Introduce new optimizer MatMul + BatchNormalization (#17915)

### Description
Introduce new ORT L1 optimizer under RewriteRule category to fuse MatMul
+ BatchNormalization node. This optimizer look for a specific pattern
observed in one of the impacting customer models and fuse the Matmul and
Batchnormalization node into a Gemm node. For details on the pattern
matching and fusion please refer to the comment section of
`matmul_bn_fusion.cc`.

To visualize, this optimizer will replace following subgraph to a Gemm
node.
<pre>
               MatMul                  GEMM 
                 |                       |     
              Reshape ^     --->      Reshape ^
                 |                       |
            Transpose ^             Transpose ^
                 |
       BatchNormalization
Note: ^ means there can be >=0 occurrence(s) of that node. 
Few example fusable pattern:
* - MatMul -> Reshape -> Transpose -> BatchNormalization ---> GEMM ->
Reshape -> Transpose
* - MatMul -> Reshape -> BatchNormalization ---> GEMM -> Reshape
* - MatMul -> Transpose -> BatchNormalization ---> GEMM -> Transpose
* - MatMul -> Reshape -> Reshape -> BatchNormalization ---> GEMM ->
Reshape -> Reshape
* - MatMul -> Reshape -> Transpose -> Reshape -> BatchNormalization --->
GEMM -> Reshape -> Transpose -> Reshape
* - MatMul -> BatchNormalization ---> GEMM
</pre>

Note: This optimizer may evolve in the future to be more generic in
terms of the pattern matching.

### Motivation and Context
- Why is this change required? What problem does it solve?
One of the user of ORT+DML ep needs this to better target the model to
DML. But this transformation applies more broadly, so added L1
optimizer.
<!-- - If it fixes an open issue, please link to the issue here. -->

onnxruntime/core/optimizer/graph_transformer_utils.cc

@@ -50,6 +50,7 @@
 #include "core/optimizer/matmul_integer_to_float.h"
 #include "core/optimizer/matmul_scale_fusion.h"
 #include "core/optimizer/matmul_transpose_fusion.h"
+#include "core/optimizer/matmul_bn_fusion.h"
 #include "core/optimizer/nchwc_transformer.h"
 #include "core/optimizer/noop_elimination.h"
 #include "core/optimizer/not_where_fusion.h"
@@ -127,6 +128,7 @@ InlinedVector<std::unique_ptr<RewriteRule>> GenerateRewriteRules(
       rules.push_back(std::make_unique<ConvAddFusion>());
       rules.push_back(std::make_unique<ConvMulFusion>());
       rules.push_back(std::make_unique<ConvBNFusion>());
+      rules.push_back(std::make_unique<MatmulBNFusion>());
       rules.push_back(std::make_unique<ClipQuantFusion>());
       rules.push_back(std::make_unique<ReluQuantFusion>());
       break;

onnxruntime/core/optimizer/initializer.cc

@@ -291,7 +291,11 @@ Initializer& Initializer::sqrt() {
 namespace {
 template <typename T>
 struct ScaleByAxis {
-  void operator()(Tensor& data, const Tensor& scalers, const size_t block_size, const size_t num_blocks) const {
+  void operator()(Tensor& data,
+                  const Tensor& scalers,
+                  const size_t block_size,
+                  const size_t num_blocks,
+                  const bool column_major) const {
     ToNumeric<T> to_numeric;
     const auto scaler_size = scalers.Shape().Size();
     T* dst = data.MutableData<T>();
@@ -303,24 +307,32 @@ struct ScaleByAxis {
       }
     } else {
       for (size_t block_offset = 0, i = 0; i < num_blocks; i++) {
-        const auto numeric_scaler = to_numeric(scalers_data[i]);
-        for (size_t j = 0; j < block_size; ++j, ++block_offset) {
-          dst[block_offset] = T(to_numeric(dst[block_offset]) * numeric_scaler);
+        if (column_major) {
+          for (size_t j = 0; j < block_size; ++j, ++block_offset) {
+            const auto numeric_scaler = to_numeric(scalers_data[j]);
+            dst[block_offset] = T(to_numeric(dst[block_offset]) * numeric_scaler);
+          }
+        } else {
+          const auto numeric_scaler = to_numeric(scalers_data[i]);
+          for (size_t j = 0; j < block_size; ++j, ++block_offset) {
+            dst[block_offset] = T(to_numeric(dst[block_offset]) * numeric_scaler);
+          }
         }
       }
     }
   }
 };
-
 }  // namespace
 
-void Initializer::scale_by_axis(const Initializer& scalers, int axis) {
+void Initializer::scale_by_axis(const Initializer& scalers, int axis, bool column_major) {
   ORT_ENFORCE(axis >= 0, "Axis must be non-negative");
   const size_t block_size = narrow<size_t>(data_.Shape().SizeFromDimension(gsl::narrow_cast<size_t>(axis)));
   const size_t num_blocks = size() / block_size;
-  ORT_ENFORCE(scalers.size() == 1 || scalers.size() == num_blocks, "Invalid other(scalers) size");
+  ORT_ENFORCE(scalers.size() == 1 ||
+                  (column_major ? scalers.size() == block_size : scalers.size() == num_blocks),
+              "Invalid other(scalers) size");
   utils::MLTypeCallDispatcher<MLFloat16, BFloat16, float, double, int32_t, int64_t> t_disp(data_.GetElementType());
-  t_disp.Invoke<ScaleByAxis>(data_, scalers.data_, block_size, num_blocks);
+  t_disp.Invoke<ScaleByAxis>(data_, scalers.data_, block_size, num_blocks, column_major);
 }
 #endif  // ORT_EXTENDED_MINIMAL_BUILD
 }  // namespace onnxruntime

onnxruntime/core/optimizer/initializer.h

@@ -86,7 +86,7 @@ class Initializer final {
 
   Initializer& sqrt();
 
-  void scale_by_axis(const Initializer& other, int axis);
+  void scale_by_axis(const Initializer& other, int axis, bool column_major = false);
 #endif  // ORT_EXTENDED_MINIMAL_BUILD
  private:
   std::string name_;

onnxruntime/core/optimizer/matmul_bn_fusion.cc

@@ -0,0 +1,230 @@
+// Copyright (c) Microsoft Corporation. All rights reserved.
+// Licensed under the MIT License.
+
+#include "core/optimizer/matmul_bn_fusion.h"
+#include "core/graph/graph_utils.h"
+#include "core/optimizer/initializer.h"
+#include "core/optimizer/utils.h"
+
+namespace onnxruntime {
+
+namespace {
+const std::vector<std::pair<std::string, InlinedVector<ONNX_NAMESPACE::OperatorSetVersion>>> ignorable_nodes{
+    {"Reshape", {1, 5, 13, 14, 19}},
+    {"Transpose", {1, 13}}};
+const std::pair<std::string, InlinedVector<ONNX_NAMESPACE::OperatorSetVersion>> dest = {"BatchNormalization", {1, 6, 7, 9, 14, 15}};
+}  // namespace
+
+bool NodeIsIgnorable(const Graph& graph, const Node& root_node, NodeIndex curr_node_index) {
+  const Node* curr_node = graph.GetNode(curr_node_index);
+
+  // curr_node has different execution provider then it's parent or
+  // has output edge != 1 (this condition will handle the case when ignorable node
+  // is graph output i.e. a graph like this "MatMul->Transpose")
+  if (curr_node->GetExecutionProviderType() != root_node.GetExecutionProviderType() ||
+      curr_node->GetOutputEdgesCount() != 1) {
+    return false;
+  }
+
+  // curr_node can be any of the ignorable_nodes.
+  for (size_t index = 0; index < ignorable_nodes.size(); index++) {
+    if (graph_utils::IsSupportedOptypeVersionAndDomain(*curr_node, ignorable_nodes[index].first, ignorable_nodes[index].second)) {
+      return true;
+    }
+  }
+
+  return false;
+}
+
+std::optional<NodeIndex> MatchPath(const Graph& graph, const Node& root_node, NodeIndex curr_node_index) {
+  while (NodeIsIgnorable(graph, root_node, curr_node_index)) {
+    curr_node_index = graph.GetNode(curr_node_index)->OutputNodesBegin()->Index();
+  }
+
+  // curr_node is neither ignorable nor dest
+  const Node* curr_node = graph.GetNode(curr_node_index);
+  if (curr_node->OpType() != dest.first) {
+    return std::nullopt;
+  }
+
+  if (curr_node->GetExecutionProviderType() == root_node.GetExecutionProviderType() &&
+      graph_utils::IsSupportedOptypeVersionAndDomain(*curr_node, dest.first, dest.second)) {
+    return curr_node_index;
+  }
+
+  // either curr_node has different execution provider or
+  // has invalid opset.
+  return std::nullopt;
+}
+
+/*
+ *   Given a MatMul node, it will verify the following pattern.
+ *                MatMul                  GEMM
+ *                  |                       |
+ *               Reshape ^     --->      Reshape ^
+ *                  |                       |
+ *             Transpose ^             Transpose ^
+ *                  |
+ *        BatchNormalization
+ * Note: ^ means there can be 0 or any occurrences of that node.
+ * Few example fusable pattern:
+ *  - MatMul -> Reshape -> Transpose -> BatchNormalization              ---> GEMM -> Reshape -> Transpose
+ *  - MatMul -> Reshape -> BatchNormalization                           ---> GEMM -> Reshape
+ *  - MatMul -> Transpose -> BatchNormalization                         ---> GEMM -> Transpose
+ *  - MatMul -> Reshape -> Reshape -> BatchNormalization                ---> GEMM -> Reshape -> Reshape
+ *  - MatMul -> Reshape -> Transpose -> Reshape -> BatchNormalization   ---> GEMM -> Reshape -> Transpose -> Reshape
+ *  - MatMul -> BatchNormalization                                      ---> GEMM
+ * Other Conditions:
+ *   - B tensor of MatMul should be constant.
+ *   - scale, B, mean, var tensors of BatchNormalization should be constant.
+ *   - Every node in the path, except the BatchNormalization, should have only 1 output edge.
+ */
+bool MatmulBNFusion::SatisfyCondition(const Graph& graph, const Node& node, const logging::Logger&) const {
+  if (!graph_utils::IsSupportedOptypeVersionAndDomain(node, "MatMul", {1, 9, 13}) ||
+      node.GetOutputEdgesCount() != 1) {
+    return false;
+  }
+
+  if (graph.NodeProducesGraphOutput(node)) {
+    return false;
+  }
+
+  // because <node> is not producing graph output, it means it will have a child node
+  NodeIndex child_node_index = node.OutputNodesBegin()->Index();
+  std::optional<NodeIndex> batch_norm_index = MatchPath(graph, node, child_node_index);
+  if (!batch_norm_index.has_value()) {
+    return false;
+  }
+
+  const Node* batch_norm_node = graph.GetNode(*batch_norm_index);
+
+  // Check that the appropriate inputs to the Matmul and BN nodes are constants.
+  if (!graph_utils::NodeArgIsConstant(graph, *node.InputDefs()[1]) ||
+      !graph_utils::NodeArgIsConstant(graph, *batch_norm_node->InputDefs()[1]) ||
+      !graph_utils::NodeArgIsConstant(graph, *batch_norm_node->InputDefs()[2]) ||
+      !graph_utils::NodeArgIsConstant(graph, *batch_norm_node->InputDefs()[3]) ||
+      !graph_utils::NodeArgIsConstant(graph, *batch_norm_node->InputDefs()[4])) {
+    return false;
+  }
+
+  // First output from BN is required. Others are optional. If any optional outputs exist we can't fuse.
+  const auto& output_defs = batch_norm_node->OutputDefs();
+  if (output_defs.size() > 1) {
+    for (size_t i = 1, end = output_defs.size(); i < end; ++i) {
+      if (output_defs[i] != nullptr && output_defs[i]->Exists()) {
+        return false;
+      }
+    }
+  }
+
+  return true;
+}
+
+/*
+ * BatchNormalization: [https://learn.microsoft.com/en-us/windows/win32/api/directml/ns-directml-dml_batch_normalization_operator_desc]
+ *   Scale * ((Input - Mean) / sqrt(Variance + Epsilon)) + Bias // ignore the FusedActivation in the above definition, that's very specific to DML
+ * Expanding out the terms:
+ *   Output = (Scale / sqrt(Variance + Epsilon)) * Input + (Scale / sqrt(Variance + Epsilon)) * -Mean + Bias
+ * Here,
+ *   [Scale/sqrt(Variance + Epsilon)] is constant, and let's call it `alpha`
+ *   [(Scale / sqrt(Variance + Epsilon)) * -Mean + Bias] is also constant, and let's call it `beta`
+ * Output = alpha * Input + beta, Input = B tensor of MatMul.
+ *
+ */
+Status MatmulBNFusion::Apply(Graph& graph, Node& matmul_node, RewriteRuleEffect& rule_effect, const logging::Logger&) const {
+  NodeIndex child_node_index = matmul_node.OutputNodesBegin()->Index();
+  NodeIndex batch_norm_node_index = MatchPath(graph, matmul_node, child_node_index).value();
+
+  Node& batch_norm_node = *graph.GetNode(batch_norm_node_index);  // need mutable node, that's why extracting node from graph
+
+  // only perform fusion if epsilon is present and is of float_32 type
+  auto epsilon_attribute = batch_norm_node.GetAttributes().find("epsilon");
+  if (epsilon_attribute == batch_norm_node.GetAttributes().end() ||
+      epsilon_attribute->second.type() != ONNX_NAMESPACE::AttributeProto_AttributeType_FLOAT) {
+    return Status::OK();
+  }
+  const float epsilon = epsilon_attribute->second.f();
+
+  const onnx::TensorProto* scale_tensor = graph_utils::GetConstantInitializer(graph, batch_norm_node.InputDefs()[1]->Name());
+  ORT_ENFORCE(scale_tensor);
+  const onnx::TensorProto* bias_tensor = graph_utils::GetConstantInitializer(graph, batch_norm_node.InputDefs()[2]->Name());
+  ORT_ENFORCE(bias_tensor);
+  const onnx::TensorProto* mean_tensor = graph_utils::GetConstantInitializer(graph, batch_norm_node.InputDefs()[3]->Name());
+  ORT_ENFORCE(mean_tensor);
+  const onnx::TensorProto* var_tensor = graph_utils::GetConstantInitializer(graph, batch_norm_node.InputDefs()[4]->Name());
+  ORT_ENFORCE(var_tensor);
+  const onnx::TensorProto* matmul_b_tensor = graph_utils::GetConstantInitializer(graph, matmul_node.InputDefs()[1]->Name());
+  ORT_ENFORCE(matmul_b_tensor);
+
+  if (!optimizer_utils::IsFloatingPointDataType(*matmul_b_tensor) ||
+      !optimizer_utils::IsFloatingPointDataType(*scale_tensor) ||
+      !optimizer_utils::IsFloatingPointDataType(*bias_tensor) ||
+      !optimizer_utils::IsFloatingPointDataType(*mean_tensor) ||
+      !optimizer_utils::IsFloatingPointDataType(*var_tensor) ||
+      scale_tensor->dims_size() != 1 ||
+      bias_tensor->dims_size() != 1 ||
+      mean_tensor->dims_size() != 1 ||
+      var_tensor->dims_size() != 1 ||
+      scale_tensor->dims(0) != matmul_b_tensor->dims(1) ||
+      bias_tensor->dims(0) != matmul_b_tensor->dims(1) ||
+      mean_tensor->dims(0) != matmul_b_tensor->dims(1) ||
+      var_tensor->dims(0) != matmul_b_tensor->dims(1)) {
+    return Status::OK();
+  }
+
+  /*
+   * temp = scale / sqrt(var + epsilon)
+   * output = (temp * Input) - ((temp * mean) + bias)
+   */
+  Initializer scale(*scale_tensor, graph.ModelPath());
+  Initializer bias(*bias_tensor, graph.ModelPath());
+  Initializer mean(*mean_tensor, graph.ModelPath());
+  Initializer var(*var_tensor, graph.ModelPath());
+  Initializer matmul_b(*matmul_b_tensor, graph.ModelPath());
+
+  var.add(epsilon);
+  var.sqrt();
+  scale.div(var);  // this is the temp
+  matmul_b.scale_by_axis(scale, 1, true);
+
+  mean.mul(scale);
+  bias.sub(mean);
+
+  // create B tensorProto for new Gemm node from <matmulB> initializer.
+  ONNX_NAMESPACE::TensorProto new_gemm_b_tensor(*matmul_b_tensor);
+  matmul_b.ToProto(new_gemm_b_tensor);
+  const std::string new_gemm_b_name = graph.GenerateNodeArgName("MatMulBnFusion_GemmB_" + matmul_b_tensor->name());
+  new_gemm_b_tensor.set_name(new_gemm_b_name);
+  NodeArg& new_gemm_b_node_arg = graph_utils::AddInitializer(graph, new_gemm_b_tensor);
+
+  // create bias tensorProto for new Gemm node from <bias> initializer.
+  ONNX_NAMESPACE::TensorProto new_gemm_bias_tensor(*bias_tensor);
+  bias.ToProto(new_gemm_bias_tensor);
+  const std::string new_gemm_bias_name = graph.GenerateNodeArgName("MatMulBnFusion_GemmBias");
+  new_gemm_bias_tensor.set_name(new_gemm_bias_name);
+  NodeArg& new_gemm_bias_node_arg = graph_utils::AddInitializer(graph, new_gemm_bias_tensor);
+
+  Node& gemm_node = graph.AddNode(
+      graph.GenerateNodeArgName("MatMulBnFusion_Gemm"),
+      "Gemm",
+      "Generated from Matmul BatchNormalization fusion",
+      {matmul_node.MutableInputDefs()[0], &new_gemm_b_node_arg, &new_gemm_bias_node_arg},
+      matmul_node.MutableOutputDefs(),
+      nullptr,
+      kOnnxDomain);
+
+  // Remove MatMul node.
+  Node* node = graph.GetNode(matmul_node.Index());
+  graph_utils::RemoveNodeOutputEdges(graph, *node);
+  graph.RemoveNode(matmul_node.Index());
+
+  // Delete optional empty output defs.
+  // Delete BatchNormalization node and update the input of the child of BatchNormalization
+  batch_norm_node.MutableOutputDefs().resize(1);
+  NodeIndex batch_norm_parent_index = graph.GetNode(child_node_index)->OpType() == "BatchNormalization" ? gemm_node.Index() : batch_norm_node.InputNodesBegin()->Index();
+  graph_utils::FinalizeNodeFusion(graph, *graph.GetNode(batch_norm_parent_index), batch_norm_node);
+
+  rule_effect = RewriteRuleEffect::kRemovedCurrentNode;
+  return Status::OK();
+}
+}  // namespace onnxruntime

onnxruntime/core/optimizer/matmul_bn_fusion.h

@@ -0,0 +1,27 @@
+// Copyright (c) Microsoft Corporation. All rights reserved.
+// Licensed under the MIT License.
+
+#pragma once
+
+#include "core/optimizer/rewrite_rule.h"
+
+namespace onnxruntime {
+/*
+ *   This fusion submerges a BatchNormalization operator to it's super
+ *   precedding MatMul operator, if and only if MatmulBNFusion::SatisfyCondition()
+ *   is true.
+ */
+class MatmulBNFusion : public RewriteRule {
+ public:
+  MatmulBNFusion() : RewriteRule("MatMul_BatchNormalization_Fusion") {}
+
+  std::vector<std::string> TargetOpTypes() const noexcept override {
+    return {"MatMul"};
+  }
+
+ private:
+  bool SatisfyCondition(const Graph& graph, const Node& node, const logging::Logger& logger) const override;
+
+  Status Apply(Graph& graph, Node& matmul_node, RewriteRuleEffect& rule_effect, const logging::Logger& logger) const override;
+};
+}  // namespace onnxruntime