built, one thread, the record's run
Replicareplicalarmorx-nc, FSL Licence · fix (the default)
Clean roomclean roomlarmorx, Apache-2.0 · reproduce, 14 threads · against FSL topup 6.0.7.23

Bit-identical20,480 of 20,480 voxels equal

the replica against the originalreplica − FSL topup 6.0.7.23
the replica against itself: fix and reproducereplica, fix − reproducescale ±0.0194

Axial slice 10 of 20; the crosshair where the default, fix, differs most from reproduce (FSL's bytes) in this slice. The interactive viewer needs JavaScript; the metrics below are complete without it.

Every comparison, computed in float64 over every voxel of the outputs

Comparison (A − B), verdict and reasonmax |Δ|Values that differRMS of Δmax |Δ| ÷ range1 − NCC
The replica against the original
replica − FSL topup 6.0.7.23Bit-identicalThe same values, every one, in every image topup writes (the record compares every file, the messages and the exit status too).identical: 0 of 20,480 values differ
The replica against itself: fix and reproduce
replica, fix − reproduceDeliberate correctionFSL-117, corrected by default with a notice: with membrane energy, Levenberg-Marquardt’s Hessian of the regularisation divides each derivative by the voxel size in mm while the cost and its gradient count in voxels; the default takes the Hessian in the cost’s units, so the steps follow the cost they minimise (the final cost is never higher: −0.00 to −1.66 % on the measured cases). reproduce gives FSL’s bytes. FSL-201, corrected by default with a notice (since 2026-11-03): with the movements estimated, FSL’s derivatives of the cost with respect to the rotations are forward differences (steps of 10⁻⁵ rad) of rotation matrices built from a single-precision angle, cosine and sine, so the rounding divided by the step makes its gradient and its Gauss-Newton Hessian of the movements other than the cost’s (their rotation entries 0.25 to 39 % off on ds005040’s pair at half resolution) and the minimisers end elsewhere than the cost’s minimum; the default computes the exact derivatives. Measured on 22 runs that estimate the movements, the default’s final cost was lower or equal in 18 (the other four, runs cut short, 0.004 to 0.13 % higher), lower in all 8 random cases whose volumes moved (by 0.01 to 44 %), and 38 % lower after 60 iterations of scaled conjugate gradients on the half-resolution pair, where FSL’s stall; fMRIPrep’s calls estimate no movements and stay the same. reproduce gives FSL’s bytes. FSL-117 FSL-2010.019420,480 of 20,480 (100 %)0.004880.2210.00659
The clean room against the original
clean room − FSL topup 6.0.7.23Open differenceWithin the record’s thresholds (specs/topup.md §12): o_movpar.txt: rotations 1.48e-05 (§12: 0.001); o_movpar.txt: translations 1.04e-05 (§12: 0.05); --fout (Hz): |d| inside the brain, mean 8.05e-07 (§12: 0.1). 3 of 4 files differ in their values (most: f.nii.gz: 20,468 of 20,480 values, at most 3.18e-06): the clean room computes in its own order of operations (written from the specification, not from FSL’s code). FSL’s own band is not measured on this case (its thread band is, on fMRIPrep’s calls). findings3.18e−620,468 of 20,480 (99.9 %)6.98e−73.62e−51.48e−9
The clean room against itself: thread counts, fix and reproduce
clean room, 1 thread − 14 threadsBit-identicalThe thread count changes nothing: the clean room’s parallel loops (fixed blocks of lines, planes and coefficients, sums in a fixed order) give its one-thread result bit for bit at 14 threads.identical: 0 of 20,480 values differ
clean room, fix − reproduceDeliberate correctionFSL-117, corrected by default with a notice: with membrane energy, Levenberg-Marquardt’s Hessian of the regularisation divides each derivative by the voxel size in mm while the cost and its gradient count in voxels; the default takes the Hessian in the cost’s units, so the steps follow the cost they minimise. Its reproduce follows FSL. FSL-201, corrected by default with a notice: with the movements estimated, FSL’s derivatives of the cost with respect to the rotations are forward differences (steps of 10⁻⁵ rad) of rotation matrices built from single-precision angles, cosines and sines, so its gradient and Gauss-Newton Hessian of the movements are not the cost’s, and its minimisers end elsewhere than the cost’s minimum; the default computes the exact derivatives. Its reproduce takes FSL’s forward differences (since 2026-11-05: the first gradients of FSL’s --debug files to 2 × 10⁻⁵ and 10⁻⁶), but FSL’s own path depends on its inputs’ last bits, so where the movements do not converge the record holds that mode to FSL’s own spread. FSL-117 FSL-2010.019420,480 of 20,480 (100 %)0.004880.2210.00659

Δ is A − B, computed in float64 over all 20,480 voxels of the outputs, not only those shown. The replica’s record’s threshold: every file FSL writes bit-identical (images decompressed, descrip left out), with FSL’s exit status and messages. The clean room’s record (fsl-topup) holds it to specs/topup.md §12’s thresholds instead: FSL’s exit status, files and messages (numbers to 4 significant digits), the headers and the log identical; inside the record’s brain the field within 0.5 Hz at the 99th percentile and 0.1 Hz on average (correlation at least 0.999), the corrected images within 1e-3 relative at the 99th percentile and 1e-3 of the mean intensity on average, the movements within 0.05 mm and 0.001 rad, the Jacobians and displacements within the field’s bounds times the read-out time; a metric beyond its threshold passes within FSL’s own band where the record measures it.

Every file topup writes, both ports against the original: FSL topup 6.0.7.23 on one thread

FileTypeSizereplica, reproducereplica, fixclean room, reproduceclean room, fix
f.nii.gzfloat3232 × 32 × 20identicaldiffers20,468 of 20,480 values differ20,480 of 20,480 values differ
half.topup_logtext40 linesidenticalidenticalidenticalidentical
o_fieldcoef.nii.gzfloat3213 × 13 × 9identicaldiffers1,518 of 1,521 values differ1,521 of 1,521 values differ
o_movpar.txttext2 linesidenticaldiffersdiffersdiffers

Each image decompressed and compared byte for byte with FSL’s: identical, the header (its descrip, FSL’s build string, aside) and every value; data identical, every value but not the header; differs, values that differ (see the comparisons above); text files byte for byte; – where the run keeps no such file. For the clean room, where the values are not the same, how many of the file’s values differ from FSL’s (its record holds it to specs/topup.md §12’s thresholds; the comparisons above give the sizes).

Provenance of this example
  • Data: OpenNeuro ds005040, sub-002, CC0-1.0
  • Inputs: half.nii.gz (sha256 33b8882c4fc51ed3…)
  • Runs: figures cache entry fsl-topup/opt-membrane-lm, computed 2026-11-03T06:20:34+00:00; 1 of its 3 runs made again since, last on 2026-11-06T19:56:31+00:00, with larmorx b2a352b8b3ba; FSL 6.0.7.23 (fsl-topup 2203.5; ~/fsl-6.0.7.23, never on the PATH); larmorx-nc 0.0.1; every command, its environment, threads and exit status are in its RUN_INFO.json
  • Runs: figures cache entry fsl-topup/original-opt-membrane-lm, computed 2026-11-03T14:29:52+00:00; its runs made again since, last on 2026-11-07T02:06:50+00:00, with larmorx 80cc208fa4a3; FSL 6.0.7.23 (fsl-topup 2203.5; ~/fsl-6.0.7.23, never on the PATH); “larmorx 0.0.1 (the clean-room original, crate larmorx-fsl, Apache-2.0)”, as the cache entry records it; every command, its environment, threads and exit status are in its RUN_INFO.json

Every number above is in fsl-topup.json, exported by python -m larmorx_validation evidence fsl-topup.

Movements estimated, and a correction by default

With --estmov=1 topup also estimates each volume’s rigid-body movement against the first. Here it is the record’s case opt/levels: two levels on the half-resolution pair, the first at subsampling 2 by Levenberg-Marquardt with the movements, the second at full resolution by scaled conjugate gradients; every output written. fMRIPrep’s configuration estimates none; FSL’s own b02b0.cnf does, in its first five levels.

FSL takes the cost’s derivatives with respect to the rotations as forward differences of rotation matrices built in single precision. So its gradient and its Hessian of the movements are not the cost’s, and its minimisers end elsewhere than the cost’s minimum. That is FSL-201, registered on 2026-11-03 with the specification’s answers to the clean room’s questions, after measuring that the exact derivatives end at a lower or equal cost in 18 of 22 runs (the other four, cut short, 0.004 to 0.13 % higher) and at a lower one in all 8 random cases whose volumes moved. The replica’s default computes the exact derivatives (deliberate correction); its reproduce gives FSL’s bytes.

The clean room’s default computes them too, and since 2026-11-05 its reproduce takes FSL’s forward differences. The movement table below the viewer shows the two defaults within 6 × 10−6 mm and 3 × 10−7 rad of each other, where FSL’s movements lie 3 × 10−4 mm and 8 × 10−5 rad from them. The clean room’s reproduce ends 5 × 10−4 mm and 1.3 × 10−4 rad from FSL. Its first gradients are FSL’s, but FSL’s path here is not determined beyond its inputs’ last bits: a rotation’s derivative jumps by about 2 % when its angle’s rounding to single precision flips. That is why the record holds this case as an expected divergence, measured against FSL’s own spread.

Worked example

Movements estimated, and a correction by default

What this shows: With the movements estimated, the replica gives FSL’s files, every value. In fix, their default, both take the exact rotation derivatives (FSL-201): the replica’s movements lie 3.1 × 10⁻⁴ mm and 8.4 × 10⁻⁵ rad from FSL’s, the clean room’s within 6.4 × 10⁻⁶ mm and 3.3 × 10⁻⁷ rad of the replica’s. In reproduce the clean room takes FSL’s forward differences but ends 4.9 × 10⁻⁴ mm and 1.3 × 10⁻⁴ rad from FSL, whose path here depends on its inputs’ last bits: an expected divergence.

The record’s case opt/levels: two levels on the half-resolution pair of ds005040, the first at subsampling 2 by Levenberg-Marquardt with the movements estimated (the second volume’s rigid-body movement against the first; fMRIPrep’s configuration estimates none, FSL’s own b02b0.cnf does in its first five levels), the second at full resolution by scaled conjugate gradients; every output written. FSL 6.0.7.23 and the replica’s reproduce give the same bytes; since 2026-11-03 the replica’s default corrects FSL-201 (FSL’s rotation derivatives are not its cost’s) and so takes other steps: the difference replica, fix − reproduce shows what that changes, and the movement table below the viewer the movements each run estimated. The clean room’s default computes the cost’s own derivatives too; since 2026-11-05 its reproduce takes FSL’s forward differences, but FSL’s path here depends on its inputs’ last bits, so its record holds this case as an expected divergence from FSL, measured against FSL’s own spread. LARMORX’s clean room runs the same call in reproduce (at 14 threads and at one) and in fix, its default: FSL’s forward differences in the first, the exact derivatives in the second.

Data
OpenNeuro ds005040, sub-002 (CC0-1.0): the record’s inputs derived from fMRIPrep’s merged pair (every second voxel; oracles/inputs/fsl-topup-replica/); shown inside sdcflows’ brain mask of the corrected reference at the same resolution, dilated by one voxel
Case
opt/levels of the records (fsl-topup-replica, fsl-topup)
Command
topup --imain=<half.nii.gz: derived from ds005040's fMRIPrep pair> --datain=<acq.txt: acquisition parameters written by the suite> --out=o --warpres=20,14 --subsamp=2,1 --fwhm=8,4 --miter=3,2 --lambda=0.005,0.0001 --estmov=1,0 --minmet=0,1 --iout=c --fout=f --dfout=d --rbmout=r --jacout=j --featout=feat -v
Programs
FSL 6.0.7.23 (fsl-topup 2203.5; ~/fsl-6.0.7.23, never on the PATH) (its run served from the cache of the original’s runs, as the record ran it); larmorx-nc 0.0.1 (the replica), from LARMORX at commit b2a352b8b3ba; LARMORX 0.0.1 (the clean room, crate larmorx-fsl, Apache-2.0), from LARMORX at commit 80cc208fa4a3
Grid
32 × 32 × 20 × 2 voxels of 6 × 6 × 6.6 mm; shown: 28 × 32 × 20 around the brain
Inputds005040 pair, half resolution, volume 1phase encoding +y; derived from fMRIPrep's merged pair (every second voxel: 6 × 6 × 6.6 mm)
Inputds005040 pair, half resolution, volume 2phase encoding −y; derived from fMRIPrep's merged pair (every second voxel: 6 × 6 × 6.6 mm)
OriginalFSL topup 6.0.7.23the conda installation, as built, one thread, the record's run
Replicareplicalarmorx-nc, FSL Licence · fix (the default)
Clean roomclean roomlarmorx, Apache-2.0 · reproduce, 14 threads · against FSL topup 6.0.7.23

Bit-identical20,480 of 20,480 voxels equal

the replica against the originalreplica − FSL topup 6.0.7.23
the replica against itself: fix and reproducereplica, fix − reproducescale ±16.7

Axial slice 10 of 20; the crosshair where the default, fix, differs most from reproduce (FSL's bytes) in this slice. The interactive viewer needs JavaScript; the metrics below are complete without it.

Every comparison, computed in float64 over every voxel of the outputs

Comparison (A − B), verdict and reasonOutputmax |Δ|Values that differRMS of Δmax |Δ| ÷ range1 − NCC
The replica against the original
replica − FSL topup 6.0.7.23Bit-identicalThe same values, every one, in every image topup writes (the record compares every file, the messages and the exit status too).every outputidentical: 0 of 20,480 values differ
The replica against itself: fix and reproduce
replica, fix − reproduceDeliberate correctionFSL-201, corrected by default with a notice (since 2026-11-03): with the movements estimated, FSL’s derivatives of the cost with respect to the rotations are forward differences (steps of 10⁻⁵ rad) of rotation matrices built from a single-precision angle, cosine and sine, so the rounding divided by the step makes its gradient and its Gauss-Newton Hessian of the movements other than the cost’s (their rotation entries 0.25 to 39 % off on ds005040’s pair at half resolution) and the minimisers end elsewhere than the cost’s minimum; the default computes the exact derivatives. Measured on 22 runs that estimate the movements, the default’s final cost was lower or equal in 18 (the other four, runs cut short, 0.004 to 0.13 % higher), lower in all 8 random cases whose volumes moved (by 0.01 to 44 %), and 38 % lower after 60 iterations of scaled conjugate gradients on the half-resolution pair, where FSL’s stall; fMRIPrep’s calls estimate no movements and stay the same. reproduce gives FSL’s bytes. FSL-201Corrected: volume 1 (+y)16.717,255 of 20,480 (84.3 %)1.250.0001449.88e−10
Corrected: volume 2 (−y)74.717,278 of 20,480 (84.4 %)5.860.0005882.16e−8
Corrected: mean of the volumes40.520,463 of 20,480 (99.9 %)3.110.0003396.08e−9
Field (Hz)0.032920,480 of 20,480 (100 %)0.009160.0003192.91e−7
Jacobian: volume 1 (+y)0.00014120,418 of 20,480 (99.7 %)2.29e−50.0002231.25e−7
Jacobian: volume 2 (−y)0.00014120,419 of 20,480 (99.7 %)2.29e−50.0002231.25e−7
Displacement along y: volume 1 (+y)0.0027220,480 of 20,480 (100 %)0.0007580.0003192.91e−7
Displacement along y: volume 2 (−y)0.0027220,480 of 20,480 (100 %)0.0007580.0003192.91e−7
--featout: magnitude40.520,463 of 20,480 (99.9 %)3.110.0003396.08e−9
--featout: field (rad/s)0.20720,480 of 20,480 (100 %)0.05750.0003192.91e−7
The clean room against the original
clean room − FSL topup 6.0.7.23Expected divergenceFSL-201 reproduced, the movements estimated and not converged: FSL’s rotation derivatives are forward differences of single-precision matrices, so each gradient entry moves by about 2 % of its lever-weighted sum when its angle’s rounding to single precision f…. The clean room’s reproduce takes FSL’s forward differences since 2026-11-05 (specs/topup.md §4.6, FSL-201); its default computes the cost’s own derivatives, as the replica’s does. record findingsCorrected: volume 1 (+y)19.417,263 of 20,480 (84.3 %)1.310.0001681.09e−9
Corrected: volume 2 (−y)95.317,276 of 20,480 (84.4 %)7.590.000753.61e−8
Corrected: mean of the volumes44.320,475 of 20,480 (100 %)3.950.0003719.81e−9
Field (Hz)0.033520,480 of 20,480 (100 %)0.009680.0003254.21e−7
Jacobian: volume 1 (+y)0.00016620,443 of 20,480 (99.8 %)2.57e−50.0002621.6e−7
Jacobian: volume 2 (−y)0.00016620,444 of 20,480 (99.8 %)2.57e−50.0002621.6e−7
Displacement along y: volume 1 (+y)0.0027720,480 of 20,480 (100 %)0.0008010.0003254.21e−7
Displacement along y: volume 2 (−y)0.0027720,480 of 20,480 (100 %)0.0008010.0003254.21e−7
--featout: magnitude44.320,475 of 20,480 (100 %)3.950.0003719.81e−9
--featout: field (rad/s)0.2120,480 of 20,480 (100 %)0.06080.0003254.21e−7
The clean room against itself: thread counts, fix and reproduce
clean room, 1 thread − 14 threadsBit-identicalThe thread count changes nothing: the clean room’s parallel loops (fixed blocks of lines, planes and coefficients, sums in a fixed order) give its one-thread result bit for bit at 14 threads.every outputidentical: 0 of 20,480 values differ
clean room, fix − reproduceDeliberate correctionFSL-201, corrected by default with a notice: with the movements estimated, FSL’s derivatives of the cost with respect to the rotations are forward differences (steps of 10⁻⁵ rad) of rotation matrices built from single-precision angles, cosines and sines, so its gradient and Gauss-Newton Hessian of the movements are not the cost’s, and its minimisers end elsewhere than the cost’s minimum; the default computes the exact derivatives. Its reproduce takes FSL’s forward differences (since 2026-11-05: the first gradients of FSL’s --debug files to 2 × 10⁻⁵ and 10⁻⁶), but FSL’s own path depends on its inputs’ last bits, so where the movements do not converge the record holds that mode to FSL’s own spread. FSL-201Corrected: volume 1 (+y)19.417,261 of 20,480 (84.3 %)1.190.0001688.97e−10
Corrected: volume 2 (−y)67.117,277 of 20,480 (84.4 %)4.70.0005281.38e−8
Corrected: mean of the volumes35.720,470 of 20,480 (100 %)2.550.0002994.08e−9
Field (Hz)0.030620,480 of 20,480 (100 %)0.009090.0002972.49e−7
Jacobian: volume 1 (+y)0.00011920,428 of 20,480 (99.7 %)2.2e−50.0001881.17e−7
Jacobian: volume 2 (−y)0.00011920,438 of 20,480 (99.8 %)2.2e−50.0001881.17e−7
Displacement along y: volume 1 (+y)0.0025320,480 of 20,480 (100 %)0.0007520.0002972.49e−7
Displacement along y: volume 2 (−y)0.0025320,480 of 20,480 (100 %)0.0007520.0002972.49e−7
--featout: magnitude35.720,470 of 20,480 (100 %)2.550.0002994.08e−9
--featout: field (rad/s)0.19220,480 of 20,480 (100 %)0.05710.0002972.49e−7

Δ is A − B, computed in float64 over all 20,480 voxels of the outputs, not only those shown. The replica’s record’s threshold: every file FSL writes bit-identical (images decompressed, descrip left out), with FSL’s exit status and messages. The clean room’s record (fsl-topup) holds it to specs/topup.md §12’s thresholds instead: FSL’s exit status, files and messages (numbers to 4 significant digits), the headers and the log identical; inside the record’s brain the field within 0.5 Hz at the 99th percentile and 0.1 Hz on average (correlation at least 0.999), the corrected images within 1e-3 relative at the 99th percentile and 1e-3 of the mean intensity on average, the movements within 0.05 mm and 0.001 rad, the Jacobians and displacements within the field’s bounds times the read-out time; a metric beyond its threshold passes within FSL’s own band where the record measures it.

Every file topup writes, both ports against the original: FSL topup 6.0.7.23 on one thread

FileTypeSizereplica, reproducereplica, fixclean room, reproduceclean room, fix
c.nii.gzfloat3232 × 32 × 20 × 2identicaldiffers34,539 of 40,960 values differ34,543 of 40,960 values differ
c_mean.nii.gzfloat3232 × 32 × 20identicaldiffers20,475 of 20,480 values differ20,468 of 20,480 values differ
d_01.nii.gzfloat3232 × 32 × 20 × 3identicaldiffers20,480 of 61,440 values differ20,480 of 61,440 values differ
d_02.nii.gzfloat3232 × 32 × 20 × 3identicaldiffers20,480 of 61,440 values differ20,480 of 61,440 values differ
f.nii.gzfloat3232 × 32 × 20identicaldiffers20,480 of 20,480 values differ20,479 of 20,480 values differ
feat_fieldmap_mag.nii.gzfloat3232 × 32 × 20identicaldiffers20,475 of 20,480 values differ20,468 of 20,480 values differ
feat_fieldmap_rads.nii.gzfloat3232 × 32 × 20identicaldiffers20,480 of 20,480 values differ20,480 of 20,480 values differ
half.topup_logtext44 linesidenticalidenticalidenticalidentical
j_01.nii.gzfloat3232 × 32 × 20identicaldiffers20,443 of 20,480 values differ20,429 of 20,480 values differ
j_02.nii.gzfloat3232 × 32 × 20identicaldiffers20,444 of 20,480 values differ20,432 of 20,480 values differ
o_fieldcoef.nii.gzfloat3219 × 19 × 13identicaldiffers4,693 of 4,693 values differ4,693 of 4,693 values differ
o_movpar.txttext2 linesidenticaldiffersdiffersdiffers
r_01.mattext4 linesidenticalidenticalidenticalidentical
r_02.mattext4 linesidenticaldiffersdiffersdiffers

Each image decompressed and compared byte for byte with FSL’s: identical, the header (its descrip, FSL’s build string, aside) and every value; data identical, every value but not the header; differs, values that differ (see the comparisons above); text files byte for byte; – where the run keeps no such file. For the clean room, where the values are not the same, how many of the file’s values differ from FSL’s (its record holds it to specs/topup.md §12’s thresholds; the comparisons above give the sizes).

The movements estimated, both ports against the original: FSL topup 6.0.7.23 (the moving volume against the first)

Runx (mm)y (mm)z (mm)rotation x (rad)rotation y (rad)rotation z (rad)|Δ| translation vs FSL|Δ| rotation vs FSL|Δ| translation vs the replica's default|Δ| rotation vs the replica's default
FSL topup 6.0.7.23−0.5832530−0.2924770.00918729−0.001637460.000675056000.0003078.43e−5
replica, reproduce−0.5832530−0.2924770.00918729−0.001637460.000675056000.0003078.43e−5
replica, fix−0.583560−0.2925330.00914028−0.001553190.0006967440.0003078.43e−500
clean room, reproduce−0.583740−0.292480.00919684−0.001509630.0007129610.0004860.0001280.0001795.66e−5
clean room, fix−0.5835550−0.2925390.00913994−0.001553340.000696840.0003028.41e−56.4e−63.31e−7

Each run’s --out movement file (_movpar.txt: one line per volume, the first volume the reference), as written; the last four columns the largest difference of the translations (mm) and of the rotations (rad) from FSL’s and from the replica’s default’s.

Provenance of this example
  • Data: OpenNeuro ds005040, sub-002, CC0-1.0
  • Inputs: half.nii.gz (sha256 33b8882c4fc51ed3…)
  • Runs: figures cache entry fsl-topup/opt-levels, computed 2026-11-03T22:11:09+00:00; 1 of its 3 runs made again since, last on 2026-11-06T19:56:31+00:00, with larmorx b2a352b8b3ba; FSL 6.0.7.23 (fsl-topup 2203.5; ~/fsl-6.0.7.23, never on the PATH); larmorx-nc 0.0.1; every command, its environment, threads and exit status are in its RUN_INFO.json
  • Runs: figures cache entry fsl-topup/original-opt-levels, computed 2026-11-03T22:11:10+00:00; its runs made again since, last on 2026-11-07T02:06:50+00:00, with larmorx 80cc208fa4a3; FSL 6.0.7.23 (fsl-topup 2203.5; ~/fsl-6.0.7.23, never on the PATH); “larmorx 0.0.1 (the clean-room original, crate larmorx-fsl, Apache-2.0)”, as the cache entry records it; every command, its environment, threads and exit status are in its RUN_INFO.json

Every number above is in fsl-topup.json, exported by python -m larmorx_validation evidence fsl-topup.

applytopup: least-squares restoration

applytopup’s default method (the record’s case at/lsr-y): the two images of opposite phase encodings restored into one, line by line along the phase encoding, with the field and movements of FSL’s topup on the same pair. FSL resamples the first image with trilinear interpolation whatever --interp asks (FSL-132). The replica’s default uses the method asked for, spline by default, and its reproduce gives FSL’s bytes. On this call the clean room’s fix and reproduce give the same values.

Worked example

applytopup: least-squares restoration of the pair (the default method)

What this shows: For applytopup’s least-squares restoration, the replica gives FSL’s image, every value; its default, fix, resamples the first image with the spline interpolation asked for (FSL-132), which changes 5,136 values by at most 0.0078. The clean room is within its record’s thresholds (1.2 × 10⁻⁵ relative at the 99th percentile of the brain, its threshold 0.001), and its fix and reproduce give the same values on this call.

The record’s case at/lsr-y: applytopup’s default method combines the two images of opposite phase encodings into one restored image, line by line along the phase encoding, with the field and movements of FSL’s topup on the same pair (the record’s case opt/levels, from the cache of the original’s runs). FSL 6.0.7.23 and the replica’s reproduce give the same bytes; the replica’s default corrects FSL-132 (the first image resampled with trilinear rather than spline interpolation). LARMORX’s clean room runs the same call in reproduce (at 14 threads and at one) and in fix, its default: on this call the two give the same values.

Data
OpenNeuro ds005040, sub-002 (CC0-1.0): the record’s inputs derived from fMRIPrep’s merged pair (every second voxel; oracles/inputs/fsl-topup-replica/); shown inside sdcflows’ brain mask of the corrected reference at the same resolution, dilated by one voxel
Case
at/lsr-y of the records (fsl-topup-replica, fsl-topup)
Command
applytopup --imain=<half_up.nii.gz: derived from ds005040's fMRIPrep pair>,<half_dn.nii.gz: derived from ds005040's fMRIPrep pair> --inindex=1,2 --datain=<acq.txt: acquisition parameters written by the suite> --topup=tu --out=hifi (tu_*: FSL's topup outputs of the case opt/levels)
Programs
FSL 6.0.7.23 (fsl-topup 2203.5; ~/fsl-6.0.7.23, never on the PATH) (its run served from the cache of the original’s runs, as the record ran it); larmorx-nc 0.0.1 (the replica), from LARMORX at commit b2a352b8b3ba; LARMORX 0.0.1 (the clean room, crate larmorx-fsl, Apache-2.0), from LARMORX at commit 80cc208fa4a3
Grid
32 × 32 × 20 voxels of 6 × 6 × 6.6 mm; shown: 28 × 32 × 20 around the brain
Inputds005040, +y image, half resolutionphase encoding +y; derived from fMRIPrep's merged pair (every second voxel: 6 × 6 × 6.6 mm)
Inputds005040, −y image, half resolutionphase encoding −y; derived from fMRIPrep's merged pair (every second voxel: 6 × 6 × 6.6 mm)
Inputthe field applied (Hz)FSL's topup on the half-resolution pair (the record's opt/levels), its --fout: the field its coefficients give
OriginalFSL applytopup 6.0.7.23the conda installation, as built, one thread, the record's run
Replicareplicalarmorx-nc, FSL Licence · fix (the default)
Clean roomclean roomlarmorx, Apache-2.0 · reproduce, 14 threads · against FSL applytopup 6.0.7.23

Bit-identical20,480 of 20,480 voxels equal

the replica against the originalreplica − FSL applytopup 6.0.7.23
the replica against itself: fix and reproducereplica, fix − reproducescale ±0.00781

Axial slice 10 of 20; the crosshair where the default, fix, differs most from reproduce (FSL's bytes) in this slice. The interactive viewer needs JavaScript; the metrics below are complete without it.

Every comparison, computed in float64 over every voxel of the outputs

Comparison (A − B), verdict and reasonmax |Δ|Values that differRMS of Δmax |Δ| ÷ range1 − NCC
The replica against the original
replica − FSL applytopup 6.0.7.23Bit-identicalThe same values, every one, in every image applytopup writes (the record compares every file, the messages and the exit status too).identical: 0 of 20,480 values differ
The replica against itself: fix and reproduce
replica, fix − reproduceDeliberate correctionFSL-132, corrected by default with a notice: applytopup resamples the first volume of the first image with trilinear interpolation whatever --interp asks (the default, spline, reaches every other volume); the default resamples every volume with the method asked for. reproduce gives FSL’s bytes. FSL-1320.007815,136 of 20,480 (25.1 %)0.0007345.26e−83.28e−16
The clean room against the original
clean room − FSL applytopup 6.0.7.23Open differenceWithin the record’s thresholds (specs/topup.md §12): hifi (applytopup’s output): relative difference inside the brain, 99th percentile 1.19e-05 (§12: 0.001); hifi (applytopup’s output): mean |d| / mean intensity inside the brain 6.84e-07 (§12: 0.001). 1 of 1 files differ in their values (most: hifi.nii.gz: 16,831 of 20,480 values, at most 0.813): the clean room computes in its own order of operations (written from the specification, not from FSL’s code). FSL’s own band is not measured on this case (its thread band is, on fMRIPrep’s calls). findings0.81316,831 of 20,480 (82.2 %)0.05475.48e−61.78e−12
The clean room against itself: thread counts, fix and reproduce
clean room, 1 thread − 14 threadsBit-identicalThe thread count changes nothing: the clean room’s parallel loops (fixed blocks of lines, planes and coefficients, sums in a fixed order) give its one-thread result bit for bit at 14 threads.identical: 0 of 20,480 values differ
clean room, fix − reproduceBit-identicalNo correction of the register applies to this call: fix and reproduce give the same values (the smoothing’s exp, glibc’s in reproduce and correctly rounded by default, GLIBC-07, gives the same result here).identical: 0 of 20,480 values differ

Δ is A − B, computed in float64 over all 20,480 voxels of the outputs, not only those shown. The replica’s record’s threshold: every file FSL writes bit-identical (images decompressed, descrip left out), with FSL’s exit status and messages. The clean room’s record (fsl-topup) holds it to specs/topup.md §12’s thresholds instead: FSL’s exit status, files and messages (numbers to 4 significant digits), the headers and the log identical; inside the record’s brain the field within 0.5 Hz at the 99th percentile and 0.1 Hz on average (correlation at least 0.999), the corrected images within 1e-3 relative at the 99th percentile and 1e-3 of the mean intensity on average, the movements within 0.05 mm and 0.001 rad, the Jacobians and displacements within the field’s bounds times the read-out time; a metric beyond its threshold passes within FSL’s own band where the record measures it.

Every file applytopup writes, both ports against the original: FSL applytopup 6.0.7.23

FileTypeSizereplica, reproducereplica, fixclean room, reproduceclean room, fix
hifi.nii.gzfloat3232 × 32 × 20identicaldiffers16,831 of 20,480 values differ16,831 of 20,480 values differ

Each image decompressed and compared byte for byte with FSL’s: identical, the header (its descrip, FSL’s build string, aside) and every value; data identical, every value but not the header; differs, values that differ (see the comparisons above); text files byte for byte; – where the run keeps no such file. For the clean room, where the values are not the same, how many of the file’s values differ from FSL’s (its record holds it to specs/topup.md §12’s thresholds; the comparisons above give the sizes).

Provenance of this example
  • Data: OpenNeuro ds005040, sub-002, CC0-1.0
  • Inputs: half_up.nii.gz (sha256 1ae5fd56967afd07…); half_dn.nii.gz (sha256 8e49385dc2270b50…); tu_field.nii.gz (sha256 d055f946cfb0adc9…); tu_fieldcoef.nii.gz (sha256 698a5839514d8d17…); tu_movpar.txt (sha256 9d69af8d9e2fd149…)
  • Runs: figures cache entry fsl-topup/at-lsr-y, computed 2026-11-03T06:20:35+00:00; 1 of its 3 runs made again since, last on 2026-11-06T19:56:31+00:00, with larmorx b2a352b8b3ba; FSL 6.0.7.23 (fsl-topup 2203.5; ~/fsl-6.0.7.23, never on the PATH); larmorx-nc 0.0.1; every command, its environment, threads and exit status are in its RUN_INFO.json
  • Runs: figures cache entry fsl-topup/original-at-lsr-y, computed 2026-11-03T14:29:52+00:00; its runs made again since, last on 2026-11-07T02:06:50+00:00, with larmorx 80cc208fa4a3; FSL 6.0.7.23 (fsl-topup 2203.5; ~/fsl-6.0.7.23, never on the PATH); “larmorx 0.0.1 (the clean-room original, crate larmorx-fsl, Apache-2.0)”, as the cache entry records it; every command, its environment, threads and exit status are in its RUN_INFO.json

Every number above is in fsl-topup.json, exported by python -m larmorx_validation evidence fsl-topup.

applytopup: Jacobian modulation

--method=jac (the record’s case at/jac-y): one image corrected by itself, resampled along the phase encoding and multiplied by the displacement’s Jacobian. FSL-132 reaches it too, and both ports’ defaults correct it. With a phase encoding along z, which topup does not estimate, FSL leaves the image uncorrected without a word; the clean room’s default refuses such a call (FSL-144), and its reproduce gives FSL’s output.

Worked example

applytopup: Jacobian modulation of one image

What this shows: For applytopup’s Jacobian modulation, the replica gives FSL’s image, every value, and the clean room is within its record’s thresholds (2.0 × 10⁻⁶ relative at the 99th percentile of the brain, its threshold 0.001). In fix, their default, both resample the image with the spline interpolation asked for (FSL-132), which changes all 20,480 values of the replica’s image.

The record’s case at/jac-y: --method=jac corrects one image by itself, resampled along the phase encoding and multiplied by the displacement’s Jacobian, with the result of FSL’s topup on the half-resolution pair. FSL 6.0.7.23 and the replica’s reproduce give the same bytes; the replica’s default corrects FSL-132 (that image’s interpolation). LARMORX’s clean room runs the same call in reproduce (at 14 threads and at one) and in fix, its default, which corrects FSL-132 as the replica’s does.

Data
OpenNeuro ds005040, sub-002 (CC0-1.0): the record’s inputs derived from fMRIPrep’s merged pair (every second voxel; oracles/inputs/fsl-topup-replica/); shown inside sdcflows’ brain mask of the corrected reference at the same resolution, dilated by one voxel
Case
at/jac-y of the records (fsl-topup-replica, fsl-topup)
Command
applytopup --imain=<half_up.nii.gz: derived from ds005040's fMRIPrep pair> --inindex=1 --datain=<acq.txt: acquisition parameters written by the suite> --topup=tu --out=hifi --method=jac (tu_*: FSL's topup outputs of the case opt/levels)
Programs
FSL 6.0.7.23 (fsl-topup 2203.5; ~/fsl-6.0.7.23, never on the PATH) (its run served from the cache of the original’s runs, as the record ran it); larmorx-nc 0.0.1 (the replica), from LARMORX at commit b2a352b8b3ba; LARMORX 0.0.1 (the clean room, crate larmorx-fsl, Apache-2.0), from LARMORX at commit 80cc208fa4a3
Grid
32 × 32 × 20 voxels of 6 × 6 × 6.6 mm; shown: 28 × 32 × 20 around the brain
Inputds005040, +y image, half resolutionphase encoding +y; derived from fMRIPrep's merged pair (every second voxel: 6 × 6 × 6.6 mm)
Inputthe field applied (Hz)FSL's topup on the half-resolution pair (the record's opt/levels), its --fout: the field its coefficients give
OriginalFSL applytopup 6.0.7.23the conda installation, as built, one thread, the record's run
Replicareplicalarmorx-nc, FSL Licence · fix (the default)
Clean roomclean roomlarmorx, Apache-2.0 · reproduce, 14 threads · against FSL applytopup 6.0.7.23

Bit-identical20,480 of 20,480 voxels equal

the replica against the originalreplica − FSL applytopup 6.0.7.23
the replica against itself: fix and reproducereplica, fix − reproducescale ±9.23e3

Axial slice 10 of 20; the crosshair where the default, fix, differs most from reproduce (FSL's bytes) in this slice. The interactive viewer needs JavaScript; the metrics below are complete without it.

Every comparison, computed in float64 over every voxel of the outputs

Comparison (A − B), verdict and reasonmax |Δ|Values that differRMS of Δmax |Δ| ÷ range1 − NCC
The replica against the original
replica − FSL applytopup 6.0.7.23Bit-identicalThe same values, every one, in every image applytopup writes (the record compares every file, the messages and the exit status too).identical: 0 of 20,480 values differ
The replica against itself: fix and reproduce
replica, fix − reproduceDeliberate correctionFSL-132, corrected by default with a notice: applytopup resamples the first volume of the first image with trilinear interpolation whatever --interp asks (the default, spline, reaches every other volume); the default resamples every volume with the method asked for. reproduce gives FSL’s bytes. FSL-1321.07e420,480 of 20,480 (100 %)8920.09790.000467
The clean room against the original
clean room − FSL applytopup 6.0.7.23Open differenceWithin the record’s thresholds (specs/topup.md §12): hifi (applytopup’s output): relative difference inside the brain, 99th percentile 2.05e-06 (§12: 0.001); hifi (applytopup’s output): mean |d| / mean intensity inside the brain 1.36e-07 (§12: 0.001). 1 of 1 files differ in their values (most: hifi.nii.gz: 13,708 of 20,480 values, at most 0.242): the clean room computes in its own order of operations (written from the specification, not from FSL’s code). FSL’s own band is not measured on this case (its thread band is, on fMRIPrep’s calls). findings0.24213,708 of 20,480 (66.9 %)0.009012.21e−65.02e−14
The clean room against itself: thread counts, fix and reproduce
clean room, 1 thread − 14 threadsBit-identicalThe thread count changes nothing: the clean room’s parallel loops (fixed blocks of lines, planes and coefficients, sums in a fixed order) give its one-thread result bit for bit at 14 threads.identical: 0 of 20,480 values differ
clean room, fix − reproduceDeliberate correctionFSL-132, corrected by default with a notice: applytopup resamples the first volume of the first image with trilinear interpolation whatever --interp asks (the default, spline, reaches every other volume); the default resamples every volume with the method asked for. Its reproduce follows FSL. FSL-1321.07e420,480 of 20,480 (100 %)8920.09790.000467

Δ is A − B, computed in float64 over all 20,480 voxels of the outputs, not only those shown. The replica’s record’s threshold: every file FSL writes bit-identical (images decompressed, descrip left out), with FSL’s exit status and messages. The clean room’s record (fsl-topup) holds it to specs/topup.md §12’s thresholds instead: FSL’s exit status, files and messages (numbers to 4 significant digits), the headers and the log identical; inside the record’s brain the field within 0.5 Hz at the 99th percentile and 0.1 Hz on average (correlation at least 0.999), the corrected images within 1e-3 relative at the 99th percentile and 1e-3 of the mean intensity on average, the movements within 0.05 mm and 0.001 rad, the Jacobians and displacements within the field’s bounds times the read-out time; a metric beyond its threshold passes within FSL’s own band where the record measures it.

Every file applytopup writes, both ports against the original: FSL applytopup 6.0.7.23

FileTypeSizereplica, reproducereplica, fixclean room, reproduceclean room, fix
hifi.nii.gzfloat3232 × 32 × 20identicaldiffers13,708 of 20,480 values differ20,480 of 20,480 values differ

Each image decompressed and compared byte for byte with FSL’s: identical, the header (its descrip, FSL’s build string, aside) and every value; data identical, every value but not the header; differs, values that differ (see the comparisons above); text files byte for byte; – where the run keeps no such file. For the clean room, where the values are not the same, how many of the file’s values differ from FSL’s (its record holds it to specs/topup.md §12’s thresholds; the comparisons above give the sizes).

Provenance of this example
  • Data: OpenNeuro ds005040, sub-002, CC0-1.0
  • Inputs: half_up.nii.gz (sha256 1ae5fd56967afd07…); tu_field.nii.gz (sha256 d055f946cfb0adc9…); tu_fieldcoef.nii.gz (sha256 698a5839514d8d17…); tu_movpar.txt (sha256 9d69af8d9e2fd149…)
  • Runs: figures cache entry fsl-topup/at-jac-y, computed 2026-11-03T06:20:35+00:00; 1 of its 3 runs made again since, last on 2026-11-06T19:56:31+00:00, with larmorx b2a352b8b3ba; FSL 6.0.7.23 (fsl-topup 2203.5; ~/fsl-6.0.7.23, never on the PATH); larmorx-nc 0.0.1; every command, its environment, threads and exit status are in its RUN_INFO.json
  • Runs: figures cache entry fsl-topup/original-at-jac-y, computed 2026-11-03T14:29:52+00:00; its runs made again since, last on 2026-11-07T02:06:50+00:00, with larmorx 80cc208fa4a3; FSL 6.0.7.23 (fsl-topup 2203.5; ~/fsl-6.0.7.23, never on the PATH); “larmorx 0.0.1 (the clean-room original, crate larmorx-fsl, Apache-2.0)”, as the cache entry records it; every command, its environment, threads and exit status are in its RUN_INFO.json

Every number above is in fsl-topup.json, exported by python -m larmorx_validation evidence fsl-topup.

applytopup --method=vb3D, and a correction by default

One of applytopup’s undocumented variational Bayes restorations (the record’s case at/vb3d-y): a smoothness prior within and between slices whose weights are estimated over 20 iterations. FSL builds the prior’s coupling between slices with half its weight, so the precision it assembles is not positive definite and its iterations diverge. FSL’s output spans about ±4 × 1010 for an input of at most 1.1 × 105, and the replica’s reproduce gives the same bytes. The replica’s default gives the prior its full weight (FSL-143), and its restoration stays in the input’s range, as vb2D’s does; so does the clean room’s default. FSL’s panel is shown in the default’s window; choose each image’s own window to read it.

Worked example

applytopup --method=vb3D, and a correction by default

What this shows: With vb3D, the replica gives FSL’s image, every value, diverging as FSL’s does, and the clean room’s reproduce follows it within its record’s thresholds (6.5 × 10⁻⁵ relative at the 99th percentile of the brain, its threshold 0.001). In fix, their default, both give the prior its full weight (FSL-143), and the restoration stays in the input’s range; the replica’s also resamples the first image with the spline interpolation asked for (FSL-132).

The record’s case at/vb3d-y: one of applytopup’s undocumented variational Bayes restorations (a smoothness prior within and between slices, its weights estimated over 20 iterations). FSL 6.0.7.23 and the replica’s reproduce give the same bytes, and both diverge: the prior FSL builds couples the slices with half its weight, which makes it indefinite (FSL-143). The replica’s default gives the prior its full weight (and resamples the first image with the spline interpolation asked for, FSL-132): its restoration stays in the input’s range. Choose each image’s own window to read FSL’s output. LARMORX’s clean room runs the same call in reproduce (at 14 threads and at one) and in fix, its default, which corrects FSL-143.

Data
OpenNeuro ds005040, sub-002 (CC0-1.0): the record’s inputs derived from fMRIPrep’s merged pair (every second voxel; oracles/inputs/fsl-topup-replica/); shown inside sdcflows’ brain mask of the corrected reference at the same resolution, dilated by one voxel
Case
at/vb3d-y of the records (fsl-topup-replica, fsl-topup)
Command
applytopup --imain=<half_up.nii.gz: derived from ds005040's fMRIPrep pair>,<half_dn.nii.gz: derived from ds005040's fMRIPrep pair> --inindex=1,2 --datain=<acq.txt: acquisition parameters written by the suite> --topup=tu --out=hifi --method=vb3D (tu_*: FSL's topup outputs of the case opt/levels)
Programs
FSL 6.0.7.23 (fsl-topup 2203.5; ~/fsl-6.0.7.23, never on the PATH) (its run served from the cache of the original’s runs, as the record ran it); larmorx-nc 0.0.1 (the replica), from LARMORX at commit b2a352b8b3ba; LARMORX 0.0.1 (the clean room, crate larmorx-fsl, Apache-2.0), from LARMORX at commit 80cc208fa4a3
Grid
32 × 32 × 20 voxels of 6 × 6 × 6.6 mm; shown: 28 × 32 × 20 around the brain
Inputds005040, +y image, half resolutionphase encoding +y; derived from fMRIPrep's merged pair (every second voxel: 6 × 6 × 6.6 mm)
Inputds005040, −y image, half resolutionphase encoding −y; derived from fMRIPrep's merged pair (every second voxel: 6 × 6 × 6.6 mm)
Inputthe field applied (Hz)FSL's topup on the half-resolution pair (the record's opt/levels), its --fout: the field its coefficients give
OriginalFSL applytopup 6.0.7.23the conda installation, as built, one thread, the record's run
Replicareplicalarmorx-nc, FSL Licence · fix (the default)
Clean roomclean roomlarmorx, Apache-2.0 · reproduce, 14 threads · against FSL applytopup 6.0.7.23

Bit-identical20,480 of 20,480 voxels equal

the replica against the originalreplica − FSL applytopup 6.0.7.23
the replica against itself: fix and reproducereplica, fix − reproducescale ±3.62e10

Axial slice 10 of 20; the crosshair where the default, fix, differs most from reproduce (FSL's bytes) in this slice. The interactive viewer needs JavaScript; the metrics below are complete without it.

Every comparison, computed in float64 over every voxel of the outputs

Comparison (A − B), verdict and reasonmax |Δ|Values that differRMS of Δmax |Δ| ÷ range1 − NCC
The replica against the original
replica − FSL applytopup 6.0.7.23Bit-identicalThe same values, every one, in every image applytopup writes (the record compares every file, the messages and the exit status too).identical: 0 of 20,480 values differ
The replica against itself: fix and reproduce
replica, fix − reproduceDeliberate correctionFSL-132, corrected by default with a notice: applytopup resamples the first volume of the first image with trilinear interpolation whatever --interp asks (the default, spline, reaches every other volume); the default resamples every volume with the method asked for. reproduce gives FSL’s bytes. FSL-143, corrected by default with a notice: vb3D’s and vb4D’s prior couples neighbouring slices with half the weight its own construction gives them, so the precision it builds is not positive definite and the iterations diverge (FSL’s output here spans about ±4 × 10¹⁰ for an input of 0 to 1.1 × 10⁵); the default gives the prior its full weight, and the restoration stays in the input’s range, as vb2D’s does. FSL-132 FSL-1433.85e1018,852 of 20,480 (92.1 %)1.47e90.5210.997
The clean room against the original
clean room − FSL applytopup 6.0.7.23Open differenceWithin the record’s thresholds (specs/topup.md §12): hifi (applytopup’s output): relative difference inside the brain, 99th percentile 6.54e-05 (§12: 0.001); hifi (applytopup’s output): mean |d| / mean intensity inside the brain 3.3e-05 (§12: 0.001). 1 of 1 files differ in their values (most: hifi.nii.gz: 18,753 of 20,480 values, at most 1.05e+06): the clean room computes in its own order of operations (written from the specification, not from FSL’s code). FSL’s own band is not measured on this case (its thread band is, on fMRIPrep’s calls). findings1.05e618,753 of 20,480 (91.6 %)4.34e41.42e−53.2e−11
The clean room against itself: thread counts, fix and reproduce
clean room, 1 thread − 14 threadsBit-identicalThe thread count changes nothing: the clean room’s parallel loops (fixed blocks of lines, planes and coefficients, sums in a fixed order) give its one-thread result bit for bit at 14 threads.identical: 0 of 20,480 values differ
clean room, fix − reproduceDeliberate correctionFSL-143, corrected by default with a notice: vb3D’s and vb4D’s prior couples neighbouring slices with half the weight its own construction gives them, so the precision it builds is not positive definite and the iterations diverge (FSL’s output here spans about ±4 × 10¹⁰ for an input of 0 to 1.1 × 10⁵); the default gives the prior its full weight, and the restoration stays in the input’s range, as vb2D’s does. FSL-1433.85e1018,851 of 20,480 (92 %)1.47e90.5210.997

Δ is A − B, computed in float64 over all 20,480 voxels of the outputs, not only those shown. The replica’s record’s threshold: every file FSL writes bit-identical (images decompressed, descrip left out), with FSL’s exit status and messages. The clean room’s record (fsl-topup) holds it to specs/topup.md §12’s thresholds instead: FSL’s exit status, files and messages (numbers to 4 significant digits), the headers and the log identical; inside the record’s brain the field within 0.5 Hz at the 99th percentile and 0.1 Hz on average (correlation at least 0.999), the corrected images within 1e-3 relative at the 99th percentile and 1e-3 of the mean intensity on average, the movements within 0.05 mm and 0.001 rad, the Jacobians and displacements within the field’s bounds times the read-out time; a metric beyond its threshold passes within FSL’s own band where the record measures it.

Every file applytopup writes, both ports against the original: FSL applytopup 6.0.7.23

FileTypeSizereplica, reproducereplica, fixclean room, reproduceclean room, fix
hifi.nii.gzfloat3232 × 32 × 20identicaldiffers18,753 of 20,480 values differ18,852 of 20,480 values differ

Each image decompressed and compared byte for byte with FSL’s: identical, the header (its descrip, FSL’s build string, aside) and every value; data identical, every value but not the header; differs, values that differ (see the comparisons above); text files byte for byte; – where the run keeps no such file. For the clean room, where the values are not the same, how many of the file’s values differ from FSL’s (its record holds it to specs/topup.md §12’s thresholds; the comparisons above give the sizes).

Provenance of this example
  • Data: OpenNeuro ds005040, sub-002, CC0-1.0
  • Inputs: half_up.nii.gz (sha256 1ae5fd56967afd07…); half_dn.nii.gz (sha256 8e49385dc2270b50…); tu_field.nii.gz (sha256 d055f946cfb0adc9…); tu_fieldcoef.nii.gz (sha256 698a5839514d8d17…); tu_movpar.txt (sha256 9d69af8d9e2fd149…)
  • Runs: figures cache entry fsl-topup/at-vb3d-y, computed 2026-11-03T06:20:43+00:00; 1 of its 3 runs made again since, last on 2026-11-06T19:56:36+00:00, with larmorx b2a352b8b3ba; FSL 6.0.7.23 (fsl-topup 2203.5; ~/fsl-6.0.7.23, never on the PATH); larmorx-nc 0.0.1; every command, its environment, threads and exit status are in its RUN_INFO.json
  • Runs: figures cache entry fsl-topup/original-at-vb3d-y, computed 2026-11-03T14:29:52+00:00; its runs made again since, last on 2026-11-07T02:06:50+00:00, with larmorx 80cc208fa4a3; FSL 6.0.7.23 (fsl-topup 2203.5; ~/fsl-6.0.7.23, never on the PATH); “larmorx 0.0.1 (the clean-room original, crate larmorx-fsl, Apache-2.0)”, as the cache entry records it; every command, its environment, threads and exit status are in its RUN_INFO.json

Every number above is in fsl-topup.json, exported by python -m larmorx_validation evidence fsl-topup.

Random testing

The replica’s random-testing run (2026-11-03), 27 minutes long, drew 14,460 random calls, each run by FSL 6.0.7.23 (through the cache of the original’s runs) and by the replica, every file compared. Every case agreed bit for bit. The record’s refusals, failures and thread counts are left to the record.

Random testing, the replica against the original: FSL topup and applytopup 6.0.7.23 (14,460 random calls, 2026-11-03)

Cases
Random cases, each compared with FSL 6.0.7.2314,460
bit-identical (every file, the messages, the exit status)14,460
differing0
fMRIPrep’s call (sdcflows’ b02b0.cnf, every output)1,440
a configuration file (--config)3,079
FSL’s shipped configurations by name (b02b0, _1, _2, _4, _7T, test_acqp)521
one, two or three levels2,784 / 3,863 / 1,385
--acqp instead of --datain864
--debug, --trace241 / 231
applytopup --method=jac, lsr1,731 / 2,268
applytopup vb2D, vb3D, vb4D161 / 180 / 127
distinct options and settings drawn64
length (minutes; 12 jobs at one thread each)27

A random-testing run of the replica, 27 minutes long (seed topup-trial, larmorx 30e7838d4f0b): EPI-like phantom heads on grids of 16 to 24 voxels per side, each volume distorted along its phase encoding by one off-resonance field (a bump of 25 to 120 Hz and a gradient) with small movements in some; pairs along y or x, three or four volumes; float, 16-bit and 64-bit images; FSL’s four output types; applytopup on FSL’s topup result of the case’s own pair (through the cache of the original’s runs, use mode). Every case in a memory box of 1.5 GiB (peaks below 0.27 GiB); the record’s own refusals, failures and thread counts are left to the record.

The clean room’s first random-testing run (2026-11-03) drew the replica’s cases again, so every run of FSL was served by the cache of the original’s runs: 14,400 random calls in 11 minutes, each compared as the record compares. 7,401 agreed within the thresholds. Each of the 6,999 that differed is in one category below:

  • most are causes already named: FSL’s movement derivatives in cases that estimate the movements (4,462; registered since as FSL-201), FSL-133’s lines (484), FSL-143’s divergence (51), and other paths of fits the data hardly determine (of the differing cases that print their costs, about four in five end within 0.1 % of FSL’s final cost or below it);
  • two defects of FSL, registered and corrected by default with reproduce switches: FSL-145 (with --splineorder=2, an axis shorter than two knot spacings gets two coefficients, one when it is shorter than one spacing, which force the field towards 0 at its far end) and FSL-146 (with volumes encoded along different axes, such as AP, PA, LR and RL, FSL’s gradient is not its cost’s, so its minimisers step in a wrong direction; 895 cases);
  • the clean room’s own mistakes (the quadratic B-splines’ count, the scaled conjugate gradients’ weight) and the suite’s (385 cases), each fixed and kept as a permanent case of the record;
  • two categories stayed open, with their reproducers: the quadratic splines’ regulariser Hessian at an even knot spacing (47 cases) and least-squares restorations on lines whose length divides 232 (8 cases). The specification answers both since 2026-11-03 (specs/topup.md §4.6 and §7.4), for the clean room’s next part.

Random testing, the clean room against the original: FSL topup and applytopup 6.0.7.23 (14,400 random calls, 2026-11-03)

Cases
Random cases, each compared with FSL 6.0.7.2314,400
within the record’s thresholds (§12), every exact check passed7,401
differing (each one triaged below)6,999
fMRIPrep’s call (sdcflows’ b02b0.cnf, every output)1,433
a configuration file (--config)3,066
one, two or three levels2,772 / 3,847 / 1,379
--debug, --trace241 / 230
applytopup --method=jac, lsr1,723 / 2,259
applytopup vb2D, vb3D, vb4D161 / 179 / 127
distinct options and settings drawn64
length (minutes; 12 jobs at one thread each)10.9

A random-testing run of the clean room, 11 minutes long (seed topup-trial: the cases of the replica’s run drawn again, so every FSL run was served by the cache of the original’s runs): the same EPI-like phantom pairs, applytopup calls and options as the replica’s run above, each compared as the record compares (specs/topup.md §12’s thresholds, the exact checks, FSL’s band of a last-bit change of its read-out times where a metric is beyond its threshold). The build: larmorx 4117c2b657d6 with the changes committed as f1288d8 (the run started before the commit). Every case in a memory box of 1.5 GiB (peaks below 0.67 GiB).

The clean room’s differences from the original in the run of 2026-11-03, by category

CategoryCasesReproducer
explained: FSL’s movement derivatives are not its cost’s (findings 14; the record’s expected divergences), in cases that estimate the movements; registered since 2026-11-03 as FSL-201 (rule A), the exact derivatives the default of both ports, the clean room’s reproduce to follow FSL’s4,462topup-trial-00006-original
a defect of FSL, registered: FSL-146, volumes encoded along different axes (rule A: corrected by default; reproduce now gives FSL’s path), with movements also findings 14895topup-trial-10280-original
explained: FSL-133, applytopup’s lines whose length does not divide 2^32 (FSL’s wrap, not reproduced; the specification states it since 2026-11-03)484topup-trial-00005-original
a LARMORX mistake, fixed: the quadratic B-splines’ count (a permanent case; FSL-145 found with it)429topup-trial-11095-original
the suite’s own mistake, fixed: an acquisition file of four rows276topup-trial-00027-original
in part a LARMORX convention, fixed (the scaled conjugate gradients’ weight, a permanent case); the rest explained: other paths of fits the data hardly determine, some just past a threshold216topup-trial-12118-original
the suite’s own mistake, fixed: outputs named by a configuration file62topup-trial-00012-original
explained: FSL-143, vb3D’s and vb4D’s indefinite prior (reproduced: its divergence amplifies last bits)51topup-trial-00275-original
open: the quadratic splines’ regulariser Hessian at an even knot spacing (the specification’s answer of 2026-11-03, §4.6, for the clean room’s next part)47topup-trial-10753-original
the suite’s own mistake, fixed: --trace with -v47topup-trial-01142-original
explained: --debug’s attempt files (how many Levenberg-Marquardt attempts a level makes: the minimisation’s path)17topup-trial-02150-original
open: applytopup --method=lsr on lines whose length divides 2^32 (99th percentile 1.4e-3 to 0.13; the specification’s answer of 2026-11-03, §7.4, for the clean room’s next part)8topup-trial-08089-original
explained: integer outputs (values a rounding apart land on neighbouring integers)4topup-trial-07321-original
explained: vb2D’s transients (as the record’s at/vb2d-aniso)1topup-trial-13346-original

Every differing case of the run in one category by its recipe and its first failed check (validation/investigations/cleanroom-topup/triage_counts.py, its output validation/investigations/cleanroom-topup/trial_triage.json), each category’s cause established by replaying samples with each fix and by probes (docs/findings/fsl-topup-clean-room.md, findings 29 to 36); the mistakes of LARMORX and of the suite are fixed, with permanent cases in the record; the reproducer is a finding of the run (difftest --replay).

The clean-room catch-up’s random-testing run (2026-11-05) drew the earlier run’s cases again and further, on the clean room with the specification’s answers: 17,535 random calls in 30 minutes. 13,300 agreed within the thresholds, among them 3,695 of the earlier run’s 6,999 differences. Each of the 4,223 that differed is in one category below:

  • three in four estimate the movements and lie within 10 times FSL’s own band of a last-bit change of its read-out times: FSL-201’s path, which reproduce follows to its first gradients but not along a path FSL’s own last bits move as far (an expected divergence, as in the record);
  • the run found FSL-242, and that FSL counts every iteration of its scaled conjugate gradients toward --miter, both fixed in the clean room after it, each with a permanent case of the record; and the suite’s own mistake with a NIfTI-2 quaternion (fixed);
  • left open, with their reproducers: the movements beyond 10 times FSL’s band (179 of them run away, most with scaled conjugate gradients, which FSL warns against: measured with --debug, FSL-201’s forward differences divided by those gradients’ small step), fits without movements that the data hardly determine, and the quadratic splines’ bending energy on FSL-145’s layout, a question for the specification.

Random testing after the catch-up, the clean room against the original: FSL topup and applytopup 6.0.7.23 (17,535 random calls, 2026-11-05)

Cases
Random cases, each compared with FSL 6.0.7.2317,535
within the record’s thresholds (§12), every exact check passed13,300
differing (each one in a category below)4,223
a register entry named in one mode only (a notice)12
fMRIPrep’s call (sdcflows’ b02b0.cnf, every output)1,738
a configuration file (--config)3,717
one, two or three levels3,378 / 4,672 / 1,686
--debug, --trace294 / 284
applytopup --method=jac, lsr2,111 / 2,731
applytopup vb2D, vb3D, vb4D200 / 214 / 158
distinct options and settings drawn64
length (minutes; 14 jobs at one thread each)30.2

A random-testing run of the clean room, 30 minutes long, after the specification’s answers of 2026-11-03 (seed topup-trial: the earlier runs’ cases drawn again and further, FSL’s runs served by the cache of the original’s runs where they had run), each compared as the record compares (specs/topup.md §12’s thresholds, the exact checks, FSL’s band of a last-bit change of its read-out times where a metric is beyond its threshold). The build: larmorx c798a295f9ca; FSL-242 and the scaled conjugate gradients’ iteration count were fixed after it, from its findings. The notices: FSL-145 1, FSL-146 12, FSL-201 11 cases. Every case in a memory box of 1.5 GiB (peaks below 0.47 GiB).

The clean room’s differences from the original in the run of 2026-11-05 (after the catch-up), by category

CategoryCasesVerdictReproducer
FSL-201’s path: the movements estimated, within 10 times FSL’s own band (FSL’s run with the read-out times changed in their last float bit; median 2.2 times). FSL’s forward differences of single-precision rotation matrices make its path depend on its inputs’ last bits; the clean room’s reproduce takes the same differences, and FSL’s runs with 16 and 256 last bits changed spread as far on sampled findings (the record’s expected divergences)3,213Expected divergencetopup-trial-00015-original
The movements estimated, beyond 10 times FSL’s own band: of 14 sampled, 6 within FSL’s spread at 16 and 256 last bits; 179 run away (159 with a scaled-conjugate-gradient level, which FSL warns against): measured with --debug, FSL-201’s forward differences part by 1.5 % at the second gradient, which the scaled conjugate gradients divide by their small step sigma608Open differencetopup-trial-00035-original
FSL-143 reproduced: applytopup vb3D’s and vb4D’s indefinite prior makes the iterations diverge, which amplifies the moved images’ last bits (the record’s expected divergences)236Expected divergencetopup-trial-00039-original
Without movements, within 10 times FSL’s own band, most just past a threshold (0.00102 for 0.001): fits the data hardly determine, whose path FSL’s own last bits move as far88Open differencetopup-trial-00063-original
--debug’s attempt files: how many Levenberg-Marquardt attempts a level makes follows the minimisation’s path, so the files written differ25Open differencetopup-trial-02087-original
Cubic splines without movements, beyond 10 times FSL’s own band: of 10 sampled, 6 within FSL’s spread at 16 and 256 last bits; two (FSL-242 and the iteration count) fixed during the run’s triage; open: topup-trial-07557 and 1061924Open differencetopup-trial-00535-original
Quadratic splines without movements on an axis shorter than two knot spacings at some level (FSL-145’s layout), deterministic (FSL’s band 10⁻⁷ to 10⁻⁵), which differ far more often than the other quadratic cases (trial_quadratic.py): the bending energy on that layout, a question for the specification (specs/topup.md §4.6)21Open differencetopup-trial-01082-original
Quadratic splines without movements, no axis that short, deterministic too: the same question7Open differencetopup-trial-01765-original
The suite’s own mistake, fixed with a test: a NIfTI-2 quaternion a float rounding above 1, refused by nibabel’s double tolerance (an exception in the comparison); it agrees on replay with the final build1fixed: agrees on replay with the final buildtopup-trial-04796-original

Every differing case of the run in one category by its recipe and how far its failed metrics lie from FSL’s own one-bit band (validation/investigations/cleanroom-topup/trial_categories.py, its output validation/investigations/cleanroom-topup/trial_catchup_triage.json), each category’s cause established by probes and samples (trial_spread.py, trial_fsl_verbose.py, trial_quadratic.py; docs/findings/fsl-topup-clean-room.md, findings 45 to 47); the triage’s entries: validation/difftest/triage.toml. The reproducer is a finding of the run (difftest --replay).

Every case

Each case of the two records, the replica’s and the clean room’s, both against FSL 6.0.7.23 (a column each), with what it tests, its command line, each port’s verdict and the reason. The ten cases random testing gave the clean room are its own. Search, filter by category or verdict (or select a segment of a chart above), sort by a column; #case=<name> links to one case.

187 cases

CaseCategoryWhat it testsreplicaclean roomChecks
fmriprep/ds001600-sub-1-noreconall worked examplefMRIPrep's callsfMRIPrep's field-map estimation call (ds001600-sub-1; sdcflows' b02b0.cnf)
command linetopup --config=<sdcflows' b02b0.cnf> --datain=<nipype's encoding file> --imain=<fMRIPrep's merged pair> --out=sub-1_dir-PA_epi_average_regrid001_merged_sliced_volreg_base --iout=sub-1_dir-PA_epi_average_regrid001_merged_sliced_volreg_corrected.nii.gz --fout=sub-1_dir-PA_epi_average_regrid001_merged_sliced_volreg_field.nii.gz --jacout=jac --logout=sub-1_dir-PA_epi_average_regrid001_merged_sliced_volreg_topup.log --rbmout=xfm --dfout=warpfield
Bit-identical

every file FSL writes (15/15) byte for byte, with FSL’s exit status and messages; fMRIPrep’s own files (its image’s topup): 9/9 the same

Within the original's own variability

In every metric of the record (specs/topup.md §12, inside the record’s brain) the clean room is no further from FSL at one thread than FSL at another thread count is (its own thread band, FSL-116): sub-1_dir-PA_epi_average_regrid001_merged_sliced_volreg_corrected: relative difference inside the brain, 99th percentile 0.00108 (§12: 0.001; FSL’s own 0.00155); sub-1_dir-PA_epi_average_regrid001_merged_sliced_volreg_corrected.nii.gz_mean: relative difference inside the brain, 99th percentile 0.000608 (§12: 0.001; FSL’s own 0.000903); sub-1_dir-PA_epi_average_regrid001_merged_sliced_volreg_corrected: mean |d| / mean intensity inside the brain 0.000126 (§12: 0.001; FSL’s own 0.000175).

7/7 · 38/38
fmriprep/ds005040-sub-002-noreconall worked examplefMRIPrep's callsfMRIPrep's field-map estimation call (ds005040-sub-002; sdcflows' b02b0.cnf)
command linetopup --config=<sdcflows' b02b0.cnf> --datain=<nipype's encoding file> --imain=<fMRIPrep's merged pair> --out=sub-002_dir-A_epi_average_merged_sliced_volreg_base --iout=sub-002_dir-A_epi_average_merged_sliced_volreg_corrected.nii.gz --fout=sub-002_dir-A_epi_average_merged_sliced_volreg_field.nii.gz --jacout=jac --logout=sub-002_dir-A_epi_average_merged_sliced_volreg_topup.log --rbmout=xfm --dfout=warpfield
Bit-identical

every file FSL writes (15/15) byte for byte, with FSL’s exit status and messages; fMRIPrep’s own files (its image’s topup): 9/9 the same

Open difference

Within the record’s thresholds (specs/topup.md §12), but further from FSL than FSL’s own thread band in 2 metric(s): _fieldcoef (its field, Hz): correlation inside the brain 0.999958 (§12: 0.999; FSL’s own 0.999966); --fout (Hz): correlation inside the brain 0.999958 (§12: 0.999; FSL’s own 0.999966).

7/7 · 38/38
debug/one-level--debug, --trace and --profileone level (subsampling 2, 3 Levenberg-Marquardt iterations) with --debug=2 and -v: every cost-function evaluation's debug images, gradients and the Hessian's weights
command linetopup --imain=<fMRIPrep's merged pair (ds005040)> --datain=<its encoding file> --out=dbg --warpres=14 --subsamp=2 --fwhm=6 --miter=3 --lambda=0.001 --ssqlambda=1 --regmod=bending_energy --estmov=0 --minmet=0 --splineorder=3 --numprec=double --interp=spline --scale=1 --debug=2 --iout=dbg_corr --fout=dbg_field -v
Bit-identical

every file FSL writes (42/42) byte for byte, with FSL’s exit status and messages

Open difference

Within the record’s thresholds (specs/topup.md §12): dbg_corr: relative difference inside the brain, 99th percentile 5.8e-06 (§12: 0.001); dbg_corr_mean: relative difference inside the brain, 99th percentile 5.14e-06 (§12: 0.001); dbg_corr: mean |d| / mean intensity inside the brain 3.59e-07 (§12: 0.001). 4 of 6 files differ in their values (most: dbg_corr.nii.gz: 280,371 of 327,680 values, at most 0.559): the clean room computes in its own order of operations (written from the specification, not from FSL’s code). FSL’s own band is not measured on this case (its thread band is, on fMRIPrep’s calls).

6/6 · 25/25
debug/movements--debug, --trace and --profileone level with the movements estimated (five parameters of the reversed scan: the translation along the phase encoding left out) by Levenberg-Marquardt, --debug=1 and -v; the rigid-body matrices
command linetopup --imain=<fMRIPrep's merged pair (ds005040)> --datain=<its encoding file> --out=mov --warpres=14 --subsamp=2 --fwhm=6 --miter=3 --lambda=0.001 --estmov=1 --minmet=0 --debug=1 --iout=mov_corr --rbmout=mov_xfm -v
Bit-identical

every file FSL writes (29/29) byte for byte, with FSL’s exit status and messages; in fix, its default, the replica corrects FSL-201 and names it: 779,994 values change

Open difference

Within the record’s thresholds (specs/topup.md §12): mov_corr: relative difference inside the brain, 99th percentile 0.000458 (§12: 0.001); mov_corr_mean: relative difference inside the brain, 99th percentile 0.000274 (§12: 0.001); _fieldcoef (its field, Hz): |d| inside the brain, mean 0.00337 (§12: 0.1). 5 of 7 files differ in their values (most: mov_corr.nii.gz: 301,203 of 327,680 values, at most 36.6): the clean room computes in its own order of operations (written from the specification, not from FSL’s code). FSL’s own band is not measured on this case (its thread band is, on fMRIPrep’s calls); in fix, its default, the clean room corrects FSL-201 and names it: 772,643…

6/6 · 25/25
debug/hessian--debug, --trace and --profile--debug=3 on the half-resolution pair: one iteration's every debug file, the Hessian's text included
command linetopup --imain=<half.nii.gz: derived from ds005040's fMRIPrep pair> --datain=<acq.txt: acquisition parameters written by the suite> --out=o --warpres=20 --miter=1 --estmov=0 --debug=3
Bit-identical

every file FSL writes (33/33) byte for byte, with FSL’s exit status and messages

Open difference

Within the record’s thresholds (specs/topup.md §12): _fieldcoef (its field, Hz): |d| inside the brain, 99th percentile 4.16e-06 (§12: 0.5); _fieldcoef (its field, Hz): |d| inside the brain, mean 5.28e-07 (§12: 0.1); _fieldcoef (its field, Hz): correlation inside the brain 1 (§12: 0.999). 1 of 3 files differ in their values (most: o_fieldcoef.nii.gz: 1,455 of 1,521 values, at most 0.0142): the clean room computes in its own order of operations (written from the specification, not from FSL’s code). FSL’s own band is not measured on this case (its thread band is, on fMRIPrep’s calls).

6/6 · 12/12
debug/trace--debug, --trace and --profile--trace: every routine's entry and exit printed, two levels
command linetopup --imain=<half.nii.gz: derived from ds005040's fMRIPrep pair> --datain=<acq.txt: acquisition parameters written by the suite> --out=o --warpres=20,14 --subsamp=2,1 --fwhm=8,4 --miter=3,2 --lambda=0.005,0.0001 --estmov=1,0 --minmet=0,1 --trace
Bit-identical

every file FSL writes (5/5) byte for byte, with FSL’s exit status and messages; in fix, its default, the replica corrects FSL-201 and names it: 4,698 values change

Open difference

Within the record’s thresholds (specs/topup.md §12): o_movpar.txt: rotations 0.000128 (§12: 0.001); _fieldcoef (its field, Hz): |d| inside the brain, mean 0.0067 (§12: 0.1); _fieldcoef (its field, Hz): |d| inside the brain, 99th percentile 0.0233 (§12: 0.5). 2 of 3 files differ in their values (most: o_fieldcoef.nii.gz: 4,693 of 4,693 values, at most 0.046): the clean room computes in its own order of operations (written from the specification, not from FSL’s code). FSL’s own band is not measured on this case (its thread band is, on fMRIPrep’s calls); in fix, its default, the clean room corrects FSL-201 and names it: 4,693 values change.

6/6 · 13/13
debug/profile--debug, --trace and --profile--profile: the levels' durations in <out>.topup_profile_topup.cpp_main.txt (the clock's numbers masked)
command linetopup --imain=<half.nii.gz: derived from ds005040's fMRIPrep pair> --datain=<acq.txt: acquisition parameters written by the suite> --out=o --warpres=20,14 --subsamp=2,1 --fwhm=8,4 --miter=3,2 --lambda=0.005,0.0001 --estmov=1,0 --minmet=0,1 --profile
Bit-identical

every file FSL writes (6/6) byte for byte, with FSL’s exit status and messages; its default, fix, names FSL-201, with the same values; the profile’s durations (the clock’s) masked before comparing

Open difference

Within the record’s thresholds (specs/topup.md §12): o_movpar.txt: rotations 0.000128 (§12: 0.001); _fieldcoef (its field, Hz): |d| inside the brain, mean 0.0067 (§12: 0.1); _fieldcoef (its field, Hz): |d| inside the brain, 99th percentile 0.0233 (§12: 0.5). 2 of 3 files differ in their values (most: o_fieldcoef.nii.gz: 4,693 of 4,693 values, at most 0.046): the clean room computes in its own order of operations (written from the specification, not from FSL’s code). FSL’s own band is not measured on this case (its thread band is, on fMRIPrep’s calls).

7/7 · 11/11
opt/defaultsoptionsevery default (warpres 10, fwhm 8, five LM iterations, movements)
command linetopup --imain=<half.nii.gz: derived from ds005040's fMRIPrep pair> --datain=<acq.txt: acquisition parameters written by the suite> --out=o
Bit-identical

every file FSL writes (5/5) byte for byte, with FSL’s exit status and messages; its default, fix, names FSL-201, with the same values

Bit-identical

The same values, every one, in all 3 files FSL writes on this case, with FSL’s exit status and messages (the record checks the headers and the log); its default, fix, names FSL-201, with the same values.

6/6 · 12/12
opt/levels worked exampleoptionstwo levels: subsampling 2 with Levenberg-Marquardt and the movements, then full resolution with scaled conjugate gradients; every output
command linetopup --imain=<half.nii.gz: derived from ds005040's fMRIPrep pair> --datain=<acq.txt: acquisition parameters written by the suite> --out=o --warpres=20,14 --subsamp=2,1 --fwhm=8,4 --miter=3,2 --lambda=0.005,0.0001 --estmov=1,0 --minmet=0,1 --iout=c --fout=f --dfout=d --rbmout=r --jacout=j --featout=feat -v
Bit-identical

every file FSL writes (16/16) byte for byte, with FSL’s exit status and messages; in fix, its default, the replica corrects FSL-201 and names it: 202,926 values change

Expected divergence

FSL-201 reproduced, the movements estimated and not converged: FSL’s rotation derivatives are forward differences of single-precision matrices, so each gradient entry moves by about 2 % of its lever-weighted sum when its angle’s rounding to single precision f…

6/6 · 46/47
opt/numprec-floatoptions--numprec=float
command linetopup --imain=<half.nii.gz: derived from ds005040's fMRIPrep pair> --datain=<acq.txt: acquisition parameters written by the suite> --out=o --warpres=20 --fwhm=8 --miter=3 --lambda=0.005 --numprec=float --fout=f
Bit-identical

every file FSL writes (6/6) byte for byte, with FSL’s exit status and messages; in fix, its default, the replica corrects FSL-201 and names it: 22,005 values change

Open difference

Within the record’s thresholds (specs/topup.md §12): o_movpar.txt: rotations 3.02e-05 (§12: 0.001); _fieldcoef (its field, Hz): |d| inside the brain, mean 0.00288 (§12: 0.1); --fout (Hz): |d| inside the brain, mean 0.00288 (§12: 0.1). 3 of 4 files differ in their values (most: f.nii.gz: 20,477 of 20,480 values, at most 0.0347): the clean room computes in its own order of operations (written from the specification, not from FSL’s code). FSL’s own band is not measured on this case (its thread band is, on fMRIPrep’s calls); in fix, its default, the clean room corrects FSL-201 and names it: 21,998 values change.

6/6 · 16/16
opt/interp-linearoptions--interp=linear
command linetopup --imain=<half.nii.gz: derived from ds005040's fMRIPrep pair> --datain=<acq.txt: acquisition parameters written by the suite> --out=o --warpres=20 --fwhm=8 --miter=3 --lambda=0.005 --interp=linear --iout=c
Bit-identical

every file FSL writes (7/7) byte for byte, with FSL’s exit status and messages; in fix, its default, the replica corrects FSL-201 and names it: 56,519 values change

Expected divergence

FSL-201 reproduced, the movements estimated and not converged: FSL’s rotation derivatives are forward differences of single-precision matrices, so each gradient entry moves by about 2 % of its lever-weighted sum when its angle’s rounding to single precision f…

6/6 · 17/18
opt/membrane-lm worked exampleoptions--regmod=membrane_energy with Levenberg-Marquardt (the energy's Hessian)
command linetopup --imain=<half.nii.gz: derived from ds005040's fMRIPrep pair> --datain=<acq.txt: acquisition parameters written by the suite> --out=o --warpres=20 --fwhm=8 --miter=3 --lambda=0.005 --regmod=membrane_energy --fout=f
Bit-identical

every file FSL writes (6/6) byte for byte, with FSL’s exit status and messages; in fix, its default, the replica corrects FSL-117, FSL-201 and names it: 22,006 values change

Open difference

Within the record’s thresholds (specs/topup.md §12): o_movpar.txt: rotations 1.48e-05 (§12: 0.001); o_movpar.txt: translations 1.04e-05 (§12: 0.05); --fout (Hz): |d| inside the brain, mean 8.05e-07 (§12: 0.1). 3 of 4 files differ in their values (most: f.nii.gz: 20,468 of 20,480 values, at most 3.18e-06): the clean room computes in its own order of operations (written from the specification, not from FSL’s code). FSL’s own band is not measured on this case (its thread band is, on fMRIPrep’s calls); in fix, its default, the clean room corrects FSL-117, FSL-201 and names it: 22,001 values change.

6/6 · 16/16
opt/membrane-scgoptions--regmod=membrane_energy with scaled conjugate gradients (the energy's gradient)
command linetopup --imain=<half.nii.gz: derived from ds005040's fMRIPrep pair> --datain=<acq.txt: acquisition parameters written by the suite> --out=o --warpres=20 --fwhm=8 --miter=3 --lambda=0.005 --regmod=membrane_energy --minmet=1 --estmov=0 --fout=f
Bit-identical

every file FSL writes (6/6) byte for byte, with FSL’s exit status and messages

Open difference

Within the record’s thresholds (specs/topup.md §12): _fieldcoef (its field, Hz): |d| inside the brain, mean 7.2e-09 (§12: 0.1); --fout (Hz): |d| inside the brain, mean 7.18e-09 (§12: 0.1); --fout (Hz): |d| inside the brain, 99th percentile 2.84e-08 (§12: 0.5). 2 of 4 files differ in their values (most: f.nii.gz: 20,117 of 20,480 values, at most 3.68e-08): the clean room computes in its own order of operations (written from the specification, not from FSL’s code). FSL’s own band is not measured on this case (its thread band is, on fMRIPrep’s calls).

6/6 · 16/16
opt/splineorder-2options--splineorder=2
command linetopup --imain=<half.nii.gz: derived from ds005040's fMRIPrep pair> --datain=<acq.txt: acquisition parameters written by the suite> --out=o --warpres=20 --fwhm=8 --miter=3 --lambda=0.005 --splineorder=2 --fout=f
Bit-identical

every file FSL writes (6/6) byte for byte, with FSL’s exit status and messages; in fix, its default, the replica corrects FSL-201 and names it: 22,006 values change

Open difference

Within the record’s thresholds (specs/topup.md §12): o_movpar.txt: rotations 4.84e-05 (§12: 0.001); _fieldcoef (its field, Hz): |d| inside the brain, mean 0.00177 (§12: 0.1); --fout (Hz): |d| inside the brain, mean 0.00177 (§12: 0.1). 3 of 4 files differ in their values (most: f.nii.gz: 20,477 of 20,480 values, at most 0.00756): the clean room computes in its own order of operations (written from the specification, not from FSL’s code). FSL’s own band is not measured on this case (its thread band is, on fMRIPrep’s calls); in fix, its default, the clean room corrects FSL-201 and names it: 21,998 values change.

6/6 · 16/16
opt/fsl145-short-axisoptions--splineorder=2 with z shorter than two knot spacings (FSL-145: 2 coefficients)
command linetopup --imain=<half.nii.gz: derived from ds005040's fMRIPrep pair> --datain=<acq.txt: acquisition parameters written by the suite> --out=o --warpres=80 --fwhm=8 --miter=3 --lambda=0.005 --estmov=0 --splineorder=2 --fout=f
Bit-identical

every file FSL writes (6/6) byte for byte, with FSL’s exit status and messages; in fix, its default, the replica corrects FSL-145 and names it: 20,580 values change

Open difference

Within the record’s thresholds (specs/topup.md §12): _fieldcoef (its field, Hz): |d| inside the brain, 99th percentile 4.68e-06 (§12: 0.5); --fout (Hz): |d| inside the brain, 99th percentile 4.53e-06 (§12: 0.5); --fout (Hz): |d| inside the brain, mean 8.83e-07 (§12: 0.1). 2 of 4 files differ in their values (most: f.nii.gz: 16,131 of 20,480 values, at most 6.68e-06): the clean room computes in its own order of operations (written from the specification, not from FSL’s code). FSL’s own band is not measured on this case (its thread band is, on fMRIPrep’s calls); in fix, its default, the clean room corrects FSL-145 and names it: 20,580 values change.

6/6 · 16/16
opt/fsl145-one-coefficientoptions--splineorder=2 at --subsamp=2 with z shorter than one knot spacing (FSL-145: one coefficient, whose refit stops FSL; the default runs)
command linetopup --imain=<half.nii.gz: derived from ds005040's fMRIPrep pair> --datain=<acq.txt: acquisition parameters written by the suite> --out=o --warpres=140 --subsamp=2 --fwhm=8 --miter=1 --lambda=0.005 --estmov=0 --splineorder=2 --fout=f
Rejected by both

topup exits 1; the same from the replica. Error occurred when preparing to run topup Exception thrown with message: Mat::operator(): index out of bounds The replica’s default, fix, finishes instead (exit status 0), naming FSL-145.

Rejected by both

–splineorder=2 at --subsamp=2 with z shorter than one knot spacing, where FSL-145 gives one coefficient whose refit stops FSL; the clean room’s reproduce stops alike since 2026-11-05 (outside the record before). topup exits 1; the same from the clean room. Error occurred when preparing to run topup Exception thrown with message: Mat::operator(): index out of bounds The clean room’s default, fix, finishes instead (exit status 0), naming FSL-145.

6/6
opt/scaleoptions--scale=1: each volume scaled to the mean
command linetopup --imain=<half.nii.gz: derived from ds005040's fMRIPrep pair> --datain=<acq.txt: acquisition parameters written by the suite> --out=o --warpres=20 --fwhm=8 --miter=3 --lambda=0.005 --scale=1 --iout=c
Bit-identical

every file FSL writes (7/7) byte for byte, with FSL’s exit status and messages; in fix, its default, the replica corrects FSL-201 and names it: 56,423 values change

Open difference

Within the record’s thresholds (specs/topup.md §12): c: relative difference inside the brain, 99th percentile 0.000469 (§12: 0.001); c_mean: relative difference inside the brain, 99th percentile 0.00028 (§12: 0.001); o_movpar.txt: rotations 6.47e-05 (§12: 0.001). 4 of 5 files differ in their values (most: c.nii.gz: 34,473 of 40,960 values, at most 21.5): the clean room computes in its own order of operations (written from the specification, not from FSL’s code). FSL’s own band is not measured on this case (its thread band is, on fMRIPrep’s calls); in fix, its default, the clean room corrects FSL-201 and names it: 56,388 values change.

6/6 · 18/18
opt/ssqlambda-0options--ssqlambda=0: lambda as given
command linetopup --imain=<half.nii.gz: derived from ds005040's fMRIPrep pair> --datain=<acq.txt: acquisition parameters written by the suite> --out=o --warpres=20 --miter=3 --lambda=50 --ssqlambda=0 --fout=f
Bit-identical

every file FSL writes (6/6) byte for byte, with FSL’s exit status and messages; in fix, its default, the replica corrects FSL-201 and names it: 22,006 values change

Open difference

Within the record’s thresholds (specs/topup.md §12): o_movpar.txt: rotations 0.000102 (§12: 0.001); o_movpar.txt: translations 0.000348 (§12: 0.05); --fout (Hz): |d| inside the brain, mean 0.000634 (§12: 0.1). 3 of 4 files differ in their values (most: f.nii.gz: 20,477 of 20,480 values, at most 0.00304): the clean room computes in its own order of operations (written from the specification, not from FSL’s code). FSL’s own band is not measured on this case (its thread band is, on fMRIPrep’s calls); in fix, its default, the clean room corrects FSL-201 and names it: 22,001 values change.

6/6 · 16/16
opt/regrid-0options--regrid=0: the image's own grid, two levels
command linetopup --imain=<half.nii.gz: derived from ds005040's fMRIPrep pair> --datain=<acq.txt: acquisition parameters written by the suite> --out=o --warpres=20,14 --subsamp=2,1 --fwhm=8,4 --miter=3,2 --lambda=0.005,0.0001 --estmov=1,0 --minmet=0,1 --regrid=0 --fout=f
Bit-identical

every file FSL writes (6/6) byte for byte, with FSL’s exit status and messages; in fix, its default, the replica corrects FSL-201 and names it: 25,176 values change

Open difference

Within the record’s thresholds (specs/topup.md §12): o_movpar.txt: rotations 0.000134 (§12: 0.001); _fieldcoef (its field, Hz): |d| inside the brain, mean 0.00604 (§12: 0.1); --fout (Hz): |d| inside the brain, mean 0.00604 (§12: 0.1). 3 of 4 files differ in their values (most: f.nii.gz: 20,479 of 20,480 values, at most 0.0327): the clean room computes in its own order of operations (written from the specification, not from FSL’s code). FSL’s own band is not measured on this case (its thread band is, on fMRIPrep’s calls); in fix, its default, the clean room corrects FSL-201 and names it: 25,172 values change.

6/6 · 16/16
opt/scgoptions--minmet=1 without movements
command linetopup --imain=<half.nii.gz: derived from ds005040's fMRIPrep pair> --datain=<acq.txt: acquisition parameters written by the suite> --out=o --warpres=20 --fwhm=8 --miter=3 --lambda=0.005 --minmet=1 --estmov=0
Bit-identical

every file FSL writes (5/5) byte for byte, with FSL’s exit status and messages

Open difference

Within the record’s thresholds (specs/topup.md §12): _fieldcoef (its field, Hz): |d| inside the brain, mean 0.000223 (§12: 0.1); _fieldcoef (its field, Hz): |d| inside the brain, 99th percentile 0.000689 (§12: 0.5); _fieldcoef (its field, Hz): correlation inside the brain 1 (§12: 0.999). 1 of 3 files differ in their values (most: o_fieldcoef.nii.gz: 1,521 of 1,521 values, at most 0.0013): the clean room computes in its own order of operations (written from the specification, not from FSL’s code). FSL’s own band is not measured on this case (its thread band is, on fMRIPrep’s calls).

6/6 · 12/12
opt/scg-movementsoptions--minmet=1 with the movements: FSL's warning
command linetopup --imain=<half.nii.gz: derived from ds005040's fMRIPrep pair> --datain=<acq.txt: acquisition parameters written by the suite> --out=o --warpres=20 --fwhm=8 --miter=3 --lambda=0.005 --minmet=1 --estmov=1 --rbmout=r
Bit-identical

every file FSL writes (7/7) byte for byte, with FSL’s exit status and messages; in fix, its default, the replica corrects FSL-201 and names it: 1,538 values change

Open difference

Within the record’s thresholds (specs/topup.md §12): r_02.mat: rotations 0.000827 (§12: 0.001); o_movpar.txt: rotations 0.000825 (§12: 0.001); r_02.mat: translations 0.0584 (§12: 0.201). 3 of 5 files differ in their values (most: o_fieldcoef.nii.gz: 1,521 of 1,521 values, at most 7.82e-06): the clean room computes in its own order of operations (written from the specification, not from FSL’s code). FSL’s own band is not measured on this case (its thread band is, on fMRIPrep’s calls); in fix, its default, the clean room corrects FSL-201 and names it: 1,521 values change.

6/6 · 16/16
opt/acqpoptions--acqp instead of --datain
command linetopup --imain=<half.nii.gz: derived from ds005040's fMRIPrep pair> --acqp=<acq.txt: acquisition parameters written by the suite> --out=o --warpres=20 --fwhm=8 --miter=3 --lambda=0.005
Bit-identical

every file FSL writes (5/5) byte for byte, with FSL’s exit status and messages; in fix, its default, the replica corrects FSL-201 and names it: 1,526 values change

Open difference

Within the record’s thresholds (specs/topup.md §12): o_movpar.txt: rotations 6.02e-05 (§12: 0.001); _fieldcoef (its field, Hz): |d| inside the brain, mean 0.00214 (§12: 0.1); _fieldcoef (its field, Hz): |d| inside the brain, 99th percentile 0.00658 (§12: 0.5). 2 of 3 files differ in their values (most: o_fieldcoef.nii.gz: 1,521 of 1,521 values, at most 0.0215): the clean room computes in its own order of operations (written from the specification, not from FSL’s code). FSL’s own band is not measured on this case (its thread band is, on fMRIPrep’s calls); in fix, its default, the clean room corrects FSL-201 and names it: 1,521 values change.

6/6 · 12/12
opt/datain-transposedoptionsan acquisition-parameter file of four rows
command linetopup --imain=<half.nii.gz: derived from ds005040's fMRIPrep pair> --datain=<acq_t.txt: acquisition parameters written by the suite> --out=o --warpres=20 --fwhm=8 --miter=3 --lambda=0.005
Bit-identical

every file FSL writes (5/5) byte for byte, with FSL’s exit status and messages; in fix, its default, the replica corrects FSL-201 and names it: 1,526 values change

Open difference

Within the record’s thresholds (specs/topup.md §12): o_movpar.txt: rotations 6.02e-05 (§12: 0.001); _fieldcoef (its field, Hz): |d| inside the brain, mean 0.00214 (§12: 0.1); _fieldcoef (its field, Hz): |d| inside the brain, 99th percentile 0.00658 (§12: 0.5). 2 of 3 files differ in their values (most: o_fieldcoef.nii.gz: 1,521 of 1,521 values, at most 0.0215): the clean room computes in its own order of operations (written from the specification, not from FSL’s code). FSL’s own band is not measured on this case (its thread band is, on fMRIPrep’s calls); in fix, its default, the clean room corrects FSL-201 and names it: 1,521 values change.

6/6 · 12/12
opt/x-phase-encodingoptionsphase encoding along x (the translation along x left out), the movements
command linetopup --imain=<xhalf.nii.gz: derived from ds005040's fMRIPrep pair> --datain=<acq_x.txt: acquisition parameters written by the suite> --out=o --warpres=20 --fwhm=8 --miter=3 --lambda=0.005 --iout=c --rbmout=r --jacout=j --dfout=d
Bit-identical

every file FSL writes (13/13) byte for byte, with FSL’s exit status and messages; in fix, its default, the replica corrects FSL-201 and names it: 138,202 values change

Open difference

Within the record’s thresholds (specs/topup.md §12): c: relative difference inside the brain, 99th percentile 0.00086 (§12: 0.001); c_mean: relative difference inside the brain, 99th percentile 0.000568 (§12: 0.001); r_02.mat: rotations 0.000113 (§12: 0.001). 9 of 11 files differ in their values (most: c.nii.gz: 34,526 of 40,960 values, at most 39.6): the clean room computes in its own order of operations (written from the specification, not from FSL’s code). FSL’s own band is not measured on this case (its thread band is, on fMRIPrep’s calls); in fix, its default, the clean room corrects FSL-201 and names it: 138,124 values change.

6/6 · 34/34
opt/four-volumesoptionsfour volumes along x and y: six movement parameters for each moving scan
command linetopup --imain=<four.nii.gz: derived from ds005040's fMRIPrep pair> --datain=<acq_four.txt: acquisition parameters written by the suite> --out=o --warpres=20 --fwhm=8 --miter=3 --lambda=0.005 --iout=c --rbmout=r --jacout=j --dfout=d
Bit-identical

every file FSL writes (19/19) byte for byte, with FSL’s exit status and messages; in fix, its default, the replica corrects FSL-146, FSL-201 and names it: 249,766 values change

Expected divergence

FSL-201 reproduced, the movements estimated and not converged: FSL’s rotation derivatives are forward differences of single-precision matrices, so each gradient entry moves by about 2 % of its lever-weighted sum when its angle’s rounding to single precision f…

6/6 · 39/50
opt/fsl146-four-axesoptionsfour volumes along y and x without the movements (FSL-146: each axis' Jacobian term of the gradient over the volumes encoded along it only)
command linetopup --imain=<four.nii.gz: derived from ds005040's fMRIPrep pair> --datain=<acq_four.txt: acquisition parameters written by the suite> --out=o --warpres=20 --fwhm=8 --miter=3 --lambda=0.005 --estmov=0 --fout=f
Bit-identical

every file FSL writes (6/6) byte for byte, with FSL’s exit status and messages; in fix, its default, the replica corrects FSL-146 and names it: 22,001 values change

Open difference

Within the record’s thresholds (specs/topup.md §12): _fieldcoef (its field, Hz): |d| inside the brain, mean 2.95e-06 (§12: 0.1); --fout (Hz): |d| inside the brain, mean 2.94e-06 (§12: 0.1); _fieldcoef (its field, Hz): |d| inside the brain, 99th percentile 4.93e-06 (§12: 0.5). 2 of 4 files differ in their values (most: f.nii.gz: 20,451 of 20,480 values, at most 5.25e-06): the clean room computes in its own order of operations (written from the specification, not from FSL’s code). FSL’s own band is not measured on this case (its thread band is, on fMRIPrep’s calls); in fix, its default, the clean room corrects FSL-146 and names it: 22,001 values change.

6/6 · 16/16
opt/oblique-phase-encodingoptionsa phase-encoding vector along x and y (the Jacobian of both displacements)
command linetopup --imain=<half.nii.gz: derived from ds005040's fMRIPrep pair> --datain=<acq_oblique.txt: acquisition parameters written by the suite> --out=o --warpres=20 --fwhm=8 --miter=3 --lambda=0.005 --jacout=j --dfout=d
Bit-identical

every file FSL writes (9/9) byte for byte, with FSL’s exit status and messages; in fix, its default, the replica corrects FSL-201 and names it: 123,477 values change

Open difference

Within the record’s thresholds (specs/topup.md §12): _fieldcoef (its field, Hz): |d| inside the brain, mean 0.00914 (§12: 0.1); d_01 (mm): |d| inside the brain, mean 0.000605 (§12: 0.00828); d_02 (mm): |d| inside the brain, mean 0.000605 (§12: 0.00828). 6 of 7 files differ in their values (most: d_01.nii.gz: 40,959 of 61,440 values, at most 0.00303): the clean room computes in its own order of operations (written from the specification, not from FSL’s code). FSL’s own band is not measured on this case (its thread band is, on fMRIPrep’s calls); in fix, its default, the clean room corrects FSL-201 and names it: 124,217 values change.

6/6 · 24/24
opt/readout-timesoptionsdifferent read-out times
command linetopup --imain=<half.nii.gz: derived from ds005040's fMRIPrep pair> --datain=<acq_rts.txt: acquisition parameters written by the suite> --out=o --warpres=20 --fwhm=8 --miter=3 --lambda=0.005 --fout=f
Bit-identical

every file FSL writes (6/6) byte for byte, with FSL’s exit status and messages; in fix, its default, the replica corrects FSL-201 and names it: 22,004 values change

Open difference

Within the record’s thresholds (specs/topup.md §12): o_movpar.txt: rotations 3.24e-05 (§12: 0.001); _fieldcoef (its field, Hz): |d| inside the brain, mean 0.00174 (§12: 0.1); --fout (Hz): |d| inside the brain, mean 0.00174 (§12: 0.1). 3 of 4 files differ in their values (most: f.nii.gz: 20,477 of 20,480 values, at most 0.00875): the clean room computes in its own order of operations (written from the specification, not from FSL’s code). FSL’s own band is not measured on this case (its thread band is, on fMRIPrep’s calls); in fix, its default, the clean room corrects FSL-201 and names it: 22,000 values change.

6/6 · 16/16
opt/zero-readoutoptionsone read-out time of 0 (a synthesised undistorted image)
command linetopup --imain=<half.nii.gz: derived from ds005040's fMRIPrep pair> --datain=<acq_zero.txt: acquisition parameters written by the suite> --out=o --warpres=20 --fwhm=8 --miter=3 --lambda=0.005 --fout=f --iout=c
Bit-identical

every file FSL writes (8/8) byte for byte, with FSL’s exit status and messages; in fix, its default, the replica corrects FSL-201 and names it: 76,720 values change

Open difference

Within the record’s thresholds (specs/topup.md §12): c: relative difference inside the brain, 99th percentile 0.000177 (§12: 0.001); c_mean: relative difference inside the brain, 99th percentile 0.000121 (§12: 0.001); o_movpar.txt: rotations 1.94e-05 (§12: 0.001). 5 of 6 files differ in their values (most: c.nii.gz: 34,416 of 40,960 values, at most 7.52): the clean room computes in its own order of operations (written from the specification, not from FSL’s code). FSL’s own band is not measured on this case (its thread band is, on fMRIPrep’s calls); in fix, its default, the clean room corrects FSL-201 and names it: 76,901 values change.

6/6 · 22/22
opt/same-directionoptionsboth volumes with the same phase encoding
command linetopup --imain=<half.nii.gz: derived from ds005040's fMRIPrep pair> --datain=<acq_same.txt: acquisition parameters written by the suite> --out=o --warpres=20 --fwhm=8 --miter=3 --lambda=0.005 --fout=f
Bit-identical

every file FSL writes (6/6) byte for byte, with FSL’s exit status and messages; in fix, its default, the replica corrects FSL-201 and names it: 22,007 values change

Open difference

Within the record’s thresholds (specs/topup.md §12): _fieldcoef (its field, Hz): |d| inside the brain, mean 0.0048 (§12: 0.1); --fout (Hz): |d| inside the brain, mean 0.0048 (§12: 0.1); _fieldcoef (its field, Hz): |d| inside the brain, 99th percentile 0.0141 (§12: 0.5). 3 of 4 files differ in their values (most: f.nii.gz: 20,479 of 20,480 values, at most 0.0149): the clean room computes in its own order of operations (written from the specification, not from FSL’s code). FSL’s own band is not measured on this case (its thread band is, on fMRIPrep’s calls); in fix, its default, the clean room corrects FSL-201 and names it: 22,001 values change.

6/6 · 16/16
opt/z-phase-encodingoptionsphase encoding along z
command linetopup --imain=<half.nii.gz: derived from ds005040's fMRIPrep pair> --datain=<acq_z.txt: acquisition parameters written by the suite> --out=o --warpres=20 --fwhm=8 --miter=3 --lambda=0.005
Rejected by both

topup exits 1; the same from the replica. Error occurred when preparing to run topup Exception thrown with message: Topup::TopupScan::TopupScan: third element of pevec must be zero

Rejected by both

topup exits 1; the same from the clean room. Error occurred when preparing to run topup Exception thrown with message: Topup::TopupScan::TopupScan: third element of pevec must be zero The clean room’s default, fix, ends with exit status 1, naming FSL-201.

5/5
opt/int16optionsa 16-bit integer image
command linetopup --imain=<half_i16.nii.gz: derived from ds005040's fMRIPrep pair> --datain=<acq.txt: acquisition parameters written by the suite> --out=o --warpres=20 --fwhm=8 --miter=3 --lambda=0.005 --iout=c
Bit-identical

every file FSL writes (7/7) byte for byte, with FSL’s exit status and messages; in fix, its default, the replica corrects FSL-201 and names it: 56,470 values change

Expected divergence

FSL-201 reproduced, the movements estimated and not converged: FSL’s rotation derivatives are forward differences of single-precision matrices, so each gradient entry moves by about 2 % of its lever-weighted sum when its angle’s rounding to single precision f…

6/6 · 17/18
opt/one-volumeoptionsa single 3D volume
command linetopup --imain=<one.nii.gz: derived from ds005040's fMRIPrep pair> --datain=<acq_one.txt: acquisition parameters written by the suite> --out=o --warpres=20 --fwhm=8 --miter=3 --lambda=0.005
Bit-identical

every file FSL writes (5/5) byte for byte, with FSL’s exit status and messages

Bit-identical

The same values, every one, in all 3 files FSL writes on this case, with FSL’s exit status and messages (the record checks the headers and the log); its default, fix, names FSL-201, with the same values.

6/6 · 12/12
opt/no-outoptionsno --out: the base name from --imain
command linetopup --imain=<half.nii.gz: derived from ds005040's fMRIPrep pair> --datain=<acq.txt: acquisition parameters written by the suite> --warpres=20 --fwhm=8 --miter=3 --lambda=0.005
Bit-identical

every file FSL writes (5/5) byte for byte, with FSL’s exit status and messages; in fix, its default, the replica corrects FSL-201 and names it: 1,526 values change

Open difference

Within the record’s thresholds (specs/topup.md §12): half_movpar.txt: rotations 6.02e-05 (§12: 0.001); _fieldcoef (its field, Hz): |d| inside the brain, mean 0.00214 (§12: 0.1); _fieldcoef (its field, Hz): |d| inside the brain, 99th percentile 0.00658 (§12: 0.5). 2 of 3 files differ in their values (most: half_fieldcoef.nii.gz: 1,521 of 1,521 values, at most 0.0215): the clean room computes in its own order of operations (written from the specification, not from FSL’s code). FSL’s own band is not measured on this case (its thread band is, on fMRIPrep’s calls); in fix, its default, the clean room corrects FSL-201 and names it: 1,521 values change.

6/6 · 12/12
opt/intentoptionsan image with intent parameters: the field's header
command linetopup --imain=<half_intent.nii.gz: derived from ds005040's fMRIPrep pair> --datain=<acq.txt: acquisition parameters written by the suite> --out=o --warpres=20 --fwhm=8 --miter=3 --lambda=0.005 --fout=f --featout=feat
Bit-identical

every file FSL writes (8/8) byte for byte, with FSL’s exit status and messages; in fix, its default, the replica corrects FSL-119, FSL-201 and names it: 62,909 values change

Open difference

Within the record’s thresholds (specs/topup.md §12): feat_fieldmap_mag: relative difference inside the brain, 99th percentile 0.0002 (§12: 0.001); o_movpar.txt: rotations 6.02e-05 (§12: 0.001); _fieldcoef (its field, Hz): |d| inside the brain, mean 0.00214 (§12: 0.1). 5 of 6 files differ in their values (most: feat_fieldmap_rads.nii.gz: 20,480 of 20,480 values, at most 0.054): the clean room computes in its own order of operations (written from the specification, not from FSL’s code). FSL’s own band is not measured on this case (its thread band is, on fMRIPrep’s calls); in fix, its default, the clean room corrects FSL-201, FSL-119 and names it: 62,917 values change.

6/6 · 22/22
opt/niftioptionsFSLOUTPUTTYPE=NIFTI
command lineFSLOUTPUTTYPE=NIFTI topup --imain=<half.nii.gz: derived from ds005040's fMRIPrep pair> --datain=<acq.txt: acquisition parameters written by the suite> --out=o --warpres=20 --fwhm=8 --miter=3 --lambda=0.005 --fout=f
Bit-identical

every file FSL writes (6/6) byte for byte, with FSL’s exit status and messages; in fix, its default, the replica corrects FSL-201 and names it: 22,002 values change

Open difference

Within the record’s thresholds (specs/topup.md §12): o_movpar.txt: rotations 6.02e-05 (§12: 0.001); _fieldcoef (its field, Hz): |d| inside the brain, mean 0.00214 (§12: 0.1); --fout (Hz): |d| inside the brain, mean 0.00214 (§12: 0.1). 3 of 4 files differ in their values (most: f.nii: 20,479 of 20,480 values, at most 0.0086): the clean room computes in its own order of operations (written from the specification, not from FSL’s code). FSL’s own band is not measured on this case (its thread band is, on fMRIPrep’s calls); in fix, its default, the clean room corrects FSL-201 and names it: 21,999 values change.

6/6 · 16/16
opt/nifti2optionsFSLOUTPUTTYPE=NIFTI2_GZ
command lineFSLOUTPUTTYPE=NIFTI2_GZ topup --imain=<half.nii.gz: derived from ds005040's fMRIPrep pair> --datain=<acq.txt: acquisition parameters written by the suite> --out=o --warpres=20 --fwhm=8 --miter=3 --lambda=0.005 --fout=f --iout=c
Bit-identical

every file FSL writes (8/8) byte for byte, with FSL’s exit status and messages; in fix, its default, the replica corrects FSL-201 and names it: 76,879 values change

Open difference

Within the record’s thresholds (specs/topup.md §12): c: relative difference inside the brain, 99th percentile 0.00044 (§12: 0.001); c_mean: relative difference inside the brain, 99th percentile 0.000261 (§12: 0.001); o_movpar.txt: rotations 6.02e-05 (§12: 0.001). 5 of 6 files differ in their values (most: c.nii.gz: 34,465 of 40,960 values, at most 20.2): the clean room computes in its own order of operations (written from the specification, not from FSL’s code). FSL’s own band is not measured on this case (its thread band is, on fMRIPrep’s calls); in fix, its default, the clean room corrects FSL-201 and names it: 76,905 values change.

6/6 · 22/22
opt/nifti-pairoptionsFSLOUTPUTTYPE=NIFTI_PAIR
command lineFSLOUTPUTTYPE=NIFTI_PAIR topup --imain=<half.nii.gz: derived from ds005040's fMRIPrep pair> --datain=<acq.txt: acquisition parameters written by the suite> --out=o --warpres=20 --fwhm=8 --miter=3 --lambda=0.005 --fout=f
Bit-identical

every file FSL writes (8/8) byte for byte, with FSL’s exit status and messages; in fix, its default, the replica corrects FSL-201 and names it: 43,999 values change

Open difference

Within the record’s thresholds (specs/topup.md §12): o_movpar.txt: rotations 6.02e-05 (§12: 0.001); _fieldcoef (its field, Hz): |d| inside the brain, mean 0.00214 (§12: 0.1); --fout (Hz): |d| inside the brain, mean 0.00214 (§12: 0.1). 3 of 4 files differ in their values (most: f.hdr: 20,479 of 20,480 values, at most 0.0086): the clean room computes in its own order of operations (written from the specification, not from FSL’s code). FSL’s own band is not measured on this case (its thread band is, on fMRIPrep’s calls); in fix, its default, the clean room corrects FSL-201 and names it: 43,998 values change.

6/6 · 24/24
opt/iout-extensionoptions--iout with .nii.gz: the mean as <name>.nii.gz_mean
command linetopup --imain=<half.nii.gz: derived from ds005040's fMRIPrep pair> --datain=<acq.txt: acquisition parameters written by the suite> --out=o --warpres=20 --fwhm=8 --miter=3 --lambda=0.005 --iout=c.nii.gz
Bit-identical

every file FSL writes (7/7) byte for byte, with FSL’s exit status and messages; in fix, its default, the replica corrects FSL-201 and names it: 56,403 values change

Open difference

Within the record’s thresholds (specs/topup.md §12): c: relative difference inside the brain, 99th percentile 0.00044 (§12: 0.001); c.nii.gz_mean: relative difference inside the brain, 99th percentile 0.000261 (§12: 0.001); o_movpar.txt: rotations 6.02e-05 (§12: 0.001). 4 of 5 files differ in their values (most: c.nii.gz: 34,465 of 40,960 values, at most 20.2): the clean room computes in its own order of operations (written from the specification, not from FSL’s code). FSL’s own band is not measured on this case (its thread band is, on fMRIPrep’s calls); in fix, its default, the clean room corrects FSL-201 and names it: 56,427 values change.

6/6 · 18/18
opt/nthr-0options--nthr=0: the default thread count (FSL_NUM_THREADS unset: one)
command linetopup --imain=<half.nii.gz: derived from ds005040's fMRIPrep pair> --datain=<acq.txt: acquisition parameters written by the suite> --out=o --warpres=20 --fwhm=8 --miter=3 --lambda=0.005 --nthr=0
Bit-identical

every file FSL writes (5/5) byte for byte, with FSL’s exit status and messages; in fix, its default, the replica corrects FSL-201 and names it: 1,526 values change

Open difference

Within the record’s thresholds (specs/topup.md §12): o_movpar.txt: rotations 6.02e-05 (§12: 0.001); _fieldcoef (its field, Hz): |d| inside the brain, mean 0.00214 (§12: 0.1); _fieldcoef (its field, Hz): |d| inside the brain, 99th percentile 0.00658 (§12: 0.5). 2 of 3 files differ in their values (most: o_fieldcoef.nii.gz: 1,521 of 1,521 values, at most 0.0215): the clean room computes in its own order of operations (written from the specification, not from FSL’s code). FSL’s own band is not measured on this case (its thread band is, on fMRIPrep’s calls); in fix, its default, the clean room corrects FSL-201 and names it: 1,521 values change.

6/6 · 12/12
opt/warpres-0options--warpres=0: a knot every voxel
command linetopup --imain=<half.nii.gz: derived from ds005040's fMRIPrep pair> --datain=<acq.txt: acquisition parameters written by the suite> --out=o --warpres=0 --miter=2 --estmov=0 --minmet=1
Bit-identical

every file FSL writes (5/5) byte for byte, with FSL’s exit status and messages

Open difference

Within the record’s thresholds (specs/topup.md §12): _fieldcoef (its field, Hz): |d| inside the brain, 99th percentile 9.34e-06 (§12: 0.5); _fieldcoef (its field, Hz): |d| inside the brain, mean 1.83e-06 (§12: 0.1); _fieldcoef (its field, Hz): correlation inside the brain 1 (§12: 0.999). 1 of 3 files differ in their values (most: o_fieldcoef.nii.gz: 20,480 of 20,480 values, at most 2.96e-05): the clean room computes in its own order of operations (written from the specification, not from FSL’s code). FSL’s own band is not measured on this case (its thread band is, on fMRIPrep’s calls).

6/6 · 12/12
opt/lambda-negativeoptionsa negative --lambda
command linetopup --imain=<half.nii.gz: derived from ds005040's fMRIPrep pair> --datain=<acq.txt: acquisition parameters written by the suite> --out=o --warpres=20 --miter=3 --lambda=-0.005 --fout=f
Bit-identical

every file FSL writes (6/6) byte for byte, with FSL’s exit status and messages; its default, fix, refuses the call (exit 1), naming FSL-126

Expected divergence

FSL-126 reproduced: a negative weight rewards a rough field, both minimisations diverge. Outside the thresholds: --fout (Hz): |d| inside the brain, 99th percentile 222.3 (threshold 0.5); --fout (Hz): |d| inside the brain, mean 60.85 (threshold 0.1); --fout (H…

7/7 · 9/17
opt/fwhm-negativeoptionsa negative --fwhm
command linetopup --imain=<half.nii.gz: derived from ds005040's fMRIPrep pair> --datain=<acq.txt: acquisition parameters written by the suite> --out=o --warpres=20 --miter=3 --fwhm=-8 --fout=f
Bit-identical

every file FSL writes (6/6) byte for byte, with FSL’s exit status and messages; its default, fix, refuses the call (exit 1), naming FSL-125

Open difference

Within the record’s thresholds (specs/topup.md §12): --fout (Hz): |d| inside the brain, 99th percentile 0.0493 (§12: 0.5); _fieldcoef (its field, Hz): |d| inside the brain, 99th percentile 0.0493 (§12: 0.5); --fout (Hz): |d| inside the brain, mean 0.00624 (§12: 0.1). 3 of 4 files differ in their values (most: f.nii.gz: 20,478 of 20,480 values, at most 451): the clean room computes in its own order of operations (written from the specification, not from FSL’s code). FSL’s own band is not measured on this case (its thread band is, on fMRIPrep’s calls); its default, fix, refuses the call (exit 1), naming FSL-125.

7/7 · 17/17
opt/warpres-negativeoptionsa negative --warpres
command linetopup --imain=<half.nii.gz: derived from ds005040's fMRIPrep pair> --datain=<acq.txt: acquisition parameters written by the suite> --out=o --warpres=-20 --miter=2 --fout=f
Expected divergence

FSL-124: refused; FSL is not run: its knot spacing of about 4e9 voxels needs more memory than the machine holds (2026-11-03: killed in a 24 GiB memory box)

Expected divergence

FSL-124: refused; FSL is not run: its knot spacing of about 4e9 voxels needs more memory than the machine holds (2026-11-03: killed in a 24 GiB memory box)

2/2
cfg/b02b0configuration filesFSL's b02b0.cnf by name, one iteration per level
command linetopup --imain=<half.nii.gz: derived from ds005040's fMRIPrep pair> --datain=<acq.txt: acquisition parameters written by the suite> --out=o --config=b02b0 --miter=1 --iout=c --fout=f
Bit-identical

every file FSL writes (8/8) byte for byte, with FSL’s exit status and messages; in fix, its default, the replica corrects FSL-201 and names it: 95,986 values change

Open difference

Within the record’s thresholds (specs/topup.md §12): c: relative difference inside the brain, 99th percentile 0.00144 (§12: 0.001); c_mean: relative difference inside the brain, 99th percentile 0.000926 (§12: 0.001); _fieldcoef (its field, Hz): |d| inside the brain, 99th percentile 0.0877 (§12: 0.5). 5 of 6 files differ in their values (most: c.nii.gz: 34,549 of 40,960 values, at most 120): the clean room computes in its own order of operations (written from the specification, not from FSL’s code). FSL’s own band is not measured on this case (its thread band is, on fMRIPrep’s calls); in fix, its default, the clean room corrects FSL-201 and names it: 95,982 values change.

6/6 · 22/22
cfg/b02b0_1configuration filesFSL's b02b0_1.cnf by name, one iteration per level
command linetopup --imain=<half.nii.gz: derived from ds005040's fMRIPrep pair> --datain=<acq.txt: acquisition parameters written by the suite> --out=o --config=b02b0_1 --miter=1 --iout=c --fout=f
Bit-identical

every file FSL writes (8/8) byte for byte, with FSL’s exit status and messages; in fix, its default, the replica corrects FSL-201 and names it: 95,901 values change

Open difference

Within the record’s thresholds (specs/topup.md §12): c: relative difference inside the brain, 99th percentile 0.000661 (§12: 0.001); c_mean: relative difference inside the brain, 99th percentile 0.000533 (§12: 0.001); --fout (Hz): |d| inside the brain, 99th percentile 0.0314 (§12: 0.5). 5 of 6 files differ in their values (most: c.nii.gz: 34,535 of 40,960 values, at most 53.7): the clean room computes in its own order of operations (written from the specification, not from FSL’s code). FSL’s own band is not measured on this case (its thread band is, on fMRIPrep’s calls); in fix, its default, the clean room corrects FSL-201 and names it: 95,947 values change.

6/6 · 22/22
cfg/b02b0_2configuration filesFSL's b02b0_2.cnf by name, one iteration per level
command linetopup --imain=<half.nii.gz: derived from ds005040's fMRIPrep pair> --datain=<acq.txt: acquisition parameters written by the suite> --out=o --config=b02b0_2 --miter=1 --iout=c --fout=f
Bit-identical

every file FSL writes (8/8) byte for byte, with FSL’s exit status and messages; in fix, its default, the replica corrects FSL-201 and names it: 95,986 values change

Open difference

Within the record’s thresholds (specs/topup.md §12): c: relative difference inside the brain, 99th percentile 0.00144 (§12: 0.001); c_mean: relative difference inside the brain, 99th percentile 0.000926 (§12: 0.001); _fieldcoef (its field, Hz): |d| inside the brain, 99th percentile 0.0877 (§12: 0.5). 5 of 6 files differ in their values (most: c.nii.gz: 34,549 of 40,960 values, at most 120): the clean room computes in its own order of operations (written from the specification, not from FSL’s code). FSL’s own band is not measured on this case (its thread band is, on fMRIPrep’s calls); in fix, its default, the clean room corrects FSL-201 and names it: 95,982 values change.

6/6 · 22/22
cfg/b02b0_4configuration filesFSL's b02b0_4.cnf by name, one iteration per level
command linetopup --imain=<half.nii.gz: derived from ds005040's fMRIPrep pair> --datain=<acq.txt: acquisition parameters written by the suite> --out=o --config=b02b0_4 --miter=1 --iout=c --fout=f
Bit-identical

every file FSL writes (8/8) byte for byte, with FSL’s exit status and messages; in fix, its default, the replica corrects FSL-201 and names it: 95,978 values change

Expected divergence

FSL-201 reproduced, the movements estimated and not converged: FSL’s rotation derivatives are forward differences of single-precision matrices, so each gradient entry moves by about 2 % of its lever-weighted sum when its angle’s rounding to single precision f…

6/6 · 20/22
cfg/b02b0_7Tconfiguration filesFSL's b02b0_7T.cnf by name, one iteration per level
command linetopup --imain=<half.nii.gz: derived from ds005040's fMRIPrep pair> --datain=<acq.txt: acquisition parameters written by the suite> --out=o --config=b02b0_7T --miter=1 --iout=c --fout=f
Bit-identical

every file FSL writes (8/8) byte for byte, with FSL’s exit status and messages; in fix, its default, the replica corrects FSL-201 and names it: 96,000 values change

Expected divergence

FSL-201 reproduced, the movements estimated and not converged: FSL’s rotation derivatives are forward differences of single-precision matrices, so each gradient entry moves by about 2 % of its lever-weighted sum when its angle’s rounding to single precision f…

6/6 · 20/22
cfg/test_acqpconfiguration filesFSL's test_acqp.cnf by name, one iteration per level
command linetopup --imain=<half.nii.gz: derived from ds005040's fMRIPrep pair> --datain=<acq.txt: acquisition parameters written by the suite> --out=o --config=test_acqp --miter=1 --iout=c --fout=f
Bit-identical

every file FSL writes (8/8) byte for byte, with FSL’s exit status and messages

Open difference

Within the record’s thresholds (specs/topup.md §12): c: relative difference inside the brain, 99th percentile 6.32e-06 (§12: 0.001); c_mean: relative difference inside the brain, 99th percentile 5.71e-06 (§12: 0.001); c: mean |d| / mean intensity inside the brain 3.1e-07 (§12: 0.001). 4 of 6 files differ in their values (most: c.nii.gz: 31,963 of 40,960 values, at most 0.248): the clean room computes in its own order of operations (written from the specification, not from FSL’s code). FSL’s own band is not measured on this case (its thread band is, on fMRIPrep’s calls).

6/6 · 22/22
cfg/b02b0-extensionconfiguration filesFSL's b02b0.cnf by name with its extension
command linetopup --imain=<half.nii.gz: derived from ds005040's fMRIPrep pair> --datain=<acq.txt: acquisition parameters written by the suite> --out=o --config=b02b0.cnf --miter=1
Bit-identical

every file FSL writes (5/5) byte for byte, with FSL’s exit status and messages; in fix, its default, the replica corrects FSL-201 and names it: 20,485 values change

Open difference

Within the record’s thresholds (specs/topup.md §12): _fieldcoef (its field, Hz): |d| inside the brain, 99th percentile 0.0837 (§12: 0.5); _fieldcoef (its field, Hz): |d| inside the brain, mean 0.011 (§12: 0.1); o_movpar.txt: rotations 5.77e-05 (§12: 0.001). 2 of 3 files differ in their values (most: o_fieldcoef.nii.gz: 20,480 of 20,480 values, at most 1.53): the clean room computes in its own order of operations (written from the specification, not from FSL’s code). FSL’s own band is not measured on this case (its thread band is, on fMRIPrep’s calls); in fix, its default, the clean room corrects FSL-201 and names it: 20,480 values change.

6/6 · 12/12
cfg/localconfiguration filesa local configuration file
command linetopup --imain=<half.nii.gz: derived from ds005040's fMRIPrep pair> --datain=<acq.txt: acquisition parameters written by the suite> --out=o --config=<short.cnf: a configuration written by the suite> --fout=f
Bit-identical

every file FSL writes (7/7) byte for byte, with FSL’s exit status and messages; in fix, its default, the replica corrects FSL-201 and names it: 25,178 values change

Open difference

Within the record’s thresholds (specs/topup.md §12): o_movpar.txt: rotations 0.000141 (§12: 0.001); _fieldcoef (its field, Hz): |d| inside the brain, mean 0.00974 (§12: 0.1); --fout (Hz): |d| inside the brain, mean 0.00974 (§12: 0.1). 3 of 4 files differ in their values (most: f.nii.gz: 20,480 of 20,480 values, at most 0.0581): the clean room computes in its own order of operations (written from the specification, not from FSL’s code). FSL’s own band is not measured on this case (its thread band is, on fMRIPrep’s calls); in fix, its default, the clean room corrects FSL-201 and names it: 25,171 values change.

6/6 · 16/16
cfg/local-no-extensionconfiguration filesa local configuration named without .cnf
command linetopup --imain=<half.nii.gz: derived from ds005040's fMRIPrep pair> --datain=<acq.txt: acquisition parameters written by the suite> --out=o --config=short --fout=f (with <short.cnf: a configuration written by the suite> in its directory)
Bit-identical

every file FSL writes (7/7) byte for byte, with FSL’s exit status and messages; in fix, its default, the replica corrects FSL-201 and names it: 25,178 values change

Open difference

Within the record’s thresholds (specs/topup.md §12): o_movpar.txt: rotations 0.000141 (§12: 0.001); _fieldcoef (its field, Hz): |d| inside the brain, mean 0.00974 (§12: 0.1); --fout (Hz): |d| inside the brain, mean 0.00974 (§12: 0.1). 3 of 4 files differ in their values (most: f.nii.gz: 20,480 of 20,480 values, at most 0.0581): the clean room computes in its own order of operations (written from the specification, not from FSL’s code). FSL’s own band is not measured on this case (its thread band is, on fMRIPrep’s calls); in fix, its default, the clean room corrects FSL-201 and names it: 25,171 values change.

6/6 · 16/16
cfg/command-line-winsconfiguration filesoptions on the command line take precedence
command linetopup --imain=<half.nii.gz: derived from ds005040's fMRIPrep pair> --datain=<acq.txt: acquisition parameters written by the suite> --out=o --config=<short.cnf: a configuration written by the suite> --warpres=30,20 --fout=f
Bit-identical

every file FSL writes (7/7) byte for byte, with FSL’s exit status and messages; in fix, its default, the replica corrects FSL-201 and names it: 22,006 values change

Open difference

Within the record’s thresholds (specs/topup.md §12): _fieldcoef (its field, Hz): |d| inside the brain, mean 0.00255 (§12: 0.1); --fout (Hz): |d| inside the brain, mean 0.00255 (§12: 0.1); --fout (Hz): |d| inside the brain, 99th percentile 0.00994 (§12: 0.5). 3 of 4 files differ in their values (most: f.nii.gz: 20,477 of 20,480 values, at most 0.0486): the clean room computes in its own order of operations (written from the specification, not from FSL’s code). FSL’s own band is not measured on this case (its thread band is, on fMRIPrep’s calls); in fix, its default, the clean room corrects FSL-201 and names it: 22,000 values change.

6/6 · 16/16
cfg/directoryconfiguration filesa directory as the configuration (read as empty)
command linetopup --imain=<half.nii.gz: derived from ds005040's fMRIPrep pair> --datain=<acq.txt: acquisition parameters written by the suite> --out=o --config=. --miter=2
Bit-identical

every file FSL writes (5/5) byte for byte, with FSL’s exit status and messages; its default, fix, names FSL-201, with the same values

Bit-identical

The same values, every one, in all 3 files FSL writes on this case, with FSL’s exit status and messages (the record checks the headers and the log); its default, fix, names FSL-201, with the same values.

6/6 · 12/12
cfg/not-foundconfiguration filesa configuration that does not exist
command linetopup --imain=<half.nii.gz: derived from ds005040's fMRIPrep pair> --datain=<acq.txt: acquisition parameters written by the suite> --out=o --config=nothere
Rejected by both

topup exits 1; the same from the replica. Cannot find config-file nothere

Rejected by both

topup exits 1; the same from the clean room. Cannot find config-file nothere

5/5 · 4/4
cfg/not-found-pathconfiguration filesa configuration with a path that does not exist
command linetopup --imain=<half.nii.gz: derived from ds005040's fMRIPrep pair> --datain=<acq.txt: acquisition parameters written by the suite> --out=o --config=./nothere
Rejected by both

topup exits 1; the same from the replica. Cannot find config-file ./nothere

Rejected by both

topup exits 1; the same from the clean room. Cannot find config-file ./nothere

5/5 · 4/4
cfg/fsldir-unsetconfiguration filesFSLDIR unset and a name to look up there
command linetopup --imain=<half.nii.gz: derived from ds005040's fMRIPrep pair> --datain=<acq.txt: acquisition parameters written by the suite> --out=o --config=b02b0 (FSLDIR unset)
Rejected by both

topup exits 1; the same from the replica. basic_string: construction from null is not valid The replica’s default, fix, ends with exit status 1, naming FSL-121.

Rejected by both

topup exits 1; the same from the clean room. basic_string: construction from null is not valid The clean room’s default, fix, ends with exit status 1, naming FSL-121.

6/6 · 5/5
cfg/bad-optionconfiguration filesan unknown option in the configuration
command linetopup --imain=<half.nii.gz: derived from ds005040's fMRIPrep pair> --datain=<acq.txt: acquisition parameters written by the suite> --out=o --config=<bad_option.cnf: a configuration written by the suite>
Rejected by both

topup exits 1; the same from the replica. Part of FSL (ID: 2412.6-dirty) topup Usage: topup --imain=<some 4D image> --datain= --config= --out=my_topup_results Compulsory arguments (You MUST set one or mor

Rejected by both

topup exits 1; the same from the clean room. Part of FSL (ID: 2412.6-dirty) topup Usage: topup --imain=<some 4D image> --datain= --config= --out=my_topup_results Compulsory arguments (You MUST set one or mor

5/5 · 4/4
cfg/long-lineconfiguration filesa comment line of more than 1023 characters before two settings
command linetopup --imain=<half.nii.gz: derived from ds005040's fMRIPrep pair> --datain=<acq.txt: acquisition parameters written by the suite> --out=o --config=<long_line.cnf: a configuration written by the suite> --fout=f
Bit-identical

every file FSL writes (7/7) byte for byte, with FSL’s exit status and messages; in fix, its default, the replica corrects FSL-122, FSL-201 and names it: 25,178 values change

Bit-identical

The same values, every one, in all 4 files FSL writes on this case, with FSL’s exit status and messages (the record checks the headers and the log); in fix, its default, the clean room corrects FSL-122, FSL-201 and names it: 25,173 values change.

6/6 · 16/16
err/no-argumentserrors and messagesno arguments: the usage
command linetopup
Rejected by both

topup exits 1; the same from the replica. Part of FSL (ID: 2412.6-dirty) topup Usage: topup --imain=<some 4D image> --datain= --config= --out=my_topup_results Compulsory arguments (You MUST set one or mor

Rejected by both

topup exits 1; the same from the clean room. Part of FSL (ID: 2412.6-dirty) topup Usage: topup --imain=<some 4D image> --datain= --config= --out=my_topup_results Compulsory arguments (You MUST set one or mor

5/5 · 4/4
err/helperrors and messages-h: the usage
command linetopup -h
Rejected by both

topup exits 1; the same from the replica. Part of FSL (ID: 2412.6-dirty) topup Usage: topup --imain=<some 4D image> --datain= --config= --out=my_topup_results Compulsory arguments (You MUST set one or mor

Rejected by both

topup exits 1; the same from the clean room. Part of FSL (ID: 2412.6-dirty) topup Usage: topup --imain=<some 4D image> --datain= --config= --out=my_topup_results Compulsory arguments (You MUST set one or mor

5/5 · 4/4
err/help-longerrors and messages--help with the compulsory options set
command linetopup --imain=<half.nii.gz: derived from ds005040's fMRIPrep pair> --datain=<acq.txt: acquisition parameters written by the suite> --help
Rejected by both

topup exits 1; the same from the replica. Part of FSL (ID: 2412.6-dirty) topup Usage: topup --imain=<some 4D image> --datain= --config= --out=my_topup_results Compulsory arguments (You MUST set one or mor

Rejected by both

topup exits 1; the same from the clean room. Part of FSL (ID: 2412.6-dirty) topup Usage: topup --imain=<some 4D image> --datain= --config= --out=my_topup_results Compulsory arguments (You MUST set one or mor

5/5 · 4/4
err/unknown-optionerrors and messagesan unknown option
command linetopup --imain=<half.nii.gz: derived from ds005040's fMRIPrep pair> --datain=<acq.txt: acquisition parameters written by the suite> --bogus=1
Rejected by both

topup exits 1; the same from the replica. Part of FSL (ID: 2412.6-dirty) topup Usage: topup --imain=<some 4D image> --datain= --config= --out=my_topup_results Compulsory arguments (You MUST set one or mor

Rejected by both

topup exits 1; the same from the clean room. Part of FSL (ID: 2412.6-dirty) topup Usage: topup --imain=<some 4D image> --datain= --config= --out=my_topup_results Compulsory arguments (You MUST set one or mor

5/5 · 4/4
err/missing-valueerrors and messagesan option without its value
command linetopup --imain
Rejected by both

topup exits 1; the same from the replica. Part of FSL (ID: 2412.6-dirty) topup Usage: topup --imain=<some 4D image> --datain= --config= --out=my_topup_results Compulsory arguments (You MUST set one or mor

Rejected by both

topup exits 1; the same from the clean room. Part of FSL (ID: 2412.6-dirty) topup Usage: topup --imain=<some 4D image> --datain= --config= --out=my_topup_results Compulsory arguments (You MUST set one or mor

5/5 · 4/4
err/no-imainerrors and messagesno --imain
command linetopup --datain=<acq.txt: acquisition parameters written by the suite>
Rejected by both

topup exits 1; the same from the replica. Part of FSL (ID: 2412.6-dirty) topup Usage: topup --imain=<some 4D image> --datain= --config= --out=my_topup_results Compulsory arguments (You MUST set one or mor

Rejected by both

topup exits 1; the same from the clean room. Part of FSL (ID: 2412.6-dirty) topup Usage: topup --imain=<some 4D image> --datain= --config= --out=my_topup_results Compulsory arguments (You MUST set one or mor

5/5 · 4/4
err/imain-missingerrors and messagesan --imain that does not exist
command linetopup --imain=nothere.nii.gz --datain=<acq.txt: acquisition parameters written by the suite>
Rejected by both

topup exits 1; the same from the replica. Image Exception : #63 :: No image files match: nothere Image Exception : #22 :: Failed to read volume nothere.nii.gz Error : No image files match: nothere Part of FSL (ID: 2412.6-dirty) topup Usage: t

Rejected by both

topup exits 1; the same from the clean room. Image Exception : #63 :: No image files match: nothere Image Exception : #22 :: Failed to read volume nothere.nii.gz Error : No image files match: nothere Part of FSL (ID: 2412.6-dirty) topup Usage: t

5/5 · 4/4
err/imain-texterrors and messagesan --imain that is not an image
command linetopup --imain=<acq.txt: acquisition parameters written by the suite> --datain=<acq.txt: acquisition parameters written by the suite>
Rejected by both

topup exits 1; the same from the replica. Image Exception : #63 :: No image files match: acq.txt Image Exception : #22 :: Failed to read volume acq.txt Error : No image files match: acq.txt Part of FSL (ID: 2412.6-dirty) topup Usage: topup –

Rejected by both

topup exits 1; the same from the clean room. Image Exception : #63 :: No image files match: acq.txt Image Exception : #22 :: Failed to read volume acq.txt Error : No image files match: acq.txt Part of FSL (ID: 2412.6-dirty) topup Usage: topup –

5/5 · 4/4
err/datain-and-acqperrors and messagesboth --datain and --acqp
command linetopup --imain=<half.nii.gz: derived from ds005040's fMRIPrep pair> --datain=<acq.txt: acquisition parameters written by the suite> --acqp=<acq.txt: acquisition parameters written by the suite>
Rejected by both

topup exits 1; the same from the replica. Part of FSL (ID: 2412.6-dirty) topup Usage: topup --imain=<some 4D image> --datain= --config= --out=my_topup_results Compulsory arguments (You MUST set one or mor

Rejected by both

topup exits 1; the same from the clean room. Part of FSL (ID: 2412.6-dirty) topup Usage: topup --imain=<some 4D image> --datain= --config= --out=my_topup_results Compulsory arguments (You MUST set one or mor

5/5 · 4/4
err/no-datainerrors and messagesneither --datain nor --acqp
command linetopup --imain=<half.nii.gz: derived from ds005040's fMRIPrep pair>
Rejected by both

topup exits 1; the same from the replica. Part of FSL (ID: 2412.6-dirty) topup Usage: topup --imain=<some 4D image> --datain= --config= --out=my_topup_results Compulsory arguments (You MUST set one or mor

Rejected by both

topup exits 1; the same from the clean room. Part of FSL (ID: 2412.6-dirty) topup Usage: topup --imain=<some 4D image> --datain= --config= --out=my_topup_results Compulsory arguments (You MUST set one or mor

5/5 · 4/4
err/datain-missingerrors and messagesa --datain that does not exist
command linetopup --imain=<half.nii.gz: derived from ds005040's fMRIPrep pair> --datain=nothere.txt
Rejected by both

topup exits 1; the same from the replica. Could not open matrix file nothere.txt Part of FSL (ID: 2412.6-dirty) topup Usage: topup --imain=<some 4D image> --datain= --config= --out=my_topup_results Compul

Rejected by both

topup exits 1; the same from the clean room. Could not open matrix file nothere.txt Part of FSL (ID: 2412.6-dirty) topup Usage: topup --imain=<some 4D image> --datain= --config= --out=my_topup_results Compul

5/5 · 4/4
err/datain-five-columnserrors and messagesa --datain of five columns
command linetopup --imain=<half.nii.gz: derived from ds005040's fMRIPrep pair> --datain=<acq_five.txt: acquisition parameters written by the suite>
Rejected by both

topup exits 1; the same from the replica. Part of FSL (ID: 2412.6-dirty) topup Usage: topup --imain=<some 4D image> --datain= --config= --out=my_topup_results Compulsory arguments (You MUST set one or mor

Rejected by both

topup exits 1; the same from the clean room. Part of FSL (ID: 2412.6-dirty) topup Usage: topup --imain=<some 4D image> --datain= --config= --out=my_topup_results Compulsory arguments (You MUST set one or mor

5/5 · 4/4
err/datain-texterrors and messagesa --datain of words
command linetopup --imain=<half.nii.gz: derived from ds005040's fMRIPrep pair> --datain=<acq_text.txt: acquisition parameters written by the suite>
Rejected by both

topup exits 1; the same from the replica. Part of FSL (ID: 2412.6-dirty) topup Usage: topup --imain=<some 4D image> --datain= --config= --out=my_topup_results Compulsory arguments (You MUST set one or mor

Rejected by both

topup exits 1; the same from the clean room. Part of FSL (ID: 2412.6-dirty) topup Usage: topup --imain=<some 4D image> --datain= --config= --out=my_topup_results Compulsory arguments (You MUST set one or mor

5/5 · 4/4
err/datain-normerrors and messagesa phase-encoding vector not of unit length
command linetopup --imain=<half.nii.gz: derived from ds005040's fMRIPrep pair> --datain=<acq_norm.txt: acquisition parameters written by the suite>
Rejected by both

topup exits 1; the same from the replica. Part of FSL (ID: 2412.6-dirty) topup Usage: topup --imain=<some 4D image> --datain= --config= --out=my_topup_results Compulsory arguments (You MUST set one or mor

Rejected by both

topup exits 1; the same from the clean room. Part of FSL (ID: 2412.6-dirty) topup Usage: topup --imain=<some 4D image> --datain= --config= --out=my_topup_results Compulsory arguments (You MUST set one or mor

5/5 · 4/4
err/datain-rowserrors and messagesthree rows for two volumes
command linetopup --imain=<half.nii.gz: derived from ds005040's fMRIPrep pair> --datain=<acq_three.txt: acquisition parameters written by the suite>
Rejected by both

topup exits 1; the same from the replica. Part of FSL (ID: 2412.6-dirty) topup Usage: topup --imain=<some 4D image> --datain= --config= --out=my_topup_results Compulsory arguments (You MUST set one or mor

Rejected by both

topup exits 1; the same from the clean room. Part of FSL (ID: 2412.6-dirty) topup Usage: topup --imain=<some 4D image> --datain= --config= --out=my_topup_results Compulsory arguments (You MUST set one or mor

5/5 · 4/4
err/acqp-rowserrors and messagesthree rows for two volumes, given with --acqp
command linetopup --imain=<half.nii.gz: derived from ds005040's fMRIPrep pair> --acqp=<acq_three.txt: acquisition parameters written by the suite>
Rejected by both

topup exits 1; the same from the replica. Part of FSL (ID: 2412.6-dirty) topup Usage: topup --imain=<some 4D image> --datain= --config= --out=my_topup_results Compulsory arguments (You MUST set one or mor The replica’s default, fix, ends with exit status 1, naming FSL-120.

Rejected by both

topup exits 1; the same from the clean room. Part of FSL (ID: 2412.6-dirty) topup Usage: topup --imain=<some 4D image> --datain= --config= --out=my_topup_results Compulsory arguments (You MUST set one or mor The clean room’s default, fix, ends with exit status 1, naming FSL-120.

6/6 · 5/5
err/readout-longerrors and messagesa read-out time above 0.2 s
command linetopup --imain=<half.nii.gz: derived from ds005040's fMRIPrep pair> --datain=<acq_long_rt.txt: acquisition parameters written by the suite>
Rejected by both

topup exits 1; the same from the replica. Part of FSL (ID: 2412.6-dirty) topup Usage: topup --imain=<some 4D image> --datain= --config= --out=my_topup_results Compulsory arguments (You MUST set one or mor

Rejected by both

topup exits 1; the same from the clean room. Part of FSL (ID: 2412.6-dirty) topup Usage: topup --imain=<some 4D image> --datain= --config= --out=my_topup_results Compulsory arguments (You MUST set one or mor

5/5 · 4/4
err/readout-two-zeroerrors and messagestwo read-out times of 0
command linetopup --imain=<half.nii.gz: derived from ds005040's fMRIPrep pair> --datain=<acq_two_zero.txt: acquisition parameters written by the suite>
Rejected by both

topup exits 1; the same from the replica. Part of FSL (ID: 2412.6-dirty) topup Usage: topup --imain=<some 4D image> --datain= --config= --out=my_topup_results Compulsory arguments (You MUST set one or mor

Rejected by both

topup exits 1; the same from the clean room. Part of FSL (ID: 2412.6-dirty) topup Usage: topup --imain=<some 4D image> --datain= --config= --out=my_topup_results Compulsory arguments (You MUST set one or mor

5/5 · 4/4
err/levels-subsamperrors and messages--subsamp of three levels against two
command linetopup --imain=<half.nii.gz: derived from ds005040's fMRIPrep pair> --datain=<acq.txt: acquisition parameters written by the suite> --out=o --warpres=20,14 --subsamp=2,1,1
Rejected by both

topup exits 1; the same from the replica. Part of FSL (ID: 2412.6-dirty) topup Usage: topup --imain=<some 4D image> --datain= --config= --out=my_topup_results Compulsory arguments (You MUST set one or mor

Rejected by both

topup exits 1; the same from the clean room. Part of FSL (ID: 2412.6-dirty) topup Usage: topup --imain=<some 4D image> --datain= --config= --out=my_topup_results Compulsory arguments (You MUST set one or mor

5/5 · 4/4
err/levels-minmeterrors and messages--minmet of three levels against two
command linetopup --imain=<half.nii.gz: derived from ds005040's fMRIPrep pair> --datain=<acq.txt: acquisition parameters written by the suite> --out=o --warpres=20,14 --minmet=0,1,1
Rejected by both

topup exits 1; the same from the replica. Part of FSL (ID: 2412.6-dirty) topup Usage: topup --imain=<some 4D image> --datain= --config= --out=my_topup_results Compulsory arguments (You MUST set one or mor

Rejected by both

topup exits 1; the same from the clean room. Part of FSL (ID: 2412.6-dirty) topup Usage: topup --imain=<some 4D image> --datain= --config= --out=my_topup_results Compulsory arguments (You MUST set one or mor

5/5 · 4/4
err/minmet-2errors and messages--minmet=2
command linetopup --imain=<half.nii.gz: derived from ds005040's fMRIPrep pair> --datain=<acq.txt: acquisition parameters written by the suite> --out=o --minmet=2
Rejected by both

topup exits 1; the same from the replica. Part of FSL (ID: 2412.6-dirty) topup Usage: topup --imain=<some 4D image> --datain= --config= --out=my_topup_results Compulsory arguments (You MUST set one or mor

Rejected by both

topup exits 1; the same from the clean room. Part of FSL (ID: 2412.6-dirty) topup Usage: topup --imain=<some 4D image> --datain= --config= --out=my_topup_results Compulsory arguments (You MUST set one or mor

5/5 · 4/4
err/subsamp-incompatibleerrors and messages--subsamp=3 on a 32-voxel axis
command linetopup --imain=<half.nii.gz: derived from ds005040's fMRIPrep pair> --datain=<acq.txt: acquisition parameters written by the suite> --out=o --subsamp=3
Rejected by both

topup exits 1; the same from the replica. Part of FSL (ID: 2412.6-dirty) topup Usage: topup --imain=<some 4D image> --datain= --config= --out=my_topup_results Compulsory arguments (You MUST set one or mor

Rejected by both

topup exits 1; the same from the clean room. Part of FSL (ID: 2412.6-dirty) topup Usage: topup --imain=<some 4D image> --datain= --config= --out=my_topup_results Compulsory arguments (You MUST set one or mor

5/5 · 4/4
err/subsamp-0errors and messages--subsamp=0
command linetopup --imain=<half.nii.gz: derived from ds005040's fMRIPrep pair> --datain=<acq.txt: acquisition parameters written by the suite> --out=o --subsamp=0
Rejected by both

topup exits 1; the same from the replica. Part of FSL (ID: 2412.6-dirty) topup Usage: topup --imain=<some 4D image> --datain= --config= --out=my_topup_results Compulsory arguments (You MUST set one or mor

Rejected by both

topup exits 1; the same from the clean room. Part of FSL (ID: 2412.6-dirty) topup Usage: topup --imain=<some 4D image> --datain= --config= --out=my_topup_results Compulsory arguments (You MUST set one or mor

5/5 · 4/4
err/subsamp-mixederrors and messagessubsamplings 3 and 2 on a 30 x 30 x 18 image (each divides it; the regridded grid of 33 does not divide by 2)
command linetopup --imain=<crop.nii.gz: derived from ds005040's fMRIPrep pair> --datain=<acq.txt: acquisition parameters written by the suite> --out=o --warpres=20,14 --subsamp=3,2 --miter=2
Rejected by both

topup exits 1; the same from the replica. Error occurred when preparing to run level 2 of topup Exception thrown with message: Topup::TopupScan::SubSample: invalid subsampling factor The replica’s default, fix, finishes instead (exit status 0), naming FSL-118, FSL-201.

Rejected by both

topup exits 1; the same from the clean room. Error occurred when preparing to run level 2 of topup Exception thrown with message: Topup::TopupScan::SubSample: invalid subsampling factor The clean room’s default, fix, finishes instead (exit status 0), naming FSL-118, FSL-201.

6/6
err/numprecerrors and messagesan unknown --numprec
command linetopup --imain=<half.nii.gz: derived from ds005040's fMRIPrep pair> --datain=<acq.txt: acquisition parameters written by the suite> --out=o --numprec=single
Rejected by both

topup exits 1; the same from the replica. Part of FSL (ID: 2412.6-dirty) topup Usage: topup --imain=<some 4D image> --datain= --config= --out=my_topup_results Compulsory arguments (You MUST set one or mor

Rejected by both

topup exits 1; the same from the clean room. Part of FSL (ID: 2412.6-dirty) topup Usage: topup --imain=<some 4D image> --datain= --config= --out=my_topup_results Compulsory arguments (You MUST set one or mor

5/5 · 4/4
err/regmoderrors and messagesan unknown --regmod
command linetopup --imain=<half.nii.gz: derived from ds005040's fMRIPrep pair> --datain=<acq.txt: acquisition parameters written by the suite> --out=o --regmod=sobolev
Rejected by both

topup exits 1; the same from the replica. Part of FSL (ID: 2412.6-dirty) topup Usage: topup --imain=<some 4D image> --datain= --config= --out=my_topup_results Compulsory arguments (You MUST set one or mor

Rejected by both

topup exits 1; the same from the clean room. Part of FSL (ID: 2412.6-dirty) topup Usage: topup --imain=<some 4D image> --datain= --config= --out=my_topup_results Compulsory arguments (You MUST set one or mor

5/5 · 4/4
err/interperrors and messagesan unknown --interp
command linetopup --imain=<half.nii.gz: derived from ds005040's fMRIPrep pair> --datain=<acq.txt: acquisition parameters written by the suite> --out=o --interp=cubic
Rejected by both

topup exits 1; the same from the replica. Part of FSL (ID: 2412.6-dirty) topup Usage: topup --imain=<some 4D image> --datain= --config= --out=my_topup_results Compulsory arguments (You MUST set one or mor

Rejected by both

topup exits 1; the same from the clean room. Part of FSL (ID: 2412.6-dirty) topup Usage: topup --imain=<some 4D image> --datain= --config= --out=my_topup_results Compulsory arguments (You MUST set one or mor

5/5 · 4/4
err/splineorder-4errors and messages--splineorder=4
command linetopup --imain=<half.nii.gz: derived from ds005040's fMRIPrep pair> --datain=<acq.txt: acquisition parameters written by the suite> --out=o --splineorder=4
Rejected by both

topup exits 1; the same from the replica. Part of FSL (ID: 2412.6-dirty) topup Usage: topup --imain=<some 4D image> --datain= --config= --out=my_topup_results Compulsory arguments (You MUST set one or mor

Rejected by both

topup exits 1; the same from the clean room. Part of FSL (ID: 2412.6-dirty) topup Usage: topup --imain=<some 4D image> --datain= --config= --out=my_topup_results Compulsory arguments (You MUST set one or mor

5/5 · 4/4
err/splineorder-negativeerrors and messages--splineorder=-1
command linetopup --imain=<half.nii.gz: derived from ds005040's fMRIPrep pair> --datain=<acq.txt: acquisition parameters written by the suite> --out=o --splineorder=-1
Rejected by both

topup exits 1; the same from the replica. Part of FSL (ID: 2412.6-dirty) topup Usage: topup --imain=<some 4D image> --datain= --config= --out=my_topup_results Compulsory arguments (You MUST set one or mor

Rejected by both

topup exits 1; the same from the clean room. Part of FSL (ID: 2412.6-dirty) topup Usage: topup --imain=<some 4D image> --datain= --config= --out=my_topup_results Compulsory arguments (You MUST set one or mor

5/5 · 4/4
err/debug-4errors and messages--debug=4
command linetopup --imain=<half.nii.gz: derived from ds005040's fMRIPrep pair> --datain=<acq.txt: acquisition parameters written by the suite> --out=o --warpres=20 --fwhm=8 --miter=3 --lambda=0.005 --debug=4
Rejected by both

topup exits 1; the same from the replica. Error occurred when preparing to run topup Exception thrown with message: Topup::TopupCF::SetDebug: Debug level out of range

Rejected by both

topup exits 1; the same from the clean room. Error occurred when preparing to run topup Exception thrown with message: Topup::TopupCF::SetDebug: Debug level out of range The clean room’s default, fix, ends with exit status 1, naming FSL-201.

5/5
err/debug-negativeerrors and messages--debug=-1
command linetopup --imain=<half.nii.gz: derived from ds005040's fMRIPrep pair> --datain=<acq.txt: acquisition parameters written by the suite> --out=o --warpres=20 --fwhm=8 --miter=3 --lambda=0.005 --debug=-1
Rejected by both

topup exits 1; the same from the replica. Error occurred when preparing to run topup Exception thrown with message: Topup::TopupCF::SetDebug: Debug level out of range

Rejected by both

topup exits 1; the same from the clean room. Error occurred when preparing to run topup Exception thrown with message: Topup::TopupCF::SetDebug: Debug level out of range The clean room’s default, fix, ends with exit status 1, naming FSL-201.

5/5
err/miter-negativeerrors and messages--miter=-1 (about 4e9 iterations)
command linetopup --imain=<half.nii.gz: derived from ds005040's fMRIPrep pair> --datain=<acq.txt: acquisition parameters written by the suite> --out=o --warpres=20 --estmov=0 --miter=-1
Bit-identical

every file FSL writes (5/5) byte for byte, with FSL’s exit status and messages; its default, fix, refuses the call (exit 1), naming FSL-123

Bit-identical

The same values, every one, in all 3 files FSL writes on this case, with FSL’s exit status and messages (the record checks the headers and the log); its default, fix, refuses the call (exit 1), naming FSL-123.

7/7 · 13/13
err/out-no-directoryerrors and messages--out in a directory that does not exist
command linetopup --imain=<half.nii.gz: derived from ds005040's fMRIPrep pair> --datain=<acq.txt: acquisition parameters written by the suite> --out=nodir/o --warpres=20 --fwhm=8 --miter=3 --lambda=0.005
Rejected by both

topup exits 1; the same from the replica. Error occured while writing output from topup Exception thrown with message: Error: cant open file nodir/o_fieldcoef.nii.gz The replica’s default, fix, ends with exit status 1, naming FSL-201.

Rejected by both

topup exits 1; the same from the clean room. Error occured while writing output from topup Exception thrown with message: Error: cant open file nodir/o_fieldcoef.nii.gz The clean room’s default, fix, ends with exit status 1, naming FSL-201.

5/5
err/iout-no-directoryerrors and messages--iout in a directory that does not exist
command linetopup --imain=<half.nii.gz: derived from ds005040's fMRIPrep pair> --datain=<acq.txt: acquisition parameters written by the suite> --out=o --warpres=20 --fwhm=8 --miter=3 --lambda=0.005 --iout=nodir/c
Rejected by both

topup exits 1; the same from the replica. Error occured while writing output from topup Exception thrown with message: Error: cant open file nodir/c.nii.gz The replica’s default, fix, ends with exit status 1, naming FSL-201.

Rejected by both

topup exits 1; the same from the clean room. Error occured while writing output from topup Exception thrown with message: Error: cant open file nodir/c.nii.gz The clean room’s default, fix, ends with exit status 1, naming FSL-201.

5/5 · 11/11
err/logout-no-directoryerrors and messages--logout in a directory that does not exist
command linetopup --imain=<half.nii.gz: derived from ds005040's fMRIPrep pair> --datain=<acq.txt: acquisition parameters written by the suite> --out=o --warpres=20 --fwhm=8 --miter=3 --lambda=0.005 --logout=nodir/log
Bit-identical

every file FSL writes (4/4) byte for byte, with FSL’s exit status and messages; in fix, its default, the replica corrects FSL-201 and names it: 1,526 values change

Open difference

Within the record’s thresholds (specs/topup.md §12): o_movpar.txt: rotations 6.02e-05 (§12: 0.001); _fieldcoef (its field, Hz): |d| inside the brain, mean 0.00214 (§12: 0.1); _fieldcoef (its field, Hz): |d| inside the brain, 99th percentile 0.00658 (§12: 0.5). 2 of 2 files differ in their values (most: o_fieldcoef.nii.gz: 1,521 of 1,521 values, at most 0.0215): the clean room computes in its own order of operations (written from the specification, not from FSL’s code). FSL’s own band is not measured on this case (its thread band is, on fMRIPrep’s calls); in fix, its default, the clean room corrects FSL-201 and names it: 1,521 values change.

6/6 · 11/11
err/rbmout-no-directoryerrors and messages--rbmout in a directory that does not exist
command linetopup --imain=<half.nii.gz: derived from ds005040's fMRIPrep pair> --datain=<acq.txt: acquisition parameters written by the suite> --out=o --warpres=20 --fwhm=8 --miter=3 --lambda=0.005 --rbmout=nodir/r
Bit-identical

every file FSL writes (5/5) byte for byte, with FSL’s exit status and messages; in fix, its default, the replica corrects FSL-201 and names it: 1,526 values change

Open difference

Within the record’s thresholds (specs/topup.md §12): o_movpar.txt: rotations 6.02e-05 (§12: 0.001); _fieldcoef (its field, Hz): |d| inside the brain, mean 0.00214 (§12: 0.1); _fieldcoef (its field, Hz): |d| inside the brain, 99th percentile 0.00658 (§12: 0.5). 2 of 3 files differ in their values (most: o_fieldcoef.nii.gz: 1,521 of 1,521 values, at most 0.0215): the clean room computes in its own order of operations (written from the specification, not from FSL’s code). FSL’s own band is not measured on this case (its thread band is, on fMRIPrep’s calls); in fix, its default, the clean room corrects FSL-201 and names it: 1,521 values change.

6/6 · 12/12
err/profile-no-directoryerrors and messages--profile with --out in a directory that does not exist
command linetopup --imain=<half.nii.gz: derived from ds005040's fMRIPrep pair> --datain=<acq.txt: acquisition parameters written by the suite> --out=nodir/o --warpres=20 --fwhm=8 --miter=3 --lambda=0.005 --profile
Rejected by both

topup exits 1; the same from the replica.

Rejected by both

topup exits 1; the same from the clean room. The clean room’s default, fix, ends with exit status 1, naming FSL-201.

5/5
err/outputtype-invaliderrors and messagesFSLOUTPUTTYPE=BOGUS
command lineFSLOUTPUTTYPE=BOGUS topup --imain=<half.nii.gz: derived from ds005040's fMRIPrep pair> --datain=<acq.txt: acquisition parameters written by the suite> --out=o --warpres=20 --fwhm=8 --miter=3 --lambda=0.005 --fout=f
Rejected by both

topup exits 1; the same from the replica. ERROR:: Unrecognised value (BOGUS) of environment variable FSLOUTPUTTYPE Legal values are: NIFTI, NIFTI_PAIR, NIFTI2, NIFTI2_PAIR, NIFTI_GZ, NIFTI_PAIR_GZ, NIFTI2_GZ, NIFTI2_PAIR_GZ, NIFTI_ZST, NIFTI_ The replica’s default, fix, ends with exit status 1, naming FSL-201.

Rejected by both

topup exits 1; the same from the clean room. ERROR:: Unrecognised value (BOGUS) of environment variable FSLOUTPUTTYPE Legal values are: NIFTI, NIFTI_PAIR, NIFTI2, NIFTI2_PAIR, NIFTI_GZ, NIFTI_PAIR_GZ, NIFTI2_GZ, NIFTI2_PAIR_GZ, NIFTI_ZST, NIFTI_ The clean room’s default, fix, ends with exit status 1, naming FSL-201.

5/5
err/outputtype-unseterrors and messagesFSLOUTPUTTYPE unset
command linetopup --imain=<half.nii.gz: derived from ds005040's fMRIPrep pair> --datain=<acq.txt: acquisition parameters written by the suite> --out=o --warpres=20 --fwhm=8 --miter=3 --lambda=0.005 --fout=f (FSLOUTPUTTYPE unset)
Rejected by both

topup exits 1; the same from the replica. ERROR:: Environment variable FSLOUTPUTTYPE is not set! Please make sure that the appropriate configuration file is sourced by your shell (e.g. by putting it in .profile). e.g. bash or sh users add the The replica’s default, fix, ends with exit status 1, naming FSL-201.

Rejected by both

topup exits 1; the same from the clean room. ERROR:: Environment variable FSLOUTPUTTYPE is not set! Please make sure that the appropriate configuration file is sourced by your shell (e.g. by putting it in .profile). e.g. bash or sh users add the The clean room’s default, fix, ends with exit status 1, naming FSL-201.

5/5
thr/nthr-4thread counts--nthr=4 (the spline field's threaded products and the Hessians' threaded sparse products) against --nthr=1
command linetopup --imain=<half.nii.gz: derived from ds005040's fMRIPrep pair> --datain=<acq.txt: acquisition parameters written by the suite> --out=o --warpres=20,14 --subsamp=2,1 --fwhm=8,4 --miter=3,2 --lambda=0.005,0.0001 --estmov=1,0 --minmet=0,1 --fout=f --iout=c --nthr=4
Bit-identical

every file FSL writes (8/8) byte for byte, with FSL’s exit status and messages; the replica at four threads gives FSL’s one-thread result; FSL’s own run at four threads keeps 2/7 files the same (its thread band, FSL-116); its default, fix, names FSL-201, with the same values

Within the original's own variability

In every metric of the record (specs/topup.md §12, inside the record’s brain) the clean room is no further from FSL at one thread than FSL at another thread count is (its own thread band, FSL-116): c: relative difference inside the brain, 99th percentile 0.00164 (§12: 0.001; FSL’s own 0.00168); c_mean: relative difference inside the brain, 99th percentile 0.000996 (§12: 0.001; FSL’s own 0.001); o_movpar.txt: rotations 0.000128 (§12: 0.001; FSL’s own 0.000132); in fix, its default, the clean room corrects FSL-201 and names it: 80,181 values change.

7/7 · 22/22
thr/fsl-num-threadsthread countsFSL_NUM_THREADS=4 with --nthr=1 (the default thread count of the Hessian's sparse matrices) against it unset
command lineFSL_NUM_THREADS=4 topup --imain=<half.nii.gz: derived from ds005040's fMRIPrep pair> --datain=<acq.txt: acquisition parameters written by the suite> --out=o --warpres=20 --fwhm=8 --miter=3 --lambda=0.005 --fout=f --iout=c
Bit-identical

every file FSL writes (8/8) byte for byte, with FSL’s exit status and messages; the replica at four threads gives FSL’s one-thread result; FSL’s own run at four threads keeps 2/7 files the same (its thread band, FSL-116); its default, fix, names FSL-201, with the same values

Open difference

Within the record’s thresholds (specs/topup.md §12), but further from FSL than FSL’s own thread band in 12 metric(s): c: relative difference inside the brain, 99th percentile 0.00044 (§12: 0.001; FSL’s own 3.64e-06); c: mean |d| / mean intensity inside the brain 2.2e-05 (§12: 0.001; FSL’s own 1.64e-07); c_mean: relative difference inside the brain, 99th percentile 0.000261 (§12: 0.001; FSL’s own 2.32e-06); in fix, its default, the clean room corrects FSL-201 and names it: 76,905 values change.

7/7 · 22/22
thr/openblasthread countsFSL_SKIP_GLOBAL=1 and OPENBLAS_NUM_THREADS=4 (OpenBLAS's own threads: its dot product of more than 10,000 values) against OpenBLAS on one thread, with scaled conjugate gradients on a knot every voxel (about 31,000 coefficients)
command lineFSL_SKIP_GLOBAL=1 OPENBLAS_NUM_THREADS=4 topup --imain=<half.nii.gz: derived from ds005040's fMRIPrep pair> --datain=<acq.txt: acquisition parameters written by the suite> --out=o --warpres=4 --fwhm=4 --miter=2 --lambda=0.005 --estmov=0 --minmet=1 --fout=f --iout=c
Bit-identical

every file FSL writes (8/8) byte for byte, with FSL’s exit status and messages; the replica at four threads gives FSL’s one-thread result; FSL’s own run at four threads keeps 7/7 files the same (its thread band, FSL-116)

Open difference

Within the record’s thresholds (specs/topup.md §12), but further from FSL than FSL’s own thread band in 10 metric(s): c: relative difference inside the brain, 99th percentile 3.64e-06 (§12: 0.001; FSL’s own 0); c: mean |d| / mean intensity inside the brain 2.53e-07 (§12: 0.001; FSL’s own 0); c_mean: relative difference inside the brain, 99th percentile 2.72e-06 (§12: 0.001; FSL’s own 0).

7/7 · 22/22
at/lsr-y worked exampleapplytopupleast-squares restoration of the pair along y (the default method, spline interpolation, the movements)
command lineapplytopup --imain=<half_up.nii.gz: derived from ds005040's fMRIPrep pair>,<half_dn.nii.gz: derived from ds005040's fMRIPrep pair> --inindex=1,2 --datain=<acq.txt: acquisition parameters written by the suite> --topup=tu --out=hifi (tu_*: FSL's topup outputs of the case opt/levels)
Bit-identical

every file FSL writes (7/7) byte for byte, with FSL’s exit status and messages; in fix, its default, the replica corrects FSL-132 and names it: 5,136 values change

Open difference

Within the record’s thresholds (specs/topup.md §12): hifi (applytopup’s output): relative difference inside the brain, 99th percentile 1.19e-05 (§12: 0.001); hifi (applytopup’s output): mean |d| / mean intensity inside the brain 6.84e-07 (§12: 0.001). 1 of 1 files differ in their values (most: hifi.nii.gz: 16,831 of 20,480 values, at most 0.813): the clean room computes in its own order of operations (written from the specification, not from FSL’s code). FSL’s own band is not measured on this case (its thread band is, on fMRIPrep’s calls).

6/6 · 8/8
at/lsr-4dapplytopupleast-squares restoration of two volumes per phase-encoding direction
command lineapplytopup --imain=<dwi_up.nii.gz: derived from ds005040's fMRIPrep pair>,<dwi_dn.nii.gz: derived from ds005040's fMRIPrep pair> --inindex=1,2 --datain=<acq.txt: acquisition parameters written by the suite> --topup=tu --out=hifi (tu_*: FSL's topup outputs of the case opt/levels)
Bit-identical

every file FSL writes (7/7) byte for byte, with FSL’s exit status and messages; in fix, its default, the replica corrects FSL-132 and names it: 5,136 values change

Open difference

Within the record’s thresholds (specs/topup.md §12): hifi (applytopup’s output): relative difference inside the brain, 99th percentile 1.2e-05 (§12: 0.001); hifi (applytopup’s output): mean |d| / mean intensity inside the brain 6.84e-07 (§12: 0.001). 1 of 1 files differ in their values (most: hifi.nii.gz: 33,652 of 40,960 values, at most 0.813): the clean room computes in its own order of operations (written from the specification, not from FSL’s code). FSL’s own band is not measured on this case (its thread band is, on fMRIPrep’s calls).

6/6 · 8/8
at/lsr-xapplytopupleast-squares restoration along x
command lineapplytopup --imain=<xhalf_up.nii.gz: derived from ds005040's fMRIPrep pair>,<xhalf_dn.nii.gz: derived from ds005040's fMRIPrep pair> --inindex=1,2 --datain=<acq_x.txt: acquisition parameters written by the suite> --topup=tu --out=hifi (tu_*: FSL's topup outputs of the case opt/x-phase-encoding)
Bit-identical

every file FSL writes (6/6) byte for byte, with FSL’s exit status and messages; in fix, its default, the replica corrects FSL-132 and names it: 5,035 values change

Open difference

Within the record’s thresholds (specs/topup.md §12): hifi (applytopup’s output): relative difference inside the brain, 99th percentile 9.68e-06 (§12: 0.001); hifi (applytopup’s output): mean |d| / mean intensity inside the brain 8.03e-07 (§12: 0.001). 1 of 1 files differ in their values (most: hifi.nii.gz: 17,116 of 20,480 values, at most 0.727): the clean room computes in its own order of operations (written from the specification, not from FSL’s code). FSL’s own band is not measured on this case (its thread band is, on fMRIPrep’s calls).

6/6 · 8/8
at/lsr-fourapplytopupleast-squares restoration of two collections (along y and along x)
command lineapplytopup --imain=<four_0.nii.gz: derived from ds005040's fMRIPrep pair>,<four_1.nii.gz: derived from ds005040's fMRIPrep pair>,<four_2.nii.gz: derived from ds005040's fMRIPrep pair>,<four_3.nii.gz: derived from ds005040's fMRIPrep pair> --inindex=1,2,3,4 --datain=<acq_four.txt: acquisition parameters written by the suite> --topup=tu --out=hifi (tu_*: FSL's topup outputs of the case opt/four-volumes)
Bit-identical

every file FSL writes (8/8) byte for byte, with FSL’s exit status and messages; its default, fix, names FSL-132, with the same values

Bit-identical

The same values, every one, in all 1 files FSL writes on this case, with FSL’s exit status and messages (the record checks the headers and the log).

6/6 · 7/7
at/lsr-threeapplytopupleast-squares restoration of three images in one collection
command lineapplytopup --imain=<half_up.nii.gz: derived from ds005040's fMRIPrep pair>,<half_dn.nii.gz: derived from ds005040's fMRIPrep pair>,<half_up.nii.gz: derived from ds005040's fMRIPrep pair> --inindex=1,2,1 --datain=<acq.txt: acquisition parameters written by the suite> --topup=tu --out=hifi (tu_*: FSL's topup outputs of the case opt/levels)
Bit-identical

every file FSL writes (7/7) byte for byte, with FSL’s exit status and messages; in fix, its default, the replica corrects FSL-132 and names it: 4,189 values change

Open difference

Within the record’s thresholds (specs/topup.md §12): hifi (applytopup’s output): relative difference inside the brain, 99th percentile 8.31e-06 (§12: 0.001); hifi (applytopup’s output): mean |d| / mean intensity inside the brain 4.63e-07 (§12: 0.001). 1 of 1 files differ in their values (most: hifi.nii.gz: 16,428 of 20,480 values, at most 0.553): the clean room computes in its own order of operations (written from the specification, not from FSL’s code). FSL’s own band is not measured on this case (its thread band is, on fMRIPrep’s calls).

6/6 · 8/8
at/lsr-trilinearapplytopup--interp=trilinear
command lineapplytopup --imain=<half_up.nii.gz: derived from ds005040's fMRIPrep pair>,<half_dn.nii.gz: derived from ds005040's fMRIPrep pair> --inindex=1,2 --datain=<acq.txt: acquisition parameters written by the suite> --topup=tu --out=hifi --interp=trilinear (tu_*: FSL's topup outputs of the case opt/levels)
Bit-identical

every file FSL writes (7/7) byte for byte, with FSL’s exit status and messages

Open difference

Within the record’s thresholds (specs/topup.md §12): hifi (applytopup’s output): relative difference inside the brain, 99th percentile 1.01e-05 (§12: 0.001); hifi (applytopup’s output): mean |d| / mean intensity inside the brain 6.7e-07 (§12: 0.001). 1 of 1 files differ in their values (most: hifi.nii.gz: 16,746 of 20,480 values, at most 1.27): the clean room computes in its own order of operations (written from the specification, not from FSL’s code). FSL’s own band is not measured on this case (its thread band is, on fMRIPrep’s calls).

6/6 · 8/8
at/lsr-cropapplytopupleast-squares restoration of lines of 30 voxels (a length that does not divide 2^32)
command lineapplytopup --imain=<crop_up.nii.gz: derived from ds005040's fMRIPrep pair>,<crop_dn.nii.gz: derived from ds005040's fMRIPrep pair> --inindex=1,2 --datain=<acq.txt: acquisition parameters written by the suite> --topup=tu --out=hifi (tu_*: FSL's topup outputs of a short run of its own)
Bit-identical

every file FSL writes (7/7) byte for byte, with FSL’s exit status and messages; in fix, its default, the replica corrects FSL-132, FSL-133 and names it: 6,612 values change

Open difference

Within the record’s thresholds (specs/topup.md §12): hifi (applytopup’s output): relative difference inside the brain, 99th percentile 1.83e-05 (§12: 0.001); hifi (applytopup’s output): mean |d| / mean intensity inside the brain 1.09e-06 (§12: 0.001). 1 of 1 files differ in their values (most: hifi.nii.gz: 13,243 of 16,200 values, at most 0.895): the clean room computes in its own order of operations (written from the specification, not from FSL’s code). FSL’s own band is not measured on this case (its thread band is, on fMRIPrep’s calls); in fix, its default, the clean room corrects FSL-133 and names it: 4,942 values change.

6/6 · 8/8
at/lsr-zero-fieldapplytopupleast-squares restoration with a field of zeros (band systems: Armadillo's band solver)
command lineapplytopup --imain=<half_up.nii.gz: derived from ds005040's fMRIPrep pair>,<half_dn.nii.gz: derived from ds005040's fMRIPrep pair> --inindex=1,2 --datain=<acq.txt: acquisition parameters written by the suite> --topup=tu --out=hifi (tu_*: FSL's topup outputs of the case opt/levels, every coefficient set to 0)
Bit-identical

every file FSL writes (6/6) byte for byte, with FSL’s exit status and messages; in fix, its default, the replica corrects FSL-132 and names it: 4,698 values change

Open difference

Within the record’s thresholds (specs/topup.md §12): hifi (applytopup’s output): relative difference inside the brain, 99th percentile 1.22e-05 (§12: 0.001); hifi (applytopup’s output): mean |d| / mean intensity inside the brain 6.54e-07 (§12: 0.001). 1 of 1 files differ in their values (most: hifi.nii.gz: 16,769 of 20,480 values, at most 0.711): the clean room computes in its own order of operations (written from the specification, not from FSL’s code). FSL’s own band is not measured on this case (its thread band is, on fMRIPrep’s calls).

6/6 · 8/8
at/vb2d-yapplytopup--method=vb2D: phi and lambda per slice
command lineapplytopup --imain=<half_up.nii.gz: derived from ds005040's fMRIPrep pair>,<half_dn.nii.gz: derived from ds005040's fMRIPrep pair> --inindex=1,2 --datain=<acq.txt: acquisition parameters written by the suite> --topup=tu --out=hifi --method=vb2D (tu_*: FSL's topup outputs of the case opt/levels)
Bit-identical

every file FSL writes (7/7) byte for byte, with FSL’s exit status and messages; in fix, its default, the replica corrects FSL-132 and names it: 5,622 values change

Open difference

Within the record’s thresholds (specs/topup.md §12): hifi (applytopup’s output): relative difference inside the brain, 99th percentile 1.16e-05 (§12: 0.001); hifi (applytopup’s output): mean |d| / mean intensity inside the brain 6.74e-07 (§12: 0.001). 1 of 1 files differ in their values (most: hifi.nii.gz: 17,764 of 20,480 values, at most 44.4): the clean room computes in its own order of operations (written from the specification, not from FSL’s code). FSL’s own band is not measured on this case (its thread band is, on fMRIPrep’s calls).

6/6 · 9/9
at/vb3d-y worked exampleapplytopup--method=vb3D: phi and lambda per volume
command lineapplytopup --imain=<half_up.nii.gz: derived from ds005040's fMRIPrep pair>,<half_dn.nii.gz: derived from ds005040's fMRIPrep pair> --inindex=1,2 --datain=<acq.txt: acquisition parameters written by the suite> --topup=tu --out=hifi --method=vb3D (tu_*: FSL's topup outputs of the case opt/levels)
Bit-identical

every file FSL writes (7/7) byte for byte, with FSL’s exit status and messages; in fix, its default, the replica corrects FSL-132, FSL-143 and names it: 18,852 values change

Open difference

Within the record’s thresholds (specs/topup.md §12): hifi (applytopup’s output): relative difference inside the brain, 99th percentile 6.54e-05 (§12: 0.001); hifi (applytopup’s output): mean |d| / mean intensity inside the brain 3.3e-05 (§12: 0.001). 1 of 1 files differ in their values (most: hifi.nii.gz: 18,753 of 20,480 values, at most 1.05e+06): the clean room computes in its own order of operations (written from the specification, not from FSL’s code). FSL’s own band is not measured on this case (its thread band is, on fMRIPrep’s calls); in fix, its default, the clean room corrects FSL-143 and names it: 18,851 values change.

6/6 · 9/9
at/vb4d-yapplytopup--method=vb4D: one phi and lambda
command lineapplytopup --imain=<half_up.nii.gz: derived from ds005040's fMRIPrep pair>,<half_dn.nii.gz: derived from ds005040's fMRIPrep pair> --inindex=1,2 --datain=<acq.txt: acquisition parameters written by the suite> --topup=tu --out=hifi --method=vb4D (tu_*: FSL's topup outputs of the case opt/levels)
Bit-identical

every file FSL writes (7/7) byte for byte, with FSL’s exit status and messages; in fix, its default, the replica corrects FSL-132, FSL-143 and names it: 18,852 values change

Open difference

Within the record’s thresholds (specs/topup.md §12): hifi (applytopup’s output): relative difference inside the brain, 99th percentile 6.54e-05 (§12: 0.001); hifi (applytopup’s output): mean |d| / mean intensity inside the brain 3.3e-05 (§12: 0.001). 1 of 1 files differ in their values (most: hifi.nii.gz: 18,753 of 20,480 values, at most 1.05e+06): the clean room computes in its own order of operations (written from the specification, not from FSL’s code). FSL’s own band is not measured on this case (its thread band is, on fMRIPrep’s calls); in fix, its default, the clean room corrects FSL-143 and names it: 18,851 values change.

6/6 · 9/9
at/lsr-longapplytopup--method=lsr on lines of 96 voxels along y (the long-line pair)
command lineapplytopup --imain=<long_up.nii.gz: ds005040's pair cropped, its lines doubled to 96 voxels>,<long_dn.nii.gz: ds005040's pair cropped, its lines doubled to 96 voxels> --inindex=1,2 --datain=<acq.txt: acquisition parameters written by the suite> --topup=tu --out=hifi --method=lsr (tu_*: FSL's topup outputs of a short run of its own)
Bit-identical

every file FSL writes (7/7) byte for byte, with FSL’s exit status and messages; in fix, its default, the replica corrects FSL-132, FSL-133 and names it: 692 values change

Open difference

Within the record’s thresholds (specs/topup.md §12): hifi (applytopup’s output): relative difference inside the brain, 99th percentile 1.71e-06 (§12: 0.001); hifi (applytopup’s output): mean |d| / mean intensity inside the brain 2.09e-07 (§12: 0.001). 1 of 1 files differ in their values (most: hifi.nii.gz: 2,734 of 6,912 values, at most 0.242): the clean room computes in its own order of operations (written from the specification, not from FSL’s code). FSL’s own band is not measured on this case (its thread band is, on fMRIPrep’s calls); in fix, its default, the clean room corrects FSL-133 and names it: 470 values change.

6/6 · 8/8
at/vb2d-longapplytopup--method=vb2D on lines of 96 voxels along y (the long-line pair)
command lineapplytopup --imain=<long_up.nii.gz: ds005040's pair cropped, its lines doubled to 96 voxels>,<long_dn.nii.gz: ds005040's pair cropped, its lines doubled to 96 voxels> --inindex=1,2 --datain=<acq.txt: acquisition parameters written by the suite> --topup=tu --out=hifi --method=vb2D (tu_*: FSL's topup outputs of a short run of its own)
Bit-identical

every file FSL writes (7/7) byte for byte, with FSL’s exit status and messages; in fix, its default, the replica corrects FSL-132, FSL-133 and names it: 5,276 values change

Open difference

Within the record’s thresholds (specs/topup.md §12): hifi (applytopup’s output): relative difference inside the brain, 99th percentile 1.49e-06 (§12: 0.001); hifi (applytopup’s output): mean |d| / mean intensity inside the brain 1.91e-07 (§12: 0.001). 1 of 1 files differ in their values (most: hifi.nii.gz: 3,612 of 6,912 values, at most 0.219): the clean room computes in its own order of operations (written from the specification, not from FSL’s code). FSL’s own band is not measured on this case (its thread band is, on fMRIPrep’s calls); in fix, its default, the clean room corrects FSL-133 and names it: 5,277 values change.

6/6 · 9/9
at/vb3d-longapplytopup--method=vb3D on lines of 96 voxels along y (the long-line pair)
command lineapplytopup --imain=<long_up.nii.gz: ds005040's pair cropped, its lines doubled to 96 voxels>,<long_dn.nii.gz: ds005040's pair cropped, its lines doubled to 96 voxels> --inindex=1,2 --datain=<acq.txt: acquisition parameters written by the suite> --topup=tu --out=hifi --method=vb3D (tu_*: FSL's topup outputs of a short run of its own)
Bit-identical

every file FSL writes (7/7) byte for byte, with FSL’s exit status and messages; in fix, its default, the replica corrects FSL-132, FSL-133, FSL-143 and names it: 5,277 values change

Open difference

Within the record’s thresholds (specs/topup.md §12): hifi (applytopup’s output): relative difference inside the brain, 99th percentile 1.81e-06 (§12: 0.001); hifi (applytopup’s output): mean |d| / mean intensity inside the brain 2.07e-07 (§12: 0.001). 1 of 1 files differ in their values (most: hifi.nii.gz: 3,675 of 6,912 values, at most 0.266): the clean room computes in its own order of operations (written from the specification, not from FSL’s code). FSL’s own band is not measured on this case (its thread band is, on fMRIPrep’s calls); in fix, its default, the clean room corrects FSL-143, FSL-133 and names it: 5,277 values change.

6/6 · 9/9
at/vb4d-longapplytopup--method=vb4D on lines of 96 voxels along y (the long-line pair)
command lineapplytopup --imain=<long_up.nii.gz: ds005040's pair cropped, its lines doubled to 96 voxels>,<long_dn.nii.gz: ds005040's pair cropped, its lines doubled to 96 voxels> --inindex=1,2 --datain=<acq.txt: acquisition parameters written by the suite> --topup=tu --out=hifi --method=vb4D (tu_*: FSL's topup outputs of a short run of its own)
Bit-identical

every file FSL writes (7/7) byte for byte, with FSL’s exit status and messages; in fix, its default, the replica corrects FSL-132, FSL-133, FSL-143 and names it: 5,277 values change

Open difference

Within the record’s thresholds (specs/topup.md §12): hifi (applytopup’s output): relative difference inside the brain, 99th percentile 1.81e-06 (§12: 0.001); hifi (applytopup’s output): mean |d| / mean intensity inside the brain 2.07e-07 (§12: 0.001). 1 of 1 files differ in their values (most: hifi.nii.gz: 3,675 of 6,912 values, at most 0.266): the clean room computes in its own order of operations (written from the specification, not from FSL’s code). FSL’s own band is not measured on this case (its thread band is, on fMRIPrep’s calls); in fix, its default, the clean room corrects FSL-143, FSL-133 and names it: 5,277 values change.

6/6 · 9/9
at/vb3d-4dapplytopup--method=vb3D of two volumes (each on its own)
command lineapplytopup --imain=<dwi_up.nii.gz: derived from ds005040's fMRIPrep pair>,<dwi_dn.nii.gz: derived from ds005040's fMRIPrep pair> --inindex=1,2 --datain=<acq.txt: acquisition parameters written by the suite> --topup=tu --out=hifi -m vb3D (tu_*: FSL's topup outputs of the case opt/levels)
Bit-identical

every file FSL writes (7/7) byte for byte, with FSL’s exit status and messages; in fix, its default, the replica corrects FSL-132, FSL-143 and names it: 37,704 values change

Open difference

Within the record’s thresholds (specs/topup.md §12): hifi (applytopup’s output): relative difference inside the brain, 99th percentile 6.5e-05 (§12: 0.001); hifi (applytopup’s output): mean |d| / mean intensity inside the brain 3.18e-05 (§12: 0.001). 1 of 1 files differ in their values (most: hifi.nii.gz: 37,467 of 40,960 values, at most 1.05e+06): the clean room computes in its own order of operations (written from the specification, not from FSL’s code). FSL’s own band is not measured on this case (its thread band is, on fMRIPrep’s calls); in fix, its default, the clean room corrects FSL-143 and names it: 37,702 values change.

6/6 · 9/9
at/vb4d-4dapplytopup--method=vb4D of two volumes (together)
command lineapplytopup --imain=<dwi_up.nii.gz: derived from ds005040's fMRIPrep pair>,<dwi_dn.nii.gz: derived from ds005040's fMRIPrep pair> --inindex=1,2 --datain=<acq.txt: acquisition parameters written by the suite> --topup=tu --out=hifi -m vb4D (tu_*: FSL's topup outputs of the case opt/levels)
Bit-identical

every file FSL writes (7/7) byte for byte, with FSL’s exit status and messages; in fix, its default, the replica corrects FSL-132, FSL-143 and names it: 37,704 values change

Open difference

Within the record’s thresholds (specs/topup.md §12): hifi (applytopup’s output): relative difference inside the brain, 99th percentile 6.5e-05 (§12: 0.001); hifi (applytopup’s output): mean |d| / mean intensity inside the brain 3.22e-05 (§12: 0.001). 1 of 1 files differ in their values (most: hifi.nii.gz: 37,499 of 40,960 values, at most 1.02e+06): the clean room computes in its own order of operations (written from the specification, not from FSL’s code). FSL’s own band is not measured on this case (its thread band is, on fMRIPrep’s calls); in fix, its default, the clean room corrects FSL-143 and names it: 37,702 values change.

6/6 · 9/9
at/vb2d-xapplytopup--method=vb2D along x
command lineapplytopup --imain=<xhalf_up.nii.gz: derived from ds005040's fMRIPrep pair>,<xhalf_dn.nii.gz: derived from ds005040's fMRIPrep pair> --inindex=1,2 --datain=<acq_x.txt: acquisition parameters written by the suite> --topup=tu --out=hifi --method=vb2D (tu_*: FSL's topup outputs of the case opt/x-phase-encoding)
Bit-identical

every file FSL writes (6/6) byte for byte, with FSL’s exit status and messages; in fix, its default, the replica corrects FSL-132 and names it: 5,444 values change

Open difference

Within the record’s thresholds (specs/topup.md §12): hifi (applytopup’s output): relative difference inside the brain, 99th percentile 9.29e-06 (§12: 0.001); hifi (applytopup’s output): mean |d| / mean intensity inside the brain 7.81e-07 (§12: 0.001). 1 of 1 files differ in their values (most: hifi.nii.gz: 18,111 of 20,480 values, at most 0.773): the clean room computes in its own order of operations (written from the specification, not from FSL’s code). FSL’s own band is not measured on this case (its thread band is, on fMRIPrep’s calls).

6/6 · 9/9
at/vb3d-xapplytopup--method=vb3D along x
command lineapplytopup --imain=<xhalf_up.nii.gz: derived from ds005040's fMRIPrep pair>,<xhalf_dn.nii.gz: derived from ds005040's fMRIPrep pair> --inindex=1,2 --datain=<acq_x.txt: acquisition parameters written by the suite> --topup=tu --out=hifi --method=vb3D (tu_*: FSL's topup outputs of the case opt/x-phase-encoding)
Bit-identical

every file FSL writes (6/6) byte for byte, with FSL’s exit status and messages; in fix, its default, the replica corrects FSL-132, FSL-143 and names it: 18,851 values change

Open difference

Within the record’s thresholds (specs/topup.md §12): hifi (applytopup’s output): relative difference inside the brain, 99th percentile 9.7e-05 (§12: 0.001); hifi (applytopup’s output): mean |d| / mean intensity inside the brain 5.19e-05 (§12: 0.001). 1 of 1 files differ in their values (most: hifi.nii.gz: 18,600 of 20,480 values, at most 1.42e+04): the clean room computes in its own order of operations (written from the specification, not from FSL’s code). FSL’s own band is not measured on this case (its thread band is, on fMRIPrep’s calls); in fix, its default, the clean room corrects FSL-143 and names it: 18,851 values change.

6/6 · 9/9
at/vb3d-anisoapplytopup--method=vb3D with voxels of 6 x 4 mm in plane (the prior's scale factors)
command lineapplytopup --imain=<half_up_aniso.nii.gz: derived from ds005040's fMRIPrep pair>,<half_up_aniso.nii.gz: derived from ds005040's fMRIPrep pair> --inindex=1,2 --datain=<acq.txt: acquisition parameters written by the suite> --topup=tu --out=hifi --method=vb3D (tu_*: FSL's topup outputs of the case opt/levels)
Bit-identical

every file FSL writes (6/6) byte for byte, with FSL’s exit status and messages; in fix, its default, the replica corrects FSL-132, FSL-143 and names it: 18,854 values change

Expected divergence

FSL-143 reproduced: FSL’s smoothness prior is indefinite, so its iterations diverge (lambda below 0 in the first iterations, then phi and lambda fall towards 1e-23 to 1e-206, the image following their ratio): the divergence amplifies the moved images’ float-l…

6/6 · 8/9
at/vb2d-anisoapplytopup--method=vb2D with voxels of 6 x 4 mm in plane
command lineapplytopup --imain=<half_up_aniso.nii.gz: derived from ds005040's fMRIPrep pair>,<half_up_aniso.nii.gz: derived from ds005040's fMRIPrep pair> --inindex=1,2 --datain=<acq.txt: acquisition parameters written by the suite> --topup=tu --out=hifi --method=vb2D (tu_*: FSL's topup outputs of the case opt/levels)
Bit-identical

every file FSL writes (6/6) byte for byte, with FSL’s exit status and messages; in fix, its default, the replica corrects FSL-132 and names it: 6,052 values change

Expected divergence

the printed phi and lambda of the top slices (few voxels in the brain) part from FSL’s in the 4th digit during the iterations (8e-4) and meet again (their final values within 1e-5): the iterations amplify the moved images’ float-level differences there; the i…

6/6 · 8/9
at/vb4d-fourapplytopup--method=vb4D of two collections (along y and along x)
command lineapplytopup --imain=<four_0.nii.gz: derived from ds005040's fMRIPrep pair>,<four_1.nii.gz: derived from ds005040's fMRIPrep pair>,<four_2.nii.gz: derived from ds005040's fMRIPrep pair>,<four_3.nii.gz: derived from ds005040's fMRIPrep pair> --inindex=1,2,3,4 --datain=<acq_four.txt: acquisition parameters written by the suite> --topup=tu --out=hifi --method=vb4D (tu_*: FSL's topup outputs of the case opt/four-volumes)
Bit-identical

every file FSL writes (8/8) byte for byte, with FSL’s exit status and messages; in fix, its default, the replica corrects FSL-132, FSL-143 and names it: 17,263 values change

Expected divergence

FSL-143 reproduced: FSL’s smoothness prior is indefinite, so its iterations diverge (lambda below 0 in the first iterations, then phi and lambda fall towards 1e-23 to 1e-206, the image following their ratio): the divergence amplifies the moved images’ float-l…

6/6 · 6/7
at/vb3d-threeapplytopup--method=vb3D of three images in one collection (FSL-142: counted as two)
command lineapplytopup --imain=<half_up.nii.gz: derived from ds005040's fMRIPrep pair>,<half_dn.nii.gz: derived from ds005040's fMRIPrep pair>,<half_up.nii.gz: derived from ds005040's fMRIPrep pair> --inindex=1,2,1 --datain=<acq.txt: acquisition parameters written by the suite> --topup=tu --out=hifi --method=vb3D (tu_*: FSL's topup outputs of the case opt/levels)
Bit-identical

every file FSL writes (7/7) byte for byte, with FSL’s exit status and messages; in fix, its default, the replica corrects FSL-132, FSL-142, FSL-143 and names it: 18,852 values change

Expected divergence

FSL-143 reproduced: FSL’s smoothness prior is indefinite, so its iterations diverge (lambda below 0 in the first iterations, then phi and lambda fall towards 1e-23 to 1e-206, the image following their ratio): the divergence amplifies the moved images’ float-l…

6/6 · 8/9
at/vb2d-threeapplytopup--method=vb2D of three images in one collection
command lineapplytopup --imain=<half_up.nii.gz: derived from ds005040's fMRIPrep pair>,<half_dn.nii.gz: derived from ds005040's fMRIPrep pair>,<half_up.nii.gz: derived from ds005040's fMRIPrep pair> --inindex=1,2,1 --datain=<acq.txt: acquisition parameters written by the suite> --topup=tu --out=hifi --method=vb2D (tu_*: FSL's topup outputs of the case opt/levels)
Bit-identical

every file FSL writes (7/7) byte for byte, with FSL’s exit status and messages; in fix, its default, the replica corrects FSL-132 and names it: 4,454 values change

Open difference

Within the record’s thresholds (specs/topup.md §12): hifi (applytopup’s output): relative difference inside the brain, 99th percentile 7.75e-06 (§12: 0.001); hifi (applytopup’s output): mean |d| / mean intensity inside the brain 4.6e-07 (§12: 0.001). 1 of 1 files differ in their values (most: hifi.nii.gz: 17,356 of 20,480 values, at most 45.1): the clean room computes in its own order of operations (written from the specification, not from FSL’s code). FSL’s own band is not measured on this case (its thread band is, on fMRIPrep’s calls).

6/6 · 9/9
at/vb-one-directionapplytopup--method=vb3D of two images with the same phase encoding (FSL's message)
command lineapplytopup --imain=<half_up.nii.gz: derived from ds005040's fMRIPrep pair>,<half_up.nii.gz: derived from ds005040's fMRIPrep pair> --inindex=1,1 --datain=<acq.txt: acquisition parameters written by the suite> --topup=tu --out=hifi --method=vb3D (tu_*: FSL's topup outputs of the case opt/levels)
Rejected by both

applytopup exits -6; the same from the replica. terminate called after throwing an instance of ‘TOPUP::ApplyTopupException’ what(): ApplyTopup::Invalid combination of phase-encode vectors for least-squares restoration The replica’s default, fix, ends with exit status 1, naming FSL-132.

Rejected by both

applytopup exits -6; the same from the clean room. terminate called after throwing an instance of ‘TOPUP::ApplyTopupException’ what(): ApplyTopup::Invalid combination of phase-encode vectors for least-squares restoration

5/5 · 4/4
at/vb-no-movementsapplytopup--method=vb4D with a field image, which has no movements (FSL-135)
command lineapplytopup --imain=<half_up.nii.gz: derived from ds005040's fMRIPrep pair>,<half_dn.nii.gz: derived from ds005040's fMRIPrep pair> --inindex=1,2 --datain=<acq.txt: acquisition parameters written by the suite> --topup=tu_field --out=hifi --method=vb4D (tu_*: FSL's topup outputs of the case opt/levels)
Rejected by both

applytopup exits -6; the same from the replica. terminate called after throwing an instance of ‘TOPUP::TopupFileIOException’ what(): TopupFileIO::TopupFileReader::MoveMatrix: Index i out of bound The replica’s default, fix, ends with exit status 1, naming FSL-135.

Rejected by both

applytopup exits -6; the same from the clean room. terminate called after throwing an instance of ‘TOPUP::TopupFileIOException’ what(): TopupFileIO::TopupFileReader::MoveMatrix: Index i out of bound The clean room’s default, fix, ends with exit status 1, naming FSL-135.

6/6 · 5/5
at/vb-other-gridapplytopup--method=vb2D with a field on another grid (FSL-134)
command lineapplytopup --imain=<crop_up.nii.gz: derived from ds005040's fMRIPrep pair>,<crop_dn.nii.gz: derived from ds005040's fMRIPrep pair> --inindex=1,2 --datain=<acq.txt: acquisition parameters written by the suite> --topup=tu --out=hifi --method=vb2D (tu_*: FSL's topup outputs of the case opt/levels)
Expected divergence

FSL-134: refused; FSL is not run: FSL reads the field’s lines past their end (undefined behaviour)

Expected divergence

FSL-134: refused; FSL is not run: FSL reads the field’s lines past their end (undefined behaviour)

2/2
at/jac-y worked exampleapplytopupJacobian modulation of one image
command lineapplytopup --imain=<half_up.nii.gz: derived from ds005040's fMRIPrep pair> --inindex=1 --datain=<acq.txt: acquisition parameters written by the suite> --topup=tu --out=hifi --method=jac (tu_*: FSL's topup outputs of the case opt/levels)
Bit-identical

every file FSL writes (6/6) byte for byte, with FSL’s exit status and messages; in fix, its default, the replica corrects FSL-132 and names it: 20,480 values change

Open difference

Within the record’s thresholds (specs/topup.md §12): hifi (applytopup’s output): relative difference inside the brain, 99th percentile 2.05e-06 (§12: 0.001); hifi (applytopup’s output): mean |d| / mean intensity inside the brain 1.36e-07 (§12: 0.001). 1 of 1 files differ in their values (most: hifi.nii.gz: 13,708 of 20,480 values, at most 0.242): the clean room computes in its own order of operations (written from the specification, not from FSL’s code). FSL’s own band is not measured on this case (its thread band is, on fMRIPrep’s calls); in fix, its default, the clean room corrects FSL-132 and names it: 20,480 values change.

6/6 · 8/8
at/jac-pairapplytopupJacobian modulation of the pair, averaged
command lineapplytopup --imain=<half_up.nii.gz: derived from ds005040's fMRIPrep pair>,<half_dn.nii.gz: derived from ds005040's fMRIPrep pair> --inindex=1,2 --datain=<acq.txt: acquisition parameters written by the suite> --topup=tu --out=hifi --method=jac (tu_*: FSL's topup outputs of the case opt/levels)
Bit-identical

every file FSL writes (7/7) byte for byte, with FSL’s exit status and messages; in fix, its default, the replica corrects FSL-132 and names it: 18,848 values change

Open difference

Within the record’s thresholds (specs/topup.md §12): hifi (applytopup’s output): relative difference inside the brain, 99th percentile 3.62e-06 (§12: 0.001); hifi (applytopup’s output): mean |d| / mean intensity inside the brain 2.33e-07 (§12: 0.001). 1 of 1 files differ in their values (most: hifi.nii.gz: 16,915 of 20,480 values, at most 0.244): the clean room computes in its own order of operations (written from the specification, not from FSL’s code). FSL’s own band is not measured on this case (its thread band is, on fMRIPrep’s calls); in fix, its default, the clean room corrects FSL-132 and names it: 18,851 values change.

6/6 · 8/8
at/jac-4dapplytopupJacobian modulation of two volumes
command lineapplytopup --imain=<dwi_up.nii.gz: derived from ds005040's fMRIPrep pair> --inindex=1 --datain=<acq.txt: acquisition parameters written by the suite> --topup=tu --out=hifi --method=jac (tu_*: FSL's topup outputs of the case opt/levels)
Bit-identical

every file FSL writes (6/6) byte for byte, with FSL’s exit status and messages; in fix, its default, the replica corrects FSL-132 and names it: 20,480 values change

Open difference

Within the record’s thresholds (specs/topup.md §12): hifi (applytopup’s output): relative difference inside the brain, 99th percentile 2.24e-06 (§12: 0.001); hifi (applytopup’s output): mean |d| / mean intensity inside the brain 1.38e-07 (§12: 0.001). 1 of 1 files differ in their values (most: hifi.nii.gz: 29,863 of 40,960 values, at most 0.242): the clean room computes in its own order of operations (written from the specification, not from FSL’s code). FSL’s own band is not measured on this case (its thread band is, on fMRIPrep’s calls); in fix, its default, the clean room corrects FSL-132 and names it: 20,480 values change.

6/6 · 8/8
at/jac-trilinearapplytopup--method=jac --interp=trilinear
command lineapplytopup --imain=<half_up.nii.gz: derived from ds005040's fMRIPrep pair> --inindex=1 --datain=<acq.txt: acquisition parameters written by the suite> --topup=tu --out=hifi --method=jac -n trilinear (tu_*: FSL's topup outputs of the case opt/levels)
Bit-identical

every file FSL writes (6/6) byte for byte, with FSL’s exit status and messages

Open difference

Within the record’s thresholds (specs/topup.md §12): hifi (applytopup’s output): relative difference inside the brain, 99th percentile 2.05e-06 (§12: 0.001); hifi (applytopup’s output): mean |d| / mean intensity inside the brain 1.36e-07 (§12: 0.001). 1 of 1 files differ in their values (most: hifi.nii.gz: 13,708 of 20,480 values, at most 0.242): the clean room computes in its own order of operations (written from the specification, not from FSL’s code). FSL’s own band is not measured on this case (its thread band is, on fMRIPrep’s calls).

6/6 · 8/8
at/jac-xapplytopupJacobian modulation along x
command lineapplytopup --imain=<xhalf_dn.nii.gz: derived from ds005040's fMRIPrep pair> --inindex=2 --datain=<acq_x.txt: acquisition parameters written by the suite> --topup=tu --out=hifi --method=jac (tu_*: FSL's topup outputs of the case opt/x-phase-encoding)
Bit-identical

every file FSL writes (5/5) byte for byte, with FSL’s exit status and messages; in fix, its default, the replica corrects FSL-132 and names it: 18,851 values change

Open difference

Within the record’s thresholds (specs/topup.md §12): hifi (applytopup’s output): relative difference inside the brain, 99th percentile 5.26e-06 (§12: 0.001); hifi (applytopup’s output): mean |d| / mean intensity inside the brain 3.27e-07 (§12: 0.001). 1 of 1 files differ in their values (most: hifi.nii.gz: 17,354 of 20,480 values, at most 0.285): the clean room computes in its own order of operations (written from the specification, not from FSL’s code). FSL’s own band is not measured on this case (its thread band is, on fMRIPrep’s calls); in fix, its default, the clean room corrects FSL-132 and names it: 18,851 values change.

6/6 · 8/8
at/jac-quadraticapplytopupa field of quadratic splines
command lineapplytopup --imain=<half_up.nii.gz: derived from ds005040's fMRIPrep pair> --inindex=1 --datain=<acq.txt: acquisition parameters written by the suite> --topup=tu --out=hifi --method=jac (tu_*: FSL's topup outputs of the case opt/splineorder-2)
Bit-identical

every file FSL writes (6/6) byte for byte, with FSL’s exit status and messages; in fix, its default, the replica corrects FSL-132 and names it: 20,478 values change

Open difference

Within the record’s thresholds (specs/topup.md §12): hifi (applytopup’s output): relative difference inside the brain, 99th percentile 2.06e-06 (§12: 0.001); hifi (applytopup’s output): mean |d| / mean intensity inside the brain 1.33e-07 (§12: 0.001). 1 of 1 files differ in their values (most: hifi.nii.gz: 13,709 of 20,480 values, at most 0.258): the clean room computes in its own order of operations (written from the specification, not from FSL’s code). FSL’s own band is not measured on this case (its thread band is, on fMRIPrep’s calls); in fix, its default, the clean room corrects FSL-132 and names it: 20,478 values change.

6/6 · 8/8
at/jac-fieldapplytopup--topup naming a field image (topup's --fout): no movements
command lineapplytopup --imain=<half_up.nii.gz: derived from ds005040's fMRIPrep pair> --inindex=1 --datain=<acq.txt: acquisition parameters written by the suite> --topup=tu_field --out=hifi --method=jac (tu_*: FSL's topup outputs of the case opt/levels)
Bit-identical

every file FSL writes (6/6) byte for byte, with FSL’s exit status and messages; in fix, its default, the replica corrects FSL-132 and names it: 20,480 values change

Open difference

Within the record’s thresholds (specs/topup.md §12): hifi (applytopup’s output): relative difference inside the brain, 99th percentile 2.07e-06 (§12: 0.001); hifi (applytopup’s output): mean |d| / mean intensity inside the brain 1.36e-07 (§12: 0.001). 1 of 1 files differ in their values (most: hifi.nii.gz: 13,745 of 20,480 values, at most 0.242): the clean room computes in its own order of operations (written from the specification, not from FSL’s code). FSL’s own band is not measured on this case (its thread band is, on fMRIPrep’s calls); in fix, its default, the clean room corrects FSL-132 and names it: 20,480 values change.

6/6 · 8/8
at/jac-field-anisoapplytopupa field image and voxels of 6 x 4 mm in plane (the Jacobian of the y displacements)
command lineapplytopup --imain=<half_up_aniso.nii.gz: derived from ds005040's fMRIPrep pair> --inindex=1 --datain=<acq.txt: acquisition parameters written by the suite> --topup=tu_field --out=hifi --method=jac (tu_*: FSL's topup outputs of the case opt/levels)
Bit-identical

every file FSL writes (6/6) byte for byte, with FSL’s exit status and messages; in fix, its default, the replica corrects FSL-132, FSL-134, FSL-273 and names it: 20,480 values change

Open difference

Within the record’s thresholds (specs/topup.md §12): hifi (applytopup’s output): relative difference inside the brain, 99th percentile 1.01e-06 (§12: 0.001); hifi (applytopup’s output): mean |d| / mean intensity inside the brain 8.36e-08 (§12: 0.001). 1 of 1 files differ in their values (most: hifi.nii.gz: 12,694 of 20,480 values, at most 0.0781): the clean room computes in its own order of operations (written from the specification, not from FSL’s code). FSL’s own band is not measured on this case (its thread band is, on fMRIPrep’s calls); in fix, its default, the clean room corrects FSL-132, FSL-273, FSL-134 and names it: 20,480 values change.

7/7 · 9/9
at/jac-coef-anisoapplytopuptopup's coefficients (a field of 6 x 6 mm voxels in plane) on images of 6 x 4 mm voxels of the same shape: FSL scales the Jacobian's y term by 4/6 (FSL-273)
command lineapplytopup --imain=<half_up_aniso.nii.gz: derived from ds005040's fMRIPrep pair> --inindex=1 --datain=<acq.txt: acquisition parameters written by the suite> --topup=tu --out=hifi --method=jac (tu_*: FSL's topup outputs of the case opt/levels)
Bit-identical

every file FSL writes (6/6) byte for byte, with FSL’s exit status and messages; in fix, its default, the replica corrects FSL-132, FSL-273 and names it: 20,480 values change

Open difference

Within the record’s thresholds (specs/topup.md §12): hifi (applytopup’s output): relative difference inside the brain, 99th percentile 1.03e-06 (§12: 0.001); hifi (applytopup’s output): mean |d| / mean intensity inside the brain 8.49e-08 (§12: 0.001). 1 of 1 files differ in their values (most: hifi.nii.gz: 12,700 of 20,480 values, at most 0.0781): the clean room computes in its own order of operations (written from the specification, not from FSL’s code). FSL’s own band is not measured on this case (its thread band is, on fMRIPrep’s calls); in fix, its default, the clean room corrects FSL-132, FSL-273 and names it: 20,480 values change.

7/7 · 9/9
at/jac-no-movparapplytopupcoefficients without their _movpar.txt (no movements)
command lineapplytopup --imain=<half_up.nii.gz: derived from ds005040's fMRIPrep pair> --inindex=1 --datain=<acq.txt: acquisition parameters written by the suite> --topup=tu --out=hifi --method=jac (tu_*: FSL's topup outputs of the case opt/levels, without tu_movpar.txt)
Bit-identical

every file FSL writes (5/5) byte for byte, with FSL’s exit status and messages; in fix, its default, the replica corrects FSL-132 and names it: 20,480 values change

Open difference

Within the record’s thresholds (specs/topup.md §12): hifi (applytopup’s output): relative difference inside the brain, 99th percentile 2.05e-06 (§12: 0.001); hifi (applytopup’s output): mean |d| / mean intensity inside the brain 1.36e-07 (§12: 0.001). 1 of 1 files differ in their values (most: hifi.nii.gz: 13,708 of 20,480 values, at most 0.242): the clean room computes in its own order of operations (written from the specification, not from FSL’s code). FSL’s own band is not measured on this case (its thread band is, on fMRIPrep’s calls); in fix, its default, the clean room corrects FSL-132 and names it: 20,480 values change.

6/6 · 8/8
at/jac-obliqueapplytopupa phase-encoding vector along x and y
command lineapplytopup --imain=<half_up.nii.gz: derived from ds005040's fMRIPrep pair> --inindex=1 --datain=<acq_oblique.txt: acquisition parameters written by the suite> --topup=tu --out=hifi --method=jac (tu_*: FSL's topup outputs of the case opt/levels)
Bit-identical

every file FSL writes (6/6) byte for byte, with FSL’s exit status and messages; in fix, its default, the replica corrects FSL-132 and names it: 19,134 values change

Open difference

Within the record’s thresholds (specs/topup.md §12): hifi (applytopup’s output): relative difference inside the brain, 99th percentile 3.31e-06 (§12: 0.001); hifi (applytopup’s output): mean |d| / mean intensity inside the brain 2.2e-07 (§12: 0.001). 1 of 1 files differ in their values (most: hifi.nii.gz: 15,754 of 20,480 values, at most 0.203): the clean room computes in its own order of operations (written from the specification, not from FSL’s code). FSL’s own band is not measured on this case (its thread band is, on fMRIPrep’s calls); in fix, its default, the clean room corrects FSL-132 and names it: 19,134 values change.

6/6 · 8/8
at/jac-zapplytopupphase encoding along z (FSL-144: FSL leaves the z component out silently; the default refuses)
command lineapplytopup --imain=<half_up.nii.gz: derived from ds005040's fMRIPrep pair> --inindex=1 --datain=<acq_z.txt: acquisition parameters written by the suite> --topup=tu --out=hifi --method=jac (tu_*: FSL's topup outputs of the case opt/levels)
Bit-identical

every file FSL writes (6/6) byte for byte, with FSL’s exit status and messages; its default, fix, refuses the call (exit 1), naming FSL-144

Bit-identical

The same values, every one, in all 1 files FSL writes on this case, with FSL’s exit status and messages (the record checks the headers and the log); its default, fix, refuses the call (exit 1), naming FSL-144.

6/6 · 8/8
at/datatype-preserveapplytopup--datatype=preserve
command lineapplytopup --imain=<half_up.nii.gz: derived from ds005040's fMRIPrep pair> --inindex=1 --datain=<acq.txt: acquisition parameters written by the suite> --topup=tu --out=hifi --method=jac --datatype=preserve (tu_*: FSL's topup outputs of the case opt/levels)
Bit-identical

every file FSL writes (6/6) byte for byte, with FSL’s exit status and messages; in fix, its default, the replica corrects FSL-132 and names it: 20,480 values change

Open difference

Within the record’s thresholds (specs/topup.md §12): hifi (applytopup’s output): relative difference inside the brain, 99th percentile 2.05e-06 (§12: 0.001); hifi (applytopup’s output): mean |d| / mean intensity inside the brain 1.36e-07 (§12: 0.001). 1 of 1 files differ in their values (most: hifi.nii.gz: 13,708 of 20,480 values, at most 0.242): the clean room computes in its own order of operations (written from the specification, not from FSL’s code). FSL’s own band is not measured on this case (its thread band is, on fMRIPrep’s calls); in fix, its default, the clean room corrects FSL-132 and names it: 20,480 values change.

6/6 · 8/8
at/datatype-charapplytopup--datatype=char
command lineapplytopup --imain=<half_up.nii.gz: derived from ds005040's fMRIPrep pair> --inindex=1 --datain=<acq.txt: acquisition parameters written by the suite> --topup=tu --out=hifi --method=jac --datatype=char (tu_*: FSL's topup outputs of the case opt/levels)
Bit-identical

every file FSL writes (6/6) byte for byte, with FSL’s exit status and messages; in fix, its default, the replica corrects FSL-09, FSL-132 and names it: 18,118 values change

Open difference

Within the record’s thresholds (specs/topup.md §12): hifi (applytopup’s output): mean |d| / mean intensity inside the brain 2.68e-05 (§12: 0.001); hifi (applytopup’s output): relative difference inside the brain, 99th percentile 0 (§12: 0.001). 1 of 1 files differ in their values (most: hifi.nii.gz: 67 of 20,480 values, at most 255): the clean room computes in its own order of operations (written from the specification, not from FSL’s code). FSL’s own band is not measured on this case (its thread band is, on fMRIPrep’s calls); in fix, its default, the clean room corrects FSL-132, FSL-09 and names it: 18,117 values change.

6/6 · 8/8
at/datatype-shortapplytopup--datatype=short
command lineapplytopup --imain=<half_up.nii.gz: derived from ds005040's fMRIPrep pair> --inindex=1 --datain=<acq.txt: acquisition parameters written by the suite> --topup=tu --out=hifi --method=jac --datatype=short (tu_*: FSL's topup outputs of the case opt/levels)
Bit-identical

every file FSL writes (6/6) byte for byte, with FSL’s exit status and messages; in fix, its default, the replica corrects FSL-09, FSL-132 and names it: 18,349 values change

Open difference

Within the record’s thresholds (specs/topup.md §12): hifi (applytopup’s output): mean |d| / mean intensity inside the brain 6.7e-07 (§12: 0.001); hifi (applytopup’s output): relative difference inside the brain, 99th percentile 0 (§12: 0.001). 1 of 1 files differ in their values (most: hifi.nii.gz: 67 of 20,480 values, at most 1): the clean room computes in its own order of operations (written from the specification, not from FSL’s code). FSL’s own band is not measured on this case (its thread band is, on fMRIPrep’s calls); in fix, its default, the clean room corrects FSL-132, FSL-09 and names it: 18,349 values change.

6/6 · 8/8
at/datatype-intapplytopup--datatype=int
command lineapplytopup --imain=<half_up.nii.gz: derived from ds005040's fMRIPrep pair> --inindex=1 --datain=<acq.txt: acquisition parameters written by the suite> --topup=tu --out=hifi --method=jac --datatype=int (tu_*: FSL's topup outputs of the case opt/levels)
Bit-identical

every file FSL writes (6/6) byte for byte, with FSL’s exit status and messages; in fix, its default, the replica corrects FSL-09, FSL-132 and names it: 18,333 values change

Open difference

Within the record’s thresholds (specs/topup.md §12): hifi (applytopup’s output): mean |d| / mean intensity inside the brain 1.55e-07 (§12: 0.001); hifi (applytopup’s output): relative difference inside the brain, 99th percentile 0 (§12: 0.001). 1 of 1 files differ in their values (most: hifi.nii.gz: 67 of 20,480 values, at most 1): the clean room computes in its own order of operations (written from the specification, not from FSL’s code). FSL’s own band is not measured on this case (its thread band is, on fMRIPrep’s calls); in fix, its default, the clean room corrects FSL-132, FSL-09 and names it: 18,333 values change.

6/6 · 8/8
at/datatype-floatapplytopup--datatype=float
command lineapplytopup --imain=<half_up.nii.gz: derived from ds005040's fMRIPrep pair> --inindex=1 --datain=<acq.txt: acquisition parameters written by the suite> --topup=tu --out=hifi --method=jac --datatype=float (tu_*: FSL's topup outputs of the case opt/levels)
Bit-identical

every file FSL writes (6/6) byte for byte, with FSL’s exit status and messages; in fix, its default, the replica corrects FSL-132 and names it: 20,480 values change

Open difference

Within the record’s thresholds (specs/topup.md §12): hifi (applytopup’s output): relative difference inside the brain, 99th percentile 2.05e-06 (§12: 0.001); hifi (applytopup’s output): mean |d| / mean intensity inside the brain 1.36e-07 (§12: 0.001). 1 of 1 files differ in their values (most: hifi.nii.gz: 13,708 of 20,480 values, at most 0.242): the clean room computes in its own order of operations (written from the specification, not from FSL’s code). FSL’s own band is not measured on this case (its thread band is, on fMRIPrep’s calls); in fix, its default, the clean room corrects FSL-132 and names it: 20,480 values change.

6/6 · 8/8
at/datatype-doubleapplytopup--datatype=double
command lineapplytopup --imain=<half_up.nii.gz: derived from ds005040's fMRIPrep pair> --inindex=1 --datain=<acq.txt: acquisition parameters written by the suite> --topup=tu --out=hifi --method=jac --datatype=double (tu_*: FSL's topup outputs of the case opt/levels)
Bit-identical

every file FSL writes (6/6) byte for byte, with FSL’s exit status and messages; in fix, its default, the replica corrects FSL-132 and names it: 20,480 values change

Open difference

Within the record’s thresholds (specs/topup.md §12): hifi (applytopup’s output): relative difference inside the brain, 99th percentile 2.05e-06 (§12: 0.001); hifi (applytopup’s output): mean |d| / mean intensity inside the brain 1.36e-07 (§12: 0.001). 1 of 1 files differ in their values (most: hifi.nii.gz: 13,708 of 20,480 values, at most 0.242): the clean room computes in its own order of operations (written from the specification, not from FSL’s code). FSL’s own band is not measured on this case (its thread band is, on fMRIPrep’s calls); in fix, its default, the clean room corrects FSL-132 and names it: 20,480 values change.

6/6 · 8/8
at/datatype-preserve-int16applytopup--datatype=preserve of a 16-bit image
command lineapplytopup --imain=<half_up_i16.nii.gz: derived from ds005040's fMRIPrep pair> --inindex=1 --datain=<acq.txt: acquisition parameters written by the suite> --topup=tu --out=hifi --method=jac -d preserve (tu_*: FSL's topup outputs of the case opt/levels)
Bit-identical

every file FSL writes (6/6) byte for byte, with FSL’s exit status and messages; in fix, its default, the replica corrects FSL-09, FSL-132 and names it: 16,761 values change

Open difference

Within the record’s thresholds (specs/topup.md §12): hifi (applytopup’s output): mean |d| / mean intensity inside the brain 1.76e-07 (§12: 0.001); hifi (applytopup’s output): relative difference inside the brain, 99th percentile 0 (§12: 0.001). 1 of 1 files differ in their values (most: hifi.nii.gz: 19 of 20,480 values, at most 1): the clean room computes in its own order of operations (written from the specification, not from FSL’s code). FSL’s own band is not measured on this case (its thread band is, on fMRIPrep’s calls); in fix, its default, the clean room corrects FSL-132, FSL-09 and names it: 16,763 values change.

6/6 · 8/8
at/datatype-preserve-int8applytopup--datatype=preserve of an 8-bit signed image
command lineapplytopup --imain=<half_up_i8.nii.gz: derived from ds005040's fMRIPrep pair> --inindex=1 --datain=<acq.txt: acquisition parameters written by the suite> --topup=tu --out=hifi --method=jac --datatype=preserve (tu_*: FSL's topup outputs of the case opt/levels)
Bit-identical

every file FSL writes (6/6) byte for byte, with FSL’s exit status and messages; in fix, its default, the replica corrects FSL-09, FSL-132 and names it: 20,480 values change

Bit-identical

The same values, every one, in all 1 files FSL writes on this case, with FSL’s exit status and messages (the record checks the headers and the log); in fix, its default, the clean room corrects FSL-132, FSL-09 and names it: 12,788 values change.

6/6 · 8/8
at/short-optionsapplytopupthe short options
command lineapplytopup -i half_up.nii.gz,half_dn.nii.gz -x 1,2 -a acq.txt -t tu -o hifi -m lsr -v (with <half_up.nii.gz: derived from ds005040's fMRIPrep pair>, <half_dn.nii.gz: derived from ds005040's fMRIPrep pair>, <acq.txt: acquisition parameters written by the suite> in its directory) (tu_*: FSL's topup outputs of the case opt/levels)
Bit-identical

every file FSL writes (7/7) byte for byte, with FSL’s exit status and messages; in fix, its default, the replica corrects FSL-132 and names it: 5,136 values change

Open difference

Within the record’s thresholds (specs/topup.md §12): hifi (applytopup’s output): relative difference inside the brain, 99th percentile 1.19e-05 (§12: 0.001); hifi (applytopup’s output): mean |d| / mean intensity inside the brain 6.84e-07 (§12: 0.001). 1 of 1 files differ in their values (most: hifi.nii.gz: 16,831 of 20,480 values, at most 0.813): the clean room computes in its own order of operations (written from the specification, not from FSL’s code). FSL’s own band is not measured on this case (its thread band is, on fMRIPrep’s calls).

6/6 · 8/8
at/niftiapplytopupFSLOUTPUTTYPE=NIFTI
command lineFSLOUTPUTTYPE=NIFTI applytopup --imain=<half_up.nii.gz: derived from ds005040's fMRIPrep pair> --inindex=1 --datain=<acq.txt: acquisition parameters written by the suite> --topup=tu --out=hifi --method=jac (tu_*: FSL's topup outputs of the case opt/levels)
Bit-identical

every file FSL writes (6/6) byte for byte, with FSL’s exit status and messages; in fix, its default, the replica corrects FSL-132 and names it: 20,480 values change

Open difference

Within the record’s thresholds (specs/topup.md §12): hifi (applytopup’s output): relative difference inside the brain, 99th percentile 2.05e-06 (§12: 0.001); hifi (applytopup’s output): mean |d| / mean intensity inside the brain 1.36e-07 (§12: 0.001). 1 of 1 files differ in their values (most: hifi.nii: 13,708 of 20,480 values, at most 0.242): the clean room computes in its own order of operations (written from the specification, not from FSL’s code). FSL’s own band is not measured on this case (its thread band is, on fMRIPrep’s calls); in fix, its default, the clean room corrects FSL-132 and names it: 20,480 values change.

6/6 · 8/8
at/no-argumentsapplytopupno arguments: the compulsory options and the usage
command lineapplytopup (tu_*: FSL's topup outputs of the case opt/levels)
Rejected by both

applytopup exits 1; the same from the replica. *************************************************** The following COMPULSORY options have not been set: -i,–imain comma separated list of names of input image (to be corrected) -a,–datain name of te

Rejected by both

applytopup exits 1; the same from the clean room. *************************************************** The following COMPULSORY options have not been set: -i,–imain comma separated list of names of input image (to be corrected) -a,–datain name of te

5/5 · 4/4
at/helpapplytopup-h: the usage
command lineapplytopup -h (tu_*: FSL's topup outputs of the case opt/levels)
Rejected by both

applytopup exits 1; the same from the replica. Part of FSL (ID: 2412.6-dirty) applytopup (Version 1.0) Copyright© 2009, University of Oxford (Jesper Andersson) Usage: applytopup -in=topdn,botup --topup=mytu --inindex=1,2 --out=hifi applytopup -i

Rejected by both

applytopup exits 1; the same from the clean room. Part of FSL (ID: 2412.6-dirty) applytopup (Version 1.0) Copyright© 2009, University of Oxford (Jesper Andersson) Usage: applytopup -in=topdn,botup --topup=mytu --inindex=1,2 --out=hifi applytopup -i

5/5 · 4/4
at/unknown-optionapplytopupan unknown option
command lineapplytopup --imain=<half_up.nii.gz: derived from ds005040's fMRIPrep pair>,<half_dn.nii.gz: derived from ds005040's fMRIPrep pair> --inindex=1,2 --datain=<acq.txt: acquisition parameters written by the suite> --topup=tu --out=hifi --bogus=1 (tu_*: FSL's topup outputs of the case opt/levels)
Rejected by both

applytopup exits 1; the same from the replica. Part of FSL (ID: 2412.6-dirty) applytopup (Version 1.0) Copyright© 2009, University of Oxford (Jesper Andersson) Usage: applytopup -in=topdn,botup --topup=mytu --inindex=1,2 --out=hifi applytopup -i

Rejected by both

applytopup exits 1; the same from the clean room. Part of FSL (ID: 2412.6-dirty) applytopup (Version 1.0) Copyright© 2009, University of Oxford (Jesper Andersson) Usage: applytopup -in=topdn,botup --topup=mytu --inindex=1,2 --out=hifi applytopup -i

5/5 · 4/4
at/invertapplytopup--invert (declared, never added to the parser)
command lineapplytopup --imain=<half_up.nii.gz: derived from ds005040's fMRIPrep pair>,<half_dn.nii.gz: derived from ds005040's fMRIPrep pair> --inindex=1,2 --datain=<acq.txt: acquisition parameters written by the suite> --topup=tu --out=hifi --invert (tu_*: FSL's topup outputs of the case opt/levels)
Rejected by both

applytopup exits 1; the same from the replica. Part of FSL (ID: 2412.6-dirty) applytopup (Version 1.0) Copyright© 2009, University of Oxford (Jesper Andersson) Usage: applytopup -in=topdn,botup --topup=mytu --inindex=1,2 --out=hifi applytopup -i

Rejected by both

applytopup exits 1; the same from the clean room. Part of FSL (ID: 2412.6-dirty) applytopup (Version 1.0) Copyright© 2009, University of Oxford (Jesper Andersson) Usage: applytopup -in=topdn,botup --topup=mytu --inindex=1,2 --out=hifi applytopup -i

5/5 · 4/4
at/missing-outapplytopupno --out
command lineapplytopup --imain=<half_up.nii.gz: derived from ds005040's fMRIPrep pair>,<half_dn.nii.gz: derived from ds005040's fMRIPrep pair> --inindex=1,2 --datain=<acq.txt: acquisition parameters written by the suite> --topup=tu (tu_*: FSL's topup outputs of the case opt/levels)
Rejected by both

applytopup exits 1; the same from the replica. *************************************************** The following COMPULSORY options have not been set: -o,–out basename for output (warped) image ***************************************************

Rejected by both

applytopup exits 1; the same from the clean room. *************************************************** The following COMPULSORY options have not been set: -o,–out basename for output (warped) image ***************************************************

5/5 · 4/4
at/interp-unknownapplytopupan unknown --interp
command lineapplytopup --imain=<half_up.nii.gz: derived from ds005040's fMRIPrep pair>,<half_dn.nii.gz: derived from ds005040's fMRIPrep pair> --inindex=1,2 --datain=<acq.txt: acquisition parameters written by the suite> --topup=tu --out=hifi --interp=cubic (tu_*: FSL's topup outputs of the case opt/levels)
Rejected by both

applytopup exits -6; the same from the replica. terminate called after throwing an instance of ‘TOPUP::ApplyTopupException’ what(): ApplyTopup::Unknown interpolation type cubic

Rejected by both

applytopup exits -6; the same from the clean room. terminate called after throwing an instance of ‘TOPUP::ApplyTopupException’ what(): ApplyTopup::Unknown interpolation type cubic

5/5 · 4/4
at/lists-mismatchapplytopuptwo images and one index
command lineapplytopup --imain=<half_up.nii.gz: derived from ds005040's fMRIPrep pair>,<half_dn.nii.gz: derived from ds005040's fMRIPrep pair> --inindex=1 --datain=<acq.txt: acquisition parameters written by the suite> --topup=tu --out=hifi (tu_*: FSL's topup outputs of the case opt/levels)
Rejected by both

applytopup exits -6; the same from the replica. terminate called after throwing an instance of ‘TOPUP::ApplyTopupException’ what(): ApplyTopup::Mismatched --in and --inindex lists

Rejected by both

applytopup exits -6; the same from the clean room. terminate called after throwing an instance of ‘TOPUP::ApplyTopupException’ what(): ApplyTopup::Mismatched --in and --inindex lists

5/5 · 4/4
at/datatype-unknownapplytopupan unknown --datatype
command lineapplytopup --imain=<half_up.nii.gz: derived from ds005040's fMRIPrep pair> --inindex=1 --datain=<acq.txt: acquisition parameters written by the suite> --topup=tu --out=hifi --method=jac --datatype=long (tu_*: FSL's topup outputs of the case opt/levels)
Rejected by both

applytopup exits -6; the same from the replica. terminate called after throwing an instance of ‘TOPUP::ApplyTopupException’ what(): ApplyTopup::Unknown data type long

Rejected by both

applytopup exits -6; the same from the clean room. terminate called after throwing an instance of ‘TOPUP::ApplyTopupException’ what(): ApplyTopup::Unknown data type long

5/5 · 4/4
at/datatype-preserve-missingapplytopup--datatype=preserve of an image that does not exist
command lineapplytopup --imain=nothere.nii.gz --inindex=1 --datain=<acq.txt: acquisition parameters written by the suite> --topup=tu --out=hifi --datatype=preserve (tu_*: FSL's topup outputs of the case opt/levels)
Rejected by both

applytopup exits -6; the same from the replica. Image Exception : #63 :: No image files match: nothere terminate called after throwing an instance of ‘std::runtime_error’ what(): No image files match: nothere

Rejected by both

applytopup exits -6; the same from the clean room. Image Exception : #63 :: No image files match: nothere terminate called after throwing an instance of ‘std::runtime_error’ what(): No image files match: nothere

5/5 · 4/4
at/imain-missingapplytopupan image that does not exist
command lineapplytopup --imain=nothere.nii.gz --inindex=1 --datain=<acq.txt: acquisition parameters written by the suite> --topup=tu --out=hifi (tu_*: FSL's topup outputs of the case opt/levels)
Rejected by both

applytopup exits -6; the same from the replica. Image Exception : #63 :: No image files match: nothere Image Exception : #22 :: Failed to read volume nothere.nii.gz Error : No image files match: nothere terminate called after throwing an instance o

Rejected by both

applytopup exits -6; the same from the clean room. Image Exception : #63 :: No image files match: nothere Image Exception : #22 :: Failed to read volume nothere.nii.gz Error : No image files match: nothere terminate called after throwing an instance o

5/5 · 4/4
at/images-mismatchapplytopupimages of different sizes
command lineapplytopup --imain=<half_up.nii.gz: derived from ds005040's fMRIPrep pair>,<crop_dn.nii.gz: derived from ds005040's fMRIPrep pair> --inindex=1,2 --datain=<acq.txt: acquisition parameters written by the suite> --topup=tu --out=hifi (tu_*: FSL's topup outputs of the case opt/levels)
Rejected by both

applytopup exits -6; the same from the replica. terminate called after throwing an instance of ‘TOPUP::ApplyTopupException’ what(): ApplyTopup::Mismatched input files crop_dn.nii.gz and half_up.nii.gz

Rejected by both

applytopup exits -6; the same from the clean room. terminate called after throwing an instance of ‘TOPUP::ApplyTopupException’ what(): ApplyTopup::Mismatched input files crop_dn.nii.gz and half_up.nii.gz

5/5 · 4/4
at/datain-missingapplytopupa --datain that does not exist
command lineapplytopup --imain=<half_up.nii.gz: derived from ds005040's fMRIPrep pair>,<half_dn.nii.gz: derived from ds005040's fMRIPrep pair> --inindex=1,2 --datain=nothere.txt --topup=tu --out=hifi (tu_*: FSL's topup outputs of the case opt/levels)
Rejected by both

applytopup exits -6; the same from the replica. Could not open matrix file nothere.txt terminate called after throwing an instance of ‘TOPUP::TopupFileIOException’ what(): TopupFileIO::TopupDatafileReader:: error reading file nothere.txt

Rejected by both

applytopup exits -6; the same from the clean room. Could not open matrix file nothere.txt terminate called after throwing an instance of ‘TOPUP::TopupFileIOException’ what(): TopupFileIO::TopupDatafileReader:: error reading file nothere.txt

5/5 · 4/4
at/datain-normapplytopupa phase-encoding vector not of unit length
command lineapplytopup --imain=<half_up.nii.gz: derived from ds005040's fMRIPrep pair>,<half_dn.nii.gz: derived from ds005040's fMRIPrep pair> --inindex=1,2 --datain=<acq_norm.txt: acquisition parameters written by the suite> --topup=tu --out=hifi (tu_*: FSL's topup outputs of the case opt/levels)
Rejected by both

applytopup exits -6; the same from the replica. terminate called after throwing an instance of ‘TOPUP::TopupFileIOException’ what(): TopupFileIO::TopupDatafileReader:: phase-encode vectors must be unity length

Rejected by both

applytopup exits -6; the same from the clean room. terminate called after throwing an instance of ‘TOPUP::TopupFileIOException’ what(): TopupFileIO::TopupDatafileReader:: phase-encode vectors must be unity length

5/5 · 4/4
at/inindex-zeroapplytopupan index of 0
command lineapplytopup --imain=<half_up.nii.gz: derived from ds005040's fMRIPrep pair> --inindex=0 --datain=<acq.txt: acquisition parameters written by the suite> --topup=tu --out=hifi (tu_*: FSL's topup outputs of the case opt/levels)
Rejected by both

applytopup exits -6; the same from the replica. terminate called after throwing an instance of ‘TOPUP::ApplyTopupException’ what(): ApplyTopup::Invalid entry in --inindex list

Rejected by both

applytopup exits -6; the same from the clean room. terminate called after throwing an instance of ‘TOPUP::ApplyTopupException’ what(): ApplyTopup::Invalid entry in --inindex list

5/5 · 4/4
at/inindex-beyondapplytopupan index beyond --datain's rows
command lineapplytopup --imain=<half_up.nii.gz: derived from ds005040's fMRIPrep pair> --inindex=3 --datain=<acq.txt: acquisition parameters written by the suite> --topup=tu --out=hifi (tu_*: FSL's topup outputs of the case opt/levels)
Rejected by both

applytopup exits -6; the same from the replica. terminate called after throwing an instance of ‘TOPUP::ApplyTopupException’ what(): ApplyTopup::Invalid entry in --inindex list

Rejected by both

applytopup exits -6; the same from the clean room. terminate called after throwing an instance of ‘TOPUP::ApplyTopupException’ what(): ApplyTopup::Invalid entry in --inindex list

5/5 · 4/4
at/topup-extensionapplytopup--topup with an extension
command lineapplytopup --imain=<half_up.nii.gz: derived from ds005040's fMRIPrep pair>,<half_dn.nii.gz: derived from ds005040's fMRIPrep pair> --inindex=1,2 --datain=<acq.txt: acquisition parameters written by the suite> --topup=tu.nii.gz --out=hifi (tu_*: FSL's topup outputs of the case opt/levels)
Rejected by both

applytopup exits -6; the same from the replica. terminate called after throwing an instance of ‘TOPUP::TopupFileIOException’ what(): TopupFileIO::TopupFileReader::TopupFileReader: Filename must be given without extension, fname=tu.nii.gz

Rejected by both

applytopup exits -6; the same from the clean room. terminate called after throwing an instance of ‘TOPUP::TopupFileIOException’ what(): TopupFileIO::TopupFileReader::TopupFileReader: Filename must be given without extension, fname=tu.nii.gz

5/5 · 4/4
at/topup-missingapplytopupa --topup that does not exist
command lineapplytopup --imain=<half_up.nii.gz: derived from ds005040's fMRIPrep pair>,<half_dn.nii.gz: derived from ds005040's fMRIPrep pair> --inindex=1,2 --datain=<acq.txt: acquisition parameters written by the suite> --topup=nothere --out=hifi (tu_*: FSL's topup outputs of the case opt/levels)
Rejected by both

applytopup exits -6; the same from the replica. terminate called after throwing an instance of ‘TOPUP::TopupFileIOException’ what(): TopupFileIO::TopupFileReader::TopupFileReader: Unable to read file nothere

Rejected by both

applytopup exits -6; the same from the clean room. terminate called after throwing an instance of ‘TOPUP::TopupFileIOException’ what(): TopupFileIO::TopupFileReader::TopupFileReader: Unable to read file nothere

5/5 · 4/4
at/lsr-obliqueapplytopup--method=lsr with an oblique phase encoding
command lineapplytopup --imain=<half_up.nii.gz: derived from ds005040's fMRIPrep pair>,<half_dn.nii.gz: derived from ds005040's fMRIPrep pair> --inindex=1,2 --datain=<acq_oblique.txt: acquisition parameters written by the suite> --topup=tu --out=hifi (tu_*: FSL's topup outputs of the case opt/levels)
Rejected by both

applytopup exits -6; the same from the replica. terminate called after throwing an instance of ‘TOPUP::ApplyTopupException’ what(): ApplyTopup::Invalid phase-encode vector for least-squares restoration The replica’s default, fix, ends with exit status 1, naming FSL-132.

Rejected by both

applytopup exits -6; the same from the clean room. terminate called after throwing an instance of ‘TOPUP::ApplyTopupException’ what(): ApplyTopup::Invalid phase-encode vector for least-squares restoration

5/5 · 4/4
at/lsr-sameapplytopup--method=lsr of two images with the same phase encoding
command lineapplytopup --imain=<half_up.nii.gz: derived from ds005040's fMRIPrep pair>,<half_dn.nii.gz: derived from ds005040's fMRIPrep pair> --inindex=1,2 --datain=<acq_same.txt: acquisition parameters written by the suite> --topup=tu --out=hifi (tu_*: FSL's topup outputs of the case opt/levels)
Rejected by both

applytopup exits -6; the same from the replica. terminate called after throwing an instance of ‘TOPUP::ApplyTopupException’ what(): ApplyTopup::Invalid combination of phase-encode vectors for least-squares restoration The replica’s default, fix, ends with exit status 1, naming FSL-132.

Rejected by both

applytopup exits -6; the same from the clean room. terminate called after throwing an instance of ‘TOPUP::ApplyTopupException’ what(): ApplyTopup::Invalid combination of phase-encode vectors for least-squares restoration

5/5 · 4/4
at/lsr-oneapplytopup--method=lsr of one image
command lineapplytopup --imain=<half_up.nii.gz: derived from ds005040's fMRIPrep pair> --inindex=1 --datain=<acq.txt: acquisition parameters written by the suite> --topup=tu --out=hifi (tu_*: FSL's topup outputs of the case opt/levels)
Rejected by both

applytopup exits -6; the same from the replica. terminate called after throwing an instance of ‘TOPUP::ApplyTopupException’ what(): ApplyTopup::Invalid combination of phase-encode vectors for least-squares restoration The replica’s default, fix, ends with exit status 1, naming FSL-132.

Rejected by both

applytopup exits -6; the same from the clean room. terminate called after throwing an instance of ‘TOPUP::ApplyTopupException’ what(): ApplyTopup::Invalid combination of phase-encode vectors for least-squares restoration

5/5 · 4/4
at/outputtype-invalidapplytopupFSLOUTPUTTYPE=BOGUS
command lineFSLOUTPUTTYPE=BOGUS applytopup --imain=<half_up.nii.gz: derived from ds005040's fMRIPrep pair> --inindex=1 --datain=<acq.txt: acquisition parameters written by the suite> --topup=tu --out=hifi --method=jac (tu_*: FSL's topup outputs of the case opt/levels)
Rejected by both

applytopup exits 1; the same from the replica. ERROR:: Unrecognised value (BOGUS) of environment variable FSLOUTPUTTYPE Legal values are: NIFTI, NIFTI_PAIR, NIFTI2, NIFTI2_PAIR, NIFTI_GZ, NIFTI_PAIR_GZ, NIFTI2_GZ, NIFTI2_PAIR_GZ, NIFTI_ZST, NIFTI_ The replica’s default, fix, ends with exit status 1, naming FSL-132.

Rejected by both

applytopup exits 1; the same from the clean room. ERROR:: Unrecognised value (BOGUS) of environment variable FSLOUTPUTTYPE Legal values are: NIFTI, NIFTI_PAIR, NIFTI2, NIFTI2_PAIR, NIFTI_GZ, NIFTI_PAIR_GZ, NIFTI2_GZ, NIFTI2_PAIR_GZ, NIFTI_ZST, NIFTI_ The clean room’s default, fix, ends with exit status 1, naming FSL-132.

5/5 · 4/4
at/method-unknownapplytopupan unknown --method (FSL-135: FSL writes nothing and exits 0)
command lineapplytopup --imain=<half_up.nii.gz: derived from ds005040's fMRIPrep pair>,<half_dn.nii.gz: derived from ds005040's fMRIPrep pair> --inindex=1,2 --datain=<acq.txt: acquisition parameters written by the suite> --topup=tu --out=hifi --method=foo (tu_*: FSL's topup outputs of the case opt/levels)
Bit-identical

every file FSL writes (6/6) byte for byte, with FSL’s exit status and messages; its default, fix, refuses the call (exit 1), naming FSL-135

Bit-identical

FSL writes no file on this case; the exit status and messages are FSL’s; its default, fix, refuses the call (exit 1), naming FSL-135.

7/7 · 6/6
at/lsr-no-movementsapplytopup--method=lsr with a field image, which has no movements (FSL-135)
command lineapplytopup --imain=<half_up.nii.gz: derived from ds005040's fMRIPrep pair>,<half_dn.nii.gz: derived from ds005040's fMRIPrep pair> --inindex=1,2 --datain=<acq.txt: acquisition parameters written by the suite> --topup=tu_field --out=hifi (tu_*: FSL's topup outputs of the case opt/levels)
Rejected by both

applytopup exits -6; the same from the replica. terminate called after throwing an instance of ‘TOPUP::TopupFileIOException’ what(): TopupFileIO::TopupFileReader::MoveMatrix: Index i out of bound The replica’s default, fix, ends with exit status 1, naming FSL-135.

Rejected by both

applytopup exits -6; the same from the clean room. terminate called after throwing an instance of ‘TOPUP::TopupFileIOException’ what(): TopupFileIO::TopupFileReader::MoveMatrix: Index i out of bound The clean room’s default, fix, ends with exit status 1, naming FSL-135.

6/6 · 5/5
at/jac-other-gridapplytopup--method=jac with a field on another grid (FSL-134: another number of voxels)
command lineapplytopup --imain=<crop_up.nii.gz: derived from ds005040's fMRIPrep pair> --inindex=1 --datain=<acq.txt: acquisition parameters written by the suite> --topup=tu --out=hifi --method=jac (tu_*: FSL's topup outputs of the case opt/levels)
Rejected by both

applytopup exits -6; the same from the replica. Image Exception : #2 :: Attempted to copydata with non-matching sizes terminate called after throwing an instance of ‘std::runtime_error’ what(): Attempted to copydata with non-matching sizes The replica’s default, fix, ends with exit status 1, naming FSL-134.

Rejected by both

applytopup exits -6; the same from the clean room. Image Exception : #2 :: Attempted to copydata with non-matching sizes terminate called after throwing an instance of ‘std::runtime_error’ what(): Attempted to copydata with non-matching sizes The clean room’s default, fix, ends with exit status 1, naming FSL-134.

6/6 · 5/5
at/jac-other-shapeapplytopup--method=jac with a field of as many voxels as the images in another shape (FSL-134)
command lineapplytopup --imain=<half_up_64x16.nii.gz: derived from ds005040's fMRIPrep pair> --inindex=1 --datain=<acq.txt: acquisition parameters written by the suite> --topup=tu --out=hifi --method=jac (tu_*: FSL's topup outputs of the case opt/levels)
Bit-identical

every file FSL writes (6/6) byte for byte, with FSL’s exit status and messages; its default, fix, refuses the call (exit 1), naming FSL-134

Open difference

Within the record’s thresholds (specs/topup.md §12): hifi (applytopup’s output): relative difference inside the brain, 99th percentile 1.51e-06 (§12: 0.001); hifi (applytopup’s output): mean |d| / mean intensity inside the brain 1.06e-07 (§12: 0.001). 1 of 1 files differ in their values (most: hifi.nii.gz: 6,680 of 20,480 values, at most 0.102): the clean room computes in its own order of operations (written from the specification, not from FSL’s code). FSL’s own band is not measured on this case (its thread band is, on fMRIPrep’s calls); its default, fix, refuses the call (exit 1), naming FSL-134.

7/7 · 9/9
at/lsr-other-gridapplytopup--method=lsr with a field on another grid (FSL-134)
command lineapplytopup --imain=<crop_up.nii.gz: derived from ds005040's fMRIPrep pair>,<crop_dn.nii.gz: derived from ds005040's fMRIPrep pair> --inindex=1,2 --datain=<acq.txt: acquisition parameters written by the suite> --topup=tu --out=hifi (tu_*: FSL's topup outputs of the case opt/levels)
Expected divergence

FSL-134: refused; FSL is not run: FSL reads the field’s lines past their end (undefined behaviour)

Expected divergence

FSL-134: refused; FSL is not run: FSL reads the field’s lines past their end (undefined behaviour)

2/2
difftest/acquisition-transposedrandom testingan acquisition file of four rows and one column per volume (read transposed): the suite's own reading fixed (the drawn case)
command lineFSLOUTPUTTYPE=NIFTI_GZ topup --imain=<random testing's input case/fsl-topup-original/acquisition-transposed/pair> --datain=<random testing's input case/fsl-topup-original/acquisition-transposed/acq> --out=o --warpres=20,14 --subsamp=2,1 --fwhm=0,0 --miter=1,1 --lambda=0.005,5e-06 --estmov=1,1 --minmet=0,0 --interp=linear --regrid=0 --jacout=jac --logout=topup.log
–Open difference

Within the record’s thresholds (specs/topup.md §12): _fieldcoef (its field, Hz): |d| inside the brain, mean 0.00574 (§12: 0.1); _fieldcoef (its field, Hz): |d| inside the brain, 99th percentile 0.0177 (§12: 0.5); o_movpar.txt: rotations 2.3e-05 (§12: 0.001). 4 of 5 files differ in their values (most: jac_01.nii.gz: 6,398 of 6,400 values, at most 0.000102): the clean room computes in its own order of operations (written from the specification, not from FSL’s code). FSL’s own band is not measured on this case (its thread band is, on fMRIPrep’s calls); in fix, its default, the clean room corrects FSL-201 and names it: 13,042 values change.

18/18
difftest/configuration-outputsrandom testingoutputs named by a configuration file (FSL's test_acqp names --fout and --iout): the suite compares them under their configured names
command lineFSLOUTPUTTYPE=NIFTI_GZ topup --imain=<random testing's input case/fsl-topup-original/configuration-outputs/pair> --datain=<random testing's input case/fsl-topup-original/configuration-outputs/acq> --out=my_topup --config=test_acqp --miter=1
–Open difference

Within the record’s thresholds (specs/topup.md §12): iout_test: relative difference inside the brain, 99th percentile 1.11e-06 (§12: 0.001); --fout (Hz): |d| inside the brain, mean 4.97e-05 (§12: 0.1); _fieldcoef (its field, Hz): |d| inside the brain, mean 4.95e-05 (§12: 0.1). 4 of 6 files differ in their values (most: iout_test.nii.gz: 479 of 1,536 values, at most 0.000122): the clean room computes in its own order of operations (written from the specification, not from FSL’s code). FSL’s own band is not measured on this case (its thread band is, on fMRIPrep’s calls); in fix, its default, the clean room corrects FSL-146 and names it: 1,286 values change.

22/22
difftest/fsl145-short-axisrandom testingFSL-145: --splineorder=2 on an axis shorter than two knot spacings (two coefficients; reproduce FSL's, the default the full count)
command lineFSLOUTPUTTYPE=NIFTI_GZ topup --imain=<random testing's input case/fsl-topup-original/fsl145-short-axis/pair> --datain=<random testing's input case/fsl-topup-original/fsl145-short-axis/acq> --out=o --warpres=20 --subsamp=2 --fwhm=2 --miter=1 --lambda=0.005 --estmov=1 --splineorder=2 --fout=field
–Open difference

Within the record’s thresholds (specs/topup.md §12): _fieldcoef (its field, Hz): |d| inside the brain, mean 0.00244 (§12: 0.1); --fout (Hz): |d| inside the brain, mean 0.00224 (§12: 0.1); _fieldcoef (its field, Hz): |d| inside the brain, 99th percentile 0.01 (§12: 0.5). 3 of 4 files differ in their values (most: field.nii.gz: 2,700 of 3,072 values, at most 0.0263): the clean room computes in its own order of operations (written from the specification, not from FSL’s code). FSL’s own band is not measured on this case (its thread band is, on fMRIPrep’s calls); in fix, its default, the clean room corrects FSL-201, FSL-145 and names it: 3,152 values change.

16/16
difftest/fsl146-mixed-axesrandom testingFSL-146: volumes encoded along y and x (FSL's gradient leaves the volumes not encoded along an axis out of its Jacobian term; reproduce agrees with FSL; the drawn case)
command lineFSLOUTPUTTYPE=NIFTI_PAIR topup --imain=<random testing's input case/fsl-topup-original/fsl146-mixed-axes/pair> --datain=<random testing's input case/fsl-topup-original/fsl146-mixed-axes/acq> --out=o --warpres=20,14,12 --subsamp=2,1,1 --fwhm=8,3,2 --miter=2,3,2 --lambda=5e-05,5e-06,5e-06 --estmov=0,0,0 --minmet=0,1,0 --regmod=membrane_energy --fout=field
–Open difference

Within the record’s thresholds (specs/topup.md §12): --fout (Hz): |d| inside the brain, mean 1.58e-05 (§12: 0.1); _fieldcoef (its field, Hz): |d| inside the brain, mean 1.57e-05 (§12: 0.1); --fout (Hz): |d| inside the brain, 99th percentile 5.51e-05 (§12: 0.5). 2 of 4 files differ in their values (most: field.hdr: 9,199 of 9,216 values, at most 7.4e-05): the clean room computes in its own order of operations (written from the specification, not from FSL’s code). FSL’s own band is not measured on this case (its thread band is, on fMRIPrep’s calls); in fix, its default, the clean room corrects FSL-117, FSL-146 and names it: 19,020 values change.

24/24
difftest/quadratic-countrandom testingtopup --splineorder=2: FSL's quadratic B-spline count ceil((2n + k - 4) / (2k)) + 2 (the clean room had taken the cubic's; the random-testing run of 2026-11-03, the drawn case)
command lineFSLOUTPUTTYPE=NIFTI_GZ topup --imain=<random testing's input case/fsl-topup-original/quadratic-count/pair> --datain=<random testing's input case/fsl-topup-original/quadratic-count/acq> --out=o --warpres=20,16 --subsamp=2,1 --fwhm=2,0 --miter=3,4 --lambda=0.005,5e-06 --estmov=0,0 --minmet=0,0 --splineorder=2 --numprec=double --scale=1 --iout=corrected -v
–Open difference

Within the record’s thresholds (specs/topup.md §12): corrected: relative difference inside the brain, 99th percentile 3.56e-06 (§12: 0.001); corrected_mean: relative difference inside the brain, 99th percentile 1.55e-06 (§12: 0.001); _fieldcoef (its field, Hz): |d| inside the brain, 99th percentile 0.000148 (§12: 0.5). 3 of 5 files differ in their values (most: corrected.nii.gz: 11,687 of 15,360 values, at most 0.00121): the clean room computes in its own order of operations (written from the specification, not from FSL’s code). FSL’s own band is not measured on this case (its thread band is, on fMRIPrep’s calls).

21/21
difftest/scg-every-iteration-countsrandom testing--miter counts every iteration of the scaled conjugate gradients, the rejected ones too (a level of 3 iterations whose first step is rejected evaluates 4 costs, not 5)
command lineFSLOUTPUTTYPE=NIFTI_GZ topup --imain=<random testing's input case/fsl-topup-original/scg-every-iteration-counts/pair> --datain=<random testing's input case/fsl-topup-original/scg-every-iteration-counts/acq> --out=topup --config=<random testing's input case/fsl-topup-original/scg-every-iteration-counts/cfg> --logout=topup.log
–Open difference

Within the record’s thresholds (specs/topup.md §12): _fieldcoef (its field, Hz): |d| inside the brain, mean 0.0216 (§12: 0.1); _fieldcoef (its field, Hz): |d| inside the brain, 99th percentile 0.0441 (§12: 0.5); _fieldcoef (its field, Hz): correlation inside the brain 1 (§12: 0.999). 1 of 3 files differ in their values (most: topup_fieldcoef.nii.gz: 315 of 315 values, at most 0.0975): the clean room computes in its own order of operations (written from the specification, not from FSL’s code). FSL’s own band is not measured on this case (its thread band is, on fMRIPrep’s calls).

12/12
difftest/scg-regulariser-weightrandom testingscaled conjugate gradients at a level whose start has Reg > 0: the gradient at x + sigma d takes the current point's regulariser weight (FSL evaluates no cost there)
command lineFSLOUTPUTTYPE=NIFTI_GZ topup --imain=<random testing's input case/fsl-topup-original/scg-regulariser-weight/pair> --acqp=<random testing's input case/fsl-topup-original/scg-regulariser-weight/acq> --out=my_topup --warpres=20,16 --subsamp=1 --fwhm=8,4 --miter=2,1 --lambda=0.001 --estmov=0 --minmet=0,1 --regmod=membrane_energy --iout=corrected --fout=field --dfout=warpfield
–Open difference

Within the record’s thresholds (specs/topup.md §12): corrected: relative difference inside the brain, 99th percentile 1.33e-05 (§12: 0.001); corrected_mean: relative difference inside the brain, 99th percentile 6.01e-06 (§12: 0.001); corrected: mean |d| / mean intensity inside the brain 1.07e-06 (§12: 0.001). 6 of 8 files differ in their values (most: corrected.nii.gz: 11,465 of 15,360 values, at most 0.00037): the clean room computes in its own order of operations (written from the specification, not from FSL’s code). FSL’s own band is not measured on this case (its thread band is, on fMRIPrep’s calls); in fix, its default, the clean room corrects FSL-117 and names it: 4…

28/28
difftest/scg-rejected-last-steprandom testinga level of scaled conjugate gradients (--miter=1) whose one step is rejected: FSL counts it and ends the level at the rejected step's point, the next level starting from it (FSL-242, reproduce)
command lineFSLOUTPUTTYPE=NIFTI_GZ topup --imain=<random testing's input case/fsl-topup-original/scg-rejected-last-step/pair> --datain=<random testing's input case/fsl-topup-original/scg-rejected-last-step/acq> --out=topup --warpres=16,14,12 --subsamp=2,1,1 --fwhm=8,6,6 --miter=3,1,2 --lambda=0.0001,5e-05,5e-06 --estmov=0 --minmet=0,1,1 --ssqlambda=1 --regmod=membrane_energy --trace
–Open difference

Within the record’s thresholds (specs/topup.md §12): _fieldcoef (its field, Hz): |d| inside the brain, mean 0.00282 (§12: 0.1); _fieldcoef (its field, Hz): |d| inside the brain, 99th percentile 0.00564 (§12: 0.5); _fieldcoef (its field, Hz): correlation inside the brain 1 (§12: 0.999). 1 of 3 files differ in their values (most: topup_fieldcoef.nii.gz: 567 of 567 values, at most 0.00886): the clean room computes in its own order of operations (written from the specification, not from FSL’s code). FSL’s own band is not measured on this case (its thread band is, on fMRIPrep’s calls); in fix, its default, the clean room corrects FSL-117 and names it: 567 values change.

13/13
difftest/time-step-zero-seriesrandom testingtopup on a series of time step 0 (fMRIPrep's call: its --iout series compared whole): FSL writes the time step as its image layer reads it (its magnitude, 1 for a series of 0); the clean rooms wrote the stored value
command lineFSLOUTPUTTYPE=NIFTI_GZ topup --config=<random testing's input case/fsl-topup-original/time-step-zero-series/b02b0> --datain=<random testing's input case/fsl-topup-original/time-step-zero-series/acq> --imain=<random testing's input case/fsl-topup-original/time-step-zero-series/pair> --out=pair_base --iout=pair_corrected --fout=pair_field --jacout=jac --logout=pair_topup.log --rbmout=xfm --dfout=warpfield
–Open difference

Within the record’s thresholds (specs/topup.md §12): pair_corrected: relative difference inside the brain, 99th percentile 0.0209 (§12: 0.001); pair_corrected_mean: relative difference inside the brain, 99th percentile 0.0151 (§12: 0.001); pair_corrected: mean |d| / mean intensity inside the brain 0.00263 (§12: 0.001). 8 of 12 files differ in their values (most: pair_corrected.nii.gz: 7,839 of 10,240 values, at most 1.29): the clean room computes in its own order of operations (written from the specification, not from FSL’s code). FSL’s own band is not measured on this case (its thread band is, on fMRIPrep’s calls).

38/38
difftest/trace-verboserandom testing--trace with -v: FSL prints a stage's Entering line inside a cost line; the suite takes the stages out
command lineFSLOUTPUTTYPE=NIFTI_GZ topup --imain=<random testing's input case/fsl-topup-original/trace-verbose/pair> --datain=<random testing's input case/fsl-topup-original/trace-verbose/acq> --out=topup --warpres=16,14 --subsamp=1 --fwhm=2,2 --miter=2,1 --lambda=0.005,5e-06 --estmov=0 --minmet=0,1 --scale=1 --dfout=warpfield --verbose --trace
–Open difference

Within the record’s thresholds (specs/topup.md §12): warpfield_01 (mm): |d| inside the brain, mean 9.23e-05 (§12: 0.002); warpfield_02 (mm): |d| inside the brain, mean 9.23e-05 (§12: 0.002); _fieldcoef (its field, Hz): |d| inside the brain, mean 0.00461 (§12: 0.1). 3 of 5 files differ in their values (most: warpfield_01.nii.gz: 3,200 of 9,600 values, at most 0.000115): the clean room computes in its own order of operations (written from the specification, not from FSL’s code). FSL’s own band is not measured on this case (its thread band is, on fMRIPrep’s calls).

22/22

Speed

topup at one thread is how fMRIPrep runs it. With --nthr FSL runs threaded, and its result then changes (FSL-116). On fMRIPrep’s two calls the clean room takes 2.8 and 7.7 s on one thread, 10 to 12 times as fast as FSL at one thread, and 0.8 and 2.0 s on 14 threads, 39 to 41 times as fast, the same bits at both; FSL at --nthr=14 takes 12.3 and 27.4 s. applytopup’s least-squares restoration of ds005040’s pair takes the clean room 0.13 s on one thread and 0.04 s on 14, against FSL’s 0.42 s. The replica, a translation, runs at about FSL’s speed in topup (29.7 and 71.4 s) and takes 3.5 s for the restoration (the report). These timings are informal: one run each, on a machine running other work beside them. The clean room’s peak memory on fMRIPrep’s call on ds005040, as the memory test measures it at one thread, is 74 MiB, against FSL’s 88.

Speed, both ports against the original: FSL topup and applytopup 6.0.7.23

Seconds for the whole call (reading the images, estimating the field or restoring the images, writing every output), one run each, on a log scale. The factor is how many times faster the clean room is than FSL, both on one thread.

  • FSL 6.0.7.23, 1 thread (as fMRIPrep runs topup)
  • FSL 6.0.7.23, --nthr=14
  • replica, 1 thread
  • clean room, 1 thread
  • clean room, 14 threads
  1. fMRIPrep's topup call: ds005040's pairFSL 6.0.7.23, 1 thread (as fMRIPrep runs topup) 32.6 sFSL 6.0.7.23, --nthr=14 12.3 sreplica, 1 thread 29.7 sclean room, 1 thread 2.83 sclean room, 14 threads 0.8 s12×
  2. fMRIPrep's topup call: ds001600's pairFSL 6.0.7.23, 1 thread (as fMRIPrep runs topup) 77.3 sFSL 6.0.7.23, --nthr=14 27.4 sreplica, 1 thread 71.3 sclean room, 1 thread 7.7 sclean room, 14 threads 1.98 s10×
  3. applytopup --method=lsr: ds005040's pair, FSL's topup resultFSL 6.0.7.23, 1 thread (as fMRIPrep runs topup) 0.42 sreplica, 1 thread 3.47 sclean room, 1 thread 0.13 sclean room, 14 threads 0.04 s3.2×
0.01 s0.1 s1 s10 s100 s

From the benchmark report docs/benchmarks/fsl-topup.json (AMD Ryzen 9 7945HX, 32 logical CPUs, shared with other work: informal, one run each). The replica’s files are FSL’s one-thread files byte for byte; the clean room’s are within its record’s thresholds and the same bits at 1 and 14 threads. FSL at --nthr=14 changes its result (FSL-116).

Corrections by default

The replica corrects FSL’s deterministic defects by default, fix, and warns. upstream_defects="reproduce" (on the command line --upstream-defects reproduce from the next release, LARMORX_UPSTREAM_DEFECTS=reproduce in 0.1.0rc1) gives FSL’s behaviour byte for byte, where FSL’s behaviour is defined (register).

In topup:

  • Above one thread (--nthr, or FSL_NUM_THREADS for some of its matrices), FSL’s result changes with the thread count. The replica gives FSL’s one-thread result at any count (FSL-116, never reproduced).
  • The membrane energy’s Hessian in the cost’s units (FSL-117, above).
  • With the movements estimated, the exact derivatives with respect to the rotations, where FSL takes forward differences of single-precision rotation matrices (FSL-201, since 2026-11-03, above).
  • With --regrid=1 and subsamplings whose largest does not divide the others, FSL stops at the first level with the smaller one; the replica’s grid grows by their least common multiple (FSL-118).
  • The field maps --fout and --featout write keep the input’s intent parameters in place (FSL-119: FSL shifts them by one).
  • A row-count mismatch of --acqp names the --acqp file (FSL-120).
  • With FSLDIR unset and a configuration name not found, a clear error instead of FSL’s exception (FSL-121).
  • A configuration line longer than 1023 characters after its first word no longer ends the reading silently (FSL-122).
  • A negative --miter, --fwhm or --lambda, which FSL uses as given (a negative --fwhm smooths nothing), is refused (FSL-123, FSL-125, FSL-126). A negative --warpres is refused in fix and in reproduce (FSL-124: FSL’s knot spacing wraps and it exhausts the memory).

In applytopup:

  • Every volume is resampled with the interpolation asked for (FSL-132, above).
  • The least-squares restoration wraps a negative index around its line (FSL-133: FSL wraps it by 232 modulo the line’s length, wrong for lengths that do not divide 232, such as 30 or 96).
  • --method=jac with a field image scales the displacement along y by the y voxel size (FSL-134: FSL takes the x one). With a field whose voxel sizes differ from the images’, it modulates by the Jacobian of the displacement applied (FSL-273, 2026-11-06: FSL divides the field’s derivatives by the field’s own voxel sizes). A field on another grid is refused, in fix and in reproduce, where FSL reads past its lines.
  • An unknown --method, which FSL ignores silently (exit status 0, no output), and lsr or vb* without movements, where FSL aborts, are refused with a reason (FSL-135).
  • The variational methods count values in 64-bit integers (FSL-141: FSL’s 32-bit ones wrap on very large images).
  • vb3D and vb4D count the data points of every image (FSL-142) and build their prior with its full weight (FSL-143, above).

Five defects of FSL’s libraries that topup and applytopup never reach are in the register too (FSL-127 to FSL-131).

The clean room makes the same corrections by default, with the same notices, each named where it changes an output: on lsr and vb3D here, FSL-132’s correction changes none of the clean room’s values, so its notice is FSL-143’s alone on vb3D. Its reproduce follows FSL where FSL’s behaviour is defined. Since 2026-11-05 that includes FSL-201’s forward differences, FSL-133’s index on lines whose length does not divide 232 and FSL-134’s field in another shape, which the specification describes since its answers of 2026-11-03; its default restores those lines circularly and refuses that field. It adds four corrections that black-box runs found:

  • --method=jac with a phase encoding along z leaves the image uncorrected silently, which the default refuses (FSL-144);
  • --splineorder=2 lays two coefficients on an axis shorter than two knot spacings, so the field is forced towards 0 at its far end; the default lays the general count (FSL-145);
  • with volumes encoded along different axes, FSL’s gradient leaves out of each axis’ term the volumes not encoded along it; the default minimises with the cost’s own gradient (FSL-146: the default’s final cost was as low as FSL’s or lower in 86 to 89 % of the differing four-volume cases of the random-testing run of 2026-11-03);
  • with scaled conjugate gradients, a level whose last step is rejected ends at that worse point; the default ends at the lowest cost the level found (FSL-242, 2026-11-05).

The replica has the first three, with their switches, since 2026-11-03: its reproduce lays FSL’s count exactly, one coefficient on an axis shorter than one spacing, and gives FSL’s error when a level refits a single coefficient.

Known limits

Both ports are experimental. Each record’s standard comparison holds fMRIPrep’s calls on its two fMRIPrep runs. The full comparisons, the data panel’s other PEPOLAR runs, the long random-testing runs and FSL’s own band on every case come in the final validation phase.

The clean room is held to thresholds, not to FSL’s bytes. Where the movements are estimated and do not converge, it ends elsewhere than FSL even in reproduce, which takes FSL’s forward differences (FSL-201) along a path FSL’s own last bits move. On fits the data hardly determine (no regularisation, a few hundred voxels at a subsampled level) it can reach another minimum, as low as FSL’s.

The questions it left for the specification until 2026-11-03 (specs/topup.md §13) are followed since 2026-11-05. The catch-up’s random-testing run left new ones:

  • the quadratic splines’ bending energy on an axis of FSL-145’s two coefficients (§4.6);
  • the smoothing on an axis shorter than its kernel (§4.4);
  • a field image’s Jacobian scales when its voxel sizes differ from the images’ (§7.3), answered on 2026-11-07 (above: FSL-273).

The replica refuses, in fix and in reproduce and with a message, applytopup’s restorations of lines of fewer than 8 or more than 256 voxels along the phase encoding, and band systems that need LAPACK’s blocked band factorisation: FSL’s arithmetic there follows other paths of OpenBLAS and Armadillo, which the replica does not reproduce yet.

Neither port reproduces FSL’s results above one thread: they depend on the thread count (FSL-116).

The cross-platform checks (how they work) hold the other five platforms to LARMORX’s own Linux output. The clean room’s cover two levels with every output, the other options, FSL-117 and FSL-145 reproduced, applytopup’s lsr, jac and vb methods, FSL-132 reproduced, and the Python functions. The benchmark is informal: one run each, on a shared machine.

History

  • 2026-11-03: the replica, translated from FSL 6.0.7.23’s source, and its random-testing run, 14,460 calls in 27 minutes. The clean room, written from specs/topup.md, the papers, FSL’s user guides and runs of FSL, and its first random-testing run, 14,400 calls in 11 minutes. The specification’s answers to the clean room’s questions the same day (§13): FSL-201 registered and corrected by default in the replica; the replica’s switches for FSL-144, FSL-145 and FSL-146.
  • 2026-11-04: the clean room’s reproduce computes exp as glibc 2.35 does (GLIBC-07); no value of the record moved.
  • 2026-11-05: the clean-room catch-up. The specification’s answers of 2026-11-03 are in the clean room since, and every example was run again; its random-testing run of 17,535 calls in 30 minutes found FSL-242, corrected by default.
  • 2026-11-06: the replica’s record made again, 177 cases, at/jac-coef-aniso among them (FSL-273, an answer written for PRELUDE’s specification).
  • 2026-11-07: the clean room’s record made again, 187 cases: at/jac-coef-aniso, with the specification’s answer on applytopup’s Jacobian (FSL-273, §7.3), and random testing’s case difftest/time-step-zero-series.

Data credits

The data this page shows or names: each dataset with its version, DOI and licence, from LARMORX's test-data catalog. Public-domain dedications (CC0, PDDL) ask nothing; the datasets are cited all the same.