[DML EP] Add dynamic graph compilation (#17876)

Historically, DML was only able to fuse partitions when all sizes are
known in advance or when we were overriding them at session creation
time. But in practice, it should be possible to compile partitions at
compute time if the caller knows that the dimensions won't be changed
for every inference (e.g. resizing a webcam window, or padding the input
to powers of 2). This graph will be cached and reused until the sizes
change.

This is an opt-in option gated under the `enable_dynamic_graph_fusion`
option, which means that it will only be enabled when the caller
requests it since they have more context on how their model will be
called between inferences.

This PR also adds the option to disable metacommands from the python
API, which is an option for the C API but was lacking for python.

Author: PatriceVignola

onnxruntime/core/providers/dml/DmlExecutionProvider/inc/DmlExecutionProvider.h

@@ -3,6 +3,9 @@
 
 #pragma once
 interface IMLOperatorRegistry;
+interface IDMLDevice;
+interface ID3D12CommandQueue;
+interface ID3D12Resource;
 
 #include "core/common/status.h"
 #include "core/framework/data_transfer.h"
@@ -28,7 +31,8 @@ namespace Dml
     std::unique_ptr<onnxruntime::IExecutionProvider> CreateExecutionProvider(
         IDMLDevice* dmlDevice,
         ID3D12CommandQueue* commandQueue,
-        bool enableMetacommands = true);
+        bool enableMetacommands,
+        bool enableDynamicGraphFusion);
 
     ID3D12Resource* GetD3D12ResourceFromAllocation(onnxruntime::IAllocator* allocator, void* ptr);
     void FlushContext(onnxruntime::IExecutionProvider* provider);

onnxruntime/core/providers/dml/DmlExecutionProvider/inc/IWinmlExecutionProvider.h

@@ -7,11 +7,14 @@
 #include <functional>
 #include <variant>
 #include <optional>
+#include <wrl/client.h>
 
 #include "core/framework/op_kernel.h"
+#include "core/providers/dml/DmlExecutionProvider/src/DmlEdgeShapes.h"
 
 struct AbstractOperatorDesc;
 interface IMLOperatorTensor;
+interface IDMLOperator;
 struct DML_INPUT_GRAPH_EDGE_DESC;
 struct DML_OUTPUT_GRAPH_EDGE_DESC;
 struct DML_INTERMEDIATE_GRAPH_EDGE_DESC;
@@ -92,6 +95,8 @@ namespace Windows::AI::MachineLearning::Adapter
         const onnxruntime::Node& node,
         MLOperatorTensorGetter& constantInputGetter,
         const void* executionHandle,
+        const EdgeShapes* inputShapesOverrides,
+        /*out*/ EdgeShapes* outputShapes,
         /*out*/ DmlGraphNodeCreateInfo* graphNodeCreateInfo
         )>;
 

onnxruntime/core/providers/dml/DmlExecutionProvider/src/AbiCustomRegistry.cpp

@@ -491,22 +491,24 @@ HRESULT STDMETHODCALLTYPE AbiCustomRegistry::RegisterOperatorKernel(
                     const onnxruntime::Node& node,
                     MLOperatorTensorGetter& constantInputGetter,
                     const void* executionHandle,
+                    const EdgeShapes* inputShapesOverrides,
+                    /*out*/ EdgeShapes* outputShapes,
                     /*out*/ DmlGraphNodeCreateInfo* graphNodeCreateInfo
                 )
                 {
                     onnxruntime::ProtoHelperNodeContext nodeContext(node);
                     onnxruntime::OpNodeProtoHelper<onnxruntime::ProtoHelperNodeContext> protoHelper(&nodeContext);
 
                     // Use the same list of required constant inputs for the shape inferrer and the kernel.
-                    EdgeShapes outputShapes;
-                    InferAndVerifyOutputSizes(node, &defaultAttributesCapture, shapeInferrerCapture.Get(), constantCpuInputCapture, constantInputGetter, nullptr, outputShapes);
+                    InferAndVerifyOutputSizes(node, &defaultAttributesCapture, shapeInferrerCapture.Get(), constantCpuInputCapture, constantInputGetter, inputShapesOverrides, *outputShapes);
 
                     // Create the kernel while allowing input shape and output shape queries according to options
                     ComPtr<DmlGraphOpKernelInfoWrapper> kernelInfoWrapper = wil::MakeOrThrow<DmlGraphOpKernelInfoWrapper>(
                             &protoHelper,
                             executionHandle,
                             true,
-                            &outputShapes,
+                            inputShapesOverrides,
+                            outputShapes,
                             &defaultAttributesCapture,
                             graphNodeCreateInfo,
                             constantCpuInputCapture,

onnxruntime/core/providers/dml/DmlExecutionProvider/src/DmlEdgeShapes.h

@@ -0,0 +1,42 @@
+// Copyright (c) Microsoft Corporation. All rights reserved.
+// Licensed under the MIT License.
+
+#pragma once
+
+namespace Windows::AI::MachineLearning::Adapter
+{
+    // edges and unused edges have an empty array of dimensions.
+    class EdgeShapes
+    {
+    public:
+        EdgeShapes() = default;
+
+        EdgeShapes(size_t count) : m_shapes(count) {}
+
+        const std::vector<uint32_t>& GetShape(size_t edgeIndex) const
+        {
+            return m_shapes[edgeIndex];
+        }
+
+        std::vector<uint32_t>& GetMutableShape(size_t edgeIndex)
+        {
+            return m_shapes[edgeIndex];
+        }
+
+        size_t EdgeCount() const { return m_shapes.size(); }
+
+        void Reset(size_t edge_count)
+        {
+            m_shapes.clear();
+            m_shapes.resize(edge_count);
+        }
+
+        bool operator!=(const EdgeShapes& other) const noexcept
+        {
+            return (m_shapes != other.m_shapes);
+        }
+
+    private:
+        std::vector<std::vector<uint32_t>> m_shapes;
+    };
+}

onnxruntime/core/providers/dml/DmlExecutionProvider/src/DmlGraphFusionHelper.cpp

@@ -1,7 +1,7 @@
 #pragma once
 
 #include "DmlGraphFusionHelper.h"
-
+#include "DmlRuntimeFusedGraphKernel.h"
 
 namespace Dml
 {
@@ -501,5 +501,171 @@ namespace DmlGraphFusionHelper
 
         graph.FinalizeFuseSubGraph(indexedSubGraph, fusedNode);
     }
+
+    void RegisterDynamicKernel(
+        onnxruntime::Graph& graph,
+        onnxruntime::KernelRegistry* registryForPartitionKernels,
+        const ExecutionProviderImpl* providerImpl,
+        std::unordered_map<const onnxruntime::Node*, GraphNodeProperties> graphNodePropertyMap,
+        const std::unordered_set<std::string>& dynamicCpuInputMap,
+        std::shared_ptr<const onnxruntime::IndexedSubGraph> indexedSubGraph,
+        std::unordered_map<std::string, std::pair<const ONNX_NAMESPACE::TensorProto*, bool>>&& isInitializerTransferable)
+    {
+        struct NodeInfo
+        {
+            std::string name;
+            std::string opType;
+            std::string description;
+            std::string domain;
+            onnxruntime::NodeAttributes attributes;
+            std::vector<onnxruntime::NodeArg*> inputDefPointers;
+            std::vector<onnxruntime::NodeArg*> outputDefPointers;
+        };
+
+        auto partitionNodePropsMap = DmlGraphFusionHelper::CreatePartitionNodePropsMap(
+            graph,
+            *indexedSubGraph,
+            std::move(graphNodePropertyMap));
+
+        auto modelPath = graph.ModelPath();
+
+        const gsl::span<const std::string> subGraphInputArgNames = indexedSubGraph->GetMetaDef()->inputs;
+        const gsl::span<const std::string> subGraphOutputArgNames = indexedSubGraph->GetMetaDef()->outputs;
+
+        std::vector<NodeInfo> nodesInfo;
+        nodesInfo.reserve(indexedSubGraph->nodes.size());
+
+        std::vector<const onnxruntime::NodeArg*> subgraphInputs;
+        subgraphInputs.reserve(subGraphInputArgNames.size());
+
+        std::vector<const onnxruntime::NodeArg*> subgraphOutputs;
+        subgraphOutputs.reserve(subGraphOutputArgNames.size());
+
+        std::vector<onnxruntime::NodeAttributes> nodeAttributes;
+        nodeAttributes.reserve(indexedSubGraph->nodes.size());
+
+        std::vector<std::shared_ptr<onnxruntime::NodeArg>> intermediateNodeArgs;
+
+        for (size_t sortedNodeIndex : indexedSubGraph->nodes)
+        {
+            auto node = graph.GetNode(sortedNodeIndex);
+
+            nodeAttributes.push_back(node->GetAttributes());
+
+            NodeInfo nodeInfo{};
+            nodeInfo.name = node->Name();
+            nodeInfo.opType = node->OpType();
+            nodeInfo.description = node->Description();
+            nodeInfo.domain = node->Domain();
+            nodeInfo.attributes = node->GetAttributes();
+            nodeInfo.inputDefPointers.reserve(node->InputDefs().size());
+            nodeInfo.outputDefPointers.reserve(node->OutputDefs().size());
+
+            for (const onnxruntime::NodeArg* inputDef : node->InputDefs())
+            {
+                intermediateNodeArgs.emplace_back(std::make_shared<onnxruntime::NodeArg>(inputDef->Name(), inputDef->TypeAsProto()));
+                nodeInfo.inputDefPointers.push_back(intermediateNodeArgs.back().get());
+            }
+
+            for (const onnxruntime::NodeArg* outputDef : node->OutputDefs())
+            {
+                intermediateNodeArgs.emplace_back(std::make_shared<onnxruntime::NodeArg>(outputDef->Name(), outputDef->TypeAsProto()));
+                nodeInfo.outputDefPointers.push_back(intermediateNodeArgs.back().get());
+            }
+
+            nodesInfo.push_back(std::move(nodeInfo));
+        }
+
+        for (const std::string& graphInputName : subGraphInputArgNames)
+        {
+            subgraphInputs.push_back(graph.GetNodeArg(graphInputName));
+        }
+
+        for (const std::string& graphOutputName : subGraphOutputArgNames)
+        {
+            subgraphOutputs.push_back(graph.GetNodeArg(graphOutputName));
+        }
+
+        // We need to keep the initializers alive since they will be freed once the nodes are removed from the graph
+        std::vector<ONNX_NAMESPACE::TensorProto> ownedInitializers;
+        ownedInitializers.reserve(isInitializerTransferable.size());
+
+        for (auto& kvp : isInitializerTransferable)
+        {
+            ONNX_NAMESPACE::TensorProto tensorProto;
+            tensorProto.set_data_type(kvp.second.first->data_type());
+            tensorProto.set_raw_data(kvp.second.first->raw_data());
+            tensorProto.set_name(kvp.second.first->name());
+
+            for (int i = 0; i < kvp.second.first->dims_size(); ++i)
+            {
+                tensorProto.add_dims(kvp.second.first->dims(i));
+            }
+            ownedInitializers.push_back(std::move(tensorProto));
+            kvp.second.first = &ownedInitializers.back();
+        }
+
+        // lamda captures for the kernel registration
+        auto fused_kernel_func = [
+            indexedSubGraph,
+            &modelPath,
+            nodesInfo = std::move(nodesInfo),
+            intermediateNodeArgs = std::move(intermediateNodeArgs),
+            subgraphInputs = std::move(subgraphInputs),
+            subgraphOutputs = std::move(subgraphOutputs),
+            partitionNodePropsMap = std::move(partitionNodePropsMap),
+            ownedInitializers = std::move(ownedInitializers)] (onnxruntime::FuncManager& func_mgr, const onnxruntime::OpKernelInfo& info, std::unique_ptr<onnxruntime::OpKernel>& out) mutable ->onnxruntime::Status
+        {
+            std::vector<std::shared_ptr<onnxruntime::Node>> subgraphNodes;
+            subgraphNodes.reserve(nodesInfo.size());
+
+            for (const NodeInfo& nodeInfo : nodesInfo)
+            {
+                subgraphNodes.emplace_back(std::make_shared<onnxruntime::Node>(
+                    nodeInfo.name,
+                    nodeInfo.opType,
+                    nodeInfo.description,
+                    nodeInfo.inputDefPointers,
+                    nodeInfo.outputDefPointers,
+                    &nodeInfo.attributes,
+                    nodeInfo.domain));
+            }
+
+            out.reset(CreateRuntimeFusedGraphKernel(
+                info,
+                indexedSubGraph,
+                modelPath,
+                std::move(subgraphNodes),
+                std::move(subgraphInputs),
+                std::move(subgraphOutputs),
+                std::move(intermediateNodeArgs),
+                std::move(partitionNodePropsMap),
+                std::move(ownedInitializers)));
+            return Status::OK();
+        };
+
+        // build the kernel definition on the fly, and register it to the fused_kernel_regisitry.
+        onnxruntime::KernelDefBuilder builder;
+        builder.SetName(indexedSubGraph->GetMetaDef()->name)
+            .SetDomain(indexedSubGraph->GetMetaDef()->domain)
+            .SinceVersion(indexedSubGraph->GetMetaDef()->since_version)
+            .Provider(onnxruntime::kDmlExecutionProvider);
+
+        // Force the CPU inputs to be allocated on the CPU
+        for (int i = 0; i < subGraphInputArgNames.size(); ++i)
+        {
+            if (dynamicCpuInputMap.find(subGraphInputArgNames[i]) != dynamicCpuInputMap.end())
+            {
+                builder.InputMemoryType(OrtMemTypeCPUInput, i);
+            }
+        }
+
+        ORT_THROW_IF_ERROR(registryForPartitionKernels->Register(builder, fused_kernel_func));
+
+        auto& fusedNode = graph.BeginFuseSubGraph(*indexedSubGraph, indexedSubGraph->GetMetaDef()->name);
+        fusedNode.SetExecutionProviderType(onnxruntime::kDmlExecutionProvider);
+
+        graph.FinalizeFuseSubGraph(*indexedSubGraph, fusedNode);
+    }
 }
 }

onnxruntime/core/providers/dml/DmlExecutionProvider/src/DmlGraphFusionHelper.h

@@ -80,5 +80,14 @@ namespace DmlGraphFusionHelper
         std::vector<uint8_t>&& isInputsUploadedByDmlEP,
         const GraphDescBuilder::GraphDesc& graphDesc,
         Microsoft::WRL::ComPtr<IDMLCompiledOperator> compiledExecutionPlanOperator);
+
+    void RegisterDynamicKernel(
+        onnxruntime::Graph& graph,
+        onnxruntime::KernelRegistry* registryForPartitionKernels,
+        const ExecutionProviderImpl* providerImpl,
+        std::unordered_map<const onnxruntime::Node*, GraphNodeProperties> graphNodePropertyMap,
+        const std::unordered_set<std::string>& dynamicCpuInputMap,
+        std::shared_ptr<const onnxruntime::IndexedSubGraph> indexedSubGraph,
+        std::unordered_map<std::string, std::pair<const ONNX_NAMESPACE::TensorProto*, bool>>&& isInitializerTransferable);
 }
 }