cherry-pick PR 16547,16736,16739 test=release/1.4 (#16748)

* fix the bug of reusing different types of variables in memory_optimiz… (#16547)

* fix the bug of reusing different types of variables in memory_optimize_pass, test=develop

* disable SELECTED_ROWS AND LOD_TENSOR_ARRAY reusage, test=develop

* only use the latest version variable for inplace strategy (#16736)

* bug-fix, test=develop

* tweak code, test=develop

* cherry-pick PR 16547,16736,16739 test=release/1.4

Author: liupluswei

paddle/fluid/framework/details/inplace_op_pass.cc

@@ -305,6 +305,12 @@ void InplacePass::TryInplaceOpInputOutput(ir::Node* op,
 
     VLOG(4) << "Try to inplace " << in_var_name << " with " << out_var_name;
 
+    if (var_nodes_[in_var_name].back() != in_node) {
+      VLOG(4) << "SKIP since " << in_var_name
+              << " is also used as output by other ops";
+      continue;
+    }
+
     bool can_replace = true;
     if (in_var_name == out_var_name) {
       can_replace = false;
@@ -527,6 +533,9 @@ void GraphView::Build(ir::Graph* g) {
   };
   for (auto& node : g->Nodes()) {
     if (!node->IsOp()) continue;
+    // avoid optimize the variable used in sub-blocks
+    if (OpHasSubBlock(node->Op())) update_skip_set(node);
+
     if (node->Name() == "send") update_skip_set(node);
     if (node->Name() == "recv") update_skip_set(node);
     if (node->Name() == "prefetch") update_skip_set(node);

paddle/fluid/framework/details/memory_optimize_helper.cc

@@ -131,16 +131,7 @@ size_t NodeSize(const VarDesc& node) {
   return type_size * std::abs(size);
 }
 
-size_t NodeSize(ir::Node* n) {
-  VarDesc* desc = nullptr;
-  // some op do not have block pointer
-  if (n->inputs[0]->Op() != nullptr) {
-    desc = FindVarDescInBlock(n);
-  } else {
-    desc = n->Var();
-  }
-  return NodeSize(*desc);
-}
+size_t NodeSize(ir::Node* n) { return NodeSize(*(n->Var())); }
 
 std::string DebugStringImpl(VarDesc* var) {
   std::stringstream ss;
@@ -163,24 +154,22 @@ std::string DebugStringImpl(VarDesc* var) {
 }
 
 std::string DebugString(ir::Node* var) {
-  return DebugStringImpl(FindVarDescInBlock(var));
+  return DebugStringImpl(GetVarDesc(var));
 }
 
 // NOTE(dzh): based ir node, if a large node has been reused
 // by a small size node, then next time it appear in pool, it will
 // have the small size. Find the original node shap from blockdesc.
-VarDesc* FindVarDescInBlock(ir::Node* n) {
+VarDesc* GetVarDesc(ir::Node* n) {
   PADDLE_ENFORCE(n->IsVar() && !n->IsCtrlVar() && n->inputs.size() == 1);
-  BlockDesc* block = n->inputs[0]->Op()->Block();
-  PADDLE_ENFORCE(block->HasVar(n->Name()),
-                 string::Sprintf("Block do not has var %s", n->Name()));
-  return block->FindVar(n->Name());
+  return n->Var();
 }
 
 struct NodeComparator {
   bool operator()(ir::Node* lhs, ir::Node* rhs) const {
-    auto* lhs_desc = FindVarDescInBlock(lhs);
-    auto* rhs_desc = FindVarDescInBlock(rhs);
+    if (lhs->Var()->GetType() != rhs->Var()->GetType()) return false;
+    auto* lhs_desc = GetVarDesc(lhs);
+    auto* rhs_desc = GetVarDesc(rhs);
     // match data type
     if (lhs_desc->GetDataType() != rhs_desc->GetDataType()) {
       return false;
@@ -204,15 +193,15 @@ void OrderedSet::Insert(ir::Node* var) {
     return;
   }
 
-  auto* var_desc = FindVarDescInBlock(var);
+  auto* var_desc = var->Var();
   auto var_shape = var_desc->GetShape();
   int batch_size = static_cast<int>(var_shape[0]);
 
   NodeComparator functor;
   Iter it = nodes_.begin();
   while (it != nodes_.end()) {
     auto& prev = it->front();
-    auto* cache_desc = FindVarDescInBlock(prev);
+    auto* cache_desc = GetVarDesc(prev);
     int cache_batch_size = cache_desc->GetShape()[0];
     if ((cache_batch_size == -1 && batch_size == -1) ||
         (cache_batch_size != -1 && batch_size != -1)) {
@@ -336,10 +325,16 @@ int MinChunkSize() {
 bool NodeCanReused(const VarDesc& node) {
   auto type = node.GetType();
   // only these types holds bulk of gpu memory
-  if (!(type == proto::VarType::LOD_TENSOR ||
-        type == proto::VarType::LOD_TENSOR_ARRAY)) {
-    return false;
-  }
+  // FIXME(liuwei1031) did not find good ways to test SELECTED_ROWS and
+  // LOD_TENSOR_ARRAY re-use logic,
+  // disable them in version 1.4
+  // if (!(type == proto::VarType::LOD_TENSOR ||
+  //       type == proto::VarType::SELECTED_ROWS ||
+  //       type == proto::VarType::LOD_TENSOR_ARRAY)) {
+  //   return false;
+  // }
+  if (type != proto::VarType::LOD_TENSOR) return false;
+
   // persistable variable is parameter
   if (node.Persistable()) {
     return false;

paddle/fluid/framework/details/memory_optimize_helper.h

@@ -20,6 +20,7 @@
 #include <map>
 #include <set>
 #include <string>
+#include <unordered_map>
 #include <utility>
 #include <vector>
 #include "paddle/fluid/framework/data_type.h"
@@ -140,11 +141,7 @@ size_t NodeSize(const VarDesc&);
 
 std::string DebugString(ir::Node* var);
 
-// NOTE(dzhwinter)
-// after node reuse, the replaced node shape is
-// different with its VarDesc. So need to find the
-// correct VarDesc in Block.
-VarDesc* FindVarDescInBlock(ir::Node* n);
+VarDesc* GetVarDesc(ir::Node* n);
 
 static inline bool IsSameDesc(OpDesc* op1, OpDesc* op2) {
   return op1->Type() == op2->Type() && op1->Inputs() == op2->Inputs() &&