maedoc / libtvb

TVB C library

SDDEKit

SDDEKit is a C rewrite and generalization of numerical algorithms backing the simulator of The Virtual Brain project, though it is intended to support any sort of modeling project involving stochastic delay differential equations.

The current focus is nailing down 'software carpentry' aspects of building a scientific library, and only preliminary sets of mathematical components are implemented; as such, it is not for general use until the upcoming v0.1 release.

Build the library & benchmark with

python -m dev build

(if you work in Bash, you may want to alias dev='python -m dev ').

We're currently rewriting the test & benchmark suite using upcoming Python bindings.

See our docs for more information.

Many thanks to our contributors

• Abheek Ghosh [email protected]

This work is partially funded by the French Medical Foundation.

libtvb's Issues

rework interpretation of c in/out

{cache, call}grind (preliminary tests done)

sort net.M to evaluate models block-wise

other bindings (Java, R?)

extern "C" headers

profiling / benchmarks (preliminary tests done)

wrap various structs & functions

LL DDE tests from Jimenez 2006

obviate makefile

reorg src/ include/ test/

eval randn only for nnz g? test w/ benchmark

network sys (connectivity, heterogeneous model sets & couplings)

test main --test-name --test --bench (if whole suite gets too slow)

impl Jf Jg Jc + taylor o(1)

coverage

lorenz chaos, test lyapunov expo

Review memory copying

When receiving heap-allocated arguments kept internally, make internal copy. When returning heap-allocated data, write to address à la memcpy.

int init(void *obj, int n, double *x);

On return, x's contents may be freed.

int recv(void *obj, int n, double *x);

On return,x points to at least n elements of etc.

callbacks for hist fill, sys, scheme & out

allow specializations to SDE, ignoring delays, idem for DDE, ignore g

Test malloc failures explicitly

Dev machine has tons of memory malloc never fails and significant error handling code is not covered by tests.

This requires adding a custom allocator which counts calls and intentionally fails after a specific number of calls. The unit tests should cover malloc failures by checking return codes.

inspired by http://stackoverflow.com/questions/3868716

clean up cmake build & move to gtest

hist linear interp, cache lookup vector between time steps

drift & diffusion integrals can be handled independently?

add restrict everywhere

specialize hist for integer delay case (use gist impl)

Consistent nbytes method

Every sd_ struct interface should have an uint32_t (*nbytes)(sd_ *) method which returns the number of bytes used by that interface. This should include the internal implementation storage as well as the interface struct(s). sd_hist provides such a method & implementation and other structs should have the same.

solv deleg common cases to dedicated code path (if profiling says it make sense!)

const correctness (not really a thing in c)

(REJECTED) models loop over nx as appropriate (too complicated)

error handler and log system

• implement log msg handlers
• error handler
• library calls return 0 for success or non-zero for some error
• err handler called with string reason
• user can set own handler or no handler
• macro for 'raising' error

move to cmake (cython: github.com/thewtex/cython-cmake-example)

hide structs into opaque pointers, impl get/set methods (or wait for a v1.0 to stabilize)

implement neural mass models

specialize out sfilt handle differet nfilt for x & c

test gcc profile guided optimizations?

general temporal filter for eg hrf

move repo to ins-amu and develop via pull request, travis labels each with test results

test problems from Kuechler Platen 2000 references (page 1)

compare stability, convergence, memory, speed

use py/matlab/java test frameworks over bindings

exact history test from TVB

obviate test collector (using gtest)

cover history with more difficult cases (double updates, etc)

add lint checks to lint.sh (e.g. all allocations use MALLOCHECK)

provide finite diff scheme (verif analyt jac)

move this todo to github

e.g. convergence test systems from Biscay et al

out fn spatio temporal filtering, plain file binary/ascii, in-mem buffer

doxygen

model knows its index within overall system (for e.g. spatial parameters)