PaddlePaddle
diff --git a/‎paddle/fluid/framework/new_executor/CMakeLists.txt
Copy file name to clipboard
+9-2Lines changed: 9 additions & 2 deletions b/‎paddle/fluid/framework/new_executor/CMakeLists.txt
Copy file name to clipboard
+9-2Lines changed: 9 additions & 2 deletions
diff --git a/‎paddle/fluid/framework/new_executor/interpreter/dependency_builder.cc
Copy file name to clipboard
+8-4Lines changed: 8 additions & 4 deletions b/‎paddle/fluid/framework/new_executor/interpreter/dependency_builder.cc
Copy file name to clipboard
+8-4Lines changed: 8 additions & 4 deletions
diff --git a/‎paddle/fluid/framework/new_executor/interpreter/interpreter_util.cc
Copy file name to clipboard
+78Lines changed: 78 additions & 0 deletions b/‎paddle/fluid/framework/new_executor/interpreter/interpreter_util.cc
Copy file name to clipboard
+78Lines changed: 78 additions & 0 deletions
diff --git a/‎paddle/fluid/framework/new_executor/interpreter/interpreter_util.h
Copy file name to clipboard
+9Lines changed: 9 additions & 0 deletions b/‎paddle/fluid/framework/new_executor/interpreter/interpreter_util.h
Copy file name to clipboard
+9Lines changed: 9 additions & 0 deletions
diff --git a/‎paddle/fluid/framework/new_executor/interpreter/stream_analyzer.cc
Copy file name to clipboard
+6-1Lines changed: 6 additions & 1 deletion b/‎paddle/fluid/framework/new_executor/interpreter/stream_analyzer.cc
Copy file name to clipboard
+6-1Lines changed: 6 additions & 1 deletion
@@ -5,8 +5,15 @@ add_subdirectory(workqueue)
 set(STANDALONE_EXECUTOR_SRCS interpretercore.cc new_executor_defs.cc
                              standalone_executor.cc)
 
-set(STANDALONE_EXECUTOR_DEPS interpreter interpretercore_garbage_collector
-                             workqueue)
+set(STANDALONE_EXECUTOR_DEPS
+    interpreter
+    interpretercore_garbage_collector
+    workqueue
+    pd_dialect
+    pd_op_to_kernel_pass
+    phi_kernel_adaptor
+    program_translator
+    new_ir)
 
 cc_library(
   standalone_executor
 
@@ -117,7 +117,8 @@ const std::map<size_t, std::set<size_t>>& DependencyBuilder::OpDownstreamMap()
 
 void DependencyBuilder::AddDependencyForCoalesceTensorOp() {
   for (size_t op_idx = 0; op_idx < op_num_; ++op_idx) {
-    if (instructions_->at(op_idx).OpBase()->Type() == kCoalesceTensor) {
+    if (instructions_->at(op_idx).OpBaseValid() &&
+        instructions_->at(op_idx).OpBase()->Type() == kCoalesceTensor) {
       VLOG(4) << "Add depend for " << kCoalesceTensor << " " << op_idx;
       auto fused_out = instructions_->at(op_idx).Outputs().at("FusedOutput")[0];
       auto outputs = instructions_->at(op_idx).Outputs().at("Output");
@@ -224,7 +225,8 @@ void DependencyBuilder::AddDependencyForCommunicationOp() {
   const std::string kSyncComm = "c_sync_comm_stream";
   dependence_op_idx = ULLONG_MAX;
   for (size_t op_idx = 0; op_idx < op_num_; ++op_idx) {
-    if (instructions_->at(op_idx).OpBase()->Type() == kSyncComm) {
+    if (instructions_->at(op_idx).OpBaseValid() &&
+        instructions_->at(op_idx).OpBase()->Type() == kSyncComm) {
       dependence_op_idx = op_idx;
     } else {
       if (dependence_op_idx != ULLONG_MAX) {
@@ -251,7 +253,8 @@ void DependencyBuilder::AddDependencyForRandomOp() {
 
   size_t dependence_op_idx = ULLONG_MAX;
   for (size_t op_idx = 0; op_idx < op_num_; ++op_idx) {
-    if (random_op_set.count(instructions_->at(op_idx).OpBase()->Type())) {
+    if (instructions_->at(op_idx).OpBaseValid() &&
+        random_op_set.count(instructions_->at(op_idx).OpBase()->Type())) {
       if (dependence_op_idx != ULLONG_MAX) {
         AddDownstreamOp(dependence_op_idx, op_idx);
       }
@@ -275,7 +278,8 @@ void DependencyBuilder::AddDependencyForReadOp() {
   std::vector<size_t> read_ops;
   std::vector<size_t> startup_ops;
   for (size_t op_idx = 0; op_idx < op_num_; ++op_idx) {
-    if (instructions_->at(op_idx).OpBase()->Type() == "read") {
+    if (instructions_->at(op_idx).OpBaseValid() &&
+        instructions_->at(op_idx).OpBase()->Type() == "read") {
       read_ops.push_back(op_idx);
     }
 
 
@@ -22,6 +22,9 @@
 #include "paddle/fluid/framework/new_executor/interpreter/data_transfer.h"
 #include "paddle/fluid/framework/new_executor/interpreter/execution_config.h"
 #include "paddle/fluid/framework/new_executor/interpreter/static_build.h"
+#include "paddle/fluid/ir/dialect/pd_dialect.h"
+#include "paddle/fluid/ir/interface/op_yaml_info.h"
+#include "paddle/fluid/ir/phi_kernel_adaptor/phi_kernel_util.h"
 #include "paddle/fluid/memory/stats.h"
 #include "paddle/fluid/operators/controlflow/conditional_block_op_helper.h"
 #include "paddle/fluid/operators/controlflow/recurrent_op_helper.h"
@@ -138,6 +141,9 @@ bool IsCommunicationOp(const std::string& op_name) {
 }
 
 bool IsCommunicationOp(const Instruction& instr) {
+  if (!instr.OpBaseValid()) {
+    return false;
+  }
   return IsCommunicationOp(instr.OpBase()->Type());
 }
 
@@ -924,6 +930,78 @@ void BuildOpFuncList(const platform::Place& place,
   delete garbages;
 }
 
+void BuildOpFuncList(
+    const platform::Place& place,
+    ::ir::Block* block,
+    std::vector<OpFuncNode>* vec_func_list,
+    framework::Scope* scope,
+    const std::unordered_map<::ir::Value, std::string>& value_2_name_map,
+    const ExecutionConfig& execution_config) {
+  vec_func_list->reserve(block->size());
+  ::ir::IrContext* ctx = ir::IrContext::Instance();
+
+  ctx->GetOrRegisterDialect<paddle::dialect::PaddleDialect>();
+
+  for (auto it = block->begin(); it != block->end(); ++it) {
+    OpFuncNode op_func_node;
+    auto attr_map = (*it)->attributes();
+
+    auto op_name = attr_map.at("op_name").dyn_cast<::ir::StrAttribute>().data();
+
+    if (op_name == "pd.fetch") {
+      VLOG(6) << "skip process pd.fetch op";
+      continue;
+    }
+    op_func_node.phi_op_name_ = op_name;
+
+    ::ir::OpInfo op_info = ctx->GetRegisteredOpInfo(op_name);
+
+    auto impl =
+        op_info.GetInterfaceImpl<paddle::dialect::OpYamlInfoInterface>();
+    auto yaml_info = impl->get_op_info_();
+
+    auto attr_info = std::get<1>(yaml_info);
+
+    op_func_node.infer_shape_interface_ =
+        op_info.GetInterfaceImpl<paddle::dialect::InferShapeInterface>();
+
+    ::ir::BuildInferMetaContext((*it),
+                                value_2_name_map,
+                                scope,
+                                yaml_info,
+                                &(op_func_node.infer_meta_context_));
+
+    auto kernel_name =
+        attr_map.at("kernel_name").dyn_cast<ir::StrAttribute>().data();
+    auto kernel_key = attr_map.at("kernel_key")
+                          .dyn_cast<paddle::dialect::KernelAttribute>()
+                          .data();
+    auto t1 =
+        phi::KernelFactory::Instance().SelectKernel(kernel_name, kernel_key);
+    op_func_node.phi_kernel_ = new phi::Kernel(t1);
+
+    PADDLE_ENFORCE_EQ(op_func_node.phi_kernel_->IsValid(),
+                      true,
+                      "not found kernel for [%s]",
+                      kernel_name);
+    ::ir::BuildPhiKernelContext((*it),
+                                value_2_name_map,
+                                scope,
+                                yaml_info,
+                                &(op_func_node.kernel_context_),
+                                &(op_func_node.input_index),
+                                &(op_func_node.output_index));
+
+    op_func_node.kernel_context_.SetDeviceContext(
+        phi::DeviceContextPool::Instance().Get(
+            phi::TransToPhiPlace(kernel_key.backend())));
+    op_func_node.dev_ctx_ = phi::DeviceContextPool::Instance().Get(
+        phi::TransToPhiPlace(kernel_key.backend()));
+
+    vec_func_list->emplace_back(op_func_node);
+  }
+}
+
 void BuildVariableScope(const framework::BlockDesc& block,
                         const ExecutionConfig& execution_config,
                         VariableScope* var_scope) {
 
@@ -38,6 +38,7 @@
 #include "paddle/fluid/framework/variable_helper.h"
 #include "paddle/fluid/platform/device_context.h"
 #include "paddle/fluid/platform/init.h"
+#include "paddle/ir/core/program.h"
 
 using AtomicVectorSizeT = std::vector<std::atomic<size_t>>;
 
@@ -93,6 +94,14 @@ void BuildOpFuncList(const platform::Place& place,
                      bool use_local_scope = true,
                      bool static_build = false);
 
+void BuildOpFuncList(
+    const platform::Place& place,
+    ::ir::Block* block,
+    std::vector<OpFuncNode>* vec_func_list,
+    framework::Scope* scope,
+    const std::unordered_map<::ir::Value, std::string>& value_2_name_map,
+    const ExecutionConfig& execution_config);
+
 void BuildVariableScope(const framework::BlockDesc& block,
                         const ExecutionConfig& execution_config,
                         VariableScope* var_scope);
 
@@ -143,6 +143,9 @@ void StreamAnalyzer::ConstructEvents(
 DeviceContext* StreamAnalyzer::ParseDeviceContext(
     const OpFuncNode& op_func_node) const {
   auto& op = op_func_node.operator_base_;
+  if (op == nullptr) {
+    return op_func_node.dev_ctx_;
+  }
   auto& op_type = op->Type();
   const std::string& execution_stream = op_func_node.execution_stream_;
   const int stream_priority = op_func_node.stream_priority_;
@@ -201,7 +204,9 @@ DeviceContext* StreamAnalyzer::ParseDeviceContext(
 #endif
   }
 
-  SetDeviceCommContext(op.get(), op_func_node.dev_ctx_);
+  if (op != nullptr) {
+    SetDeviceCommContext(op.get(), op_func_node.dev_ctx_);
+  }
   return op_func_node.dev_ctx_;
 }