[IR] Support inplace pass (#56672)

* add code

* add code

* refine code

* add code

* fix bug

* fix bug

* fix bug

* add code

* add ut

* polish code

* fix bug

* refine code

* fix bug

* refine code

* fix bug

* refine code

* fix bug

* refine code

* fix bug

* refine code

* add code

* fix bug

* fix bug

* fix bug

* fix bug

* fix bug

* refine code

Author: zhangbo9674

paddle/fluid/framework/new_executor/CMakeLists.txt

@@ -16,6 +16,7 @@ set(STANDALONE_EXECUTOR_DEPS
     phi_kernel_adaptor
     program_translator
     instruction_base
+    pd_inplace_pass
     ir)
 
 cc_library(

paddle/fluid/framework/new_executor/standalone_executor.cc

@@ -16,16 +16,24 @@
 #include "paddle/fluid/framework/new_executor/feed_fetch_utils.h"
 #include "paddle/fluid/framework/new_executor/interpreter/interpreter_util.h"
 #include "paddle/fluid/framework/new_executor/program_interpreter.h"
+#include "paddle/fluid/platform/flags.h"
 #include "paddle/fluid/platform/profiler/event_tracing.h"
 
 #include "paddle/fluid/ir/transforms/pd_op_to_kernel_pass.h"
 
+#include "paddle/fluid/ir/transforms/inplace_pass.h"
 #include "paddle/fluid/ir_adaptor/translator/translate.h"
 #include "paddle/ir/core/program.h"
+#include "paddle/ir/pass/pass.h"
+#include "paddle/ir/pass/pass_manager.h"
 
 PHI_DECLARE_bool(enable_new_ir_in_executor);
 PHI_DECLARE_bool(enable_new_ir_api);
 
+PADDLE_DEFINE_EXPORTED_bool(new_ir_apply_inplace_pass,
+                            true,
+                            "new ir kernel program apply inplace pass.");
+
 namespace paddle {
 namespace framework {
 StandaloneExecutor::StandaloneExecutor(const platform::Place& place,
@@ -101,6 +109,13 @@ StandaloneExecutor::StandaloneExecutor(const platform::Place& place,
       }
       auto kernel_program =
           paddle::dialect::PdOpLowerToKernelPass(base_program.get(), place);
+
+      if (FLAGS_new_ir_apply_inplace_pass) {
+        ir::PassManager pm(ir::IrContext::Instance(), 3);
+        pm.AddPass(ir::CreateInplacePass());
+        pm.Run(kernel_program.get());
+      }
+
       interpretercores_.emplace_back(
           std::make_shared<InterpreterCore>(place_,
                                             fetch_var_names_,

paddle/fluid/ir/dialect/op_generator/op_gen.py

@@ -856,7 +856,7 @@ def OpGenerator(
                     op_infer_meta_map,
                     muta_attr_is_input=False,
                 )
-                if len(op_attribute_name_list) > 1:
+                if len(op_attribute_name_list) > 0:
                     (
                         build_args_with_attr_is_map_for_declare,
                         build_func_with_attr_is_map,

paddle/fluid/ir/dialect/paddle_dialect/utils/op_yaml_info_parser.cc

@@ -84,15 +84,15 @@ const OpRunTimeInfo& OpYamlInfoParser::OpRuntimeInfo() const {
   return std::get<3>(op_info_tuple_);
 }
 
-const std::map<std::string, int>& OpYamlInfoParser::InputName2Id() const {
+const std::map<std::string, uint32_t>& OpYamlInfoParser::InputName2Id() const {
   return input_name2id_;
 }
 
-const std::map<std::string, int>& OpYamlInfoParser::OutputName2Id() const {
+const std::map<std::string, uint32_t>& OpYamlInfoParser::OutputName2Id() const {
   return output_name2id_;
 }
 
-const std::vector<int>& OpYamlInfoParser::NoNeedBufferIds() const {
+const std::vector<uint32_t>& OpYamlInfoParser::NoNeedBufferIds() const {
   return no_need_buffer_ids_;
 }
 
@@ -118,6 +118,17 @@ const std::string& OpYamlInfoParser::InplaceName(
       "Can not find inplace input of [%s].", out_name));
 }
 
+std::unordered_map<uint32_t, uint32_t> OpYamlInfoParser::GetInplaceIdMap()
+    const {
+  std::unordered_map<uint32_t, uint32_t> inplace_id_map;
+  auto& inplace_info = std::get<3>(op_info_tuple_).inplace;
+  for (const auto& info : inplace_info) {
+    inplace_id_map[OutputName2Id().at(info.first)] =
+        InputName2Id().at(info.second);
+  }
+  return inplace_id_map;
+}
+
 bool OpYamlInfoParser::HasView(const std::string& out_name) const {
   auto& view_info = std::get<3>(op_info_tuple_).view;
   for (size_t i = 0; i < view_info.size(); i++) {

paddle/fluid/ir/dialect/paddle_dialect/utils/op_yaml_info_parser.h

@@ -34,10 +34,10 @@ class OpYamlInfoParser {
   const std::vector<std::string>& TensorParams(bool is_kernel = false) const;
   const std::vector<std::string>& AttrParams(bool is_kernel = false) const;
   const OpRunTimeInfo& OpRuntimeInfo() const;
-  const std::map<std::string, int>& InputName2Id() const;
-  const std::map<std::string, int>& OutputName2Id() const;
+  const std::map<std::string, uint32_t>& InputName2Id() const;
+  const std::map<std::string, uint32_t>& OutputName2Id() const;
 
-  const std::vector<int>& NoNeedBufferIds() const;
+  const std::vector<uint32_t>& NoNeedBufferIds() const;
 
   const std::vector<std::string>& InputNames() const {
     return input_name_list_;
@@ -53,6 +53,8 @@ class OpYamlInfoParser {
 
   const std::string& InplaceName(const std::string& out_name) const;
 
+  std::unordered_map<uint32_t, uint32_t> GetInplaceIdMap() const;
+
   bool HasView(const std::string& out_name) const;
 
   const std::string& ViewName(const std::string& out_name) const;
@@ -68,20 +70,20 @@ class OpYamlInfoParser {
   OpInfoTuple op_info_tuple_;
 
   // input info
-  std::map<std::string, int> input_name2id_;
+  std::map<std::string, uint32_t> input_name2id_;
   std::vector<std::string> input_name_list_;
   std::map<std::string, OpInputInfo> input_info_;
-  int input_tensor_number_{0};
+  uint32_t input_tensor_number_{0};
 
   // no_need_buffer_ids
-  std::vector<int> no_need_buffer_ids_;
+  std::vector<uint32_t> no_need_buffer_ids_;
 
   // attribute info
   std::vector<std::string> attribute_name_list_;
   std::map<std::string, OpAttributeInfo> attr_info_;
 
   // output info
-  std::map<std::string, int> output_name2id_;
+  std::map<std::string, uint32_t> output_name2id_;
   std::vector<std::string> output_name_list_;
   std::map<std::string, OpOutputInfo> output_info_;
 

paddle/fluid/ir/transforms/CMakeLists.txt

@@ -12,3 +12,8 @@ cc_library(
   _constant_folding_pass
   SRCS constant_folding_pass.cc
   DEPS standalone_executor pd_op_to_kernel_pass transform_general_functions)
+
+cc_library(
+  pd_inplace_pass
+  SRCS inplace_pass.cc
+  DEPS pd_dialect_core op_yaml_info_parser)