* This release brings both general improvements and new features, as well
as a slew of maintenance updates and bug fixes. The first few items below
will discuss the main highlights.
* General speedup in a lot of the code. In particular the Fast Fourier
Transform code, used when expanding phonon density of states to full
scattering kernels for inelastic modelling, was improved by as much as
40%. This will make a noticeable difference when initialising many
materials, and is an improvement independent from the multi-threading
option mentioned in the previous item.
* If requested, NCrystal can now initialise materials even faster, by
performing the initialisation calculations in multiple threads. This can
for instance be enabled by setting the environment variable
NCRYSTAL_FACTORY_THREADS to the number of threads you wish to use (e.g. 8
if you have 8 CPU cores available, or 9999 to have NCrystal pick a number
suitable for your machine). Alternatively, you can call functions in the
Python, C, or C++ APIs to enable this:
NC.enableFactoryThreads() Python
NC::FactoryThreadPool::enable(); //C++
ncrystal_enable_factory_threadpool(9999); //C
export NCRYSTAL_FACTORY_THREADS=9999 BASH
* NCrystal now includes the first iteration of a "Mini Monte Carlo"
framework (for lack of a better word), which makes it possible to insert
an NCrystal material into a simple geometric "sample" shape, subject it
to a source of neutrons, and obtain a spectrum or pattern of the outgoing
neutrons. Thus, one can easily produce a simplistic diffraction pattern,
including effects of multiple scattering and both elastic and inelastic
physics. For now, the focus has been on developing the framework itself,
rather than preparing a library of interesting shapes, sources, and
tallies. Thus, the only readily available setup in this release is that
of a monochromatic pencil beam impinging on a sphere, tallying the
outgoing angle of the neutrons as they leave that sphere. Nonetheless,
the results provides a very convenient view of how NCrystal modelling
translates into a particular pattern -- without the need for an
additional software package like McStas, Geant4, or OpenMC. The `nctool`
command will now by default show such scattering pattern, or they can be
explicitly requested, as shown in this example where 1.8Aa neutrons
impinge on an aluminium sphere with a diameter of 2cm:
$> nctool --mc 1.8Aa 2cm Al_sg225.ncmat
The scope of this "Mini Monte Carlo" framework is of course not to try to
rival existing applications in terms of ability to simulate complicated
geometries or beamlines, but rather to provide another convenient way to
quickly investigate NCrystal materials, as well as potentially serve as a
way for NCrystal to be used in a regression-loop while analysing actual
neutron scattering data. For that reason, speed has been a major concern,
and both multithreading and vectorisation has been utilised to bring the
time it takes to go from defined NCMAT data to having a high-statistics
diffraction pattern available, down to the sub-second scale.
* Internally NCrystal code no longer uses raw `std::cout` or `printf`
statements. Rather, all messages are now emitted via a centralised
infrastructure. This most importantly fixes an issue where output from
the Python and C++ layers could occasionally be emitted in incorrect
order - now, when using the NCrystal Python API, all output will be
emitted through Python's own print buffer. Secondly, several NCrystal
objects now gets a `.dump_str()` method, meaning for instance that
material dumps can be captured in strings rather than emitted on
std-output. Finally, this new feature could in principle be used for
various output redirection purposes (e.g. a GUI application might wish to
show all output from NCrystal in some dedicated text box).
* NCrystal has been updated to work with latest versions of Numpy, Clang,
GCC, Python, Gemmi, and Spglib. Additionally work has started to support
Intel OneAPI compilers and Microsoft VisualStudio, although more work is
needed to finalise this. As part of these efforts, the ABI had to be
broken in some places, since for instance Microsoft VisualStudio did not
support all the data structures used previously (for instance,
VisualStudio `std::vector` does not accept move-only objects like
`std::unique_ptr`).
* This concludes the highlights in this changelog entry - refer to the
following items for more detailed information about specific changes.
* Several speedups of the phonon convolution calculation code was
implemented, resulting in noticable speed increases (20-40%): the FFT
phase factor calculation by a factor of two, swap indices were cached,
and an integer division was rewritten as a bit shift.
* Reduce luxury in some VDOS integrations slightly for a more reasonable
amount of calculations, speeding up Info material creation.
* Add internal utilities to allow usage of multiple threads for material
loading, allowing for sizeable speedups. This required many changes to
ensure good behaviour (i.e. using a fixed number of threads only), and no
dead-locks of various kinds even though the multi-threading can take
place at various levels during the loading. For now, the number of
available threads can be controlled with the NCRYSTAL_FACTORY_THREADS
environment variable, the NCrystal::FactoryThreadPool::enable() C++
function, the NCrystal.enableFactoryThreads Python function, or the
ncrystal_enable_factory_threadpool C function.
* Build NCrystal library with -fno-math-errno (github issue 169).
* Several non-trivial fixes were implemented to prepare for MS VisualStudio
and Intel oneapi compilation support. This breaks ABI.
* Apply -fp-model=precise for intel oneapi compiler.
* Change C++ DynamicInfoList type from
std::vector<std::unique_ptr<DynamicInfo>> to typedef
SmallVector<std::unique_ptr<DynamicInfo>,4>, since VisualStudio
apparently does not support std::vector's with move-only objects. This in
principle breaks the ABI and API, but is deemed necessary. A few other
types in internal non-API headers were changed as well for the same
reason.
* Fix build error due to ODR violation in the TopLvlVar type.
* Removed noexcept from C++ functions where it should not have been added
(e.g. those used for streaming to std::ostream).
* Fix cfg-string bug (github issue 182). This only affected multiphase
cfg-strings in which the first phase contained either a density or
phasechoice setting, so is not believed to have affected many users.
* ProcComposition::consumeAndCombine(..) fct is now better at removing null
components.
* Adding the infrastructure for better control of NCrystal output, through
usage of a new setMessageHandler(..) function (related to github issues
53 and 122).
* Rewrote the NCrystal::dump C++ function (which provides printouts/"dumps"
of material Info objects) to stop using printf entirely and now only use
C++ output streams, but notably while preserving existing output
exactly. This makes it possible to capture the dumps into strings and
streams, rather than printing them out. Along with the general message
handling changes, this should prevent Python output buffering causing
incorrect output ordering (seen in e.g. Jupyter notebooks).
* Python objects with .dump(..) methods gets new .dump_str() methods as
well, which simply returns the information as a string instead of
printing it directly.
* Update cifutils to follow changes in upstream Gemmi project.
* Updates to reflect upstream API changes in spglib.
* Stop using numpy.asfarray which was removed in numpy 2.0.0 (github issue
185).
* More helpful error message when loading huge DFT-generated unit cells
(thousands of atoms, no symmetries) without raising dcutoff (github issue
160).
* DI_VDOSDebye Python objects now also get an analyseVDOS() method, for
consistency (github issue 174).
* Update NCMAT documentation to clarify the ordering of sab entries (github
issue 170).
* Add convenience .xsect(..) and .macroscopic_xsect(..) methods to
LoadedMaterial instances in the Python API.
* Add experimental material loading benchmark utility to nctool. This
feature is for now kept out of the documentation, but can be tested with:
"nctool --bench --help".
* Updated year in copyright notice in all files.
* Info objects gets a new .hklObjects() method which returns HKL list
information in a more object oriented and convenient format than the
existing .hklList(). See issue 164 for more details.