Pass Creation

This is a practitioner’s guide to using the Execution Framework. Before continuing, it is recommended you first review the Execution Framework Overview along with basic topics such as Graphs Concepts, Pass Concepts, and Execution Concepts.

There are two parts to adding a new pass to the pass pipeline. First, we need to define the new pass and then we need to register it.

There are multiple types of passes supported in EF. A full list can be found at PassType. To create a pass, we start by implementing the relevant pass type’s interface. In practice, this will be either IPopulatePass or IPartitionPass. In rare cases, a pass may inherit from IGlobalPass, but such a pass is discouraged for performance reasons. Implementing population and partition passes is covered below.

Once a pass is implemented, it must be registered with one of the following macros: OMNI_GRAPH_EXEC_REGISTER_POPULATE_PASS(), OMNI_GRAPH_EXEC_REGISTER_PARTITION_PASS(), OMNI_GRAPH_EXEC_REGISTER_GLOBAL_PASS().

Listing 17 Registration code for different pass types. These macros should be called at global scope.
 // Register MyPopulatePass against a symbol (node or node graph def) name
 OMNI_GRAPH_EXEC_REGISTER_POPULATE_PASS(MyPopulatePass, "symbol_name");

 // Register MyPartitionPass with a given priority
 OMNI_GRAPH_EXEC_REGISTER_PARTITION_PASS(MyPartitionPass, 1);

 // Register MyGlobalPass
 OMNI_GRAPH_EXEC_REGISTER_GLOBAL_PASS(MyGlobalPass);

Implementing a Populate Pass

The example below implements a new populate pass that will instantiate a new MyInstancedDef definition and attach it to any node or definition that matches the name given during registration.

Listing 18 Example populate pass
class TestPopulateInstances : public Implements<IPopulatePass>
{
public:
    TestPopulateInstances(IGraphBuilder*) noexcept
    {
    }

    void run_abi(IGraphBuilder* builder, INode* node) noexcept override
    {
        omni::core::ObjectPtr<TestNodeGraphDef> nodeGraphDef{ TestNodeGraphDef::create(
            builder, "MyInstancedDef") };
        builder->setNodeGraphDef(node, nodeGraphDef);
    }
};

Here, we register the pass, telling EF that this pass should be run on nodes named “my.special.node”.

Listing 19 Register TestPopulateInstances pass to run on every node named “my.special.node”
 OMNI_GRAPH_EXEC_REGISTER_POPULATE_PASS(TestPopulateInstances, "my.special.node");

We can also register the pass to run on nodes with a definition named “my.def.special” already attached to the node.

Listing 20 Register TestPopulateInstances pass to run on every node with “my.def.special” definition attached
 OMNI_GRAPH_EXEC_REGISTER_POPULATE_PASS(TestPopulateInstances, "my.def.special");

Above, we see that the name given during registration can match either the node name or definition name. See Pass Concepts to better understand EF’s algorithm for running population passes.

Implementing a Partition Pass

In this example, we will be partitioning the following graph using either manual or automatic partition generation

Listing 21 An example input graph for partitioning passes
//                   ┌───────────────────┐
//                   │┌─────── B ───────┐│
//                   ││┌────┐     ┌────┐││
//                   │││ Ax ├─┬──►│ Bx │││
//                   ││└────┘ │   └────┘││
//                   ││       │   ┌────┐││
//                   ││       └──►│ Cx │││
//                   ││           └────┘││                         ┌────┐
//           ┌──────►│└─────────────────┘│                      ┌─►│ I  ├─┐
//           │       └───────────────────┘                      │  └────┘ │
// ┌────┐    │       ┌────┐     ┌────┐        ┌────┐    ┌────┐  │         ▼  ┌────┐
// │ A  ├──┬─┴──────►│ Cx ├─┬──►│ Dx │────┬──►│ A2 ├───►│ H  ├─┬┴─────────┴─►│ J  │
// └────┘  │         └─┬──┘ │   └────┘    │   └────┘    └────┘ │             └────┘
//         │           │    │   ┌────┐    │                    │             ┌────┐
//         │           │    └──►│ Ex │    │                    └────────────►│ K  │
//         │           │        └────┘    │                                  └────┘
//         │           └──────────────────┘
//         │      ┌────┐     ┌────┐
//         └─────►│ Fy ├────►│ Gy │
//                └────┘     └────┘
//
// Nodes with matching definition (B/Ax, Cx, Fy, H), (B/Bx, Dx, Gy, J), (B/Cx, Ex, K)
//

The task for our first partition pass will be to recognize two clusters of nodes (B/Ax, B/Bx, B/Cx) and (Cx, Dx, Ex), but not recognize (H, J, K) since node I causes the cluster to form a cycle after replacing. The following implementation will do this manually from the run_abi() method.

Listing 22 An example implementation of a partitioning pass forming clusters manually
class TestPartitionPassX : public Implements<IPartitionPass>
{
public:
    TestPartitionPassX(IGraphBuilder*) noexcept
    {
    }

    bool initialize_abi(ITopology* topology) noexcept override
    {
        return true;
    }

    void run_abi(INode* node) noexcept override
    {
        if (!isMatch(node, TestDefType::eType1))
            return;

        INode* nodeType2{ nullptr };
        INode* nodeType3{ nullptr };

        for (auto* child : node->getChildren())
        {
            if (isMatch(child, TestDefType::eType2))
                nodeType2 = child;
            else if (isMatch(child, TestDefType::eType3))
                nodeType3 = child;
            else if (!isMatch(child, TestDefType::eType4))
                return;
        }

        if (nodeType2 && nodeType3)
        {
            std::array<INode*, 3> partition{ node, nodeType2, nodeType3 };
            m_manualPartitions.emplace_back(std::move(partition));
        }
    }

    void commit_abi(IGraphBuilder* builder) noexcept override
    {
        for (const auto& partition : m_manualPartitions)
        {
            // Verify partition pattern (trivial check in this case)
            if (partition.size() != 3)
                continue; // LCOV_EXCL_LINE

            // Create override definition. Note that we exclude the definition from code coverage
            // because the graph is never executed in this unit test.
            NodeDefPtr nodeDef{ NodeDefLambda::create(
                "PatternX", [](ExecutionTask& info) -> Status { return Status::eSuccess; }, // LCOV_EXCL_LINE
                SchedulingInfo::eSerial) };

            // Commit the change
            builder->replacePartition(NodePartition(partition.data(), partition.size()), nodeDef.get());
        }
    }

private:
    std::vector<std::array<INode*, 3>> m_manualPartitions;
};

Now let’s implement the same partitioning strategy but leverage the quickPartitioning() utility to forming clusters.

Listing 23 An example implementation of a partitioning pass forming clusters manually
class TestPartitionPassX : public Implements<IPartitionPass>
{
public:
    TestPartitionPassX(IGraphBuilder*) noexcept
    {
    }

    bool initialize_abi(ITopology* topology) noexcept override
    {
        m_selectedNodes.reserve(topology->getMaxNodeIndex());
        return true;
    }

    void run_abi(INode* node) noexcept override
    {
        if (isMatch(node, TestDefType::eType1) || isMatch(node, TestDefType::eType2) ||
            isMatch(node, TestDefType::eType3))
        {
            m_selectedNodes.push_back(node);
        }
    }

    void commit_abi(IGraphBuilder* builder) noexcept override
    {
        quickPartitioning(
            builder->getTopology(), Span<INode*>(m_selectedNodes.data(), m_selectedNodes.size()),
            [builder](INode** begin, uint64_t size)
            {
                // For the test, we will only allow partitions of size 3
                if (size != 3)
                    return;

                // Create override definition. Note that we exclude the definition from code coverage
                // because the graph is never executed in this unit test.
                NodeDefPtr nodeDef{ NodeDefLambda::create(
                    "PatternX", [](ExecutionTask& info) -> Status { return Status::eSuccess; }, // LCOV_EXCL_LINE
                    SchedulingInfo::eSerial) };

                // Mutate the graph
                builder->replacePartition(NodePartition(begin, size), nodeDef.get());
            });
    }

private:
    std::vector<INode*> m_selectedNodes;
};

Finally, we have to remember to register our new pass.

Listing 24 Register TestPartitionPassX pass with priority 1
 OMNI_GRAPH_EXEC_REGISTER_PARTITION_PASS(TestPartitionPassX, 1);

Implementing a Global Pass

Users are discouraged from authoring global passes as they can severely slow down graph construction and/or create ordering issues if a specific execution order is required. For all these reasons, users should try implementing new functionality with the pass types above rather than global passes.

The following is a complete implementation of a global pass that EF useds to detect cycles in the graph.

Listing 25 The global pass used for detecting cycles in the execution graph.
class PassStronglyConnectedComponents : public Implements<IGlobalPass>
{
public:
    static omni::core::ObjectPtr<PassStronglyConnectedComponents> create(
        omni::core::ObjectParam<exec::unstable::IGraphBuilder> builder)
    {
        return omni::core::steal(new PassStronglyConnectedComponents(builder.get()));
    }

protected:
    PassStronglyConnectedComponents(IGraphBuilder*)
    {
    }

    void run_abi(IGraphBuilder* builder) noexcept override
    {
        _detectCycles(builder, builder->getTopology());
    }

private:
    void _detectCycles(IGraphBuilder* builder, ITopology* topology)
    {
        struct SCC_NodeData
        {
            size_t index{ 0 };
            size_t lowLink{ 0 };
            uint32_t cycleParentCount{ 0 };
            bool onStack{ false };
        };

        size_t globalIndex = 0;
        std::stack<INode*> globalStack;

        traverseDepthFirst<VisitAll, SCC_NodeData>(
            topology->getRoot(),
            [this, builder, &globalIndex, &globalStack](auto info, INode* prev, INode* curr)
            {
                auto pushStack = [&globalStack](INode* node, SCC_NodeData& data)
                {
                    data.onStack = true;
                    globalStack.push(node);
                };

                auto popStack = [builder, &info, &globalStack]()
                {
                    auto* top = globalStack.top();
                    globalStack.pop();

                    auto& userData = info.userData(top);
                    userData.onStack = false;

                    auto node = exec::unstable::cast<exec::unstable::IGraphBuilderNode>(top);
                    node->setCycleParentCount(userData.cycleParentCount);

                    return top;
                };

                auto& userData = info.userData(curr);
                auto& userDataPrev = info.userData(prev);
                if (info.isFirstVisit())
                {
                    userData.index = userData.lowLink = globalIndex++;
                    pushStack(curr, userData);

                    info.continueVisit(curr);

                    userDataPrev.lowLink = std::min(userDataPrev.lowLink, userData.lowLink);

                    if (userData.lowLink == userData.index)
                    {
                        auto* top = popStack();
                        if (top != curr)
                        {
                            while (top != curr)
                            {
                                top = popStack();
                            }
                        }
                    }

                    auto nodeGraph = curr->getNodeGraphDef();
                    if (nodeGraph)
                    {
                        this->_detectCycles(builder, nodeGraph->getTopology());
                    }
                }
                else if (userData.onStack)
                {
                    userDataPrev.lowLink = std::min(userDataPrev.lowLink, userData.index);
                    userData.cycleParentCount++;
                }
            });
    }
};

Next Steps

Pass creation is often the first step to adding new functionality to EF. The next logical steps often involve adding custom definitions, graph traversals, and executors. See Definition Creation, Graph Traversal Guide, and Executor Creation for details.