Update Conv-Add-Relu Fusion Transformation (#15834)

…n name to make it more intuitive. ### Description Update Conv-Add-Relu Fusion Transformation to handle additional case where NhwcFusedConv is present. ### Motivation and Context Handle additional case where NhwcFusedConv is present.
2026-07-30 20:18:08 +00:00 · 2023-05-09 11:12:49 -07:00 · 2023-05-09 11:12:49 -07:00 · c2e631d356
commit c2e631d356
parent 02d94bcc8e
4 changed files with 110 additions and 18 deletions
--- a/onnxruntime/core/optimizer/conv_activation_fusion.cc
+++ b/onnxruntime/core/optimizer/conv_activation_fusion.cc
@ -71,9 +71,9 @@ bool ConvFusionDataTypeCheck(const Node& conv_node) {
  return true;
 }

-class ConvActivation : public NodeSelector {
+class ConvActivationSelector : public NodeSelector {
 public:
-  ConvActivation() = default;
+  ConvActivationSelector() = default;

  std::optional<NodesToOptimizeIndices> Select(const GraphViewer& graph_viewer, const Node& node) const override {
    const std::string_view node_ep = node.GetExecutionProviderType();
@ -172,7 +172,7 @@ class ConvAddRelu : public NodeSelector {
 namespace actions {
 using NTO = NodesToOptimize;

-class FuseConvActivation : public ReplaceWithNew {
+class FuseConvActivationAction : public ReplaceWithNew {
 private:
  std::string OpType(const RuntimeState&) const override { return "FusedConv"; }

@ -258,9 +258,9 @@ class FuseConvAddRelu : public ReplaceWithNew {

 void RegisterConvActivationFusionRules(SelectorActionRegistry& registry) {
  const auto name = "ConvAct";
-  auto action = std::make_unique<actions::FuseConvActivation>();
+  auto action = std::make_unique<actions::FuseConvActivationAction>();
 #if !defined(ORT_MINIMAL_BUILD)
-  auto selector = std::make_unique<selectors::ConvActivation>();
+  auto selector = std::make_unique<selectors::ConvActivationSelector>();
  registry.RegisterSelectorAndAction(name, {{"Conv", {1, 11}}},
                                     std::move(selector), std::move(action));
 #else
--- a/onnxruntime/core/optimizer/conv_add_act_fusion.cc
+++ b/onnxruntime/core/optimizer/conv_add_act_fusion.cc
@ -40,10 +40,9 @@ const Node* GetLoneConsumerNode(const GraphViewer& graph_viewer, const Node& nod
  return &*node.OutputNodesBegin();
 }

-class ConvAddActivation : public NodeSelector {
+class ConvAddActivationSelector : public NodeSelector {
 public:
-  ConvAddActivation() = default;
-
+  ConvAddActivationSelector() = default;
  std::optional<NodesToOptimizeIndices> Select(const GraphViewer& graph_viewer, const Node& node) const override {
    const std::string_view node_ep = node.GetExecutionProviderType();
 #ifdef MLAS_F16VEC_INTRINSICS_SUPPORTED
@ -62,7 +61,7 @@ class ConvAddActivation : public NodeSelector {
    // 1. Its type is 'conv', 2. it has to satisfy the other requirements,like shape, please refer to SelectConvProducer for more info
    const Node* conv_node = nullptr;
    const auto* add_node = GetLoneConsumerNode(graph_viewer, node);
-    if (!add_node) {
+    if (add_node == nullptr) {
      return std::nullopt;
    }
    // Let's support addition first, leave any-element-wise-op fusion in the future.
@ -73,13 +72,13 @@ class ConvAddActivation : public NodeSelector {
    if (graph_utils::IsSupportedOptypeVersionAndDomain(*add_node, "Add", {7, 13, 14})) {
      conv_node = SelectProducerConv(*add_node);
    }
-    if (!conv_node) {
+    if (conv_node == nullptr) {
      return std::nullopt;
    }
    // GetLoneConsumerNode will ensure outputedge_count is 1
    const auto* act_node = GetLoneConsumerNode(graph_viewer, *add_node);
    // even the next node is not a activation node, it's also fine.
-    if (!act_node) {
+    if (act_node == nullptr) {
      // we can't fuse add-activation when add_node has multiple consumer nodes
      act_node = nullptr;
    } else if (SelectActivation(graph_viewer, *act_node)) {
@ -91,7 +90,7 @@ class ConvAddActivation : public NodeSelector {
    NodesToOptimizeIndicesBuilder builder{};
    builder.target_node = conv_node->Index();
    builder.output_nodes = {add_node->Index()};
-    if (act_node) {
+    if (act_node != nullptr) {
      builder.output_nodes.push_back(act_node->Index());
    }
    return builder.Build();
@ -176,8 +175,8 @@ class ConvAddActivation : public NodeSelector {
      // Check if this is a single use convolution that hasn't already
      // been fused with another Add/Sum node. The Add/Sum can also only be
      // fused if the convolution isn't itself fused with an activation.
-      if ((inputs_node[n]->OpType() == "Conv") && (pre_input_defs_count < 4) && (producer_input_args_count.size() < 4) &&
-          (graph_utils::GetNodeAttribute(*inputs_node[n], "activation") == nullptr) && (inputs_node[n]->GetOutputEdgesCount() == 1)) {
+      if (
+          (inputs_node[n]->OpType() == "Conv" || inputs_node[n]->OpType() == "NhwcFusedConv") && (pre_input_defs_count < 4) && (producer_input_args_count.size() < 4) && (graph_utils::GetNodeAttribute(*inputs_node[n], "activation") == nullptr) && (inputs_node[n]->GetOutputEdgesCount() == 1)) {
        if (pre_input_defs_count < 3) {
          // The optional bias parameter is empty so set to an empty string.
          // TODO, add a new null arguments for bias
@ -196,9 +195,14 @@ class ConvAddActivation : public NodeSelector {
 namespace actions {
 using NTO = NodesToOptimize;

-class FuseConvAddActivation : public ReplaceWithNew {
+class FuseConvAddActivationAction : public ReplaceWithNew {
+ public:
+  FuseConvAddActivationAction() = default;
+
 private:
-  std::string OpType(const RuntimeState&) const override { return "FusedConv"; }
+  std::string OpType(const RuntimeState& runtimeState) const override {
+    return (runtimeState.selected_nodes.Target().OpType() == "Conv") ? "FusedConv" : "NhwcFusedConv";
+  }

  std::string Domain(const RuntimeState&) const override { return kMSDomain; }

@ -272,8 +276,8 @@ class FuseConvAddActivation : public ReplaceWithNew {

 void RegisterConvAddActivationFusionRules(SelectorActionRegistry& registry) {
  const auto name = "ConvAddAct";
-  auto action = std::make_unique<actions::FuseConvAddActivation>();
-  auto selector = std::make_unique<selectors::ConvAddActivation>();
+  auto action = std::make_unique<actions::FuseConvAddActivationAction>();
+  auto selector = std::make_unique<selectors::ConvAddActivationSelector>();
  registry.RegisterSelectorAndAction(name, {{"Conv", {1, 11}}},
                                     std::move(selector), std::move(action));
 }
--- a/onnxruntime/test/optimizer/resnet50_fusion_test.cc
+++ b/onnxruntime/test/optimizer/resnet50_fusion_test.cc
@ -112,6 +112,23 @@ TEST_F(ResNet50FusionTests, FuseConvAddUnitTest) {
  ASSERT_STATUS_OK(Model::Save(*p_model, "conv_add_fp16_fused.onnx"));
  ASSERT_TRUE(op_to_count["Add"] == 0);  // Add removed from graph
 }
+TEST_F(ResNet50FusionTests, FuseConvReluUnitTest) {
+  constexpr const ORTCHAR_T* model_uri = MODEL_FOLDER "fusion/conv_relu_fp16.onnx";
+  std::shared_ptr<Model> p_model;
+  ASSERT_STATUS_OK(Model::Load(model_uri, p_model, nullptr, *logger));
+  Graph& graph = p_model->MainGraph();
+  for (auto& node : p_model->MainGraph().Nodes()) {
+    node.SetExecutionProviderType(kCpuExecutionProvider);
+  }
+  std::map<std::string, int> op_to_count = CountOpsInGraph(graph);
+  ASSERT_TRUE(op_to_count["Relu"] == 1);
+  onnxruntime::GraphTransformerManager graph_transformation_mgr{5};
+  ASSERT_STATUS_OK(graph_transformation_mgr.Register(std::make_unique<ConvActivationFusion>(), TransformerLevel::Level3));
+  ASSERT_STATUS_OK(graph_transformation_mgr.ApplyTransformers(graph, TransformerLevel::Level3, *logger));
+  op_to_count = CountOpsInGraph(graph);
+  ASSERT_STATUS_OK(Model::Save(*p_model, "conv_relu_fp16_fused.onnx"));
+  ASSERT_TRUE(op_to_count["Relu"] == 0);  // Add removed from graph
+}
 #endif  // defined(MLAS_F16VEC_INTRINSICS_SUPPORTED) && !defined(DISABLE_CONTRIB_OPS)
 }  // namespace test
 }  // namespace onnxruntime
--- a/onnxruntime/test/testdata/transform/fusion/conv_relu_fp16.onnx
+++ b/onnxruntime/test/testdata/transform/fusion/conv_relu_fp16.onnx
@ -0,0 +1,71 @@
+onnx_conv_relu:Ë
+;
+Wgraph_input_cast_1graph_input_cast1"Cast*	
+to
+ 
+;
+Xgraph_input_cast_0graph_input_cast0"Cast*	
+to
+ 
+H
+graph_input_cast_0
+graph_input_cast_1Y"Conv*
+kernel_shape@@ 
+
+Ygraph_output_cast_0"Relu
+=
+graph_output_cast_0Zgraph_output_cast0"Cast*	
+to 
+test-modelZ
+X
+
+
+
+
+Z
+W
+
+
+
+
+b
+Z
+
+
+
+
+j
+Y
+
+
+
+
+
+j
+Z
+
+
+
+
+j,
+graph_input_cast_0
+
+
+
+
+
+j,
+graph_input_cast_1
+
+
+
+
+
+j-
+graph_output_cast_0
+
+
+
+
+
+B