Dump subgraph ID and fused graph ID (#2607)

* Dump subgraph ID and fused graph ID

Dump subgraph ID and fused graph ID for better debugging

* Remove local static fused_count

added a field global_fused_count_ to NupharExecutionProvider class

Author: yangchen-MS

onnxruntime/core/codegen/common/common.cc

@@ -117,10 +117,10 @@ const onnxruntime::Node* GetInputNode(const Node& node, const NodeArg* def) {
 
 // create capacity from subgraph
 std::unique_ptr<ComputeCapability> ToCapacity(const onnxruntime::GraphViewer& graph,
+                                              int fused_count,
                                               std::unique_ptr<IndexedSubGraph>& subgraph) {
   auto meta_def = onnxruntime::make_unique<::onnxruntime::IndexedSubGraph::MetaDef>();
-  static int fuse_count = 0;
-  meta_def->name = "Fuse" + std::to_string(fuse_count++);
+  meta_def->name = "Fuse" + std::to_string(fused_count);
   meta_def->domain = "Fuse";
 
   std::set<NodeIndex> node_indices(subgraph->nodes.begin(), subgraph->nodes.end());

onnxruntime/core/codegen/common/common.h

@@ -113,6 +113,7 @@ bool IsAliasNode(const onnxruntime::Node& node);
 
 // Helper function that creates ComputeCapability for subgraphs
 std::unique_ptr<ComputeCapability> ToCapacity(const onnxruntime::GraphViewer& graph,
+                                              int fused_count,
                                               std::unique_ptr<IndexedSubGraph>& subgraph);
 
 bool IsFusedNode(const Node& node);

onnxruntime/core/providers/nuphar/nuphar_execution_provider.cc

@@ -25,6 +25,7 @@ namespace onnxruntime {
 thread_local int64_t NupharSubgraphUnit::counter = 0;
 
 thread_local std::unique_ptr<std::unordered_map<std::string, int64_t>> NupharExecutionProvider::tls_realized_dims_;
+int NupharExecutionProvider::global_fused_count_ = 0;
 
 static std::string GetCurrentHostTargetString() {
 #if USE_TVM_WITH_LLVM
@@ -311,7 +312,12 @@ NupharExecutionProvider::GetCapability(const onnxruntime::GraphViewer& graph_vie
   };
   GraphPartitioner graph_partitioner(is_supported_func);
 
-  ORT_ENFORCE(graph_partitioner.Partition(graph_viewer, results).IsOK());
+  ORT_ENFORCE(graph_partitioner.Partition(graph_viewer, global_fused_count_, results).IsOK());
+
+  // reset global_fused_count_ for main graph, since there might be multiple sessions for subgraphs,
+  // this is the time all graph cut should be finished as ORT handles main graph last
+  if (!graph_viewer.IsSubgraph())
+    global_fused_count_ = 0;
 
   // for any node being fused in results, save initializer tensors
   // because IExecutionProvider::Compile would be called without OpKernelInfo

onnxruntime/core/providers/nuphar/nuphar_execution_provider.h

@@ -153,6 +153,10 @@ class NupharExecutionProvider : public IExecutionProvider {
 
   mutable std::unordered_map<std::string, std::unique_ptr<Tensor>> constant_initializers_used_in_compiled_nodes_;
   mutable std::unordered_map<std::string, int> domain_versions_;
+
+  // used to create unique fused node name, make it static because
+  // subsession may create multiple instances of EPs
+  static int global_fused_count_;
 };
 
 }  // namespace onnxruntime

onnxruntime/core/providers/nuphar/partition/graph_partitioner.cc

@@ -140,6 +140,7 @@ bool GraphPartitioner::ForcePartition(
 
 // Partition the graph (fusing ops) based on the dependency and whether ops are supported:
 Status GraphPartitioner::Partition(const onnxruntime::GraphViewer& graph,
+                                   int& fused_count,
                                    std::vector<std::unique_ptr<ComputeCapability>>& result) {
   // call partition
   ORT_RETURN_IF_ERROR(Evaluate(graph, /*distinguish_subgraph*/ true));
@@ -170,9 +171,9 @@ Status GraphPartitioner::Partition(const onnxruntime::GraphViewer& graph,
     if (codegen::CodeGenSettings::Instance().HasOption(kNupharDumpPartition)) {
       std::ostringstream stream;
       if (graph.IsSubgraph()) {
-        stream << "[NUPHAR_DUMP_PARTITION] ## Subgraph ## " << std::endl;
+        stream << "[NUPHAR_DUMP_PARTITION] ## Subgraph ## Fused graph ID " << fused_count << std::endl;
       } else {
-        stream << "[NUPHAR_DUMP_PARTITION]" << std::endl;
+        stream << "[NUPHAR_DUMP_PARTITION] ## Fused graph ID " << fused_count << std::endl;
       }
       stream << "Partition of size " << iter.second.nodes.size() << " [";
       for (const auto& node_index : partition->nodes) {
@@ -186,6 +187,7 @@ Status GraphPartitioner::Partition(const onnxruntime::GraphViewer& graph,
     result.emplace_back(
         ToCapacity(
             graph,
+            fused_count++,
             partition));
   }
 

onnxruntime/core/providers/nuphar/partition/graph_partitioner.h

@@ -22,6 +22,7 @@ class GraphPartitioner : public Partitioner {
       : Partitioner(), is_op_type_supported_func_(is_op_type_supported_func) {}
 
   Status Partition(const onnxruntime::GraphViewer& graph,
+                   int& fused_count,
                    std::vector<std::unique_ptr<ComputeCapability>>& result);
 
  private:

onnxruntime/core/providers/nuphar/partition/subgraph_partitioner.cc

@@ -387,7 +387,7 @@ Status SubgraphPartitioner::Partition(
 
     if (codegen::CodeGenSettings::Instance().HasOption(kNupharDumpFusedNodes)) {
       std::ostringstream stream;
-      stream << "[NUPHAR_DUMP_FUSED_NODES]" << std::endl;
+      stream << "[NUPHAR_DUMP_FUSED_NODES] ID " << subgraph.UniqueId() << std::endl;
       stream << "NupharSubgraphUnit of size " << results.back().nodes.size() << " [";
       for (const auto& n : results.back().nodes) {
         stream << "(" << n->Name() << ", " << n->OpType() << ") ";