Evidence · AFNI
3dUnifize
3dUnifize makes a T1w’s white matter uniform (or, with -T2, a T2w’s or an EPI image’s). LARMORX has two ports of it, both in larmorx itself. The replica gives AFNI 26.3.05’s bytes in all 97 compared cases of its record that write outputs, fMRIPrep’s own calls included. The clean room gives them in 44 of its 99, and in the other 55 once AFNI uses correctly rounded maths: a few squashed values differ in their last bit where glibc’s tanhf is not correctly rounded.
- Ports
- replica
larmorxApache-2.0, with CC BY 4.0's attribution for the Medical College of Wisconsin's code: translated from AFNI's source; inpip install larmorxsince 2026-10-29, until then the separate GPL packagelarmorx-gpl - clean room
larmorxApache-2.0: written from specs/3dUnifize.md
- replica
- Status
- replica validated; clean room validated (marked before the statuses' definition)
- Records
- the replica's: standard comparison (AFNI 26.3.05) · the clean room's: standard comparison (AFNI 26.3.05) · findings afni-masks-clean-room
- Reference
3dUnifize· API page afni-unifize
What is compared
3dUnifize divides each voxel by the local white-matter level, a mean of the upper percentiles of the masked values in a ball around it (-Urad, 18.3 voxels across by default, whatever the voxel size; on a half-size copy above a million voxels), scales white matter to 1000 and squashes the brightest values. With -T2 it inverts the contrast first and back after; -GM also scales grey matter; -EPI works volume by volume. fMRIPrep calls it once per BOLD run, with -T2, on its skull-stripped BOLD reference, before the second skull strip (3dAutomask).
The two ports run beside AFNI 26.3.05’s 3dUnifize and AFNI 25.2.09’s on the same files with the same arguments, both AFNI releases the repository’s builds, at one and six OpenMP threads and with correctly rounded maths; fMRIPrep’s own output is the file it wrote in its run.
larmorx afni 3dUnifize -clfrac 0.2 -rbt 18.3 65.0 90.0 -prefix uni.nii.gz -T2 -input ref_brain.nii.gzimport larmorx as lx
uni = lx.afni.unifize("ref_brain.nii.gz", **lx.afni.fmriprep_unifize_options(), n_threads=6)
t1 = lx.afni.unifize("sub-01_T1w.nii.gz", save_scale=True)The calls above leave implementation at "auto". The API page maps every option. How to read what follows: how to read the evidence.
Its two ports
3dUnifize’s results depend on AFNI library routines copyrighted by the Medical College of Wisconsin, GPL-2 until AFNI 25.2.09, so LARMORX has two ports of it. Since 2026-10-29 both are part of LARMORX itself (pip install larmorx). AFNI 26.3.05 releases the Medical College of Wisconsin’s code, which these programs’ results depend on, under CC BY 4.0. So the replica moved from the separate GPL package larmorx-gpl into LARMORX, under the Apache License with CC BY 4.0’s attribution (NOTICE, licensing), and runs in the same process. The replica is translated from AFNI’s source. The clean room is written without it, from a description of what the program does (specs/3dUnifize.md).
implementation="auto" | "replica" | "clean-room" chooses between them ("original" in 0.1.0rc1). On the command line, larmorx --implementation clean-room afni 3dUnifize ... does the same from the next release; in 0.1.0rc1 it is the environment variable LARMORX_IMPLEMENTATION. auto takes the replica (before 2026-10-29 only where larmorx-gpl was installed, the clean room otherwise). The examples below were computed with the replica as the separate program it was then; its records, made again in the same process on 2026-10-29, agree case for case.
The C library’s functions. The squash calls tanhf: the replica carries fdlibm’s, as glibc 2.35 runs it, checked against glibc on all 232 float inputs. -EPI’s radius calls cbrtf, whose glibc code is not part of LARMORX: the replica computes it correctly rounded and warns on the voxel sizes where the cube root’s last bit could change the result (2,953 of the float voxel volumes between 1 and 100 mm³; on 80 of them glibc’s last bit changes the ball of neighbours). The clean room uses a correctly rounded tanhf (a port of CORE-MATH’s), as LARMORX’s own maths does wherever it is not following glibc on purpose.
The clean room’s speed. Its ball percentiles are its speed: instead of gathering and sorting some 25,000 values at every voxel, it slides a window of counts per rank from voxel to voxel, changing only the ball’s two surfaces, and reads the values between the two percentiles in ascending order, which gives the specification’s sum exactly. Planes run in parallel, with the same bytes at any thread count.
The verdicts at a glance
3dUnifize 26.3.05 (standard comparison, 127 cases of the record), by category and verdictSelect a segment to list its cases below.
3dUnifize 26.3.05 (standard comparison, 129 cases of the record), by category and verdict- Bit-identical 49
- Identical with correctly rounded maths 55
- Deliberate correction 2
- Expected divergence 4
- Rejected by both 19
Select a segment to list its cases below.
The clean room’s record runs 129 cases and the replica’s 127, both against AFNI 26.3.05 (the version step below), with AFNI’s exit status and messages in every case. The replica gives AFNI’s bytes in every case that writes an output, 97 of 97. The clean room gives them in 44 of the 99 compared cases that write outputs. In the other 55 its headers are AFNI’s and a few squashed values (above 1300) differ in their last bit, at most 1,633 values of a T1w and 8.7 × 10−8 of the largest value, against a threshold of 10−5: AFNI’s squash calls glibc’s tanhf, which is not correctly rounded on some arguments, the clean room’s is. Run with CORE-MATH’s correctly rounded functions preloaded, AFNI gives the clean room’s bytes in all 55 (identical with correctly rounded maths). AFNI computes every voxel’s local level independently in its OpenMP loop: its result does not depend on the thread count, and neither does either port’s. Where AFNI crashes (AFNI-40, AFNI-46), both stop with an error instead; where AFNI reads past the end of its half-size scale image (AFNI-47, a volume of more than a million voxels only two slices thick), AFNI crashed in our runs and both define the result.
Since 2026-10-29 both ports follow AFNI 26.3.05, AFNI’s newest release, and their records run against it, with AFNI 25.2.09 run beside it on every case (AFNI’s own step): the replica since the version update, the clean room since its alignment the same day. Between the two releases these programs changed in exactly two places, both corrections LARMORX already made by default: header slice code 2 read as NIfTI defines it (AFNI-03), and 3dTshift’s -no_detrend de-meaning every voxel (AFNI-05). In reproduce both ports now give 26.3.05’s result there and name neither. The examples below show both versions’ outputs: on fMRIPrep’s calls they are the same bytes.
In fMRIPrep: the uniformization of a BOLD reference
fMRIPrep’s own call on one of its runs, given to AFNI at one and six OpenMP threads and to the replica at one and six threads in fix and reproduce, beside the output fMRIPrep produced in its run; then to the clean room at one and six threads in fix and reproduce, and to AFNI with CORE-MATH’s correctly rounded functions preloaded.
Worked example
fMRIPrep's uniformization of a real BOLD reference
What this shows: On fMRIPrep’s call the replica gives AFNI’s values, every one, in both releases and on 1 and 6 threads, and fMRIPrep’s own output. The clean room gives AFNI’s values with correctly rounded maths; AFNI as built differs from it only where glibc’s tanhf is not correctly rounded (largest difference 0.00012). No correction by default changes this call.
- Replica against AFNI 3dUnifize 26.3.05Bit-identical
- Clean room against AFNI 3dUnifize 26.3.05Identical with correctly rounded maths
The 3dUnifize node fMRIPrep 25.2.6 ran on the BOLD reference of OpenNeuro ds005040 sub-002 (task clip, run 1), N4-corrected and skull-stripped once: its arguments (-T2: the contrast inverted, unifized, inverted back) given to AFNI 25.2.09 at 1 and 6 OpenMP threads, to AFNI 26.3.05 at 1 (both ports’ reference since 2026-10-29), to the replica and to LARMORX’s clean room at 1 and 6 threads in fix and in reproduce. fMRIPrep’s own output is the panel it produced in its run.
- Data
- OpenNeuro ds005040, sub-002 (CC0-1.0): 64 × 64 × 40 float32 values; shown inside the input’s own mask, dilated by one voxel
- Node
- fMRIPrep 25.2.6,
bold_fit_wf/enhance_and_skullstrip_bold_wf/unifize - Command
3dUnifize -clfrac 0.2 -rbt 18.3 65.0 90.0 -prefix uni.nii.gz -T2 -input …/sub-002_task-clip_run-1_desc-hmc_boldref_corrected_brain_masked.nii.gz- Programs
- AFNI_25.2.09, gcc (Ubuntu 11.4.0-1ubuntu1~22.04.3) 11.4.0 (the repository’s build; OpenMP); AFNI AFNI_25.2.09 (Aug 24 2025) [64-bit] in fMRIPrep’s image; larmorx-gpl 0.0.1 (the replica, with fdlibm’s tanhf as glibc 2.35 runs it), from LARMORX at commit
ef4d5685a291; LARMORX 0.0.1 (the clean room) at commitca2145f7951e(a correctly rounded tanhf) - Output
- 43,338 non-zero values (AFNI’s final automask), up to 1499.47
- Grid
- 64 × 64 × 40 voxels of 3 × 3 × 3.3 mm; shown: 53 × 62 × 40 around the brain
- AFNI versions
- AFNI 26.3.05 (both ports’ reference since 2026-10-29) on the same arguments as AFNI 25.2.09: the same bytes in every output of fMRIPrep’s call. What changed between the two releases in these programs (header slice code 2, AFNI-03;
-no_detrend, AFNI-05; both corrected by LARMORX before AFNI fixed them) does not meet this call.
Data shown
- Political Moralization (OpenNeuro ds005040, version 1.2.0, doi:10.18112/openneuro.ds005040.v1.2.0), CC0 1.0 public-domain dedication.




Bit-identical163,840 of 163,840 voxels equal
Bit-identical163,840 of 163,840 voxels equal

Axial slice 19 of 40; the crosshair at the brain's centre. 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 reason | max |Δ| | Values that differ | RMS of Δ | max |Δ| ÷ range | 1 − NCC |
|---|---|---|---|---|---|
| The original against itself: the yardstick | |||||
| AFNI 3dUnifize 25.2.09, 6 threads − 1 threadreferenceAFNI computes every voxel’s local mean independently in its OpenMP loop: the same bytes at 1 and 6 threads. | identical: 0 of 163,840 values differ | ||||
| AFNI 3dUnifize 26.3.05 − 25.2.09, 1 thread eachreferenceAFNI 26.3.05 gives AFNI 25.2.09’s bytes on this call: the version step changes nothing here (its two changes, AFNI-03 and AFNI-05, need header slice code 2 or 3dTshift’s -no_detrend). | identical: 0 of 163,840 values differ | ||||
| fMRIPrep's output − AFNI 3dUnifize 25.2.09referencefMRIPrep’s own run (its image’s build of AFNI) gives the repository’s AFNI result, every value. | identical: 0 of 163,840 values differ | ||||
| The replica against the original and fMRIPrep's output | |||||
| replica − AFNI 3dUnifize 25.2.09, 1 thread eachBit-identicalThe same values, every one. | identical: 0 of 163,840 values differ | ||||
| replica − AFNI 3dUnifize 26.3.05, 1 thread eachBit-identicalThe same values, every one. | identical: 0 of 163,840 values differ | ||||
| replica − fMRIPrep's outputBit-identicalThe same values, every one. | identical: 0 of 163,840 values differ | ||||
| The replica against itself: thread counts, fix and reproduce | |||||
| replica, 6 threads − 1 threadBit-identicalThe same bytes on any number of threads. | identical: 0 of 163,840 values differ | ||||
replica, fix − reproduceBit-identicalNo correction of the register applies to this input: fix and reproduce give the same bytes. | identical: 0 of 163,840 values differ | ||||
| The clean room against the original, fMRIPrep's output and the replica | |||||
| clean room − AFNI 3dUnifize 25.2.09, 1 thread eachIdentical with correctly rounded mathsAFNI’s squash calls glibc’s tanhf, which is not correctly rounded on some arguments; the clean room’s is. With CORE-MATH’s correctly rounded functions preloaded, AFNI gives the clean room’s bytes, every value. | 0.000122 | 7 of 163,840 (0.00427 %) | 7.98e−7 | 8.14e−8 | 2.79e−18 |
| clean room − AFNI 3dUnifize 26.3.05, 1 thread eachIdentical with correctly rounded mathsAFNI 26.3.05 gives AFNI 25.2.09’s bytes on this call, so this is the comparison with 25.2.09: AFNI’s squash calls glibc’s tanhf, which is not correctly rounded on some arguments; the clean room’s is. With CORE-MATH’s correctly rounded functions preloaded, AFNI gives the clean room’s bytes, every value. | 0.000122 | 7 of 163,840 (0.00427 %) | 7.98e−7 | 8.14e−8 | 2.79e−18 |
| clean room − AFNI 3dUnifize with correctly rounded mathsBit-identicalThe same values, every one: AFNI with correctly rounded maths. | identical: 0 of 163,840 values differ | ||||
| clean room − fMRIPrep's outputIdentical with correctly rounded mathsfMRIPrep’s own output is AFNI’s; the clean room differs from it only where glibc’s tanhf is not correctly rounded (AFNI with correctly rounded maths gives the clean room’s bytes). | 0.000122 | 7 of 163,840 (0.00427 %) | 7.98e−7 | 8.14e−8 | 2.79e−18 |
| clean room − replica, 1 thread eachIdentical with correctly rounded mathsThe replica carries glibc 2.35’s tanhf, the clean room a correctly rounded one: they differ exactly where AFNI and AFNI with correctly rounded maths do. | 0.000122 | 7 of 163,840 (0.00427 %) | 7.98e−7 | 8.14e−8 | 2.79e−18 |
| The clean room against itself: thread counts, fix and reproduce | |||||
clean room, fix − reproduceBit-identicalNo correction of the register applies to this input: fix and reproduce give the same bytes. | identical: 0 of 163,840 values differ | ||||
| clean room, 6 threads − 1 threadBit-identicalThe clean room’s result does not depend on its thread count. | identical: 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 range is B’s maximum minus its minimum. The records’ thresholds: every output bit-identical to AFNI’s (the replica), or float32 values within 1e-5 of AFNI’s largest with bit-identity reported (the clean room); with their exit status and messages. The rows the original against itself are AFNI’s own variability on this input: its thread counts and builds.
Provenance of this example
- Data: OpenNeuro ds005040, sub-002, CC0-1.0
- Inputs: BOLD reference, masked (sha256 76db33db4206571d…); fMRIPrep’s output (sha256 d6d86af0e97d73b7…)
- Runs: figures cache entry
afni-unifize/fmriprep-ds005040-clip-1, computed 2026-10-23T11:06:34+00:00 with larmorxef4d5685a291; AFNI_25.2.09, gcc (Ubuntu 11.4.0-1ubuntu1~22.04.3) 11.4.0; AFNI_25.2.09 (Aug 24 2025) [64-bit]; larmorx-gpl 0.0.1; 7f84379b40f399876909ecf7c236105d6b7c43fdec5111fb7b03f8dee204ac10; AFNI_26.3.05; every command, its environment, threads and exit status are in itsRUN_INFO.json - Runs of the clean room: figures cache entry
afni-unifize/fmriprep-ds005040-clip-1-original, computed 2026-10-23T23:46:10+00:00 with larmorxca2145f7951e(fix, the default at 1 and 6 threads,reproduceat 1)
Every number above is in afni-unifize.json, exported by python -m larmorx_validation evidence afni-unifize.
A T1w, on the half-size path
Above a million voxels, 3dUnifize computes the local white-matter level on a half-size copy and brings it back; choose Scale image above the panels for the factor each voxel is multiplied by.
Worked example
AFNI's default on a real T1w
What this shows: On a real T1w, with no options, the replica gives AFNI’s values, every one, in both releases and on 1 and 6 threads, its output and its scale image. The clean room gives AFNI’s values with correctly rounded maths; AFNI as built differs from it only where glibc’s tanhf is not correctly rounded (largest difference 0.00012). No correction by default changes this call.
- Replica against AFNI 3dUnifize 26.3.05Bit-identical
- Clean room against AFNI 3dUnifize 26.3.05Identical with correctly rounded maths
3dUnifize’s main use: a T1w’s white matter made uniform and scaled to 1000. The T1w of OpenNeuro ds000005 sub-01 has 5,898,240 voxels, more than a million: the local white-matter level is computed on a half-size copy and brought back (AFNI’s duplo path). AFNI 25.2.09 at 1 and 6 threads, fMRIPrep’s build of AFNI, the replica and the clean room in fix and in reproduce; -ssave writes the scale image each voxel is multiplied by.
- Data
- OpenNeuro ds000005 (Tom et al., 2007), snapshot 00001, sub-01’s T1w (PDDL-1.0); shown inside fMRIPrep’s brain mask, dilated by 2 voxels
- Command
3dUnifize -ssave scale.nii.gz -prefix uni.nii.gz -input sub-01_T1w.nii.gz- Programs
- AFNI_25.2.09, gcc (Ubuntu 11.4.0-1ubuntu1~22.04.3) 11.4.0; AFNI AFNI_25.2.09 (Aug 24 2025) [64-bit] in fMRIPrep’s image; larmorx-gpl 0.0.1; LARMORX 0.0.1 (the clean room)
- Grid
- 160 × 192 × 192 voxels of 1 × 1.33 × 1.33 mm; shown: 151 × 125 × 111 around the brain
- AFNI versions
- AFNI 26.3.05 (both ports’ reference since 2026-10-29) on the same arguments as AFNI 25.2.09: the same bytes in every output of the call. What changed between the two releases in these programs (header slice code 2, AFNI-03;
-no_detrend, AFNI-05; both corrected by LARMORX before AFNI fixed them) does not meet this call.
Data shown
- Mixed-gambles task (OpenNeuro ds000005, version 00001), Open Data Commons Public Domain Dedication and Licence 1.0.




Bit-identical5,898,240 of 5,898,240 voxels equal
Bit-identical5,898,240 of 5,898,240 voxels equal

Axial slice 67 of 111; the crosshair at the brain's centre. 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 reason | Output | max |Δ| | Values that differ | RMS of Δ | max |Δ| ÷ range | 1 − NCC |
|---|---|---|---|---|---|---|
| The original against itself: the yardstick | ||||||
| AFNI 3dUnifize 25.2.09, 6 threads − 1 threadreferenceAFNI computes every voxel’s local mean independently in its OpenMP loop: the same bytes at 1 and 6 threads. | both outputs | identical: 0 of 5,898,240 values differ | ||||
| AFNI 3dUnifize 26.3.05 − 25.2.09, 1 thread eachreferenceAFNI 26.3.05 gives AFNI 25.2.09’s bytes on this call: the version step changes nothing here (its two changes, AFNI-03 and AFNI-05, need header slice code 2 or 3dTshift’s -no_detrend). | both outputs | identical: 0 of 5,898,240 values differ | ||||
| AFNI 3dUnifize 25.2.09, fMRIPrep's build − the repository's build, 6 threads eachreferenceTwo builds of AFNI 25.2.09 (fMRIPrep’s image and the repository’s) give the same bytes. | both outputs | identical: 0 of 5,898,240 values differ | ||||
| The replica against the original | ||||||
| replica − AFNI 3dUnifize 25.2.09, 1 thread eachBit-identicalThe same values, every one. | both outputs | identical: 0 of 5,898,240 values differ | ||||
| replica − AFNI 3dUnifize 26.3.05, 1 thread eachBit-identicalThe same values, every one. | both outputs | identical: 0 of 5,898,240 values differ | ||||
| The replica against itself: thread counts, fix and reproduce | ||||||
| replica, 6 threads − 1 threadBit-identicalThe same bytes on any number of threads. | both outputs | identical: 0 of 5,898,240 values differ | ||||
replica, fix − reproduceBit-identicalNo correction of the register applies to this input: fix and reproduce give the same bytes. | both outputs | identical: 0 of 5,898,240 values differ | ||||
| The clean room against the original and the replica | ||||||
| clean room − AFNI 3dUnifize 25.2.09, 1 thread eachIdentical with correctly rounded mathsAFNI’s squash calls glibc’s tanhf, which is not correctly rounded on some arguments; the clean room’s is. With CORE-MATH’s correctly rounded functions preloaded, AFNI gives the clean room’s bytes, every value. | Unifized image | 0.000122 | 1,633 of 5,898,240 (0.0277 %) | 2.03e−6 | 8.14e−8 | 1.32e−17 |
| Scale image (-ssave) | 0 | 0 of 5,898,240 | 0 | 0 | 0 | |
| clean room − AFNI 3dUnifize 26.3.05, 1 thread eachIdentical with correctly rounded mathsAFNI 26.3.05 gives AFNI 25.2.09’s bytes on this call, so this is the comparison with 25.2.09: AFNI’s squash calls glibc’s tanhf, which is not correctly rounded on some arguments; the clean room’s is. With CORE-MATH’s correctly rounded functions preloaded, AFNI gives the clean room’s bytes, every value. | Unifized image | 0.000122 | 1,633 of 5,898,240 (0.0277 %) | 2.03e−6 | 8.14e−8 | 1.32e−17 |
| Scale image (-ssave) | 0 | 0 of 5,898,240 | 0 | 0 | 0 | |
| clean room − AFNI 3dUnifize with correctly rounded mathsBit-identicalThe same values, every one: AFNI with correctly rounded maths. | both outputs | identical: 0 of 5,898,240 values differ | ||||
| clean room − replica, 1 thread eachIdentical with correctly rounded mathsThe replica carries glibc 2.35’s tanhf, the clean room a correctly rounded one: they differ exactly where AFNI and AFNI with correctly rounded maths do. | Unifized image | 0.000122 | 1,633 of 5,898,240 (0.0277 %) | 2.03e−6 | 8.14e−8 | 1.32e−17 |
| Scale image (-ssave) | 0 | 0 of 5,898,240 | 0 | 0 | 0 | |
| The clean room against itself: thread counts, fix and reproduce | ||||||
clean room, fix − reproduceBit-identicalNo correction of the register applies to this input: fix and reproduce give the same bytes. | both outputs | identical: 0 of 5,898,240 values differ | ||||
| clean room, 6 threads − 1 threadBit-identicalThe clean room’s result does not depend on its thread count. | both outputs | identical: 0 of 5,898,240 values differ | ||||
Δ is A − B. Computed in float64 over all 5,898,240 voxels of the outputs, not only those shown; the range is B’s maximum minus its minimum. The records’ thresholds: every output bit-identical to AFNI’s (the replica), or float32 values within 1e-5 of AFNI’s largest with bit-identity reported (the clean room); with their exit status and messages. The rows the original against itself are AFNI’s own variability on this input: its thread counts and builds.
Provenance of this example
- Data: OpenNeuro ds000005 (Tom et al., 2007), snapshot 00001, sub-01, PDDL-1.0
- Inputs: T1w (sha256 75c615d7cce397be…)
- Runs: figures cache entry
afni-unifize/ds000005-t1w, computed 2026-10-23T11:09:11+00:00 with larmorxef4d5685a291; AFNI_25.2.09, gcc (Ubuntu 11.4.0-1ubuntu1~22.04.3) 11.4.0; AFNI_25.2.09 (Aug 24 2025) [64-bit]; larmorx-gpl 0.0.1; 7f84379b40f399876909ecf7c236105d6b7c43fdec5111fb7b03f8dee204ac10; AFNI_26.3.05; every command, its environment, threads and exit status are in itsRUN_INFO.json - Runs of the clean room: figures cache entry
afni-unifize/ds000005-t1w-original, computed 2026-10-23T23:47:01+00:00 with larmorxca2145f7951e(fix, the default at 1 and 6 threads,reproduceat 1)
Every number above is in afni-unifize.json, exported by python -m larmorx_validation evidence afni-unifize.
A deliberate correction: -T2up 100 (AFNI-45)
At -T2up 100, the upper percentile 3dUnifize inverts about is read one past the end of the sorted values (AFNI-45): what AFNI finds there is memory it never wrote. Both ports’ default, fix, inverts about the largest value, and warns; reproduce fills the read with 0, as AFNI read it here.
Worked example
AFNI-45: -T2up 100 reads past the sorted values
What this shows: At -T2up 100 AFNI reads one past the sorted values, memory it never wrote, which held 0 here: every value inverts to 0, and so does the output, in both releases (AFNI-45). In fix, their default, the replica and the clean room invert about the largest value, the 100th percentile, and warn; in reproduce both give AFNI’s result, the read filled with 0 as AFNI read it on this machine.
- Replica against AFNI 3dUnifize 25.2.09Deliberate correction
- Clean room against AFNI 3dUnifize 25.2.09Deliberate correction
With -T2, 3dUnifize inverts the contrast about an upper percentile of the masked values (-T2up, 98.5 by default, 90 to 100 allowed). At 100 the percentile’s index is one past the end of the sorted values. fMRIPrep’s call with -T2up 100 added, given to AFNI 25.2.09, to AFNI 26.3.05, to the replica and to the clean room in fix and in reproduce, beside fMRIPrep’s own call.
- Input
- the masked BOLD reference of the first example
- Command
3dUnifize -clfrac 0.2 -rbt 18.3 65.0 90.0 -prefix uni.nii.gz -T2 -T2up 100 -input ref.nii.gz- Defect
- AFNI-45: -T2up near 100 reads past the sorted values; corrected by default, with a warning
- AFNI's output
- all 163,840 values 0
- AFNI versions
- AFNI 26.3.05 (both ports’ reference since 2026-10-29) on the same arguments as AFNI 25.2.09: the same bytes in every output of the call with -T2up 100. What changed between the two releases in these programs (header slice code 2, AFNI-03;
-no_detrend, AFNI-05; both corrected by LARMORX before AFNI fixed them) does not meet this call.
Data shown
- Political Moralization (OpenNeuro ds005040, version 1.2.0, doi:10.18112/openneuro.ds005040.v1.2.0), CC0 1.0 public-domain dedication.






Bit-identical163,840 of 163,840 voxels equal
Axial slice 19 of 40; the crosshair at the brain's centre. 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 reason | max |Δ| | Values that differ | RMS of Δ | max |Δ| ÷ range | 1 − NCC |
|---|---|---|---|---|---|
| The original against itself: the yardstick | |||||
| AFNI 3dUnifize 26.3.05 − 25.2.09, -T2up 100referenceAFNI 26.3.05 gives AFNI 25.2.09’s bytes on this call: the read one past the sorted values (AFNI-45) is in both releases, and both read 0 there on this machine. AFNI-45 | identical: 0 of 163,840 values differ | ||||
| The replica against the original | |||||
replica (fix) − AFNI 3dUnifize 25.2.09, -T2up 100Deliberate correctionAFNI-45, corrected by default with a warning: at -T2up 100 the percentile’s index falls one past the sorted values, where AFNI reads memory it never wrote (here 0: every value inverts to 0, and so does the output); the default inverts about the largest value, the 100th percentile. reproduce fills that read with 0. AFNI-45 | 1.46e3 | 47,016 of 163,840 (28.7 %) | 318 | 0 | 1 |
replica (reproduce) − AFNI 3dUnifize 25.2.09, -T2up 100Bit-identicalIn reproduce, the replica gives AFNI’s result: the read past the end filled with 0, as AFNI read it on this machine. AFNI-45 | identical: 0 of 163,840 values differ | ||||
replica (reproduce) − AFNI 3dUnifize 26.3.05, -T2up 100Bit-identicalIn reproduce, the replica gives AFNI’s result: the read past the end filled with 0, as AFNI read it on this machine. AFNI-45 | identical: 0 of 163,840 values differ | ||||
| The clean room against the original | |||||
clean room (fix) − AFNI 3dUnifize 25.2.09, -T2up 100Deliberate correctionAFNI-45, corrected by default with a warning: at -T2up 100 the percentile’s index falls one past the sorted values, where AFNI reads memory it never wrote (here 0: every value inverts to 0, and so does the output); the default inverts about the largest value, the 100th percentile. reproduce fills that read with 0. AFNI-45 | 1.46e3 | 47,016 of 163,840 (28.7 %) | 318 | 0 | 1 |
clean room (reproduce) − AFNI 3dUnifize 25.2.09, -T2up 100Bit-identicalIn reproduce, the clean room gives AFNI’s result: the read past the end filled with 0, as AFNI read it on this machine. AFNI-45 | identical: 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 range is B’s maximum minus its minimum. The records’ thresholds: every output bit-identical to AFNI’s (the replica), or float32 values within 1e-5 of AFNI’s largest with bit-identity reported (the clean room); with their exit status and messages. The rows the original against itself are AFNI’s own variability on this input: its thread counts and builds.
Provenance of this example
- Data: OpenNeuro ds005040, sub-002, CC0-1.0
- Inputs: BOLD reference, masked (sha256 76db33db4206571d…); fMRIPrep’s output (sha256 d6d86af0e97d73b7…)
- Runs: figures cache entry
afni-unifize/fmriprep-ds005040-clip-1, computed 2026-10-23T11:06:34+00:00 with larmorxef4d5685a291; AFNI_25.2.09, gcc (Ubuntu 11.4.0-1ubuntu1~22.04.3) 11.4.0; AFNI_25.2.09 (Aug 24 2025) [64-bit]; larmorx-gpl 0.0.1; 7f84379b40f399876909ecf7c236105d6b7c43fdec5111fb7b03f8dee204ac10; AFNI_26.3.05; every command, its environment, threads and exit status are in itsRUN_INFO.json - Runs of the clean room: figures cache entry
afni-unifize/fmriprep-ds005040-clip-1-original, computed 2026-10-23T23:46:10+00:00 with larmorxca2145f7951e(fix, the default at 1 and 6 threads,reproduceat 1)
Every number above is in afni-unifize.json, exported by python -m larmorx_validation evidence afni-unifize.
Every case
Each case of the two records, with what it tests, its command line, each port’s verdict and the reason. Search, filter by category or verdict (or select a segment of a chart above), sort by a column; #case=<name> links to one case.
131 cases
| Case | Category | What it tests | clean room | replica | Checks |
|---|---|---|---|---|---|
options/default | options | the defaults on the T1-like phantomcommand line3dUnifize -prefix uni.nii <generated: head-f32> | Identical with correctly rounded maths uni.nii: 1 values differ from AFNI’s in their last bits (AFNI’s squash calls glibc’s tanhf, which is not correctly rounded on some arguments; the clean room’s is correctly rounded); with CORE-MATH’s correctly rounded functions preloaded into AFNI, every value is the clean room’s | Bit-identical | 10/10 · 8/8 |
options/fmriprep-call | options | fMRIPrep's options on the T2-like phantom: -clfrac 0.2 -rbt 18.3 65 90 -T2command line3dUnifize -clfrac 0.2 -rbt 18.3 65.0 90.0 -prefix uni.nii -T2 -input <generated: head-t2-f32> | Bit-identical | Bit-identical | 8/8 |
options/Urad-5 | options | -Urad 5command line3dUnifize -Urad 5 -prefix uni.nii <generated: head-f32> | Identical with correctly rounded maths uni.nii: 3 values differ from AFNI’s in their last bits (AFNI’s squash calls glibc’s tanhf, which is not correctly rounded on some arguments; the clean room’s is correctly rounded); with CORE-MATH’s correctly rounded functions preloaded into AFNI, every value is the clean room’s | Bit-identical | 10/10 · 8/8 |
options/Urad-12 | options | -Urad 12command line3dUnifize -Urad 12 -prefix uni.nii <generated: head-f32> | Bit-identical | Bit-identical | 7/7 |
options/Urad-30 | options | -Urad 30command line3dUnifize -Urad 30 -prefix uni.nii <generated: head-f32> | Bit-identical | Bit-identical | 7/7 |
options/rbt | options | -rbt 10 30 60command line3dUnifize -rbt 10 30 60 -prefix uni.nii <generated: head-f32> | Identical with correctly rounded maths uni.nii: 326 values differ from AFNI’s in their last bits (AFNI’s squash calls glibc’s tanhf, which is not correctly rounded on some arguments; the clean room’s is correctly rounded); with CORE-MATH’s correctly rounded functions preloaded into AFNI, every value is the clean room’s | Bit-identical | 10/10 · 8/8 |
options/param | options | -param, the same as -rbtcommand line3dUnifize -param 14 75 90 -prefix uni.nii <generated: head-f32> | Identical with correctly rounded maths uni.nii: 1 values differ from AFNI’s in their last bits (AFNI’s squash calls glibc’s tanhf, which is not correctly rounded on some arguments; the clean room’s is correctly rounded); with CORE-MATH’s correctly rounded functions preloaded into AFNI, every value is the clean room’s | Bit-identical | 9/9 · 7/7 |
options/clfrac-0.1 | options | -clfrac 0.1command line3dUnifize -clfrac 0.1 -prefix uni.nii <generated: head-f32> | Identical with correctly rounded maths uni.nii: 1 values differ from AFNI’s in their last bits (AFNI’s squash calls glibc’s tanhf, which is not correctly rounded on some arguments; the clean room’s is correctly rounded); with CORE-MATH’s correctly rounded functions preloaded into AFNI, every value is the clean room’s | Bit-identical | 9/9 · 7/7 |
options/clfrac-0.5 | options | -clfrac 0.5command line3dUnifize -clfrac 0.5 -prefix uni.nii <generated: head-f32> | Bit-identical | Bit-identical | 7/7 |
options/clfrac-0.9 | options | -clfrac 0.9command line3dUnifize -clfrac 0.9 -prefix uni.nii <generated: head-f32> | Identical with correctly rounded maths uni.nii: 4 values differ from AFNI’s in their last bits (AFNI’s squash calls glibc’s tanhf, which is not correctly rounded on some arguments; the clean room’s is correctly rounded); with CORE-MATH’s correctly rounded functions preloaded into AFNI, every value is the clean room’s | Bit-identical | 9/9 · 7/7 |
options/mfrac | options | -mfrac, the same as -clfraccommand line3dUnifize -mfrac 0.3 -prefix uni.nii <generated: head-f32> | Identical with correctly rounded maths uni.nii: 1 values differ from AFNI’s in their last bits (AFNI’s squash calls glibc’s tanhf, which is not correctly rounded on some arguments; the clean room’s is correctly rounded); with CORE-MATH’s correctly rounded functions preloaded into AFNI, every value is the clean room’s | Bit-identical | 9/9 · 7/7 |
options/T2 | options | -T2 on the T2-like phantomcommand line3dUnifize -T2 -prefix uni.nii <generated: head-t2-f32> | Bit-identical | Bit-identical | 8/8 |
options/T2-T2 | options | -T2 -T2: no second inversion, bright edges clippedcommand line3dUnifize -T2 -T2 -prefix uni.nii <generated: head-t2-f32> | Bit-identical | Bit-identical | 8/8 |
options/T2-T2-T2 | options | -T2 three times: AFNI neither inverts back nor clips; refused by default (AFNI-48)command line3dUnifize -T2 -T2 -T2 -prefix uni.nii <generated: head-t2-f32> | Deliberate correction Refused by default (AFNI-48, decided 2026-10-23, rule B: the input makes no sense and AFNI silently does something arbitrary): the help documents -T2 once and twice; given a third time it neither inverts the output back nor clips its edges, so AFNI writes an inverted image. The clean room’s default stops with a fatal error naming the entry (exit status 1); | Deliberate correction Refused by default (AFNI-48, decided 2026-10-23, rule B: the input makes no sense and AFNI silently does something arbitrary): the help documents -T2 once and twice; given a third time it neither inverts the output back nor clips its edges, so AFNI writes an inverted image. The replica’s default stops with a fatal error naming the entry (exit status 1); | 7/7 |
options/T2-x4 | options | -T2 four times: as three (AFNI-48)command line3dUnifize -T2 -T2 -T2 -T2 -prefix uni.nii <generated: head-t2-f32> | Deliberate correction Refused by default (AFNI-48, decided 2026-10-23, rule B: the input makes no sense and AFNI silently does something arbitrary): the help documents -T2 once and twice; given a third time it neither inverts the output back nor clips its edges, so AFNI writes an inverted image. The clean room’s default stops with a fatal error naming the entry (exit status 1); | Deliberate correction Refused by default (AFNI-48, decided 2026-10-23, rule B: the input makes no sense and AFNI silently does something arbitrary): the help documents -T2 once and twice; given a third time it neither inverts the output back nor clips its edges, so AFNI writes an inverted image. The replica’s default stops with a fatal error naming the entry (exit status 1); | 7/7 |
options/T2-x3-EPI | options | -T2 three times with -EPI: -EPI sets one -T2, nothing refusedcommand line3dUnifize -T2 -T2 -T2 -EPI -prefix uni.nii <generated: series-f32> | Bit-identical | Bit-identical | 7/7 |
options/T2up-95 | options | -T2 -T2up 95command line3dUnifize -T2 -T2up 95 -prefix uni.nii <generated: head-t2-f32> | Bit-identical | Bit-identical | 8/8 |
options/GM | options | -GMcommand line3dUnifize -GM -prefix uni.nii <generated: head-f32> | Identical with correctly rounded maths uni.nii: 1 values differ from AFNI’s in their last bits (AFNI’s squash calls glibc’s tanhf, which is not correctly rounded on some arguments; the clean room’s is correctly rounded); with CORE-MATH’s correctly rounded functions preloaded into AFNI, every value is the clean room’s | Bit-identical | 10/10 · 8/8 |
options/GM-T2 | options | -GM -T2command line3dUnifize -GM -T2 -prefix uni.nii <generated: head-t2-f32> | Bit-identical | Bit-identical | 7/7 |
options/nosquash | options | -nosquashcommand line3dUnifize -nosquash -prefix uni.nii <generated: head-f32> | Bit-identical | Bit-identical | 8/8 |
options/oink | options | -oink, the same as -nosquashcommand line3dUnifize -oink -prefix uni.nii <generated: head-f32> | Bit-identical | Bit-identical | 7/7 |
options/ssave-amsave | options | -ssave and -amsavecommand line3dUnifize -ssave scale.nii -amsave am.nii -prefix uni.nii <generated: head-f32> | Identical with correctly rounded maths uni.nii: 1 values differ from AFNI’s in their last bits (AFNI’s squash calls glibc’s tanhf, which is not correctly rounded on some arguments; the clean room’s is correctly rounded); with CORE-MATH’s correctly rounded functions preloaded into AFNI, every value is the clean room’s | Bit-identical | 14/14 · 12/12 |
options/quiet | options | -quietcommand line3dUnifize -quiet -prefix uni.nii <generated: head-f32> | Identical with correctly rounded maths uni.nii: 1 values differ from AFNI’s in their last bits (AFNI’s squash calls glibc’s tanhf, which is not correctly rounded on some arguments; the clean room’s is correctly rounded); with CORE-MATH’s correctly rounded functions preloaded into AFNI, every value is the clean room’s | Bit-identical | 9/9 · 7/7 |
options/verb | options | -verb: the automask's messages toocommand line3dUnifize -verb -prefix uni.nii <generated: head-f32> | Identical with correctly rounded maths uni.nii: 1 values differ from AFNI’s in their last bits (AFNI’s squash calls glibc’s tanhf, which is not correctly rounded on some arguments; the clean room’s is correctly rounded); with CORE-MATH’s correctly rounded functions preloaded into AFNI, every value is the clean room’s | Bit-identical | 9/9 · 7/7 |
options/verb-verb | options | -verb -verbcommand line3dUnifize -verb -verb -prefix uni.nii <generated: head-f32> | Identical with correctly rounded maths uni.nii: 1 values differ from AFNI’s in their last bits (AFNI’s squash calls glibc’s tanhf, which is not correctly rounded on some arguments; the clean room’s is correctly rounded); with CORE-MATH’s correctly rounded functions preloaded into AFNI, every value is the clean room’s | Bit-identical | 9/9 · 7/7 |
options/useall | options | -useall: disabled, a warningcommand line3dUnifize -useall -prefix uni.nii <generated: head-f32> | Identical with correctly rounded maths uni.nii: 1 values differ from AFNI’s in their last bits (AFNI’s squash calls glibc’s tanhf, which is not correctly rounded on some arguments; the clean room’s is correctly rounded); with CORE-MATH’s correctly rounded functions preloaded into AFNI, every value is the clean room’s | Bit-identical | 9/9 · 7/7 |
options/inset | options | -insetcommand line3dUnifize -inset <generated: head-f32> -prefix uni.nii | Identical with correctly rounded maths uni.nii: 1 values differ from AFNI’s in their last bits (AFNI’s squash calls glibc’s tanhf, which is not correctly rounded on some arguments; the clean room’s is correctly rounded); with CORE-MATH’s correctly rounded functions preloaded into AFNI, every value is the clean room’s | Bit-identical | 9/9 · 7/7 |
options/case-insensitive | options | -t2 -gm -NOSQUASH: AFNI's strcasecmp optionscommand line3dUnifize -t2 -gm -NOSQUASH -prefix uni.nii <generated: head-t2-f32> | Bit-identical | Bit-identical | 7/7 |
options/duplo | options | 1.2 million voxels: the half-size pathcommand line3dUnifize -prefix uni.nii.gz <generated: big> | Identical with correctly rounded maths uni.nii.gz: 5 values differ from AFNI’s in their last bits (AFNI’s squash calls glibc’s tanhf, which is not correctly rounded on some arguments; the clean room’s is correctly rounded); with CORE-MATH’s correctly rounded functions preloaded into AFNI, every value is the clean room’s | Bit-identical | 9/9 · 7/7 |
options/noduplo | options | 1.2 million voxels with -noduplocommand line3dUnifize -noduplo -prefix uni.nii.gz <generated: big> | Identical with correctly rounded maths uni.nii.gz: 10 values differ from AFNI’s in their last bits (AFNI’s squash calls glibc’s tanhf, which is not correctly rounded on some arguments; the clean room’s is correctly rounded); with CORE-MATH’s correctly rounded functions preloaded into AFNI, every value is the clean room’s | Bit-identical | 10/10 · 8/8 |
options/duplo-T2 | options | the half-size path with -T2 and -ssavecommand line3dUnifize -T2 -ssave scale.nii.gz -prefix uni.nii.gz <generated: big-t2> | Bit-identical | Bit-identical | 9/9 |
options/EPI | options | -EPI on a 4D seriescommand line3dUnifize -EPI -prefix uni.nii <generated: series-i16> | Bit-identical | Bit-identical | 8/8 |
options/EPI-GM | options | -EPI -GM: -GM turned offcommand line3dUnifize -EPI -GM -prefix uni.nii <generated: series-f32> | Bit-identical | Bit-identical | 7/7 |
options/EPI-3D | options | -EPI on one volume: a warning, -T2 keptcommand line3dUnifize -EPI -prefix uni.nii <generated: head-t2-f32> | Bit-identical | Bit-identical | 7/7 |
options/4D-without-EPI | options | a 4D series without -EPI: sub-brick #0 onlycommand line3dUnifize -prefix uni.nii <generated: series-i16> | Identical with correctly rounded maths uni.nii: 9 values differ from AFNI’s in their last bits (AFNI’s squash calls glibc’s tanhf, which is not correctly rounded on some arguments; the clean room’s is correctly rounded); with CORE-MATH’s correctly rounded functions preloaded into AFNI, every value is the clean room’s | Bit-identical | 9/9 · 7/7 |
datatypes/head-i16 | data types | head-i16, with -ssavecommand line3dUnifize -ssave scale.nii -prefix uni.nii <generated: head-i16> | Identical with correctly rounded maths uni.nii: 1 values differ from AFNI’s in their last bits (AFNI’s squash calls glibc’s tanhf, which is not correctly rounded on some arguments; the clean room’s is correctly rounded); with CORE-MATH’s correctly rounded functions preloaded into AFNI, every value is the clean room’s | Bit-identical | 11/11 · 9/9 |
datatypes/head-u8 | data types | head-u8, with -ssavecommand line3dUnifize -ssave scale.nii -prefix uni.nii <generated: head-u8> | Bit-identical | Bit-identical | 9/9 |
datatypes/head-i16-slope2 | data types | head-i16-slope2, with -ssavecommand line3dUnifize -ssave scale.nii -prefix uni.nii <generated: head-i16-slope2> | Bit-identical | Bit-identical | 9/9 |
datatypes/head-i16-slope1 | data types | head-i16-slope1, with -ssavecommand line3dUnifize -ssave scale.nii -prefix uni.nii <generated: head-i16-slope1> | Identical with correctly rounded maths uni.nii: 1 values differ from AFNI’s in their last bits (AFNI’s squash calls glibc’s tanhf, which is not correctly rounded on some arguments; the clean room’s is correctly rounded); with CORE-MATH’s correctly rounded functions preloaded into AFNI, every value is the clean room’s | Bit-identical | 11/11 · 9/9 |
datatypes/head-f64 | data types | head-f64, with -ssavecommand line3dUnifize -ssave scale.nii -prefix uni.nii <generated: head-f64> | Identical with correctly rounded maths uni.nii: 1 values differ from AFNI’s in their last bits (AFNI’s squash calls glibc’s tanhf, which is not correctly rounded on some arguments; the clean room’s is correctly rounded); with CORE-MATH’s correctly rounded functions preloaded into AFNI, every value is the clean room’s | Bit-identical | 11/11 · 9/9 |
datatypes/head-i8 | data types | head-i8, with -ssavecommand line3dUnifize -ssave scale.nii -prefix uni.nii <generated: head-i8> | Bit-identical | Bit-identical | 9/9 |
datatypes/head-nan | data types | head-nan, with -ssavecommand line3dUnifize -ssave scale.nii -prefix uni.nii <generated: head-nan> | Bit-identical | Bit-identical | 9/9 |
datatypes/head-oblique | data types | head-oblique, with -ssavecommand line3dUnifize -ssave scale.nii -prefix uni.nii <generated: head-oblique> | Identical with correctly rounded maths uni.nii: 1 values differ from AFNI’s in their last bits (AFNI’s squash calls glibc’s tanhf, which is not correctly rounded on some arguments; the clean room’s is correctly rounded); with CORE-MATH’s correctly rounded functions preloaded into AFNI, every value is the clean room’s | Bit-identical | 11/11 · 9/9 |
datatypes/head-qform-only | data types | head-qform-only, with -ssavecommand line3dUnifize -ssave scale.nii -prefix uni.nii <generated: head-qform-only> | Identical with correctly rounded maths uni.nii: 1 values differ from AFNI’s in their last bits (AFNI’s squash calls glibc’s tanhf, which is not correctly rounded on some arguments; the clean room’s is correctly rounded); with CORE-MATH’s correctly rounded functions preloaded into AFNI, every value is the clean room’s | Bit-identical | 11/11 · 9/9 |
datatypes/head-las | data types | head-las, with -ssavecommand line3dUnifize -ssave scale.nii -prefix uni.nii <generated: head-las> | Identical with correctly rounded maths uni.nii: 1 values differ from AFNI’s in their last bits (AFNI’s squash calls glibc’s tanhf, which is not correctly rounded on some arguments; the clean room’s is correctly rounded); with CORE-MATH’s correctly rounded functions preloaded into AFNI, every value is the clean room’s | Bit-identical | 11/11 · 9/9 |
defects/head-f32-slope-inter | defects of AFNI | float32 with scl_slope and scl_inter (AFNI-02)command line3dUnifize -prefix uni.nii <generated: head-f32-slope-inter> | Bit-identical in | Bit-identical in | 7/7 |
defects/head-metre | defects of AFNI | a grid in metres (AFNI-07)command line3dUnifize -prefix uni.nii <generated: head-metre> | Identical with correctly rounded maths uni.nii: 1 values differ from AFNI’s in their last bits (AFNI’s squash calls glibc’s tanhf, which is not correctly rounded on some arguments; the clean room’s is correctly rounded); with CORE-MATH’s correctly rounded functions preloaded into AFNI, every value is the clean room’s; in | Bit-identical in | 9/9 · 7/7 |
defects/head-nearly-plumb | defects of AFNI | a grid 0.009 degrees from plumb (AFNI-16)command line3dUnifize -prefix uni.nii <generated: head-nearly-plumb> | Identical with correctly rounded maths uni.nii: 1 values differ from AFNI’s in their last bits (AFNI’s squash calls glibc’s tanhf, which is not correctly rounded on some arguments; the clean room’s is correctly rounded); with CORE-MATH’s correctly rounded functions preloaded into AFNI, every value is the clean room’s; in | Bit-identical in | 9/9 · 7/7 |
defects/nifti-scale-no | defects of AFNI | AFNI_NIFTI_SCALE=NO on int16 with an intercept (AFNI-23)command lineAFNI_NIFTI_SCALE=NO 3dUnifize -prefix uni.nii <generated: head-i16-intercept> | Bit-identical in | Bit-identical in | 7/7 |
defects/T2up-100 | defects of AFNI | -T2 -T2up 100: one past the sorted values (AFNI-45)command line3dUnifize -T2 -T2up 100 -prefix uni.nii <generated: head-t2-f32> | Bit-identical in | Bit-identical in | 8/8 |
defects/T2up-99.99 | defects of AFNI | -T2 -T2up 99.99 on a small mask: one past the values too (AFNI-45)command line3dUnifize -T2 -T2up 99.99 -prefix uni.nii <generated: slice> | Bit-identical in | Bit-identical in | 7/7 |
defects/EPI-negative-slope | defects of AFNI | -EPI on a series with scl_slope -2 (AFNI-01)command line3dUnifize -EPI -prefix uni.nii <generated: series-negative> | Bit-identical in | Bit-identical in | 7/7 |
defects/slab | defects of AFNI | 1 million voxels with an axis of 2: the half-size grid keeps 1 (AFNI-47)command line3dUnifize -prefix uni.nii.gz <generated: slab> | Expected divergence AFNI reads memory it never wrote (AFNI-47): this run crashed (Fatal Signal 11) after the same messages; the clean room fills the read with 0 in | Expected divergence AFNI reads memory it never wrote (AFNI-47): this run crashed (Fatal Signal 11) after the same messages; the replica fills the read with 0 in | 3/3 |
dimensions/slice | dimensions | a 2D slice (64 x 64 x 1)command line3dUnifize -prefix uni.nii <generated: slice> | Bit-identical | Bit-identical | 7/7 |
dimensions/two-slices | dimensions | two slicescommand line3dUnifize -prefix uni.nii <generated: two-slices> | Bit-identical | Bit-identical | 7/7 |
dimensions/tiny | dimensions | 2 x 2 x 2 voxelscommand line3dUnifize -prefix uni.nii <generated: tiny> | Bit-identical | Bit-identical | 7/7 |
dimensions/zeros | dimensions | an image of zeroscommand line3dUnifize -prefix uni.nii <generated: zeros> | Bit-identical | Bit-identical | 7/7 |
dimensions/constant | dimensions | a constant imagecommand line3dUnifize -prefix uni.nii <generated: constant> | Bit-identical | Bit-identical | 7/7 |
dimensions/series-one | dimensions | a 4D file of one volumecommand line3dUnifize -prefix uni.nii <generated: series-one> | Identical with correctly rounded maths uni.nii: 7 values differ from AFNI’s in their last bits (AFNI’s squash calls glibc’s tanhf, which is not correctly rounded on some arguments; the clean room’s is correctly rounded); with CORE-MATH’s correctly rounded functions preloaded into AFNI, every value is the clean room’s | Bit-identical | 9/9 · 7/7 |
errors/slice-small | errors | a 2D slice smaller than 16 x 16: no automaskcommand line3dUnifize -prefix uni.nii <generated: slice-small> | Rejected by both AFNI exit 1: ** FATAL ERROR: Can’t compute Unifize-d dataset for some reason :-(; the clean room exit 1: ** FATAL ERROR: Can’t compute Unifize-d dataset for some reason :-( (the same messages) | Rejected by both AFNI exit 1: ** FATAL ERROR: Can’t compute Unifize-d dataset for some reason :-(; the replica exit 1: ** FATAL ERROR: Can’t compute Unifize-d dataset for some reason :-( (the same messages) | 4/4 |
errors/slice-small-T2 | errors | -T2 on a 2D slice smaller than 16 x 16: AFNI crashes (AFNI-46)command line3dUnifize -T2 -prefix uni.nii <generated: slice-small> | Expected divergence AFNI crashes (exit 1, Fatal Signal 11 (SIGSEGV) received); the clean room refuses (exit 1: ** FATAL ERROR: the automask of a 2D image smaller than 16 x 16 is empty (NULL): AFNI crashes inverting the contrast with it (AFNI-46)), after the same messages | Expected divergence AFNI crashes (exit 1, Fatal Signal 11 (SIGSEGV) received); the replica refuses (exit 1: ** FATAL ERROR: the automask of a 2D image smaller than 16 x 16 is empty (NULL): AFNI crashes inverting the contrast with it (AFNI-46)), after the same messages | 3/3 |
errors/Urad-low | errors | -Urad 3 <input>command line3dUnifize -Urad 3 <generated: head-f32> | Rejected by both AFNI exit 1: ** FATAL ERROR: Illegal value 3.000000 after option -Urad; the clean room exit 1: ** FATAL ERROR: Illegal value 3.000000 after option -Urad (the same messages) | Rejected by both AFNI exit 1: ** FATAL ERROR: Illegal value 3.000000 after option -Urad; the replica exit 1: ** FATAL ERROR: Illegal value 3.000000 after option -Urad (the same messages) | 4/4 |
errors/Urad-high | errors | -Urad 120 <input>command line3dUnifize -Urad 120 <generated: head-f32> | Rejected by both AFNI exit 1: ** FATAL ERROR: Illegal value 120.000000 after option -Urad; the clean room exit 1: ** FATAL ERROR: Illegal value 120.000000 after option -Urad (the same messages) | Rejected by both AFNI exit 1: ** FATAL ERROR: Illegal value 120.000000 after option -Urad; the replica exit 1: ** FATAL ERROR: Illegal value 120.000000 after option -Urad (the same messages) | 4/4 |
errors/rbt-bottom | errors | -rbt 18 20 80 <input>command line3dUnifize -rbt 18 20 80 <generated: head-f32> | Rejected by both AFNI exit 1: ** FATAL ERROR: Illegal values (R pb pt) after ‘-rbt’; the clean room exit 1: ** FATAL ERROR: Illegal values (R pb pt) after ‘-rbt’ (the same messages) | Rejected by both AFNI exit 1: ** FATAL ERROR: Illegal values (R pb pt) after ‘-rbt’; the replica exit 1: ** FATAL ERROR: Illegal values (R pb pt) after ‘-rbt’ (the same messages) | 4/4 |
errors/rbt-order | errors | -rbt 18 70 60 <input>command line3dUnifize -rbt 18 70 60 <generated: head-f32> | Rejected by both AFNI exit 1: ** FATAL ERROR: Illegal values (R pb pt) after ‘-rbt’; the clean room exit 1: ** FATAL ERROR: Illegal values (R pb pt) after ‘-rbt’ (the same messages) | Rejected by both AFNI exit 1: ** FATAL ERROR: Illegal values (R pb pt) after ‘-rbt’; the replica exit 1: ** FATAL ERROR: Illegal values (R pb pt) after ‘-rbt’ (the same messages) | 4/4 |
errors/rbt-top | errors | -rbt 18 70 95 <input>command line3dUnifize -rbt 18 70 95 <generated: head-f32> | Rejected by both AFNI exit 1: ** FATAL ERROR: Illegal values (R pb pt) after ‘-rbt’; the clean room exit 1: ** FATAL ERROR: Illegal values (R pb pt) after ‘-rbt’ (the same messages) | Rejected by both AFNI exit 1: ** FATAL ERROR: Illegal values (R pb pt) after ‘-rbt’; the replica exit 1: ** FATAL ERROR: Illegal values (R pb pt) after ‘-rbt’ (the same messages) | 4/4 |
errors/rbt-missing | errors | -rbt 18 60command line3dUnifize -rbt 18 60 | Rejected by both AFNI exit 1: ** FATAL ERROR: Need 3 arguments (R pb pt) after ‘-rbt’; the clean room exit 1: ** FATAL ERROR: Need 3 arguments (R pb pt) after ‘-rbt’ (the same messages) | Rejected by both AFNI exit 1: ** FATAL ERROR: Need 3 arguments (R pb pt) after ‘-rbt’; the replica exit 1: ** FATAL ERROR: Need 3 arguments (R pb pt) after ‘-rbt’ (the same messages) | 4/4 |
errors/clfrac-low | errors | -clfrac 0.05 <input>command line3dUnifize -clfrac 0.05 <generated: head-f32> | Rejected by both AFNI exit 1: ** FATAL ERROR: -clfrac value 0.050000 is illegal!; the clean room exit 1: ** FATAL ERROR: -clfrac value 0.050000 is illegal! (the same messages) | Rejected by both AFNI exit 1: ** FATAL ERROR: -clfrac value 0.050000 is illegal!; the replica exit 1: ** FATAL ERROR: -clfrac value 0.050000 is illegal! (the same messages) | 4/4 |
errors/clfrac-missing | errors | -clfraccommand line3dUnifize -clfrac | Rejected by both AFNI exit 1: ** FATAL ERROR: Need argument after ‘-clfrac’; the clean room exit 1: ** FATAL ERROR: Need argument after ‘-clfrac’ (the same messages) | Rejected by both AFNI exit 1: ** FATAL ERROR: Need argument after ‘-clfrac’; the replica exit 1: ** FATAL ERROR: Need argument after ‘-clfrac’ (the same messages) | 4/4 |
errors/T2up-low | errors | -T2up 89 <input>command line3dUnifize -T2up 89 <generated: head-f32> | Rejected by both AFNI exit 1: ** FATAL ERROR: -T2up value is out of range 90…100 :-(; the clean room exit 1: ** FATAL ERROR: -T2up value is out of range 90…100 :-( (the same messages) | Rejected by both AFNI exit 1: ** FATAL ERROR: -T2up value is out of range 90…100 :-(; the replica exit 1: ** FATAL ERROR: -T2up value is out of range 90…100 :-( (the same messages) | 4/4 |
errors/T2up-high | errors | -T2up 101 <input>command line3dUnifize -T2up 101 <generated: head-f32> | Rejected by both AFNI exit 1: ** FATAL ERROR: -T2up value is out of range 90…100 :-(; the clean room exit 1: ** FATAL ERROR: -T2up value is out of range 90…100 :-( (the same messages) | Rejected by both AFNI exit 1: ** FATAL ERROR: -T2up value is out of range 90…100 :-(; the replica exit 1: ** FATAL ERROR: -T2up value is out of range 90…100 :-( (the same messages) | 4/4 |
errors/T2up-missing | errors | -T2upcommand line3dUnifize -T2up | Expected divergence AFNI crashes (exit 1, Fatal Signal 11 (SIGSEGV) received); the clean room refuses (exit 1: ** FATAL ERROR: -T2up needs a value: AFNI crashes (Fatal Signal 11) when an option’s value is missing (AFNI-40)), after the same messages | Expected divergence AFNI crashes (exit 1, Fatal Signal 11 (SIGSEGV) received); the replica refuses (exit 1: ** FATAL ERROR: -T2up needs a value: AFNI crashes (Fatal Signal 11) when an option’s value is missing (AFNI-40)), after the same messages | 3/3 |
errors/prefix-missing | errors | -prefixcommand line3dUnifize -prefix | Rejected by both AFNI exit 1: ** FATAL ERROR: Need argument after ‘-prefix’; the clean room exit 1: ** FATAL ERROR: Need argument after ‘-prefix’ (the same messages) | Rejected by both AFNI exit 1: ** FATAL ERROR: Need argument after ‘-prefix’; the replica exit 1: ** FATAL ERROR: Need argument after ‘-prefix’ (the same messages) | 4/4 |
errors/ssave-missing | errors | -ssavecommand line3dUnifize -ssave | Rejected by both AFNI exit 1: ** FATAL ERROR: Need argument after ‘-ssave’; the clean room exit 1: ** FATAL ERROR: Need argument after ‘-ssave’ (the same messages) | Rejected by both AFNI exit 1: ** FATAL ERROR: Need argument after ‘-ssave’; the replica exit 1: ** FATAL ERROR: Need argument after ‘-ssave’ (the same messages) | 4/4 |
errors/input-twice | errors | -input <input> -input <input>command line3dUnifize -input <generated: head-f32> -input <generated: head-f32> | Rejected by both AFNI exit 1: ** FATAL ERROR: Can’t use ‘-input’ twice; the clean room exit 1: ** FATAL ERROR: Can’t use ‘-input’ twice (the same messages) | Rejected by both AFNI exit 1: ** FATAL ERROR: Can’t use ‘-input’ twice; the replica exit 1: ** FATAL ERROR: Can’t use ‘-input’ twice (the same messages) | 4/4 |
errors/unknown-option | errors | -bogus <input>command line3dUnifize -bogus <generated: head-f32> | Rejected by both AFNI exit 1: ** FATAL ERROR: Unknown option: -bogus; the clean room exit 1: ** FATAL ERROR: Unknown option: -bogus (the same messages) | Rejected by both AFNI exit 1: ** FATAL ERROR: Unknown option: -bogus; the replica exit 1: ** FATAL ERROR: Unknown option: -bogus (the same messages) | 4/4 |
errors/no-dataset | errors | -quietcommand line3dUnifize -quiet | Rejected by both AFNI exit 1: ** FATAL ERROR: No dataset name on command line?; the clean room exit 1: ** FATAL ERROR: No dataset name on command line? (the same messages) | Rejected by both AFNI exit 1: ** FATAL ERROR: No dataset name on command line?; the replica exit 1: ** FATAL ERROR: No dataset name on command line? (the same messages) | 4/4 |
errors/missing-input | errors | -prefix uni.nii /nonexistent/x.niicommand line3dUnifize -prefix uni.nii /nonexistent/x.nii | Rejected by both AFNI exit 1: ** FATAL ERROR: Can’t open dataset ‘/nonexistent/x.nii’; the clean room exit 1: ** FATAL ERROR: Can’t open dataset ‘/nonexistent/x.nii’ (the same messages) | Rejected by both AFNI exit 1: ** FATAL ERROR: Can’t open dataset ‘/nonexistent/x.nii’; the replica exit 1: ** FATAL ERROR: Can’t open dataset ‘/nonexistent/x.nii’ (the same messages) | 4/4 |
divergences/default-prefix | expected divergences | no -prefix: AFNI writes its own format (Unifized+orig)command line3dUnifize <generated: head-f32> | Expected divergence AFNI’s default prefix ‘Unifized’ is its own dataset format, which the clean room does not write: it refuses (NIfTI only). | Expected divergence AFNI’s default prefix ‘Unifized’ is its own dataset format, which the replica does not write: it refuses (NIfTI only). | 2/2 |
global/exists | global options | an existing output (AFNI-11)command line3dUnifize -prefix uni.nii <generated: head-f32> (with uni.nii present) | Bit-identical Neither writes an output; the exit status (0) and the messages are AFNI’s. The default, | Bit-identical Neither writes an output; the exit status (0) and the messages are AFNI’s. The default, | 5/5 |
global/overwrite | global options | -overwritecommand line3dUnifize -overwrite -prefix uni.nii <generated: head-f32> (with uni.nii present) | Identical with correctly rounded maths uni.nii: 1 values differ from AFNI’s in their last bits (AFNI’s squash calls glibc’s tanhf, which is not correctly rounded on some arguments; the clean room’s is correctly rounded); with CORE-MATH’s correctly rounded functions preloaded into AFNI, every value is the clean room’s | Bit-identical | 9/9 · 7/7 |
global/deconflict-yes | global options | AFNI_DECONFLICT=YES: renamedcommand lineAFNI_DECONFLICT=YES 3dUnifize -prefix uni.nii <generated: head-f32> (with uni.nii present) | Identical with correctly rounded maths uni_AA1.nii: 1 values differ from AFNI’s in their last bits (AFNI’s squash calls glibc’s tanhf, which is not correctly rounded on some arguments; the clean room’s is correctly rounded); with CORE-MATH’s correctly rounded functions preloaded into AFNI, every value is the clean room’s | Bit-identical | 9/9 · 7/7 |
global/ssave-exists | global options | an existing -ssave output (AFNI-11)command line3dUnifize -ssave scale.nii -prefix uni.nii <generated: head-f32> (with scale.nii present) | Identical with correctly rounded maths uni.nii: 1 values differ from AFNI’s in their last bits (AFNI’s squash calls glibc’s tanhf, which is not correctly rounded on some arguments; the clean room’s is correctly rounded); with CORE-MATH’s correctly rounded functions preloaded into AFNI, every value is the clean room’s; in | Bit-identical in | 9/9 · 7/7 |
global/D-write-type-2 | global options | -DAFNI_NIFTI_WRITE_TYPE=2command line3dUnifize -DAFNI_NIFTI_WRITE_TYPE=2 -prefix uni.nii <generated: head-f32> | Identical with correctly rounded maths uni.nii: 1 values differ from AFNI’s in their last bits (AFNI’s squash calls glibc’s tanhf, which is not correctly rounded on some arguments; the clean room’s is correctly rounded); with CORE-MATH’s correctly rounded functions preloaded into AFNI, every value is the clean room’s | Bit-identical | 9/9 · 7/7 |
environment/quiet-startup | environment variables | AFNI_QUIET_STARTUP=YEScommand lineAFNI_QUIET_STARTUP=YES 3dUnifize -prefix uni.nii <generated: head-f32> | Identical with correctly rounded maths uni.nii: 1 values differ from AFNI’s in their last bits (AFNI’s squash calls glibc’s tanhf, which is not correctly rounded on some arguments; the clean room’s is correctly rounded); with CORE-MATH’s correctly rounded functions preloaded into AFNI, every value is the clean room’s | Bit-identical | 9/9 · 7/7 |
environment/message-prefix | environment variables | AFNI_MESSAGE_PREFIXcommand lineAFNI_MESSAGE_PREFIX=UN 3dUnifize -useall -prefix uni.nii <generated: head-f32> | Identical with correctly rounded maths uni.nii: 1 values differ from AFNI’s in their last bits (AFNI’s squash calls glibc’s tanhf, which is not correctly rounded on some arguments; the clean room’s is correctly rounded); with CORE-MATH’s correctly rounded functions preloaded into AFNI, every value is the clean room’s | Bit-identical | 9/9 · 7/7 |
environment/echo-commandline | environment variables | AFNI_ECHO_COMMANDLINE=YEScommand lineAFNI_ECHO_COMMANDLINE=YES 3dUnifize -prefix uni.nii <generated: head-f32> | Identical with correctly rounded maths uni.nii: 1 values differ from AFNI’s in their last bits (AFNI’s squash calls glibc’s tanhf, which is not correctly rounded on some arguments; the clean room’s is correctly rounded); with CORE-MATH’s correctly rounded functions preloaded into AFNI, every value is the clean room’s | Bit-identical | 9/9 · 7/7 |
global/prefix-with-blank | global options | a -prefix with a blank (AFNI-12)command line3dUnifize -prefix 'a b.nii' <generated: head-f32> | Rejected by both AFNI refuses it (exit 1), and so does the clean room in | Rejected by both AFNI refuses it (exit 1), and so does the replica in | 5/5 |
global/input-dir-with-blank | global options | an input in a directory with a blank (AFNI-18)command line3dUnifize -prefix uni.nii <generated: head-f32, in a directory with a blank> | Rejected by both AFNI refuses it (exit 1), and so does the clean room in | Rejected by both AFNI refuses it (exit 1), and so does the replica in | 5/5 |
global/missing-directory | global options | an output in a missing directory (AFNI-26)command line3dUnifize -prefix nodir/uni.nii <generated: head-f32> | Bit-identical Neither writes an output; the exit status (0) and the messages are AFNI’s. The default, | Bit-identical Neither writes an output; the exit status (0) and the messages are AFNI’s. The default, | 5/5 |
fmriprep/ds000005-task_mixedgamblestask_run_01 | fMRIPrep's calls | fMRIPrep's call on its BOLD reference's input (ds000005-task_mixedgamblestask_run_01)command line3dUnifize -clfrac 0.2 -rbt 18.3 65.0 90.0 -prefix uni.nii.gz -T2 -input <fMRIPrep ds000005-task_mixedgamblestask_run_01: masked BOLD reference> | Bit-identical AFNI at 1 and 6 threads and fMRIPrep 25.2.6’s AFNI give the bytes of the repository’s build; uni.nii.gz: fMRIPrep’s own output, the same bytes | Bit-identical AFNI at 1 and 6 threads and fMRIPrep 25.2.6’s AFNI give the bytes of the repository’s build; uni.nii.gz: fMRIPrep’s own output, the same bytes | 12/12 |
fmriprep/ds000005-task_mixedgamblestask_run_02 | fMRIPrep's calls | fMRIPrep's call on its BOLD reference's input (ds000005-task_mixedgamblestask_run_02)command line3dUnifize -clfrac 0.2 -rbt 18.3 65.0 90.0 -prefix uni.nii.gz -T2 -input <fMRIPrep ds000005-task_mixedgamblestask_run_02: masked BOLD reference> | Identical with correctly rounded maths uni.nii.gz: 2 values differ from AFNI’s in their last bits (AFNI’s squash calls glibc’s tanhf, which is not correctly rounded on some arguments; the clean room’s is correctly rounded); with CORE-MATH’s correctly rounded functions preloaded into AFNI, every value is the clean room’s; AFNI at 1 and 6 threads and fMRIPrep 25.2.6’s AFNI give the bytes of the repository’s build; uni.nii.gz: fMRIPrep’s own output, 2 values differ (fMRIPrep’s output is the output of the repository’s AFNI build bit for bit: the same last-bit differences) | Bit-identical AFNI at 1 and 6 threads and fMRIPrep 25.2.6’s AFNI give the bytes of the repository’s build; uni.nii.gz: fMRIPrep’s own output, the same bytes | 15/15 · 12/12 |
fmriprep/ds000005-task_mixedgamblestask_run_03 | fMRIPrep's calls | fMRIPrep's call on its BOLD reference's input (ds000005-task_mixedgamblestask_run_03)command line3dUnifize -clfrac 0.2 -rbt 18.3 65.0 90.0 -prefix uni.nii.gz -T2 -input <fMRIPrep ds000005-task_mixedgamblestask_run_03: masked BOLD reference> | Identical with correctly rounded maths uni.nii.gz: 2 values differ from AFNI’s in their last bits (AFNI’s squash calls glibc’s tanhf, which is not correctly rounded on some arguments; the clean room’s is correctly rounded); with CORE-MATH’s correctly rounded functions preloaded into AFNI, every value is the clean room’s; AFNI at 1 and 6 threads and fMRIPrep 25.2.6’s AFNI give the bytes of the repository’s build; uni.nii.gz: fMRIPrep’s own output, 2 values differ (fMRIPrep’s output is the output of the repository’s AFNI build bit for bit: the same last-bit differences) | Bit-identical AFNI at 1 and 6 threads and fMRIPrep 25.2.6’s AFNI give the bytes of the repository’s build; uni.nii.gz: fMRIPrep’s own output, the same bytes | 15/15 · 12/12 |
fmriprep/ds000228-task_pixar | fMRIPrep's calls | fMRIPrep's call on its BOLD reference's input (ds000228-task_pixar)command line3dUnifize -clfrac 0.2 -rbt 18.3 65.0 90.0 -prefix uni.nii.gz -T2 -input <fMRIPrep ds000228-task_pixar: masked BOLD reference> | Identical with correctly rounded maths uni.nii.gz: 4 values differ from AFNI’s in their last bits (AFNI’s squash calls glibc’s tanhf, which is not correctly rounded on some arguments; the clean room’s is correctly rounded); with CORE-MATH’s correctly rounded functions preloaded into AFNI, every value is the clean room’s; AFNI at 1 and 6 threads and fMRIPrep 25.2.6’s AFNI give the bytes of the repository’s build; uni.nii.gz: fMRIPrep’s own output, 4 values differ (fMRIPrep’s output is the output of the repository’s AFNI build bit for bit: the same last-bit differences) | Bit-identical AFNI at 1 and 6 threads and fMRIPrep 25.2.6’s AFNI give the bytes of the repository’s build; uni.nii.gz: fMRIPrep’s own output, the same bytes | 15/15 · 12/12 |
fmriprep/ds000256-task_restbaseline | fMRIPrep's calls | fMRIPrep's call on its BOLD reference's input (ds000256-task_restbaseline)command line3dUnifize -clfrac 0.2 -rbt 18.3 65.0 90.0 -prefix uni.nii.gz -T2 -input <fMRIPrep ds000256-task_restbaseline: masked BOLD reference> | Identical with correctly rounded maths uni.nii.gz: 1 values differ from AFNI’s in their last bits (AFNI’s squash calls glibc’s tanhf, which is not correctly rounded on some arguments; the clean room’s is correctly rounded); with CORE-MATH’s correctly rounded functions preloaded into AFNI, every value is the clean room’s; AFNI at 1 and 6 threads and fMRIPrep 25.2.6’s AFNI give the bytes of the repository’s build; uni.nii.gz: fMRIPrep’s own output, 1 values differ (fMRIPrep’s output is the output of the repository’s AFNI build bit for bit: the same last-bit differences) | Bit-identical AFNI at 1 and 6 threads and fMRIPrep 25.2.6’s AFNI give the bytes of the repository’s build; uni.nii.gz: fMRIPrep’s own output, the same bytes | 15/15 · 12/12 |
fmriprep/ds000256-task_restfixationfixed | fMRIPrep's calls | fMRIPrep's call on its BOLD reference's input (ds000256-task_restfixationfixed)command line3dUnifize -clfrac 0.2 -rbt 18.3 65.0 90.0 -prefix uni.nii.gz -T2 -input <fMRIPrep ds000256-task_restfixationfixed: masked BOLD reference> | Identical with correctly rounded maths uni.nii.gz: 2 values differ from AFNI’s in their last bits (AFNI’s squash calls glibc’s tanhf, which is not correctly rounded on some arguments; the clean room’s is correctly rounded); with CORE-MATH’s correctly rounded functions preloaded into AFNI, every value is the clean room’s; AFNI at 1 and 6 threads and fMRIPrep 25.2.6’s AFNI give the bytes of the repository’s build; uni.nii.gz: fMRIPrep’s own output, 2 values differ (fMRIPrep’s output is the output of the repository’s AFNI build bit for bit: the same last-bit differences) | Bit-identical AFNI at 1 and 6 threads and fMRIPrep 25.2.6’s AFNI give the bytes of the repository’s build; uni.nii.gz: fMRIPrep’s own output, the same bytes | 15/15 · 12/12 |
fmriprep/ds000256-task_restfixationnofeedback | fMRIPrep's calls | fMRIPrep's call on its BOLD reference's input (ds000256-task_restfixationnofeedback)command line3dUnifize -clfrac 0.2 -rbt 18.3 65.0 90.0 -prefix uni.nii.gz -T2 -input <fMRIPrep ds000256-task_restfixationnofeedback: masked BOLD reference> | Identical with correctly rounded maths uni.nii.gz: 4 values differ from AFNI’s in their last bits (AFNI’s squash calls glibc’s tanhf, which is not correctly rounded on some arguments; the clean room’s is correctly rounded); with CORE-MATH’s correctly rounded functions preloaded into AFNI, every value is the clean room’s; AFNI at 1 and 6 threads and fMRIPrep 25.2.6’s AFNI give the bytes of the repository’s build; uni.nii.gz: fMRIPrep’s own output, 4 values differ (fMRIPrep’s output is the output of the repository’s AFNI build bit for bit: the same last-bit differences) | Bit-identical AFNI at 1 and 6 threads and fMRIPrep 25.2.6’s AFNI give the bytes of the repository’s build; uni.nii.gz: fMRIPrep’s own output, the same bytes | 15/15 · 12/12 |
fmriprep/ds002422-task_arithm_run_01 | fMRIPrep's calls | fMRIPrep's call on its BOLD reference's input (ds002422-task_arithm_run_01)command line3dUnifize -clfrac 0.2 -rbt 18.3 65.0 90.0 -prefix uni.nii.gz -T2 -input <fMRIPrep ds002422-task_arithm_run_01: masked BOLD reference> | Identical with correctly rounded maths uni.nii.gz: 13 values differ from AFNI’s in their last bits (AFNI’s squash calls glibc’s tanhf, which is not correctly rounded on some arguments; the clean room’s is correctly rounded); with CORE-MATH’s correctly rounded functions preloaded into AFNI, every value is the clean room’s; AFNI at 1 and 6 threads and fMRIPrep 25.2.6’s AFNI give the bytes of the repository’s build; uni.nii.gz: fMRIPrep’s own output, 13 values differ (fMRIPrep’s output is the output of the repository’s AFNI build bit for bit: the same last-bit differences) | Bit-identical AFNI at 1 and 6 threads and fMRIPrep 25.2.6’s AFNI give the bytes of the repository’s build; uni.nii.gz: fMRIPrep’s own output, the same bytes | 15/15 · 12/12 |
fmriprep/ds002422-task_arithm_run_02 | fMRIPrep's calls | fMRIPrep's call on its BOLD reference's input (ds002422-task_arithm_run_02)command line3dUnifize -clfrac 0.2 -rbt 18.3 65.0 90.0 -prefix uni.nii.gz -T2 -input <fMRIPrep ds002422-task_arithm_run_02: masked BOLD reference> | Identical with correctly rounded maths uni.nii.gz: 7 values differ from AFNI’s in their last bits (AFNI’s squash calls glibc’s tanhf, which is not correctly rounded on some arguments; the clean room’s is correctly rounded); with CORE-MATH’s correctly rounded functions preloaded into AFNI, every value is the clean room’s; AFNI at 1 and 6 threads and fMRIPrep 25.2.6’s AFNI give the bytes of the repository’s build; uni.nii.gz: fMRIPrep’s own output, 7 values differ (fMRIPrep’s output is the output of the repository’s AFNI build bit for bit: the same last-bit differences) | Bit-identical AFNI at 1 and 6 threads and fMRIPrep 25.2.6’s AFNI give the bytes of the repository’s build; uni.nii.gz: fMRIPrep’s own output, the same bytes | 15/15 · 12/12 |
fmriprep/ds002422-task_rest | fMRIPrep's calls | fMRIPrep's call on its BOLD reference's input (ds002422-task_rest)command line3dUnifize -clfrac 0.2 -rbt 18.3 65.0 90.0 -prefix uni.nii.gz -T2 -input <fMRIPrep ds002422-task_rest: masked BOLD reference> | Identical with correctly rounded maths uni.nii.gz: 4 values differ from AFNI’s in their last bits (AFNI’s squash calls glibc’s tanhf, which is not correctly rounded on some arguments; the clean room’s is correctly rounded); with CORE-MATH’s correctly rounded functions preloaded into AFNI, every value is the clean room’s; AFNI at 1 and 6 threads and fMRIPrep 25.2.6’s AFNI give the bytes of the repository’s build; uni.nii.gz: fMRIPrep’s own output, 4 values differ (fMRIPrep’s output is the output of the repository’s AFNI build bit for bit: the same last-bit differences) | Bit-identical AFNI at 1 and 6 threads and fMRIPrep 25.2.6’s AFNI give the bytes of the repository’s build; uni.nii.gz: fMRIPrep’s own output, the same bytes | 15/15 · 12/12 |
fmriprep/ds002603-task_wm_acq_Mb4Mm3Tr550 | fMRIPrep's calls | fMRIPrep's call on its BOLD reference's input (ds002603-task_wm_acq_Mb4Mm3Tr550)command line3dUnifize -clfrac 0.2 -rbt 18.3 65.0 90.0 -prefix uni.nii.gz -T2 -input <fMRIPrep ds002603-task_wm_acq_Mb4Mm3Tr550: masked BOLD reference> | Identical with correctly rounded maths uni.nii.gz: 7 values differ from AFNI’s in their last bits (AFNI’s squash calls glibc’s tanhf, which is not correctly rounded on some arguments; the clean room’s is correctly rounded); with CORE-MATH’s correctly rounded functions preloaded into AFNI, every value is the clean room’s; AFNI at 1 and 6 threads and fMRIPrep 25.2.6’s AFNI give the bytes of the repository’s build; uni.nii.gz: fMRIPrep’s own output, 7 values differ (fMRIPrep’s output is the output of the repository’s AFNI build bit for bit: the same last-bit differences) | Bit-identical AFNI at 1 and 6 threads and fMRIPrep 25.2.6’s AFNI give the bytes of the repository’s build; uni.nii.gz: fMRIPrep’s own output, the same bytes | 15/15 · 12/12 |
fmriprep/ds002603-task_wm_acq_SeqMm3Tr2000 | fMRIPrep's calls | fMRIPrep's call on its BOLD reference's input (ds002603-task_wm_acq_SeqMm3Tr2000)command line3dUnifize -clfrac 0.2 -rbt 18.3 65.0 90.0 -prefix uni.nii.gz -T2 -input <fMRIPrep ds002603-task_wm_acq_SeqMm3Tr2000: masked BOLD reference> | Identical with correctly rounded maths uni.nii.gz: 10 values differ from AFNI’s in their last bits (AFNI’s squash calls glibc’s tanhf, which is not correctly rounded on some arguments; the clean room’s is correctly rounded); with CORE-MATH’s correctly rounded functions preloaded into AFNI, every value is the clean room’s; AFNI at 1 and 6 threads and fMRIPrep 25.2.6’s AFNI give the bytes of the repository’s build; uni.nii.gz: fMRIPrep’s own output, 10 values differ (fMRIPrep’s output is the output of the repository’s AFNI build bit for bit: the same last-bit differences) | Bit-identical AFNI at 1 and 6 threads and fMRIPrep 25.2.6’s AFNI give the bytes of the repository’s build; uni.nii.gz: fMRIPrep’s own output, the same bytes | 15/15 · 12/12 |
fmriprep/ds003592-ses_1_task_rest_echo_1 | fMRIPrep's calls | fMRIPrep's call on its BOLD reference's input (ds003592-ses_1_task_rest_echo_1)command line3dUnifize -clfrac 0.2 -rbt 18.3 65.0 90.0 -prefix uni.nii.gz -T2 -input <fMRIPrep ds003592-ses_1_task_rest_echo_1: masked BOLD reference> | Identical with correctly rounded maths uni.nii.gz: 6 values differ from AFNI’s in their last bits (AFNI’s squash calls glibc’s tanhf, which is not correctly rounded on some arguments; the clean room’s is correctly rounded); with CORE-MATH’s correctly rounded functions preloaded into AFNI, every value is the clean room’s; AFNI at 1 and 6 threads and fMRIPrep 25.2.6’s AFNI give the bytes of the repository’s build; uni.nii.gz: fMRIPrep’s own output, 6 values differ (fMRIPrep’s output is the output of the repository’s AFNI build bit for bit: the same last-bit differences) | Bit-identical AFNI at 1 and 6 threads and fMRIPrep 25.2.6’s AFNI give the bytes of the repository’s build; uni.nii.gz: fMRIPrep’s own output, the same bytes | 15/15 · 12/12 |
fmriprep/ds004884-ses_217_task_naming40_acq_epfid2_dir_AP_run_2 | fMRIPrep's calls | fMRIPrep's call on its BOLD reference's input (ds004884-ses_217_task_naming40_acq_epfid2_dir_AP_run_2)command line3dUnifize -clfrac 0.2 -rbt 18.3 65.0 90.0 -prefix uni.nii.gz -T2 -input <fMRIPrep ds004884-ses_217_task_naming40_acq_epfid2_dir_AP_run_2: masked BOLD reference> | Identical with correctly rounded maths uni.nii.gz: 2 values differ from AFNI’s in their last bits (AFNI’s squash calls glibc’s tanhf, which is not correctly rounded on some arguments; the clean room’s is correctly rounded); with CORE-MATH’s correctly rounded functions preloaded into AFNI, every value is the clean room’s; AFNI at 1 and 6 threads and fMRIPrep 25.2.6’s AFNI give the bytes of the repository’s build; uni.nii.gz: fMRIPrep’s own output, 2 values differ (fMRIPrep’s output is the output of the repository’s AFNI build bit for bit: the same last-bit differences) | Bit-identical AFNI at 1 and 6 threads and fMRIPrep 25.2.6’s AFNI give the bytes of the repository’s build; uni.nii.gz: fMRIPrep’s own output, the same bytes | 15/15 · 12/12 |
fmriprep/ds004884-ses_217_task_rest_acq_epfidp2p2_dir_AP_run_9 | fMRIPrep's calls | fMRIPrep's call on its BOLD reference's input (ds004884-ses_217_task_rest_acq_epfidp2p2_dir_AP_run_9)command line3dUnifize -clfrac 0.2 -rbt 18.3 65.0 90.0 -prefix uni.nii.gz -T2 -input <fMRIPrep ds004884-ses_217_task_rest_acq_epfidp2p2_dir_AP_run_9: masked BOLD reference> | Identical with correctly rounded maths uni.nii.gz: 2 values differ from AFNI’s in their last bits (AFNI’s squash calls glibc’s tanhf, which is not correctly rounded on some arguments; the clean room’s is correctly rounded); with CORE-MATH’s correctly rounded functions preloaded into AFNI, every value is the clean room’s; AFNI at 1 and 6 threads and fMRIPrep 25.2.6’s AFNI give the bytes of the repository’s build; uni.nii.gz: fMRIPrep’s own output, 2 values differ (fMRIPrep’s output is the output of the repository’s AFNI build bit for bit: the same last-bit differences) | Bit-identical AFNI at 1 and 6 threads and fMRIPrep 25.2.6’s AFNI give the bytes of the repository’s build; uni.nii.gz: fMRIPrep’s own output, the same bytes | 15/15 · 12/12 |
fmriprep/ds004958-task_rest | fMRIPrep's calls | fMRIPrep's call on its BOLD reference's input (ds004958-task_rest)command line3dUnifize -clfrac 0.2 -rbt 18.3 65.0 90.0 -prefix uni.nii.gz -T2 -input <fMRIPrep ds004958-task_rest: masked BOLD reference> | Identical with correctly rounded maths uni.nii.gz: 1 values differ from AFNI’s in their last bits (AFNI’s squash calls glibc’s tanhf, which is not correctly rounded on some arguments; the clean room’s is correctly rounded); with CORE-MATH’s correctly rounded functions preloaded into AFNI, every value is the clean room’s; AFNI at 1 and 6 threads and fMRIPrep 25.2.6’s AFNI give the bytes of the repository’s build; uni.nii.gz: fMRIPrep’s own output, 1 values differ (fMRIPrep’s output is the output of the repository’s AFNI build bit for bit: the same last-bit differences) | Bit-identical AFNI at 1 and 6 threads and fMRIPrep 25.2.6’s AFNI give the bytes of the repository’s build; uni.nii.gz: fMRIPrep’s own output, the same bytes | 15/15 · 12/12 |
fmriprep/ds005003-ses_1_task_motor | fMRIPrep's calls | fMRIPrep's call on its BOLD reference's input (ds005003-ses_1_task_motor)command line3dUnifize -clfrac 0.2 -rbt 18.3 65.0 90.0 -prefix uni.nii.gz -T2 -input <fMRIPrep ds005003-ses_1_task_motor: masked BOLD reference> | Identical with correctly rounded maths uni.nii.gz: 3 values differ from AFNI’s in their last bits (AFNI’s squash calls glibc’s tanhf, which is not correctly rounded on some arguments; the clean room’s is correctly rounded); with CORE-MATH’s correctly rounded functions preloaded into AFNI, every value is the clean room’s; AFNI at 1 and 6 threads and fMRIPrep 25.2.6’s AFNI give the bytes of the repository’s build; uni.nii.gz: fMRIPrep’s own output, 3 values differ (fMRIPrep’s output is the output of the repository’s AFNI build bit for bit: the same last-bit differences) | Bit-identical AFNI at 1 and 6 threads and fMRIPrep 25.2.6’s AFNI give the bytes of the repository’s build; uni.nii.gz: fMRIPrep’s own output, the same bytes | 15/15 · 12/12 |
fmriprep/ds005003-ses_1_task_rest | fMRIPrep's calls | fMRIPrep's call on its BOLD reference's input (ds005003-ses_1_task_rest)command line3dUnifize -clfrac 0.2 -rbt 18.3 65.0 90.0 -prefix uni.nii.gz -T2 -input <fMRIPrep ds005003-ses_1_task_rest: masked BOLD reference> | Identical with correctly rounded maths uni.nii.gz: 8 values differ from AFNI’s in their last bits (AFNI’s squash calls glibc’s tanhf, which is not correctly rounded on some arguments; the clean room’s is correctly rounded); with CORE-MATH’s correctly rounded functions preloaded into AFNI, every value is the clean room’s; AFNI at 1 and 6 threads and fMRIPrep 25.2.6’s AFNI give the bytes of the repository’s build; uni.nii.gz: fMRIPrep’s own output, 8 values differ (fMRIPrep’s output is the output of the repository’s AFNI build bit for bit: the same last-bit differences) | Bit-identical AFNI at 1 and 6 threads and fMRIPrep 25.2.6’s AFNI give the bytes of the repository’s build; uni.nii.gz: fMRIPrep’s own output, the same bytes | 15/15 · 12/12 |
fmriprep/ds005040-task_clip_run_1 worked example | fMRIPrep's calls | fMRIPrep's call on its BOLD reference's input (ds005040-task_clip_run_1)command line3dUnifize -clfrac 0.2 -rbt 18.3 65.0 90.0 -prefix uni.nii.gz -T2 -input <fMRIPrep ds005040-task_clip_run_1: masked BOLD reference> | Identical with correctly rounded maths uni.nii.gz: 7 values differ from AFNI’s in their last bits (AFNI’s squash calls glibc’s tanhf, which is not correctly rounded on some arguments; the clean room’s is correctly rounded); with CORE-MATH’s correctly rounded functions preloaded into AFNI, every value is the clean room’s; AFNI at 1 and 6 threads and fMRIPrep 25.2.6’s AFNI give the bytes of the repository’s build; uni.nii.gz: fMRIPrep’s own output, 7 values differ (fMRIPrep’s output is the output of the repository’s AFNI build bit for bit: the same last-bit differences) | Bit-identical AFNI at 1 and 6 threads and fMRIPrep 25.2.6’s AFNI give the bytes of the repository’s build; uni.nii.gz: fMRIPrep’s own output, the same bytes | 15/15 · 12/12 |
fmriprep/ds005454-task_rest | fMRIPrep's calls | fMRIPrep's call on its BOLD reference's input (ds005454-task_rest)command line3dUnifize -clfrac 0.2 -rbt 18.3 65.0 90.0 -prefix uni.nii.gz -T2 -input <fMRIPrep ds005454-task_rest: masked BOLD reference> | Identical with correctly rounded maths uni.nii.gz: 14 values differ from AFNI’s in their last bits (AFNI’s squash calls glibc’s tanhf, which is not correctly rounded on some arguments; the clean room’s is correctly rounded); with CORE-MATH’s correctly rounded functions preloaded into AFNI, every value is the clean room’s; AFNI at 1 and 6 threads and fMRIPrep 25.2.6’s AFNI give the bytes of the repository’s build; uni.nii.gz: fMRIPrep’s own output, 14 values differ (fMRIPrep’s output is the output of the repository’s AFNI build bit for bit: the same last-bit differences) | Bit-identical AFNI at 1 and 6 threads and fMRIPrep 25.2.6’s AFNI give the bytes of the repository’s build; uni.nii.gz: fMRIPrep’s own output, the same bytes | 15/15 · 12/12 |
fmriprep/ds005752-ses_01_task_rest_dir_forward | fMRIPrep's calls | fMRIPrep's call on its BOLD reference's input (ds005752-ses_01_task_rest_dir_forward)command line3dUnifize -clfrac 0.2 -rbt 18.3 65.0 90.0 -prefix uni.nii.gz -T2 -input <fMRIPrep ds005752-ses_01_task_rest_dir_forward: masked BOLD reference> | Identical with correctly rounded maths uni.nii.gz: 1 values differ from AFNI’s in their last bits (AFNI’s squash calls glibc’s tanhf, which is not correctly rounded on some arguments; the clean room’s is correctly rounded); with CORE-MATH’s correctly rounded functions preloaded into AFNI, every value is the clean room’s; AFNI at 1 and 6 threads and fMRIPrep 25.2.6’s AFNI give the bytes of the repository’s build; uni.nii.gz: fMRIPrep’s own output, 1 values differ (fMRIPrep’s output is the output of the repository’s AFNI build bit for bit: the same last-bit differences) | Bit-identical AFNI at 1 and 6 threads and fMRIPrep 25.2.6’s AFNI give the bytes of the repository’s build; uni.nii.gz: fMRIPrep’s own output, the same bytes | 15/15 · 12/12 |
fmriprep/ds005752-ses_01_task_rest_dir_reverse | fMRIPrep's calls | fMRIPrep's call on its BOLD reference's input (ds005752-ses_01_task_rest_dir_reverse)command line3dUnifize -clfrac 0.2 -rbt 18.3 65.0 90.0 -prefix uni.nii.gz -T2 -input <fMRIPrep ds005752-ses_01_task_rest_dir_reverse: masked BOLD reference> | Identical with correctly rounded maths uni.nii.gz: 2 values differ from AFNI’s in their last bits (AFNI’s squash calls glibc’s tanhf, which is not correctly rounded on some arguments; the clean room’s is correctly rounded); with CORE-MATH’s correctly rounded functions preloaded into AFNI, every value is the clean room’s; AFNI at 1 and 6 threads and fMRIPrep 25.2.6’s AFNI give the bytes of the repository’s build; uni.nii.gz: fMRIPrep’s own output, 2 values differ (fMRIPrep’s output is the output of the repository’s AFNI build bit for bit: the same last-bit differences) | Bit-identical AFNI at 1 and 6 threads and fMRIPrep 25.2.6’s AFNI give the bytes of the repository’s build; uni.nii.gz: fMRIPrep’s own output, the same bytes | 15/15 · 12/12 |
real/T1w-ds000005 worked example | real images | AFNI's default on the T1w of ds000005 (the half-size path)command line3dUnifize -ssave scale.nii.gz -prefix uni.nii.gz ds000005/sub-01/anat/sub-01_T1w.nii.gz | Identical with correctly rounded maths uni.nii.gz: 1,633 values differ from AFNI’s in their last bits (AFNI’s squash calls glibc’s tanhf, which is not correctly rounded on some arguments; the clean room’s is correctly rounded); with CORE-MATH’s correctly rounded functions preloaded into AFNI, every value is the clean room’s; AFNI at 6 threads and fMRIPrep 25.2.6’s AFNI give the bytes of the repository’s build | Bit-identical AFNI at 6 threads and fMRIPrep 25.2.6’s AFNI give the bytes of the repository’s build | 13/13 · 11/11 |
real/EPI-ds000005 | real images | -EPI on the BOLD series of ds000005command line3dUnifize -EPI -prefix uni.nii.gz ds000005/sub-01/func/sub-01_task-mixedgamblestask_run-01_bold.nii.gz | Bit-identical AFNI at 6 threads and fMRIPrep 25.2.6’s AFNI give the bytes of the repository’s build | Bit-identical AFNI at 6 threads and fMRIPrep 25.2.6’s AFNI give the bytes of the repository’s build | 9/9 |
guard/epi-um3 | AFNI-75's guard | Bit-identical | Bit-identical | 3/3 | |
guard/epi-um1 | AFNI-75's guard | Bit-identical | Bit-identical | 3/3 | |
guard/epi-mm0.05 | AFNI-75's guard | Bit-identical | Bit-identical | 3/3 | |
guard/epi-um3-one | AFNI-75's guard | Bit-identical | Bit-identical | 3/3 | |
guard/epi-mm0.3 | AFNI-75's guard | Bit-identical | Bit-identical | 7/7 | |
guard/epi-mm3 | AFNI-75's guard | Bit-identical | Bit-identical | 7/7 | |
glibc-05/epi-ball | GLIBC-05's cube root | Bit-identical in | Bit-identical in | 11/11 | |
glibc-05/epi-ball-one | GLIBC-05's cube root | Bit-identical in | Bit-identical in | 10/10 | |
glibc-05/epi-printed | GLIBC-05's cube root | Bit-identical | Bit-identical | 10/10 | |
glibc-05/epi-same-result | GLIBC-05's cube root | Bit-identical | Bit-identical | 10/10 | |
glibc-05/epi-ordinary | GLIBC-05's cube root | Bit-identical | Bit-identical | 10/10 | |
difftest/epi-ball-beyond-any-machine-original | random testing | Bit-identical | – | 1/1 | |
difftest/epi-slice-code-2-fix-notice | random testing | Bit-identical | – | 7/7 | |
difftest/epi-slice-code-2-reproduce-notices | random testing | Bit-identical | – | 7/7 | |
difftest/tiny-oblique-header | random testing | Bit-identical | – | 7/7 | |
difftest/epi-ball-beyond-any-machine | random testing | – | Bit-identical | 1/1 | |
difftest/epi-slice-code-2-afni26 | random testing | – | Bit-identical | 7/7 |
Speed
3dUnifize 25.2.09Seconds for the whole program, informal timings of 2026-10-24.
- AFNI 3dUnifize 25.2.09, 1 thread
- AFNI 3dUnifize 25.2.09, 6 threads
- replica, 1 thread
- replica, 6 threads
- clean room, 1 thread
- clean room, 6 threads
- fMRIPrep's call, ds000005 BOLD referenceAFNI 3dUnifize 25.2.09, 1 thread 93.3 sAFNI 3dUnifize 25.2.09, 6 threads 21.6 sreplica, 1 thread 39.4 sreplica, 6 threads 8.68 sclean room, 1 thread 3.76 sclean room, 6 threads 0.94 s25×
- fMRIPrep's call, ds005454 BOLD reference (half-size path)AFNI 3dUnifize 25.2.09, 1 thread 18.7 sAFNI 3dUnifize 25.2.09, 6 threads 7.45 sreplica, 1 thread 13.3 sreplica, 6 threads 4.85 sclean room, 1 thread 4.58 sclean room, 6 threads 1.21 s4.1×
- a T1w, ds000005 (half-size path)AFNI 3dUnifize 25.2.09, 1 thread 67.1 sAFNI 3dUnifize 25.2.09, 6 threads 23.6 sreplica, 1 thread 35.4 sreplica, 6 threads 14.9 sclean room, 1 thread 9.26 sclean room, 6 threads 2.66 s7.2×
Each program run once (the replica after one warm-up run), reading and writing compressed NIfTI included, on a 12-CPU machine shared with another job throughout (load 17 to 22): indications, not a benchmark report. The ratio is AFNI’s time over the clean room’s, at one thread each. Tool: validation/investigations/afni-masks-original/timing.py.
The clean room is 4 to 25 times faster than AFNI at the same thread count, and 3 to 10 times faster than the replica, which computes as AFNI does: its ball percentiles slide from voxel to voxel instead of being gathered and sorted at every voxel. On the half-size path the ball works on an eighth of the voxels, and reading and writing the full-size compressed images takes a larger share of the time. These are informal timings on a shared machine (findings); there is no benchmark report yet.
Corrections by default
Where AFNI has a deterministic defect that changes a result, both ports correct it by default, fix, and warn. upstream_defects="reproduce" (on the command line --upstream-defects reproduce from the next release, LARMORX_UPSTREAM_DEFECTS=reproduce in 0.1.0rc1) gives AFNI’s behaviour, and the records compare with it. The records’ lists of what fix changed meet AFNI-45 (-T2up near 100), AFNI-47 (a volume of more than a million voxels with an axis of 2 or 3), AFNI-01 (-EPI’s median of a series with a negative scale factor) and the dataset defects 3dUnifize shares with 3dTshift (AFNI-02, AFNI-07, AFNI-11, AFNI-12, AFNI-16, AFNI-18, AFNI-23, AFNI-26) (register). AFNI-48 (-T2 given three times without -EPI, which leaves the output inverted) is refused by default, in the same words by both ports (decided 2026-10-23), with reproduce giving AFNI’s output.
Random testing
Random testing runs random valid inputs and option combinations through the original and both ports, in long runs of several hours: inputs from seeded plain arithmetic (up to 32³ voxels; orientations, data types, scale factors and value ranges of every kind), with options drawn in random combinations from those the program accepts (how random testing works). A difference is shrunk to the smallest input and fewest options that still show it, then fixed and kept as a case of the record, registered as a defect of the original, or reported. The latest long run:
Random testing, both ports against the original: AFNI 3dUnifize 26.3.05, the long run of 2026-10-25 (seed 20261025, 6.2 h for 21 programs): random valid inputs and option combinations, each run through all three
| Program | Random cases | Extreme | Options covered | Agree | glibc | Notices | Undefined | Differ | Heavy | Guard | Unclear |
|---|---|---|---|---|---|---|---|---|---|---|---|
| 3dUnifizereplica, clean room | 568 | 0 | 18 | 551 | · | 2 | · | 9 | 5 | · | 1 |
- fixed in LARMORX 3dUnifize (clean room), 7 case(s): the
fixpass, the notices infixand inreproduce: AFNI-03 named in both modes and AFNI-08 inreproduce, in the replica’s words (2026-10-29) (AFNI-08) - fixed in LARMORX 3dUnifize (clean room), 1 case(s):
am.nii: header bytes,uni.nii.gz: header bytes: the quaternion’s polar decomposition: AFNI takes1/detin x87 extended precision (observed on 1,200 matrices through NIfTI-2 outputs); fixed in the clean room (2026-10-29) (details) - fixed in LARMORX 3dUnifize (clean room), 1 case(s):
uni.nii: header bytes: the quaternion’s polar decomposition: AFNI takes1/detin x87 extended precision (observed on 1,200 matrices through NIfTI-2 outputs); fixed in the clean room (2026-10-29) (details)
Notices: a correction named in a case where it changed nothing the input shows (defect, cases): 3dUnifize: AFNI-08 (11).
Heavy cases (a run over the memory limit of a case, queued and run to its end later, one at a time, in a large box): afni-unifize:20261025:89 (beyond any machine: predicted, not run); afni-unifize:20261025:118 (beyond any machine: predicted, not run); afni-unifize:20261025:226 (beyond any machine: predicted, not run); afni-unifize:20261025:282 (beyond any machine: predicted, not run); afni-unifize:20261025:326 (beyond any machine: predicted, not run).
Every case is drawn from its tool, the run's seed and its number, and replays anywhere: how random testing works; each difference's triage: validation/difftest/triage.toml.
Known limits
LARMORX writes NIfTI only: AFNI’s default prefix Unifized is its own dataset format, which both ports refuse (an expected divergence). The clean room’s squashed values (above 1300) can differ from AFNI’s in their last bit where glibc’s tanhf is not correctly rounded; AFNI with correctly rounded maths gives the clean room’s bytes. There is no benchmark report yet. The cross-platform checks cannot reach the half-size path (more than a million voxels, beyond their 32³ budget); the records cover it on real T1w images. The cross-platform checks (how they work) hold the other five platforms to LARMORX’s own Linux output.
History
- 2026-10-23: AFNI-48 refused by default, by the project’s rule for an input that makes no sense.
- 2026-10-29: AFNI 26.3.05 releases the Medical College of Wisconsin’s code under CC BY 4.0; the replica moves from
larmorx-gplintolarmorx, and both ports follow AFNI 26.3.05. - 2026-10-30: both records made again against AFNI 26.3.05: 129 and 127 cases.
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.
- Mixed-gambles task (OpenNeuro ds000005, version 00001), Open Data Commons Public Domain Dedication and Licence 1.0.
- Children and adults watching a short film (OpenNeuro ds000228, version 1.1.1, doi:10.18112/openneuro.ds000228.v1.1.1), CC0 1.0 public-domain dedication.
- Behavioral interventions for reducing head motion in children (OpenNeuro ds000256, version 00002), Open Data Commons Public Domain Dedication and Licence 1.0.
- fMRI: resting state and arithmetic task (OpenNeuro ds002422, version 1.1.0, doi:10.18112/openneuro.ds002422.v1.1.0), CC0 1.0 public-domain dedication.
- MB-EPI-comparison (OpenNeuro ds002603, version 1.0.0, doi:10.18112/openneuro.ds002603.v1.0.0), CC0 1.0 public-domain dedication.
- Neurocognitive aging data release (OpenNeuro ds003592, version 1.0.13, doi:10.18112/openneuro.ds003592.v1.0.13), CC0 1.0 public-domain dedication.
- Aphasia Recovery Cohort (OpenNeuro ds004884, version 1.0.2, doi:10.18112/openneuro.ds004884.v1.0.2), CC0 1.0 public-domain dedication.
- The FreeSurfer Maintenance Dataset (OpenNeuro ds004958, version 1.0.0, doi:10.18112/openneuro.ds004958.v1.0.0), CC0 1.0 public-domain dedication.
- Retrospective Task/Rest fMRI Data from Tumor Patients (OpenNeuro ds005003, version 2.0.0, doi:10.18112/openneuro.ds005003.v2.0.0), CC0 1.0 public-domain dedication.
- Political Moralization (OpenNeuro ds005040, version 1.2.0, doi:10.18112/openneuro.ds005040.v1.2.0), CC0 1.0 public-domain dedication.
- fMRIPrep boot camp 7T training dataset (OpenNeuro ds005454, version 1.0.0, doi:10.18112/openneuro.ds005454.v1.0.0), Creative Commons Attribution 4.0 (CC BY 4.0). Attribution: Sleight EC, Esteban O, Grouiller F. fMRIPrep boot camp -- 7T training dataset. OpenNeuro ds005454, version 1.0.0. doi:10.18112/openneuro.ds005454.v1.0.0 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.
- NIMH Healthy Research Volunteer Dataset (OpenNeuro ds005752, version 2.1.0, doi:10.18112/openneuro.ds005752.v2.1.0), CC0 1.0 public-domain dedication.