https://github.com/vitay/ANNarchy

sort by:
Revision Author Date Message Commit Date
16ab4d6 Release 4.7.0.2. 19 January 2022, 09:07:18 UTC
2e2977d Bugfix (openMP): the spike generation of SpikeSourceArray was not implemented correctly. 18 January 2022, 15:49:32 UTC
fc5a909 Bugfix (CUDA): code generated for delays and spiking was not correct. 18 January 2022, 15:48:04 UTC
8521886 Small adjustment of commit 4ae3cc9: the compiler flags are only shown if verbose (--verbose/-v) enabled. 11 January 2022, 10:29:51 UTC
58a7aef Added check on the parameters of the random distributions. 10 January 2022, 11:07:01 UTC
0f3949a Reverted the cpdefs in Connector.pyx 07 January 2022, 12:33:57 UTC
4ae3cc9 MacOS: now supports openMP with gcc 07 January 2022, 12:15:49 UTC
a7498bb Merge branch 'master' of https://ai.informatik.tu-chemnitz.de/gogs/annarchy/annarchy 06 January 2022, 18:08:40 UTC
0b4ad83 MacOS: attempt at using openMP with gcc 06 January 2022, 18:08:29 UTC
6e5dfb4 The <sched.h> import is only valid for linux systems. 06 January 2022, 16:58:31 UTC
5fd9f11 Merge branch 'master' of https://ai.informatik.tu-chemnitz.de/gogs/annarchy/annarchy 06 January 2022, 14:46:01 UTC
ab279af Adapted for mac m1. 06 January 2022, 14:45:23 UTC
bf64e4f Script does not need to be executable. 15 December 2021, 16:16:18 UTC
faa1653 Bugfix/Enhancement (commit 281e272): constants (in this case the uniform delays) are stored as 0-dimensional ndarrays. To detect them, ndim check should be favoured instead of size. 09 December 2021, 13:49:29 UTC
a46d305 Release 4.7.0.1 19 November 2021, 08:20:39 UTC
5d62a1e BoldModel: allow strings as argument for *inputs* if only one variable should be considered. 11 November 2021, 13:46:12 UTC
796c473 Updated changelog. 11 November 2021, 09:36:16 UTC
3b22830 Added openMP support for the BOLD monitor extension. 11 November 2021, 09:30:08 UTC
88975af addition to commit 281e272 21 October 2021, 08:00:38 UTC
79ce864 Removed test code from example. 11 October 2021, 14:15:58 UTC
7717815 Commit 281e272 ignored single weight constants. 11 October 2021, 14:13:22 UTC
adf386e Some test on bitbucket pipeline 11 October 2021, 13:13:22 UTC
281e272 Critical bugfix: save/load of networks was not working. Disabled CPP-side connectors as the issue is not fixed for them yet. 11 October 2021, 12:56:24 UTC
842803d Bugfixes/Improvements: - Replaced deprecated numpy.random.random_integers by numpy.random.randint - fixed a code generation bug related to delays in custom connector functions - added new unittests to cover the previous described bug in the future 11 October 2021, 08:28:43 UTC
e1db0ac travis-ci stopped working 04 October 2021, 08:37:03 UTC
1855475 Updated bitbucket pipeline. 04 October 2021, 08:25:35 UTC
40f4d45 Release 4.7.0 04 October 2021, 08:18:56 UTC
4291e9c bold extension: BOLD is the default output of a BOLD model 01 October 2021, 13:30:39 UTC
ac4eb5a The destruction of sparse matrix containers is now performed by ProjStruct::clear() instead of destructors. 30 September 2021, 08:56:05 UTC
77d13d9 BOLD: changed I to I_CBF 29 September 2021, 10:07:10 UTC
5bb1c52 Merge branch 'develop' of bitbucket.org:annarchy/annarchy into develop 29 September 2021, 08:20:31 UTC
f3e4d9c Bugfix (a65cd0f): did not consider the fact that a network might not be created. 29 September 2021, 08:18:35 UTC
54b8970 Merge branch 'develop' of bitbucket.org:annarchy/annarchy into develop 29 September 2021, 08:01:13 UTC
b14a714 Small changes to the doc. 29 September 2021, 08:00:38 UTC
a65cd0f Bugfix: in some cases the generated shared library was not loaded correctly. 29 September 2021, 07:56:35 UTC
5d80df0 Merge branch 'develop' of https://ai.informatik.tu-chemnitz.de/gogs/ANNarchy/ANNarchy into develop 29 September 2021, 06:21:08 UTC
16e2dda Minor bugfixes. - floating constants used in numerical methods: some methods like midpoint or RK4 have floating values in their evaluation, more precise dt/2.0 or dt/6.0. For single precision, this values need to be casted into dt/2.0f or dt/6.0f otherwise appearantly implicit type conversion is performed which increases computation time. - missing rng argument in CSRC_T::fixed_probability_pattern 29 September 2021, 06:19:06 UTC
e12059d Added documentation of the BOLD extension. 28 September 2021, 16:42:25 UTC
a562d5a Disabled CPP-side connectors by default. 22 September 2021, 08:41:54 UTC
6888143 Addition (commit a8da945): PartitionedMatrix need to implement this function too. 10 September 2021, 07:13:48 UTC
c629fce Improved performance of fixed_number_pre connector functions for CSR-type formats. 09 September 2021, 17:09:37 UTC
d53b60d Enabled CPP-side construction of fixed_number_pre pattern for the CSR (includes CSRC and CSRC_T) format. 09 September 2021, 11:53:15 UTC
a8da945 Added a function to count the number of efferent connections in a projection (spiking models only). 09 September 2021, 08:04:27 UTC
b01b847 Bugfix: code generation for Runge-Kutta 4th order was not executioning the statements in the correct order. 03 September 2021, 13:12:12 UTC
0f6c3d3 Removed chunks_ which was intended to mimic an openMP static scheduling. This should be replaced by a thread-local implementation in the future. 02 September 2021, 10:04:21 UTC
4300c69 Minor bugfixes and code improvements + LIL-SIMD: loop unroll for adding up partial sums + fixed ELLRMatrixCUDA inheritance error 01 September 2021, 08:41:37 UTC
cfa1575 Moved structural plasticity template codes into format specific files. + implemented reset(synapses==True) 01 September 2021, 07:00:16 UTC
eab381a Refactoring, CPP template code documentation and finalized commit 25dc72f - finalized distinct code pattern files for CSRC and transposed CSRC. - bugfixes in distinct code pattern for sliced LIL - resolved naming incorrectness as the "ell" format was indeed an ELLPACK-R. - added ELLPACK format for CPUs. - Hybrid format depends on ELLPACK now. - ST parameter is now part of all CPP templates - changed memory transfer mechanism for GPUs - Removed some remaining references to CudaCheck. - Transformed ParalleLIL into a more generalized class: PartitionedMatrix which can handle also other SpMV formats 23 August 2021, 07:02:24 UTC
7351ec5 Adjusted figure labels for the BOLD examples. 18 August 2021, 13:45:29 UTC
46da4dd Suggested changes by Oliver on the normalization. 12 August 2021, 10:42:34 UTC
bfdf0f0 Avoid the usage of omp_get_max_threads/omp_set_number_threads. If multiple CPP libraries in one Python-context access the global number of openMP threads, this can lead to unpredicatable segmentation faults. 02 August 2021, 11:44:09 UTC
0d59c52 bugfix in code generation. 30 July 2021, 11:06:35 UTC
c763e6b Bugfix: save_connectivity couldn't access w. 30 July 2021, 10:26:58 UTC
90839d1 Added BOLD_two_inputs example. Updated balloo_two_inputs equations. Added some missing references. 28 July 2021, 09:18:04 UTC
2532a11 Refactoring in the new BOLD monitor interface. BoldMonitor: - renamed: input_variables -> source_variables, output_variables->input_variables PredefinedModels: - hardcoded 1000. within equations replaced by a parameter - models are now implemented as classes (as done for built-in neuron/synapse models) - added literature references. 27 July 2021, 15:26:32 UTC
02d9374 bold extension: ctaching variable r is fixed. 27 July 2021, 08:20:17 UTC
19cc393 Disabled recording of parameters (Interface, code generation was already - commit 530aa07 ). Removed remaining occurences of ANN4.6-style accessors (as done in commit 5470ae1) 27 July 2021, 05:43:21 UTC
a03ee47 Introduced BoldModel for an easy-to-use definition of bold models. 26 July 2021, 17:21:58 UTC
25dc72f CPP templates: added ST template argument for member variables whose maximum value exceed number of rows/columns. 26 July 2021, 14:02:12 UTC
5470ae1 Bugfix: Projection._set_cython_attribute() still called up ANN4.6-style CPP accessors. 23 July 2021, 14:41:07 UTC
e7a7534 Some code documentation. 21 July 2021, 16:14:21 UTC
7af1e64 Merge branch 'master' into develop 21 July 2021, 16:13:32 UTC
d7ee4c0 TransposeProjection: added access to post_ranks, nb_synapses for the inverted matrix. 19 July 2021, 14:17:20 UTC
530aa07 Refactor (CPU/GPU code generation): * removed recording code for parameters Refactor (GPU code code generation): * global parameters are provided as constants to the kernel not as allocated device variables * removed device side t variable, instead provide it as function parameter * removed unnecessary cudaDeviceSynchronize() calls - currently all functions are placed in default-stream * memory transfers (device to host) are not performed by default anymore. the first access to a variable will trigger the memory transfer. * spiked index transfer only if there were events (appearantly CUDA detected this by itself, but it's better to prevent zero-copies) * CSR, COO: removed old accessor templates. CPP Templates (CUDA): * CSR: added FP, FNP pattern * CSRC: added FP pattern * ELL-R: only a debug statement Bugfix(TimedPoisson/openMP): * update was executed by all threads which lead to too many spike events. 19 July 2021, 11:32:53 UTC
a6ee07d Refactor: nb_synapses() returns overall number of synapses in a C++ projection while dendrite_size(int lil_idx) returns the number of synapses in a given dendrite. Improved performance of COO accessors under the assumption of sorted ranks. Introduced load_from_matrix_market which is intended for benchmarking. 15 July 2021, 12:56:57 UTC
8d485c6 A first draft for the BOLD monitor example. 14 July 2021, 07:02:18 UTC
3fd2224 Refactored launch configuration for CUDA device functions. Until now, we had a global define which determined the number of blocks/number of threads per block. Especially the different sparse matrix formats require a more fine-grained definition. As first step, I introduced a projection/population variable. The control of the variable will be more fine-grained in the future (maybe also extend to one variable per operation) 13 July 2021, 17:13:10 UTC
61b9e20 Bugfix (commit 48501eb): missing accessor for specific projections (i. e. conv, pool) 13 July 2021, 10:27:19 UTC
72c3c56 Some changes on the BOLD monitor (suggested by Oliver): - BOLD models: unified variable names and updated BOLDNeuron_new - ACCProjection: baseline normalization using standard deviation removed - BOLDMonitor: BoldNeuron_RBN is now the standard model 12 July 2021, 09:03:17 UTC
48501eb Removed ProjStruct.size attribute. Python/Cython should access the nb_dendrites() function which is part of the connector definition. Added missing nb_dendrites() to ParallelLIL. Tentative: introduced SizeType template parameter for C++ (only LIL format yet). 12 July 2021, 08:52:54 UTC
34c2df5 storage_format argument was not intended to be obligatory. 12 July 2021, 06:44:11 UTC
9ba8c15 Added init_from_csv to CSRMatrix (intended for benchmarking). 12 July 2021, 06:37:29 UTC
637138e Bugfix (commit d7fdf6c): missing update of COO-partition (HtoD memory transfer) 05 July 2021, 09:00:30 UTC
cdc01b6 Disabled load/store of projections using convolution. 02 July 2021, 08:12:10 UTC
3286572 Added AVX version for double precision on ELLPACK format (single thread). 30 June 2021, 15:11:45 UTC
d7fdf6c Hybrid sparse matrix format on GPUs. 30 June 2021, 06:37:46 UTC
d06c230 CUDA code generation: added const/__restrict__ modifier on continuous transmission kernel 28 June 2021, 10:41:09 UTC
ec835d5 Bugfix introduced in commit d10a59d 28 June 2021, 09:35:21 UTC
3882446 BoldMonitor: z-score is riskful if standard deviation is near 0.0. Slightly modified formula (in agreement with Oliver) 28 June 2021, 08:23:21 UTC
d10a59d FNP pattern for rate-coded models on CUDA. 28 June 2021, 07:42:46 UTC
dded804 missing return in CSRMatrix. 25 June 2021, 08:14:55 UTC
01a247b Tentative: Runge-Kutta 4th order for coupled ODEs 23 June 2021, 08:07:30 UTC
2c03fca Annotate handwritten SIMD code with profile measurement. 22 June 2021, 08:49:20 UTC
6201aca + disable hand-written SpMV on non-linux systems. + Hand-written SpMV with single weight using SSE. 21 June 2021, 09:16:52 UTC
188bb7c + added flag to disable hand-written SIMD (intended for benchmarking only) + added uniform delays for Hybrid format (linear rate-coded neuron benchmark) + ELLMatrix shows in debug some more details on the nonzero distribution 21 June 2021, 08:05:30 UTC
196c7a6 Continuation of commit 8607faf 18 June 2021, 14:06:49 UTC
7e568c3 If "show_time" is set to True the times required for code gen, compile and initialize is printed. 18 June 2021, 08:36:54 UTC
c499313 Prototype implementation for a Transpose projection applied on a spiking projection. 17 June 2021, 12:53:13 UTC
61cdf9c TransposeProjection should not export any values. Added sanity check after construction of the inverse view. 17 June 2021, 09:22:48 UTC
24158a8 + Bugfix (SpecificPopulations): pop.cyinstance.set_xxx is not working anymore due changes in commit e52f3b3 + Added clear() field to allow definition of own destructor codes for projections. 17 June 2021, 08:14:18 UTC
bb512f2 Bugfix (f533891, CUDA): missing differentiation. 16 June 2021, 11:30:41 UTC
f533891 Accessors are only generated for those categories (i. e. local, semiglobal, global) where are variables to access/return. 16 June 2021, 10:56:44 UTC
e52f3b3 Restructured accessor functions for population attributes: instead one function per attribute, we have one function per cpp data type. 16 June 2021, 09:41:19 UTC
956e4b7 Small changes in code generation. + Missing deepcopy within ProjectionGenerators lead to wrong template dictionaries. For instance, first LIL then CSR lead to the situation that a CSR with AVX code was generated (and crashed during compile). + Spiking models should use sliced matrices for post-to-pre matrix orientation to maintain an acceptable performance. 14 June 2021, 14:04:05 UTC
f3221f7 Applied tentative optimization (commit be5b53a) on openMP. Adapted SIMD-SpMV codes for sliced LIL-matrices. 14 June 2021, 10:21:53 UTC
6ea30fe Refactor: vectorized SpMV is now part of the code dictionary. instead of hard-code links to specific templates. 14 June 2021, 09:22:58 UTC
8607faf CSR on GPUs: row_ptr_ is from type size_t. 14 June 2021, 07:53:39 UTC
6f6bdf5 Refactor: removed paradigm postfix from template files. Now _T (transposed) and _P (partitioned) remain as possible postfixes. Consequently, its possible to remove doubled entries for sliced/non-sliced matrices which slightly increases readability of the generator code. 14 June 2021, 07:26:34 UTC
94949d5 Bugfix (e6acec8): missing datatype in setter used in NormProjection. 10 June 2021, 17:39:13 UTC
3acec06 ANNarchy.core.Global.clear() did not destroyed the magic network (and especially the bound C++ objects). Destructor codes for projections are now part of the templates. 10 June 2021, 07:31:35 UTC
back to top