Network message

[Robadob]

Some of the original ideas here.

Stage 1 - Static environment graph

• Description
• Host Set
• Host Get
• Host Misc (e.g. from file)
• Host API element methods to accept N=0 (I may rework this, the solution is imperfect)
  • Host API SWIG elements
• Agent Get
• Agent Traverse
  • Edges Leaving
  • Edges Joining
• Device Curve
• RTC Curve
• Submodel support
• SWIG support (This is main reason CI currently failing)
• C tests
• Py tests
• Refactor to remove nested class
• Solve Agent Traverse seatbelts
• Improve HostAPI auto sync (See comment below)
• Make ID useful? (This adds an extra layer of indirection to lookups, will need to consider perf impact)
• Remove indexable edge access? (The sort for CSR/CSC mixes up the order)
• Docs
• Add ADVANCED_API method to HostAPI to expose CUDA buffers/stream.

• Do we want host traverse? (Not immediately)
• (internal) Proper backing graph data structure/sort

Stage 2 - Messages for the graph

• Everything (or just use bucket?)

[Robadob]

I don't really like how inside EnvironmentDirectedGraph::CUDABuffers, a reference to the map of agents is held, so that when buffers are resized curve can be updated in each agent function. Something about this feels wrong.

Alternate options being:

• A central set of HostCurve objects is maintained in parallel which can be shared.
• Everytime an agent function is loaded, it re-sets the graph ptrs in curve, although they won't change beyond init in most cases.

[Robadob]

All current TestEnvironmentDirectedGraph tests are now passing. This was solved by defining edge source-dest pairs in the same order they would be sorted into.

It's assumed subsequent tests will cover the traversal of graphs, so the actual sort will be tested implicitly.

However, we may wish to remove indexable edge access in future, as the sort will lead to confusion??

[Robadob]

ID currently exists as an automatically added Vertex variable. But it's not used for anything? Whereas other libs/pete use ID inplace where we currently use index.

[Robadob]

The HostAPI auto sync could be improved.

If the user calls

std::shared_ptr<flamegpu::detail::CUDAEnvironmentDirectedGraphBuffers> map = FLAMEGPU->environment.getDirectedGraph("map").getCUDABuffers();

The graph will be sync'd, prior to them accessing the CUDA buffers (ignoring the fact this doesn't mark the graph for rebuild). This however would require some proper internal tracking to ensure it persists to the end of the host fn.

[Robadob]

I don't think it validates that edge src/dest are within range of valid vertex indices.

[Robadob]

Deprecated by #1089, to avoid breaking FRE dependency during development before the feature is merged into master.