activate spack env

.github/workflows/build.yml

@@ -5,7 +5,7 @@ jobs:
   pip-tvbk:
     runs-on: ubuntu-latest
     strategy:
-      fail-fast: false
+      fail-fast: true
       matrix:
         python-version: [ "3.10", ]
 
@@ -61,6 +61,7 @@ jobs:
         shell: spack-bash {0}
         run: |
           cd tvb_kernels
+          spack env activate .
           pip install .
           pytest
 

tvb_kernels/README.md

@@ -6,85 +6,34 @@ This is a library of computational kernels for TVB.
 
 in order of priority
 
-- [ ] sparse delay coupling functions
+- [x] sparse delay coupling functions
 - [ ] fused heun neural mass model step functions
 - [ ] neural ODE
 - [ ] bold / tavg monitors
 
-## building
-
-For now, a `make` invocation is enough, which calls `mkispc.sh` to build
-the ISPC kernels, then CMake to build the nanobind wrappers.  The next
-steps will convert this to a pure CMake based process.
-
-## variants
-
-### batch variants
-
-It may be desirable to compute batches of a kernel.
-
-### parameters
-
-Parameters may be global or "spatialized" meaning different
-values for every node.  Kernels should provide separate calls
-for both cases.
-
-## matlab
-
-Given some users in matlab, it could be useful to wrap some of the algorithms
-as MEX files.
-
-## working notes
-
-### inner loop
-
-for the inner loop, excl monitors etc, all that's needed is
-- cx_all
-- step: stochastic heun on local modelo
-- buf update
+which will then be used internally by TVB.
 
-### scalable design
-
-#### parameter sweeps
-
-separating the data from the work arrays, allows defining
-the data once and reusing it, which is important for large
-connectivities.
-
-the work arrays can then be allocated for each simulation to do.
-
-#### components
-
-enabling a multicomponent design requires some modifications.
-as a useful example consider a model with the following elements
-
-- cortical field with modified epileptor
-- subcortical regions with Jansen-Rit
-- corticortical excitatory connectivity
-- subcortical excitattory connectivity
-- cortical->subcortical inhibitory connectivity
-- subcortical->cortical inhibitory connectivity
-- dopamine region
-- dopa->cortical connectivity
-
-which implies some notions we already have. let's use the
-word *domain* for field, regions, etc and *projection* for
-the different connectivities
-
-- a domain has one or more input *ports* which sum afferent
-  values from various projections
-- a domain has one or more output *ports* which expose efferent
-  values read by various projections
-- a single projection may transport one or more values
-
-changes required
-
-- proj should not own the afferent cx buffer
-  rather take it as a operator argument like t
-  or just have a pointer
-- a given model should own the input port buffer
-- projections should own the delay buffer
-- delay buffers should have num_cvars as last dimension
+## building
 
-- once step taken, push to "listening" projections
+No stable versions are available, so users can install from
+source with `pip install .` or to just get a wheel `pip wheel .`
 
+For development, prefer
+```
+pip install nanobind scikit-build-core[pyproject]
+pip install --no-build-isolation -Ceditable.rebuild=true -ve .
+```
+
+## roadmap
+
+- improve api (ownership, checking etc)
+- variants
+  - single
+  - batches
+  - spatialized parameters
+- cuda ports for Jax, CuPy and Torch users 
+- mex functions?
+- scalable components
+  - domains
+  - projections
+  - small vm to automate inner loop