The main feature of this MR is to introduce the Select operator, allowing conditionnal graph execution.
The condition can be data dependent, yet the graph scheduling remains fully static!
As per Aidge's philosophy, sub-graph hierarchy is an optional feature, not a mandatory workaround, contrary to ONNX's If operator or PyTorch torch.cond method. The Select operator has the following advantages over them:
Allow interleaved and hierarchical conditions;
Allow pre-execution of conditional branches or not.
Two scheduling behaviors are possible, depending on wether Scheduler::tagConditionalNodes() was called or not:
Without tag: the graph is scheduled and run as is, meaning every conditional branch is run before selection. Of course, this may lead to lots of unnecessary computation. However, branches can be run in parallel, as well as in parallel with the condition determination path;
With tags: only the selected conditional branch is run. To achieve this, the condition determination path has to be scheduled and run entirely before any conditional branch.
This MR adds the following:
Operator Mod (modulus), conformant to ONNX;
Operator CryptoHash, for computing SHA256 (using OpenSSL, now an optional dependency to aidge_backend_cpu);
Operator Select, which selects one of its inputs;
The Scheduler::tagConditionalNodes() method that tags nodes with their conditions in the schedule.cond attribute;
Adaptation of the Scheduler to take into account the schedule.cond nodes tag;
Addition of analysis/ sub-folder with the StaticAnalysis and OperatorStats now in separate source files;
Addition of DynamicAnalysis class that has the same functions as StaticAnalysis but on the scheduled graph. The DynamicAnalysis will therefore report stats for the actual graph execution, which may differ from the static graph if for example some nodes are run multiple time (in recurrent/cyclic graphs) or never run (conditional node);
Add unit tests for GraphView::replace() from !281 (closed);
Registrar now also displays the list of available keys when a key is not found;
Log::error("missing or invalid registrar key: {} for registrable object {}\nDid you include/import the corresponding module?\nIf so, it is possible that the object is not yet supported.",key,typeid(C).name());
Log::info("Available registrar keys are:");
for(constauto&keyValue:C::registry()){
Log::info("- {}",keyValue.first);
}
AIDGE_THROW_OR_ABORT(std::runtime_error,"missing or invalid registrar key");
}
AIDGE_ASSERT(exists(key),
"missing or invalid registrar key: {} for registrable object {}\n"
"Did you include/import the corresponding module?\n"
"If so, it is possible that the object is not yet supported.\n\n"
