Group merged refactor

[neworderofjamie]

This has been annoying me for a while but I was faced with defining and passing through yet another callback and I decided to fix it. Basically, in GeNN 4 I split the generic code generation from the backend-specific code generation using a callback system like:

backend.genCustomUpdate(customUpdate, modelMerged,
        ...
        // Custom update handler
        [&modelMerged](CodeStream &os, const CustomUpdateGroupMerged &cg, Substitutions &popSubs)
        {
            genCustomUpdate(os, popSubs, cg, modelMerged, "id",
                            [&cg, &modelMerged](const Models::VarReference &varRef, const std::string &index)
                            {
                                return cg.getVarRefIndex(varRef.getDelayNeuronGroup() != nullptr,
                                                         modelMerged.getModel().getBatchSize(),
                                                         getVarAccessDuplication(varRef.getVar().access),
                                                         index);
                            });
        },
        ...

This made sense as there was no code-generation level object where this stuff could go. However, since adding support for kernel merging, the (slightly nightmarish) GroupMerged class hierarchy fits the bill exactly so all these callbacks can instead be added as methods to the appropriate GroupMerged classes. This tidies up a whole load of stuff and reduces the pain of adding new functionality:

• There's no need to typedef a callback type for each type of merged group in BackendBase
• The main BackendBase interface no longer needs updating and the main generating functions no longer need dozens of arguments:

virtual void genSynapseUpdate(CodeStream &os, const ModelSpecMerged &modelMerged, HostHandler preambleHandler,
                              PresynapticUpdateGroupMergedHandler wumThreshHandler, PresynapticUpdateGroupMergedHandler wumSimHandler,
                              PresynapticUpdateGroupMergedHandler wumEventHandler, PresynapticUpdateGroupMergedHandler wumProceduralConnectHandler,
                              PostsynapticUpdateGroupMergedHandler postLearnHandler, SynapseDynamicsGroupMergedHandler synapseDynamicsHandler,
                              HostHandler pushEGPHandler) const = 0;

[codecov]

Codecov Report

Merging #475 (f12694f) into master (9d43b7c) will decrease coverage by 0.05%.
The diff coverage is 95.03%.

@@            Coverage Diff             @@
##           master     #475      +/-   ##
==========================================
- Coverage   87.52%   87.46%   -0.06%     
==========================================
  Files          78       82       +4     
  Lines       17278    17266      -12     
==========================================
- Hits        15122    15102      -20     
- Misses       2156     2164       +8     

Impacted Files	Coverage Δ
include/genn/backends/cuda/backend.h	96.29% <ø> (ø)
include/genn/backends/opencl/backend.h	98.86% <ø> (ø)
...nclude/genn/backends/single_threaded_cpu/backend.h	82.35% <ø> (ø)
include/genn/genn/code_generator/backendBase.h	91.52% <ø> (ø)
include/genn/genn/code_generator/backendSIMT.h	96.51% <ø> (ø)
include/genn/genn/code_generator/groupMerged.h	84.18% <ø> (-2.47%)	⬇️
include/genn/genn/code_generator/modelSpecMerged.h	97.40% <ø> (ø)
...enn/code_generator/presynapticUpdateStrategySIMT.h	100.00% <ø> (ø)
src/genn/backends/cuda/optimiser.cc	77.55% <ø> (ø)
src/genn/generator/generator.cc	90.00% <ø> (ø)
... and 19 more

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 9d43b7c...f12694f. Read the comment docs.

[tnowotny]

Ok, I have lost track of how the code generation works after the group merging and splitting to backends etc. I have to trust you on this one.