Devices

[thesamovar]

@mstimberg, take a look and maybe do an initial merge if you think it's OK to go into master? May not play nice with the functions branch.

[coveralls]

Coverage remained the same when pulling 4700202 on devices2 into 36a55c0 on master.

[mstimberg]

Related to what we discussed via email, I just made some changes to the way variables are handled. The problem was that CodeObject.__init__ took care of preparing the namespace for the CodeObject but this transformation is target-specific. This is now done in a separate function variables_to_namespace which is implemented differently for the different code objects.

Regarding dynamic arrays: These do no longer need to be "faked" as a Variable object when they are to be resized in a code object. Instead, the runtime targets add two symbols to the namespace array and array_object, where the former refers to the underlying array (for read-only situations, e.g. synaptic propagation) and the latter refers to the DynamicArray object (for resizing the array, e.g. in synapse creation). Standalone does not deal with this issue at this stage, it adds a code line doing something equivalent in Device.build.

That's not perfect yet, in particular the constant attribute is misinterpreted in several places. For code generation/execution this should not be relevant anymore as soon as we deal specifically with t and dt (the only AttributeVariable objects we are currently using). Currently, runtime code injects the values of non-constant variables into the namespace at every timestep but this is unnecessary, since the reference to the values does not change. I think the only use of the constant flag should be in high-level code such as state updaters.

What do you think -- should I merge it as it is and you continue to work on the devices part?

One small thing: I replace id(codeobj) by codeobj.name as the dictionary key for code_object_defs -- the id differs between the object and a weakref.proxy to it.

[coveralls]

Coverage remained the same when pulling f1a4e1e on devices2 into 36a55c0 on master.

[thesamovar]

Yep, go ahead and merge. Let's deal with the constants issue separately. By the way, does it play nice with your functions work?

[thesamovar]

Before you merge, note that in devices the BrianObject.update, pre_update and post_update methods are gone, and there are now Updater objects. There is also a now pointless .code_objects attribute in BrianObject which is the leftover of an earlier abortive attempt to do what the .updaters attribute does. This should be removed at some point but it doesn't harm anything.

[mstimberg]

By the way, does it play nice with your functions work?

I didn't merge it yet, I think I'll do it the other way round, i.e. merge master into finalize_functions after this pull request is merged. But I'm not too worried, anyway :)

[thesamovar]

Restored devices2 branch because I'll keep working on it.