rm convertToSSA API,test=huawei_ascend_npu test=nvidia_tensorrt test=verisilicon_timvx (#8988) (#9233)

weishengying · web-flow · commit 73750d9d2f59 · 2022-07-09T17:35:10.000+08:00
diff --git a/CMakeLists.txt b/CMakeLists.txt
@@ -88,8 +88,6 @@ lite_option(LITE_WITH_XCODE                    "when debug in xcode, its ON."
 lite_option(LITE_WITH_ARM82_FP16               "when compile with arm v8.2 fp16, it's ON."                            OFF)
 lite_option(LITE_WITH_ARM82_INT8_SDOT          "when compile with arm v8.2 int8, it's ON."                            OFF)
 lite_option(LITE_WITH_CODE_META_INFO           "include git version in the header file."                              ON)
-# whether convert input model which is not a DAG to SSA graph
-lite_option(WITH_CONVERT_TO_SSA           "whether convert input model which is not a DAG to SSA graph"               ON)
 
 # Thirdparty
 set(THIRD_PARTY_PATH "${CMAKE_BINARY_DIR}/third_party" CACHE STRING
diff --git a/cmake/configure.cmake b/cmake/configure.cmake
@@ -301,6 +301,3 @@ if (LITE_WITH_M1)
 add_definitions("-DLITE_WITH_M1")
 endif(LITE_WITH_M1)
 
-if (WITH_CONVERT_TO_SSA STREQUAL ON)
-  add_definitions("-DWITH_CONVERT_TO_SSA")
-endif(WITH_CONVERT_TO_SSA)
diff --git a/lite/core/optimizer/mir/type_target_cast_pass.cc b/lite/core/optimizer/mir/type_target_cast_pass.cc
@@ -39,6 +39,8 @@ void TypeTargetTransformPass::Apply(const std::unique_ptr<SSAGraph>& graph) {
 
   // record the copied node.
   std::map<std::string, Node*> copied_nodes;
+  // record the origin node.
+  std::map<std::string, Node*> input_nodes;
   std::vector<std::string> skip_ops = {
       "while", "conditional_block", "write_back"};
 
@@ -48,8 +50,14 @@ void TypeTargetTransformPass::Apply(const std::unique_ptr<SSAGraph>& graph) {
     if (!node->IsStmt() || iter != skip_ops.end()) continue;
     auto inlinks = node->inlinks;
     for (auto* in : inlinks) {
+      if (!input_nodes.count(in->AsArg().name))
+        input_nodes[in->AsArg().name] = in;
       ComplementInputs(graph.get(), node, in, &copied_nodes);
     }
+    auto outlinks = node->outlinks;
+    for (auto* out : outlinks) {
+      ComplementOutputs(graph.get(), node, out, &input_nodes);
+    }
   }
 }
 
@@ -78,17 +86,174 @@ void TypeTargetTransformPass::ComplementInputs(
             << " for kernel " << inst.op()->DebugString() << " "
             << *in->AsArg().type << " -> " << *decl_arg_type;
     // Add an IoCopy instruction to make the input compatible with other dist.
-    AddIoCopyInst(*in->AsArg().type,
-                  *decl_arg_type,
-                  in,
-                  graph,
-                  inst_node,
-                  copied_nodes,
-                  valid_places_);
+    AddInputIoCopyInst(*in->AsArg().type,
+                       *decl_arg_type,
+                       in,
+                       graph,
+                       inst_node,
+                       copied_nodes,
+                       valid_places_);
+  }
+}
+
+void TypeTargetTransformPass::AddOutputIoCopyInst(
+    const Type& from,
+    const Type& to,
+    Node* out,
+    SSAGraph* graph,
+    Node* inst_node,
+    const std::vector<Place>& valid_places) {
+  CHECK(!valid_places.empty()) << "valid_place should be set";
+  // inst -> out node(new_name) -> io_copy_op -> new_var_node(out->AsArg().name)
+  // So there will be a new Argument node and a new IoCopy Statement Node.
+  CHECK(out->IsArg());
+  auto new_name = string_format("%s/target_trans", out->AsArg().name.c_str());
+  auto* new_var_node = graph->NewArgumentNode(out->AsArg().name);
+
+  // Set the place for new var node, the target should be equal to to.target()
+  // The precision and layout should be equal to from.precision(), from.layout()
+  bool is_tensor = from.IsTensor();
+  if (!is_tensor) {
+    CHECK(from.IsTensorList()) << "only support tensor or tensor_array.";
+  }
+  if (is_tensor) {
+    new_var_node->AsArg().type =
+        LiteType::GetTensorTy(to.target(), from.precision(), from.layout());
+  } else {
+    new_var_node->AsArg().type =
+        LiteType::GetTensorListTy(to.target(), from.precision(), from.layout());
+  }
+  auto* io_copy_inst = graph->NewInstructNode();
+  std::string io_copy_type = "io_copy";
+  // create Op and kernels.
+  auto io_copy_op = LiteOpRegistry::Global().Create(io_copy_type);
+  CHECK(io_copy_op) << "create op [" << io_copy_op << "] failed";
+  // CHECK(io_copy_op);
+  // Create the new var manually.
+  inst_node->AsStmt().op()->scope()->Var(new_name);
+
+  // Create IoCopy Instruction.
+  cpp::OpDesc op_desc;
+  op_desc.SetType(io_copy_type);
+  if (is_tensor) {
+    op_desc.SetInput("Input", {new_name});
+    op_desc.SetOutput("Out", {out->AsArg().name});
+  } else {
+    op_desc.SetInput("InputArray", {new_name});
+    op_desc.SetOutput("OutArray", {out->AsArg().name});
+  }
+  io_copy_op->Attach(op_desc, inst_node->AsStmt().op()->scope());
+  auto kernels = io_copy_op->CreateKernels(valid_places);
+  bool is_found = false;
+  std::vector<std::unique_ptr<KernelBase>> selected_kernels;
+  for (auto& kernel : kernels) {
+    const Type* in_arg_ty = nullptr;
+    const Type* out_arg_ty = nullptr;
+    if (is_tensor) {
+      in_arg_ty = kernel->GetInputDeclType("Input");
+      out_arg_ty = kernel->GetOutputDeclType("Out");
+    } else {
+      in_arg_ty = kernel->GetInputDeclType("InputArray");
+      out_arg_ty = kernel->GetOutputDeclType("OutArray");
+    }
+
+    VLOG(4) << "------ kernel info -------";
+    VLOG(4) << "*in_arg_ty(io_copy kernel input):" << *in_arg_ty;
+    VLOG(4) << "from(last kernel output):" << from;
+    VLOG(4) << "out_arg_ty(io_copy kernel output):" << *out_arg_ty;
+    VLOG(4) << "to:" << to << "\n";
+
+    if (TypeCompatible(*in_arg_ty, from) &&
+        TargetCompatibleTo(*out_arg_ty, to)) {
+      VLOG(4) << "picked";
+      is_found = true;
+    }
+
+    if (is_found) {
+      selected_kernels.emplace_back(std::move(kernel));
+      // we pick the kernel
+      io_copy_inst->AsStmt(
+          io_copy_type, std::move(selected_kernels), io_copy_op);
+      break;
+    }
+    VLOG(4) << "not picked";
+  }
+
+  CHECK(is_found) << "Can't find a io_copy  kernel for io_copy op: " << from
+                  << ":" << inst_node->AsStmt().op_info()->Type() << " -> "
+                  << to << ":" << out->AsArg().name;
+  // Add new link, inst -> var -> io_copy_op -> new_var_node
+  DirectedLink(out, io_copy_inst);
+  DirectedLink(io_copy_inst, new_var_node);
+
+  // Update the original instruction OpDesc.
+  // Update its output var name to the io_copy_output_name
+  auto* inst_node_op_desc = inst_node->AsStmt().op()->mutable_op_info();
+  for (auto& op_output : *inst_node_op_desc->mutable_outputs()) {
+    for (auto& var_name : op_output.second)
+      if (var_name == out->AsArg().name) var_name = new_name;
+  }
+  // Update the input name of Ops whose input var is out var node
+  for (auto& op : out->outlinks) {
+    if (!op->IsStmt()) continue;
+    auto* op_desc = op->AsStmt().op()->mutable_op_info();
+    for (auto& op_input : *op_desc->mutable_inputs())
+      for (auto& var_name : op_input.second)
+        if (var_name == out->AsArg().name) var_name = new_name;
+  }
+  // reset opdesc and update kernel information
+  out->AsArg().name = new_name;
+  auto original_selected_kernel =
+      std::move(inst_node->AsStmt().kernels().front());
+  auto update_op_info = *inst_node->AsStmt().op_info();
+  inst_node->AsStmt().ResetOp(update_op_info, graph->valid_places());
+  inst_node->AsStmt().kernels().clear();
+  inst_node->AsStmt().kernels().emplace_back(
+      std::move(original_selected_kernel));
+
+  for (auto& kernel : inst_node->AsStmt().kernels()) {
+    VLOG(4) << "kernel info: " << kernel->name();
+    inst_node->AsStmt().op()->AttachKernel(kernel.get());
+  }
+
+  graph->CheckValid();
+}
+
+void TypeTargetTransformPass::ComplementOutputs(
+    SSAGraph* graph,
+    Node* inst_node,
+    Node* out,
+    std::map<std::string, Node*>* input_nodes) {
+  // If this output is out of date.
+  if (inst_node->outlinks.end() ==
+      std::find(inst_node->outlinks.begin(), inst_node->outlinks.end(), out))
+    return;
+
+  CHECK(inst_node->IsStmt());
+  auto& inst = inst_node->AsStmt();
+  VLOG(3) << "found Target tensor: " << out->AsArg().name;
+  CHECK(out->IsRoleSet());
+  CHECK(out->IsArg());
+  CHECK(out->AsArg().type);
+  if (input_nodes->count(out->AsArg().name)) {
+    if (!TargetCompatibleTo(
+            *out->AsArg().type,
+            *input_nodes->at(out->AsArg().name)->AsArg().type)) {
+      VLOG(3) << "found Output Target unmatched tensor: " << out->AsArg().name
+              << " for kernel " << inst.op()->DebugString() << " "
+              << *out->AsArg().type << " -> "
+              << *(input_nodes->at(out->AsArg().name))->AsArg().type;
+      AddOutputIoCopyInst(*out->AsArg().type,
+                          *input_nodes->at(out->AsArg().name)->AsArg().type,
+                          out,
+                          graph,
+                          inst_node,
+                          valid_places_);
+    }
   }
 }
 
-void TypeTargetTransformPass::AddIoCopyInst(
+void TypeTargetTransformPass::AddInputIoCopyInst(
     const Type& from,
     const Type& to,
     Node* in,
diff --git a/lite/core/optimizer/mir/type_target_cast_pass.h b/lite/core/optimizer/mir/type_target_cast_pass.h
@@ -38,13 +38,25 @@ class TypeTargetTransformPass : public ProgramPass {
                         Node* in,
                         std::map<std::string, Node*>* copied_nodes);
 
-  void AddIoCopyInst(const Type& from,
-                     const Type& to,
-                     Node* in,
-                     SSAGraph* graph,
-                     Node* inst_node,
-                     std::map<std::string, Node*>* copied_nodes,
-                     const std::vector<Place>& valid_places);
+  void ComplementOutputs(SSAGraph* graph,
+                         Node* inst_node,
+                         Node* out,
+                         std::map<std::string, Node*>* input_nodes);
+
+  void AddInputIoCopyInst(const Type& from,
+                          const Type& to,
+                          Node* in,
+                          SSAGraph* graph,
+                          Node* inst_node,
+                          std::map<std::string, Node*>* copied_nodes,
+                          const std::vector<Place>& valid_places);
+
+  void AddOutputIoCopyInst(const Type& from,
+                           const Type& to,
+                           Node* out,
+                           SSAGraph* graph,
+                           Node* inst_node,
+                           const std::vector<Place>& valid_places);
 
   void SetValidPlaces(const std::vector<Place>& valid_places);
 
diff --git a/lite/model_parser/model_parser.cc b/lite/model_parser/model_parser.cc
@@ -245,9 +245,6 @@ void LoadModelPb(const std::string &model_dir,
   pb::ProgramDesc pb_prog(&pb_proto_prog);
   // Transform to cpp::ProgramDesc
   TransformProgramDescAnyToCpp(pb_prog, cpp_prog);
-#ifdef WITH_CONVERT_TO_SSA
-  general::ssa::ConvertToSSA(cpp_prog);
-#endif
 
   // Load params data from file.
   // NOTE: Only main block be used now.
