feat: support the case when the injected node is not supported in dependency analysis

Signed-off-by: Bo Wang <[email protected]>

Author: bowang007

core/compiler.cpp

@@ -274,7 +274,7 @@ torch::jit::script::Module EmbedEngineInNewModule(const std::string& engine) {
   auto new_g = std::make_shared<torch::jit::Graph>();
   AddEngineToGraph(new_mod, new_g, engine);
   auto new_method = new_mod._ivalue()->compilation_unit()->create_function("forward", new_g);
-  auto schema = GenerateGraphSchema(new_mod, new_method->name(), new_g);
+  auto schema = util::GenerateGraphSchema(new_method->name(), new_g);
   new_mod.type()->addMethod(new_method);
   new_method->setSchema(schema);
 

core/partitioning/partitioning.cpp

@@ -9,19 +9,46 @@ namespace trtorch {
 namespace core {
 namespace partitioning {
 
-inline bool isTensorOrTensorList(torch::jit::Value* val) {
-  return val->type()->isSubtypeOf(torch::jit::TensorType::get()) ||
-      val->type()->isSubtypeOf(torch::jit::ListType::ofTensors());
-}
-
 struct usage_info {
   int produce_id = -1;
   std::vector<int> torch_use_id;
   std::vector<int> tensorrt_use_id;
 };
 
+inline bool isTensorOrTensorList(torch::jit::Value* val) {
+  return val->type()->isSubtypeOf(torch::jit::TensorType::get()) ||
+      val->type()->isSubtypeOf(torch::jit::ListType::ofTensors());
+}
+
+bool isAllNodesSupported(const std::vector<torch::jit::Node*>& nodes) {
+  for (auto node : nodes) {
+    if (!conversion::OpSupported(node)) {
+      return false;
+    }
+  }
+  return true;
+}
+
+bool containNonTensorInputs(torch::jit::Node* n, const std::unordered_set<torch::jit::Value*>& target_inputs) {
+  for (auto input : n->inputs()) {
+    if (!isTensorOrTensorList(input) && target_inputs.count(input)) {
+      return true;
+    }
+  }
+  return false;
+}
+
+bool containNonTensorOutputs(torch::jit::Node* n) {
+  for (auto output : n->outputs()) {
+    if (!isTensorOrTensorList(output)) {
+      return true;
+    }
+  }
+  return false;
+}
+
 std::vector<torch::jit::Node*> getDependencyNodes(std::vector<torch::jit::Value*>& vals) {
-  // using bfs to get the DAG dependency nodes for input value
+  // use bfs to get the DAG dependency nodes for input value
   std::queue<torch::jit::Value*, std::deque<torch::jit::Value*>> q(
       std::deque<torch::jit::Value*>(vals.begin(), vals.end()));
   std::unordered_set<torch::jit::Node*> visited;
@@ -43,17 +70,50 @@ std::vector<torch::jit::Node*> getDependencyNodes(std::vector<torch::jit::Value*
   return stk;
 }
 
-SegmentedBlock injectNodesForNonTensorInputs(SegmentedBlock& seg_block) {
+std::vector<SegmentedBlock> injectNodesForNonTensorInputs(SegmentedBlock& seg_block) {
   // reconstruct segmented_block if this block requires nonTensor input
   std::vector<torch::jit::Value*> nontensor_inputs;
   for (auto input : seg_block.raw_inputs()) {
     if (!isTensorOrTensorList(input)) {
       nontensor_inputs.push_back(input);
     }
   }
-  std::vector<torch::jit::Node*> new_block_nodes = getDependencyNodes(nontensor_inputs);
-  new_block_nodes.insert(new_block_nodes.end(), seg_block.raw_nodes().begin(), seg_block.raw_nodes().end());
-  return std::move(SegmentedBlock(seg_block.target(), new_block_nodes));
+  std::vector<torch::jit::Node*> dependency_nodes = getDependencyNodes(nontensor_inputs);
+
+  std::vector<SegmentedBlock> new_seg_blocks;
+  // if current block is kTorch or current block is TensorRT and all dependent nodes are also supported, construct only
+  // one new block
+  if (seg_block.target() == SegmentedBlock::kTorch || isAllNodesSupported(dependency_nodes)) {
+    dependency_nodes.insert(dependency_nodes.end(), seg_block.raw_nodes().begin(), seg_block.raw_nodes().end());
+    new_seg_blocks.emplace_back(seg_block.target(), dependency_nodes);
+  } else {
+    // if current block is kTensorRT but the dependency nodes contain unsupported node, then we have to segment again
+    std::unordered_set<torch::jit::Value*> nontensor_inputs_set(nontensor_inputs.begin(), nontensor_inputs.end());
+    new_seg_blocks.emplace_back(SegmentedBlock::kTorch, dependency_nodes);
+    std::vector<torch::jit::Node*> tensorrt_nodes, pytorch_nodes;
+    bool prev_non_tensor_outputs = false;
+    for (auto n : seg_block.raw_nodes()) {
+      // it's a kTorch block if it uses the nonTensor input and the nonTensor input is produced in kTorch block
+      if (containNonTensorInputs(n, nontensor_inputs_set) || prev_non_tensor_outputs) {
+        if (!tensorrt_nodes.empty()) {
+          new_seg_blocks.emplace_back(SegmentedBlock::kTensorRT, tensorrt_nodes);
+        }
+        pytorch_nodes.push_back(n);
+        prev_non_tensor_outputs = containNonTensorOutputs(n);
+      } else {
+        if (!pytorch_nodes.empty()) {
+          new_seg_blocks.emplace_back(SegmentedBlock::kTorch, pytorch_nodes);
+        }
+        tensorrt_nodes.push_back(n);
+      }
+    }
+    if (!tensorrt_nodes.empty()) {
+      new_seg_blocks.emplace_back(SegmentedBlock::kTensorRT, tensorrt_nodes);
+    } else {
+      new_seg_blocks.emplace_back(SegmentedBlock::kTorch, pytorch_nodes);
+    }
+  }
+  return std::move(new_seg_blocks);
 }
 
 void resolveNonTensorInputs(PartitionedGraph& segmented_blocks, std::shared_ptr<torch::jit::Graph> g) {
@@ -80,16 +140,17 @@ void resolveNonTensorInputs(PartitionedGraph& segmented_blocks, std::shared_ptr<
     if (segmented_blocks[use_info.produce_id].target() == SegmentedBlock::kTensorRT && !use_info.torch_use_id.empty()) {
       int first_torch_id = use_info.torch_use_id.front();
       if (!updated_segments.count(first_torch_id)) {
-        auto new_torch_block = injectNodesForNonTensorInputs(segmented_blocks[first_torch_id]);
+        auto new_torch_block = injectNodesForNonTensorInputs(segmented_blocks[first_torch_id]).front();
         segmented_blocks[first_torch_id] = new_torch_block;
         updated_segments.insert(first_torch_id);
       }
     } else {
       // KTensorRT segments always need to inject nodes for the nonTensor inputs
       for (int i : use_info.tensorrt_use_id) {
         if (!updated_segments.count(i)) {
-          auto new_seg_block = injectNodesForNonTensorInputs(segmented_blocks[i]);
-          segmented_blocks[i] = new_seg_block;
+          auto to_inject_blocks = injectNodesForNonTensorInputs(segmented_blocks[i]);
+          segmented_blocks.erase(segmented_blocks.begin() + i);
+          segmented_blocks.insert(segmented_blocks.begin() + i, to_inject_blocks.begin(), to_inject_blocks.end());
           updated_segments.insert(i);
         }
       }