Skip to content

FSL 6.0.7.18, 22nd May 2025

file-tree-fsl 0.2.2 -> 0.2.3

Detailed history

  • python 3.6 and 3.7 are no longer supported by file-tree
  • Replace pkg_resources with importlib_resources
  • Bump to v0.2.3

fsleyes 1.13.0 -> 1.14.2

This MR also contains some changes which should suppress some harmless errors originating from toolbar/dialog widgets that could occur when removing overlays. - ENH: Tractogram Slice width option New Slice width option (--sliceWidth / -sw on the CLI), allowing the width of 2D tractogram slices to be adjusted. A couple of other unrelated changes: - Plot panels will now use a solid line by default for all overlays, instead of cycling through all possible line styles. - cluster_mask_* images from FEAT analyses will now be shown as mask outlines when FSLeyes is started with --autoDisplay.

Detailed history

  • RF: Add cluster mask as mask rather than label image
  • RF: Apply alpha modulation for melodic images with autoDisplay
  • DOC: Icon image for changelog in API docs
  • BF: copy-paste bug, oldOrder used in multiple branches
  • BF: Prevent (harmless) errors in widgets when overlays are removed
  • TEST: Unit test for syncing of DC.overlayOrder
  • RF: Show cluster masks as outlines. Clean up FEAT image autodisplay a bit - separate functions for each image type
  • RF: Use filled line by default for all images in plot panels - other line styles are often undesirable. Drop grey from pt lookup table, but double the number of colours used by creating darker equivalents
  • ENH: New TractogramOpts.sliceWidth property, for controlling width of 2D slices, as a percentage of Z axis bounds
  • RF: Restore behaviour of clipping in terms of ref image voxels when a ref is set
  • TEST: Test new TractogramOpts.sliceWidth option

fslpy 3.21.1 -> 3.22.0

The existing Image.voxToScaledVoxMat and Image.scaledVoxToVoxMat properties have been deprecated - the getAffine method is a nicer interface, and these properties are confusingly named, as they refer to the fsl coordinate system.

Detailed history

  • ENH: Add support for "scaled" coordinate system to getAffine - scaled voxels without L/R flip.
  • MNT,CI: Test py313, drop py39
  • BF: Mixed up affines
  • TEST: Add basic test cover for scaled coord system
  • MNT: Deprecate unfortunately named voxToScaledVoxMat and scaledVoxToVoxMat methods
  • TEST: Replace voxToScaledVox with getAffine
  • MNT: Pre-release version bump [skip-ci]
  • RF: No need for import inside function as numpy is already being imported at the top-level
  • TEST: Adjust identifyAffine test
  • DOC: Changelog [skip-ci]

fsl-armawrap 0.6.0 -> 0.7.0

Detailed history

  • TEST: Normalise signs of EigenValues/svd outputs. Default to c++17
  • MNT: Update armadillo to version 14.2.1
  • Fix CI rules path
  • CI job to rebuild docker image

fsl-base 2502.1 -> 2504.3

Adjust fslversion to better support FSL installations where $FSLDIR is a sym-link, and/or where the fsinstaller.py script was passed a sym-linked destination. - MNT: Minor updates to build system - Add support for .c++ as a source-file suffix. - Tweak to C/C++ dependency file generation - now if a dependency file has changed, the object file is not re-built. This is useful for CUDA projects where the same source file may be compiled more than once for CPU and GPU builds. In this case the dependency file will be re-generated, but we only want object files to be recompiled if the source file has changed, not the dependency file. - TEST: Make sure command output is printed to standard output/error - TEST: Make sure setuptools-scm doesn't cause a crash - RF: Improvements to update_fsl_package The update_fsl_package command can now be used to install/update third-party packages from conda-forge. - RF: Update update_fsl_package to print the underlying conda command And also fix a bug in pyproject.toml. - MNT: Update update_fsl_package to support both .conda and .tar.bz2 packages

Detailed history

  • BF: Make sure that FSLDEVDIR takes precedence
  • BF: If BUILDDIR is set but CUDABUILDDIR isn't, default the latter to the former
  • RF: Adjust fslversion to try and support sym-linked installations
  • MNT: Add support for .c++ source file suffix
  • RF: Make dep files order-only dependencies, so make doesn't rebuild object files based on their timestamp. This is useful for CUDA projects where the same source file may be compiled more than once for different builds. In this case the dependency file is re-generated, but we only want object files to be recompiled if the source file has changed, not the dep file.
  • MNT: Use setuptools-scm to automatically set version when building python package
  • TEST: Make sure command is printed to stdout/err
  • TEST: Make sure setuptools-scm doesn't cause a crash
  • RF: Use --override-channels instead of trying to patch .condarc, which is insanity. Adjust update_fsl_package so it allows new packages to be installed, including third-party packages
  • MNT: update license def
  • MNT: Fix license identiifer
  • RF: Print underlying conda command
  • MNT: .conda packages are stored in a separate area of the repodata.json to .tar.bz2 files

fsl-basisfield 2203.1 -> 2203.4

Detailed history

  • MNT: Need to link against fsl-utils on macOS
  • MNT: Need to link against fsl-utils in unit tests
  • TEST: Replace use of arma rng with cross-platform implementation
  • TEST: Rework unit tests. All outputs generated from the MNI152 are identical across the three platforms. Use a RNG for other tests to ensure results are the same across platforms.

fsl-eddy_qc v1.2.3 -> v1.3.0

There also seem to be some issue in the way that eddy_qc loads eddy acquisition parameter and volume index files (typically called acqparams.txt and index.txt). It was incapable of handling datasets with entries in acqparams.txt which were not referenced by index.txt. The relevant code has been re-written. Results should be the same, although order in which PE directions are reported in the eddy_qc output files may change.

The project has also been updated to use pyproject.toml instead of setup.py, which is necessary for modern Python build tooling.

Detailed history

  • MNT: Replace setup.py with pyproject.toml. Make eddy_quad/eddy_squad python modules. Replace pkg_resources with importlib
  • BF: Code seems to struggle when acqparams contains multiple entries, but provided data corresponds to fewer entries (e.g. acqparams describes both PA and AP, but data/index.txt only corresopnds to AP).
  • MNT: pypdf has renamed PdfFileReader to PdfReader
  • Bf: Re-write janky logic for standardising acqparams and volume indices
  • MNT: PyPDF2 ->pypdf and associated superfluous name changes
  • MNT: Clarification about acqparams/indexes standardisation
  • MNT: Bump version, update readme
  • DOC: Convert readme to markdown
  • MNT: Remove duplicate help text
  • MNT: Remove out-dated/incorrect instructions from README

fsl-fdt 2202.10 -> 2202.11

  • RF: Fixes to NEWIMAGE::volume data access in vecreg Replaced calls to NEWIMAGE::volume[t] (x, y, z) with NEWIMAGE::volume(x, y, z, t) - the former pattern causes unnecessary copies of entire 3D volumes (specificallly spline coefficients) to be made. As this was happening for every voxel, this was having a substantial impact on performance.

Detailed history

  • MNT: Use vol(x,y,z,t) rather than vol[t] (x,y,z)

fsl-first_lib 2111.0 -> 2111.1

Detailed history

  • MNT: Need to link against fsl-utils on macOS

fsl-flameo 2111.1 -> 2111.2

Detailed history

  • MNT: The newimage volume class no longer allows manipulation of state on const volumes, so this sort of ugliness needs to be handled by the client code
  • MNT: Change linking order (make sure utils comes after newimage/miscmaths)

fsl-flirt 2111.2 -> 2111.4

  • MNT: Cannot modify properties of a const NEWIMAGE::volume In this instance there is no need for the volume to be const.
  • BF: Fix bug introduced in last change Cannot pass a non-const rvalue (returned from volume.operator[]) through to final_transform. This is easily avoided by creating a local ShadowVolume instance, and passing that to final_transform.
  • RF: Update method for changing interpolation/extrapolation settings on NEWIMAGE::volume instances This MR contains some ShadowVolume-related fixes to flirt and applyxfm4D:
  • The code was calling invol[t].setpadvalue(..);. This is a throwback to an old version of newimage where, for 4D images, the volume class would store a std::vector<volume> containing a separate 3D volume for each timepoint. After the introduction of the ShadowVolume pattern, this code was having no effect, as the operator[] function would return a new ShadowVolume instance on each invocation. The logic in applyxfm4D has been changed so that the pad value is set via the NEWIMAGE::affine_transform overload (in newimagefns.cc) which accepts a set_backgnd parameter.
  • With the introduction of the ShadowVolume cache in fsl/newimage!41, the calls to invol[t] described above resulted in the creation of a ShadowVolume for each time point, which meant that the subsequent call to invol.setinterpolationmethod (and similar) were not being applied to those ShadowVolume instances. This is being addressed in fsl/newimage!42, but the reworked code here avoids this issue.

Detailed history

  • MNT: Cannot modify properties of a const volume, In this instance there is no need for the volume to be marked as const
  • BF: Fix bug introduced in last change - cannot pass a non-const rvalue (from volume.operator[]) through to final_transform. Change code to create a local ShadowVolume, and pass that through.
  • MNT: Normalise indentation
  • RF: Change method of setting per-volume pad value - we can't set properties on a ShadowVolume instance, but can do it via the newimagefns.cc:affine_transform overload which accepts a set_backgnd parameter.
  • MNT: Avoid calling set* interp/extrap functions on ShadowVolume instances

fsl-fnirt 2203.1 -> 2203.2

This MR also makes a small adjustment to the Makefile so that config files are installed at make install rather than at make.

Detailed history

  • MNT: Do not modify any properties on a const volume. Adjustment to make file so that config files are installed at "make install" rather than at "make"

fsl-installer 3.16.5 -> 3.16.8

Detailed history

  • MNT: Use --rc-file argument to try and persuade conda to only use our configuration file
  • BF: Initial micromamba implementation wouldn't work when installing into a location requiring administrative privileges
  • DOC: Readme
  • MNT: Use more standard tar incantation. Assume that tar will auto-detect compression format - works on ubuntu 22.04/macOS 11.7
  • BF: Wrapper scripts were potentially not being created when FSLDIR is set to a sym-linked location - it seems that $PREFIX gets set to the actual path
  • TEST: Unit test for create/removeFSLWrapper with sym-linked $FSLDIR

fsl-libvis 2111.0 -> 2111.2

Detailed history

  • MNT: Need to use CFLAGS, not CXXFLAGS, when compiling C code
  • BF: Fix terrible bug - pointer referencing destroyed object

fsl-miscmaths 2203.2 -> 2412.4

  • ENH: Allow Splinterpolator instances to be created from an existing set of spline coefficients, and extend extrapolation options
  • The Splinterpolator class can now accept a pointer to an array of pre-calculated coefficients - in this case, a copy of the pointer is used (the data is not copied), and pre-filtering (coefficient calculation) is skipped
  • New SoftZeros extrapolation method - out of bounds coefficients will be set to zero, but out-of-bounds coordinates will still be constructed via interpolation. This differs from the existing Zeros extrapolation method, where out-of-bounds coordinates are forced to zero.
  • Adjustment to the coefficient initialisation method used when Constant extrapolation is requested; previously, a mirroring boundary condition was used.
  • New Splinterpolator::Copy function, which allows the spline coefficients to be copied/extracted.
  • BF: Fix bug in code controlling which adjacent spline coefficients to use when interpolating This MR fixes a very subtle bug in the Splinterpolator class which was causing artifacts when extrapolating into negative coordinate space.

When using third-order cubic b-spline interpolation to interpolate a value at a particular location, the four nearest spline coefficients are combined to generate the interpolated value. For example, if interpolating a value at location -0.25, the spline coefficients at locations [-2, -1, 0, 1] would be used.

If interpolating at an integer location, e.g. 2, the choice of whether to use coefficients at [0, 1, 2, 3], or [1, 2, 3, 4] is arbitrary, but the convention in the Splinterpolator class is to use the former.

The logic which controls which spline coefficients are used is in the get_start_indicies function. While this function produces sensible values for positive coordinates, it was mis-behaving for negative coordinates. - For coordinates with a fractional component greater than 0.5, such as -3.75, the code was correctly selecting coefficients [-5, -4, -3, -2]. However, for coordinates with a fractional component smaller than 0.5, such as -3.25, the code was incorrectly selecting coefficients [-4, -3, -2, -1]. - For a negative integer coordinate such as -4, the code was selecting coefficients [-5, -4, -3, -2], which would produce the correct result, but which was not following the same convention as for positive coordinates (i.e. it should be [-6, -5, -4, -3]). - MNT: Move constructor/assignment operator for Splinterpolator, which inherits ownership of data pointer - MNT: Revert spline coefficient initialisation boundary condition This MR reverts the spline coefficient initialisation boundary condition for constant/zeros extrapolation to preserve the existing/established behaviour. The old behaviour for constant/zeros extrapolation was to use an initialisation method suitable for mirror extrapolation.

This is not strictly the best approach, which is why I changed it in !17. However, the old logic has been in place for at least 16 years, and so arguably the better option is to preserve the existing behaviour. We can revisit this in the future if needed. - TEST: Re-generate splinterpolator benchmark data - TEST: Tweak test tolerance

Detailed history

  • BF: Fix start index selection for even order interpolation
  • MNT: Normalise indentation
  • MNT: remove obsolete debugging macros and statements
  • ENH: Scaffolding for ability to create a Splinterpolator instance from an existing set of spline coefficients
  • ENH: Adjust logic so that Splinterpolator can be created with pre-calculated coefficeints
  • ENH: New function to copy spline coefficients
  • MNT: Safeguard against being passed a bad value for nthr
  • ENH,RF: Implement clamped extrapolation boundary condition for coefficient initialisation
  • ENH: NEW "SoftZeros" extrapolation method, which sets out-of-bounds locations to zero, but still interpolates out-of-bounds, to allow a smooth transition from in-to-out of bounds
  • TEST: Unit tests for splinterpolator
  • BF: Fix extrapolation indexing bug in Splinterpolator::coef(). I think this code was added in an early commit, but was then replaced with indx2indx, and never updated / used.
  • TEST: Expand splinterpolator tests to cover all extrapolation methods, and order=2 interpolation
  • TEST: link fsl-utils
  • MNT: Don't pass a value to NoOfThreads, so that it will use a default value
  • MNT: Move constructor/assignment operator for Splinterpolator, which inherits ownership of data pointer
  • MNT: Const version of kernel::kernelvals function
  • MNT: Revert spline coefficient initialisation boundary condition for constant/zeros to previous incorrect method, to preserve existing behaviour
  • MNT: Comment
  • TEST: Re-generate splinterpolator benchmark data
  • TEST: Tweak test tolerance

fsl_mrs 2.4.0 -> 2.4.3

  • Add negative voigt model Adds model which allows voigt peaks to be inverted.
  • BF: minor bugs Dealing with accumulated small bugs.
  • Handle single coil data in coil combine function
  • Allow phased peaks to be added to basis set.
  • Removed use of depreciated matplotlib function tostring_rgb.

Also added the negativevoigt model, which allows inverted peaks. - ENH: P spline baseline in dynamic fitting - P-spline baseline added as an option to dynamic fitting. - Dynamic fitting default solver changed to L-BFGS (method='quasi-newton') - Citation information added for p-spline fitting (https://doi.org/10.1002/mrm.28385). - ENH: freeshift with lorentzian - Added a freeshift_lorentzian model with inter-metabolite shifts (freeshift) and lorentzian lineshape. - Added freeshift option to dynamic fitting script and MRSI - ENH: Phase corrected report - add option to show phase corrected fit in main HTML report - ENH: Preproc options - Added option to suppress alignment step in fsl_mrs_preproc{_edit}. - Changed syntax for water removal in fsl_mrs_preproc{_edit} from --hlsvd to --remove-water - Changed syntax for truncation in fsl_mrs_preproc{_edit} from --leftshift {N} to --truncate-fid {N} - BF: basis_tools all_shift conjugation - Basis sets are automatically conjugated when using basis_tools shift_all. - BF: preproc(edit) noalign * Fixed bug introduced in 2.4.2 where the option to suppress alignment step in fsl_mrs_preproc{_edit} only suppressed some alignment. * Update minimum hlsvdpropy version to 2.0.2

Detailed history

  • Add negative voigt model
  • BF: minor bugs
  • ENH: P spline baseline in dynamic fitting
  • ENH: freeshift with lorentzian
  • ENH: Phase corrected report
  • ENH: Preproc options
  • BF: basis_tools all_shift conjugation
  • BF: preproc(edit) noalign

fsl-newimage 2203.12 -> 2501.4

This MR makes some adjustments to the NEWIMAGE::volume class to improve thread safety.

  • Several member fields of the volume class were unnecessarily declared as mutable, and several functions which modify fields were unnecessarily declared as const.

  • Functions which modify the extrapolation/interpolation settings of a const volume have been marked as deprecated, and their use will cause a warning to be emitted at compilation time.

  • The extrapolate() function previously returned a reference to the mutable T extrapval field, which was required in the unlikely event of client code writing to a out-of-bounds voxel via the operator() function (e.g. vol(-1, -1, -1) = 0;). This has been changed so that extrapolate() now returns a reference to a member of the new extrapparams class (more on this below); the extrapval field has been removed from the volume class.

  • The userinterpolation, userextrapolation, boundsassert, boundsexception interpolation/extrapolation methods have been removed, as they were not being used anywhere throughout the FSL codebase.

  • Fields related to spline interpolation have been left as mutable, as otherwise it would not be possible to have spline coefficients lazily calculated on-demand when performing spline interpolation on a const volume instance. Access to these fields were already being protected by a std::mutex, and so should be thread-safe.

Downstream uses of the volume class may have to be changed to ensure thread safety. Traditionally, a const volume has allowed its interpolation and extrapolation settings to be changed, which has led to widespread use of this coding pattern:

extrapolation oldex = vol.getextrapolationmethod();
if (/* oldex isn't suitable for the current application */) {
    vol.setextrapolationmethod(/* suitable extrapolation method */);
}

// do work which only involves reading from vol
float val = vol(            x, y, z);
val       = vol.interpolate(x, y, z);

// restore extrapolation method
vol.setextrapolationmethod(oldex);

Unfortunately, as this code mutates the state of the vol object, it is not thread-safe. Therefore a compiler warning will now be emitted whenever setextrapolationmethod is called on a const volume. Additionally, the operator(), interpolate, splineinterpolate, kernelinterpolation and extrapolate functions have been extended to accept a new optional extrapparams parameter. This allows the extrapolation settings to be changed without having to change the state of the volume instance, e.g.:

float val = vol(            x, y, z, /* extrapolation parameters */);
val       = vol.interpolate(x, y, z, /* extrapolation parameters */);

Similar changes have been made to the valid() functions with respect to the setextrapolationvalidity settings - code which previously had this structure (where vol is a const volume): 

std::vector<bool> oldvalidity = vol.getextrapolationvalidity();
vol.setextrapolationvalidity(/* suitable validity */);

// do work, reading from from vol
// ...
if (vol.valid(x, y, z)) {
  // ...
}
// ...

vol.setextrapolationvalidity(oldvalidity[0], oldvalidity[1], oldvalidity[2]);

should be adjusted to use this structure: 

bool xvalid = /* suitable validity */;
bool yvalid = /* suitable validity */;
bool zvalid = /* suitable validity */;

// do work, reading from from vol
// ...
if (vol.valid(x, y, z, {xvalid, yvalid, zvalid})) {
  // ...
}
// ...

This MR also makes some adjustments to the ShadowVolume class to ensure that interpolation/extrapolation settings on a ShadowVolume instance match that of the owning volume instance. This was previously a given, as the volume::operator[] function would return a new ShadowVolume instance on every call. However, now that we are using a ShadowVolume cache, there is a chance that the interpolation/extrapolation settings on a ShadowVolume instance could get out of sync with those of its owning volume. This has been resolved by refactoring the ShadowVolume class so that it retains a reference to its owning volume, and forwards all interpolation/extrapolation settings accesses to the owning volume. - BF: Use volume extrapolation getters rather than fields, as the fields may not have been initialised on ShadowVolume instances, whereas the getters will delegate to the owning volume - MNT: Adjust volume::invalidateSplines() to resolve performance issue Adjust the volume::invalidateSplines()function so that it accesses the p_interpmethod field directly instead of calling the virtual getinterpolationmethod() function. The latter approach has a noticeable effect on performance when many large images need to be written to via the operator() function.

In most circumstances use of getinterpolationmethod() is necessary, as there are circumstances in which the p_interpmethod field may not be set to the correct value on ShadowVolume instances. However in this situation accessing p_interpmethod is harmless - the code causes the ShadowVolume cache to be cleared, but a ShadowVolume instance will never cache other ShadowVolume instances. - ENH: Parallelise NEWIMAGE::convolve operations on a per-voxel basis The NEWIMAGE::convolve routines are trivially parallelisable - the same operation is performed independently for each voxel. This MR parallelises the inner loops using the Utilities::parfor routine, and uses source.nthreads() as the number of threads to use - this essentially means that the value of the FSL_NUM_THREADS environment variable will be used if it is set, or one thread will be used as the default.

This affects all uses of NEWIMAGE::smooth and NEWIMAGE::generic_convolve, for example the fslmaths -s flag.

When run in single-threaded mode, the new code is slightly slower than the original code, by about 5%. This is primarily due to the additional overhead of use of the vbegin()/vend() functions, which use a NEWIMAGE::voxelIterator to generate 3D voxel coordinates. - MNT: Don't use source.nthreads() to decide number of threads - just let NoOfThreads select the default value

Detailed history

  • MNT: Normalise indentation
  • MNT: Adjust raw_affine_transform so it doesn't modify properties of a const volume
  • RF: Make all get/set interp/extrap functions virtual. Override set* functions in ShadowVolume so that they throw an error. Adjust ShadowVolume class so it stores a reference to its owning volume
  • RF: Add interp/extrap setting getters to ShadowVolume which delegate to the owning volume instance
  • MNT: Remove unused variable
  • RF: Handle interp/extrap settings access prior to initialisation, just by returning the ShadowVolume settings
  • RF: Make _owner a non-const pointer-to-const to allow copy-assignment
  • RF: Change ShadowVolume creation method. Instead of using constSubVolume, create directly from the source volume, calling initialize+copyproperties directly. Also save the 4D index to allow for copy construction.
  • TEST: Test that interp/extrap settings are propagated to shadowvolume instances
  • MNT: Template instantiations for some things in newimagefns.h, and a few fixes in those things
  • RF: Revert volume.initialize visibility back to protected. No need to be public, but can't be private, as ShadowVolume calls it.
  • TEST: std::execution not supported by clang?
  • BF: Use volume extrapolation getters rather than fields, as the fields won't have been initialised on ShadowVolume instances, whereas the getters will delegate to the owning volume
  • MNT: More useful error messages
  • MNT: Access p_interpmethod field directly instead of calling virtual getinterpolationmethod() function - the latter causes a noticeable slowdown when writing large images
  • TEST: Adjust test build to make in-source building easier
  • TEST: Sample off-centre when testing interpolation
  • ENH: Parallelise NEWIMAGE::convolve operations on a per-voxel basis
  • TEST: Basic unit tests for NEWIMAGE::convolve, both standard and masked versions, parallelising with different numbers of threads, comparing results against scipy.signal.convolve
  • MNT: Don't use source.nthreads() to decide number of threads - just let NoOfThreads select the default value

fsl-newmesh 2111.0 -> 2111.2

Detailed history

  • MNT: Remove unnessary include statement - fslsurface is not required to build newmesh
  • MNT: Need to link against libfsl-utils on macOS

fsl-pipe 1.2.2 -> 1.3.0

Detailed history

  • Do not expect ignored value after --skip-missing
  • BF: linked placeholders can now be overwritten using the CLI
  • Release v1.3.0

fsl-possum 2111.2 -> 2111.3

Detailed history

  • MNT: Don't need to adjust volume extrapolation, as boundsassert/boundsexception are no longer options

fsl-pyfeeds 0.13.1 -> 0.15.0

Detailed history

  • RF: Adjust argument order for evaluation routines so they can be used by feedsRun scripts. Not sure why I didn't do it this way in the first place.
  • TEST: Update tests for new eval routine signature
  • MNT: Bump version. Expose all eval routines at package level
  • DOC: Mention eval routines
  • BF:: copy+paste typo
  • MNT: Accept pathlib.Path, normalise to string for image cache keys. All other eval routines should already be compatible
  • RF: Adjust evalHeader so it compares all valid dimensions. Replace alldims arg with ndims for comparing 3D
  • BF: Not passing pyfeeds object to loadImage
  • BF: Typos/mistakes in tempdir routine
  • TEST: Tests are useful it turns out. Use pyfeeds.common.tempdir instead of pyfeeds.tests.tempdir
  • ENH: New --printLogs flag, which prints all feedsRun.log files on success or failure. Useful for when running tests in CI, and you don't have access to the output dir

fsl-shapemodel 2111.0 -> 2111.1

Detailed history

  • MNT: Need to link against fsl-utils

fsl-sub-plugin-slurm 1.6.5 -> 1.6.7

Detailed history

  • Fix account/project enumeration
  • Add documentation about smt option
  • Remove support for 3.8 Add 3.13
  • Remove 3.8 from CI file

fsl-tirl 3.6.2 -> v3.6.3

fsl-utils 2203.5 -> 2412.1

  • ENH: Adjustments and enhancements to threading.h
  • Adjust the NoOfThreads class so that, if a value is not provided, or a value <=0 is provided, it will be set to a value read from the ${FSL_NUM_THREADS} environment variable, or 1 if that variable is not set.
  • Add new parfor and apply functions for trivially parallelisable tasks. These functions perform essentially the same role as std::for_eaach and std::transform, but allow explicit control over the number of threads.
  • ENH: New Utilities::ScopedTimer class, for lightweight profiling A new class, ScopedTimer, which allows code blocks to be timed in a lightweight manner. To use, you just need to create a ScopedTimer instance at the beginning of the code block to be timed, e.g.:
void some_func(bool some_cond) {
  if (some_cond) {
    ScopedTimer t("some_func - true branch");
    // do work
  }
  else {
    ScopedTimer t("some_func - false branch");
    // do work
  }
}

When the ScopedTimer instance goes out of scope, its destructor will calculate the elapsed time, and emit a message to standard out, e.g. if some_func(false) is called:

some_func - false branch: 2.5 seconds

Detailed history

  • ENH: Adjust NoOfThreads so that it defaults to FSL_NUM_THREADS, which is read from the environment. Add two new utility functions, parfor and apply, for trivially parallelisable tasks over vector-likes
  • ENH: parfor/apply overloads which can be called without a NoOfThreads instance, and use the default number of threads (FSL_NUM_THREADS if set, single-threaded otherwise)
  • ENH: Variants for parfor/apply - they can now be called: - with a NoOfThreads instance - with an integer - without either, in which case FSL_NUM_THREADS will be used (or 1 if it is unset)
  • ENH: New Utilities::ScopedTimer class, for lightweight profiling

fsl-warpfns 2203.0 -> 2501.0

This MR depends on fsl/utils!15 and fsl/newimage!40 - RF: Change NEWIMAGE::volume extrapolation settings in a thread-safe manner This MR builds on !7, which must be merged first.

This MR makes some adjustments to the functions in warpfns.h so that changes to the extrapolation settings of NEWIMAGE::volume instances are made via passing in extrapparams instances at the time of interpolation/extrapolation. These changes should be source-compatible with downstream projects, but they will require recompilation.

This is an alternative to using the setextrapolationmethod and similar functions which, because they change the state of the volume instance, cannot be safely used in multi-threaded applications.

I have also added a warpfns_stubs.cc file which contains instantiations of the templated functions in warpfns.h, to allow us to catch errors at build time, rather than when downstream code is compiled.

Detailed history

  • ENH,RF: Use new features in utils/newimage to parallelise voxel-wise operations. - Don't pass a value to Utilities::NoOfThreads - this allows it to take a value from the FSL_NUM_THREADS environemnt variable (or default to single-threaded) - Use Utilities::parfor to paralellise voxel-wise operations
  • ENH: Parallelise shift2warp
  • MNT: Use volume.operator(x,y,z,t) rather than volume[t] (x,y,z)
  • MNT: Normalise indentation
  • BF: Don't use auto with armadillo expressions; don't use volume.operator[] within voxel for-loops
  • MNT: Don't call setextrapolationmethod on const volumes - pass extrapolation settings into volume function calls instead
  • RF,BK: Rename set_extrapolation to calculate_extrapolation, and change it to return a pair of extrapolationparams instances. Next step is to update warpfns functions to pass these through.
  • MNT: Normalise indentation, pass input volume extrapolation params to affine_no_derivs/affine_with_derivs
  • MNT: Normalise indentation, pass extrapolationparams to displacements_[no,with]_iT
  • RF: Move extrapparams arguments alongside inputs - these are internal functions, so no need to preserve signature
  • RF,BF: Rename to extrapparams, a couple of fixes
  • RF: Adjust raw_general_transform to accept optional extrapparams for input volume and displacement field
  • RF: Adjust apply_warp to pass custom extrapolation parameters in rather than calling vol.setextrapolation..
  • MNT: Add stubs file which contains template instantiations of warpfns.h functions, to catch errors at build time
  • MNT: No longer need to adjust boundsassert/boundsexception extrapolation types, as they have been removed from newimage
  • BF: Accidentally re-declared variable
  • BF: Was calling wrong volume::valid overload - float coords silently coerced to int with no compiler warning. Fix logic in calculate_extrapolation - dfield method should be either periodic or extraslice

fsl-xtract 2.0.6 -> 2.0.7

Detailed history

  • Merge pull request #10 from SPMIC-UoN/xpy
  • minor fix to splash
  • minor fix: resolve splash escape sequence error
  • minor updates to splash

fsl-cudabasisfield 1.1.1 -> 1.1.2

Detailed history

  • MNT: Add FSL copyright placeholder
  • MNT: Use default compilation/dependency rules

fsl-mmorf v0.3.3 -> v0.4.2

The armadillo issue has been resolved upstream, but as we will keep using armadillo 14.2.1 for some time, a work-around is needed in the MMORF code. When we upgrade armadillo this work-around can be removed, although it is not particularly burdensome.

This MR also makes a few minor changes to the MMORF Makefile to work with the latest version of the FSL build system. - MNT: Revert behaviour of --version so that it does not cause MMORF to exit In version 0.4.0, the behaviour of the --version flag was changed so that it would print the MMORF version number and then exit. This breaks compatibility with existing MMORF configuration files that contain version = true, so this MR reverts the behaviour.

Detailed history

  • MNT: Use nthreads setting rather than hard-coding 6
  • DOC: Add some entries to changelog
  • DOC: More changelog updates
  • MNT: Set version number from git state
  • MNT: Print version at startup
  • MNT: formatting
  • MNT: Tweaks to Makefile to work with FSL build system behaviour - most notably: - use the default .cpp / .cu compilation rules - set "SRCDIR" before including default.mk, so that it doesn't try to take over our test compilation
  • TEST: Update to latest version of catch
  • TEST: Basic smoke test for main tensor cost function. Using pyfeeds test data
  • RF: Use temp matrix to store inverse to work around armadillo performance issue
  • MNT: CUDABUILDDIR should default to BUILDDIR if not set
  • RF: Revert behaviour of --version, as this was a breaking change