ANNarchy/ANNarchy: ANNarchy Release 5.0

Important: this release introduces changes/features which are not compatible with ANNarchy 4.x. For more information please refer to the manual: https://annarchy.github.io/

• the construction of networks using a magic network is deprecated. All ANNarchy objects, such as populations or projections, should now use the Network class. A warning is printed when the old-style interface is called (e.g., 29a3154, 64ddc5d).
• Changed the default for reshape argument for Monitor.get() from False to True (see commit b4142d1)
• timed input populations, i.e., TimedArray, TimedPoissonPopulation, HomogeneousCorrelatedSpikeTrains, have now an update method. For this specific populations, its not recommended to use the direct accessors rates, schedule or period! (commit 9c8b4c0, 0d5c2e2)

Next to the above breaking changes, we further updated/added:

• Introduced default_rng property to Network, which is a numpy random default_rng instance seeded by seed provided to constructor (commit 591bc8d)
• uniform distributions are now initialized with a rng instance (i.e., numpy.random.default_rng, default_rng attribute (see above), or comparable) instead of a seed (commit 90bad44).
• uniform distributions (int/float) expect now that min<max is fulfilled (see commit f1ba487)
• added reseed_rng attribute to Network.reset() to reset RNGs (commit 144a6e2, aeca104)
• Attributes recorded by monitors can be accessed now directly: m['r'] instead of m.get('r')
• ANNarchy uses now nanobind instead of Cython as interface between Python and C++
• We extended the implementation for various formats, e.g.:
  • CSRC on GPUs (commit 2764e2b)
  • Dense on GPUs (commit 8b46823)
  • Experimental feature: Bitmasks for Dense (commit 89e8f0e) and BSR (commit 4ead060)
  • Dense/BSR is limited to projections without PopulationViews (72dd33d)
• Restructured unittests and changed the execution of them (see ANNarchy/tests/README.txt)
• Added standalone C++ unittests for the sparse matrix implementations placed in ANNarchy/include
• Added utility functions for conversion of ms into steps (commit 672c08e)
• Reduced the number of memory transfers between GPU and Host to improve overall simulation time (e.g., c80d403,116264b)
• Reduced the number of stored RNG states on GPUs (commit 02c1919)

Bugfixes:

• Projection.save_connectivity() / Projection.from_file() ignored the dt when storing delays (commit a877043)
• Fixed a bug for models using non-uniform delays and OpenMP parallelization (commit a9a0530, 5430e94)
• Fixed wrong ordering of indices in PopulationViews, when they are constructed from IndividualNeurons (commit 6777787).
• Non-ODEs related to synaptic inputs (i.e., g_[target]) has not been executed during refractory period. (commit 01dc47f)