!1626 Support ffts memory assign

From: @tangqunzhang
Reviewed-by: @wqtshg,@xchu42,@ji_chen
Signed-off-by: @ji_chen

ge/graph/build/memory/block_mem_assigner.cc

@@ -239,6 +239,10 @@ bool MemoryBlock::IsSameBatchLabel() {
   return all_same_label;
 }
 
+bool MemoryBlock::CanReuse(int32_t thread_scope_id) const {
+  return (thread_scope_id_.find(thread_scope_id) == thread_scope_id_.end());
+}
+
 bool CanNotLifeReuse(MemoryBlock *block) {
   if ((block == nullptr) || !block->reuse_mem_ || block->deleted_block_) {
     return true;
@@ -283,6 +287,14 @@ void MemoryBlock::AddLifeReuseBlock(MemoryBlock *block, DependStreamLife &total_
   if (CanNotLifeReuse(this) || CanNotLifeReuse(block) || (batch_label_ != block->batch_label_)) {
     return;
   }
+
+  // not same thread scode id can reuse
+  for (auto thread_scope_id : ThreadScopeId()) {
+    if (!block->CanReuse(thread_scope_id)) {
+      return;
+    }
+  }
+
   if (block->continuous_block_) {
     AddContinuousLifeReuseBlock(block, total_node_depend_stream_life);
     return;
@@ -659,7 +671,12 @@ bool IsDirectOutputNode(const NodePtr &node, int idx) {
   return false;
 }
 
-bool CanReuseBlock(size_t continuous_life_begin, const MemoryBlock &reusable_block, size_t block_size) {
+bool CanReuseBlock(int32_t thread_scope_id, size_t continuous_life_begin, const MemoryBlock &reusable_block,
+                   size_t block_size) {
+  if (!reusable_block.CanReuse(thread_scope_id)) {
+    return false;
+  }
+
   bool can_reuse = false;
   if (reusable_block.Size() == block_size) {
     // in some continuous input case, continuous first input node's is not same as topo first node.
@@ -1122,6 +1139,8 @@ MemoryBlock *BlockMemAssigner::ApplyMemory(size_t block_size, size_t real_size,
   }
 
   bool is_reuse_memory = false;
+  int32_t thread_scope_id = kInvalidThreadScopeId;
+  (void)ge::AttrUtils::GetInt(node_op_desc, ATTR_NAME_THREAD_SCOPE_ID, thread_scope_id);
   if (ge_disable_reuse_mem_env_ != "1") {
     bool reuse_mem_flag = (mem_type == kOutput) ? IsPreReuse(n, out_index) :
                           !((workspace_reuse_flag.size() > out_index) && !workspace_reuse_flag[out_index]);
@@ -1141,8 +1160,8 @@ MemoryBlock *BlockMemAssigner::ApplyMemory(size_t block_size, size_t real_size,
         GE_IF_BOOL_EXEC(reusable_block->batch_label_ != batch_label, continue);
 
         // A node can reuse blocks of the same stream and preorder streams
-        if (CanReuseBlock(continuous_life_begin_, *reusable_block, block_size)) {
-          reusable_block->AddNodeTypeIndex({n, mem_type, out_index, false, continuous_life_begin_},
+        if (CanReuseBlock(thread_scope_id, continuous_life_begin_, *reusable_block, block_size)) {
+          reusable_block->AddNodeTypeIndex({n, mem_type, out_index, false, continuous_life_begin_, thread_scope_id},
                                            real_size, no_align_size);
           if (mem_type == kOutput) {
             auto iter = anchor_to_symbol_.find(NodeIndexIO(n, out_index, kOut).ToString());
@@ -1168,7 +1187,8 @@ MemoryBlock *BlockMemAssigner::ApplyMemory(size_t block_size, size_t real_size,
 
   // Data and netoutput need zero copy block
   block->is_zero_copy_ = IsZeroCopyBlock(n, continuous);
-  block->AddNodeTypeIndex({n, mem_type, out_index, false, continuous_life_begin_}, real_size, no_align_size);
+  block->AddNodeTypeIndex({n, mem_type, out_index, false, continuous_life_begin_, thread_scope_id},
+                          real_size, no_align_size);
   block->stream_id_ = node_op_desc->GetStreamId();
   block->continuous_block_ = continuous;
   block->batch_label_ = batch_label;
@@ -2062,7 +2082,13 @@ void SetOffsetSize(const NodeTypeIndex &node_type, const MemoryBlock *block,
                    size_t real_size, size_t no_align_size, int32_t child_block_level) {
   ge::OpDescPtr op_desc = node_type.node->GetOpDesc();
   GE_CHK_BOOL_TRUE_EXEC_WITH_LOG(op_desc == nullptr, return, "op_desc is null.");
-  string graph_name = node_type.node->GetOwnerComputeGraph()->GetName();
+  auto owner_graph = node_type.node->GetOwnerComputeGraph();
+  GE_CHK_BOOL_TRUE_EXEC_WITH_LOG(owner_graph == nullptr, return, "owner_graph is null.");
+  string graph_name = owner_graph->GetName();
+  if (owner_graph->GetParentGraph() != nullptr) {
+    graph_name = owner_graph->GetParentGraph()->GetName();
+  }
+
   vector<int64_t> memorys_type;
   int64_t offset = block->HeadOffset();
   size_t end = node_type.life_time_end;
@@ -2108,12 +2134,12 @@ void SetOffsetSize(const NodeTypeIndex &node_type, const MemoryBlock *block,
     op_desc->SetWorkspace(workspace_list);
   }
   GELOGI("[IMAS]Set %s name[%s] optype[%s] %s[%u] offset to [%ld] streamid[%ld] memtype[%ld] size[%zu] realsize[%zu] "
-         "noalignsize[%zu] life time begin[%s] life time end[%zu] child[%d:%d:%d:%d:%d] isref[%d] batch[%s]",
+         "noalignsize[%zu] life time begin[%s] life time end[%zu] child[%d:%d:%d:%d:%d] isref[%d] batch[%s] scope[%d]",
          graph_name.c_str(), op_desc->GetName().c_str(), node_type.node->GetType().c_str(),
          node_type.GetMemType().c_str(), node_type.index, offset, op_desc->GetStreamId(),block->memory_type_,
          block->Size(), real_size, no_align_size, node_type.GetLifeBeginDesc().c_str(), end, child_block_level,
          block->reuse_mem_, block->continuous_block_, block->is_zero_copy_, block->same_stream_, node_type.ref_input,
-         block->batch_label_.c_str());
+         block->batch_label_.c_str(), node_type.thread_scope_id);
 }
 
 void SetBlockOpMemOffset(MemoryBlock *block, int32_t child_block_level) {
@@ -2176,8 +2202,7 @@ Status BlockMemAssigner::Assign() {
 
 bool BlockMemAssigner::CheckIsZeroMemNodeType(const string &node_type) const {
   return (node_type == VARIABLE) || (node_type == CONSTANT) || (node_type == MULTISHAPE) ||
-         (node_type == CONSTANTOP) || (node_type == ASSIGNADD) || (node_type == ASSIGNSUB) ||
-         (node_type == ASSIGN) || (node_type == HVDWAIT);
+         (node_type == CONSTANTOP) || (node_type == HVDWAIT);
 }
 
 bool BlockMemAssigner::GetWorkSpaceMemoryType(const NodePtr &node, size_t index, int64_t &memory_type) {

ge/graph/build/memory/block_mem_assigner.h

@@ -33,14 +33,17 @@
 
 namespace ge {
 const size_t kMaxLifeTime = 0xffffffff;
+const int32_t kInvalidThreadScopeId = -1;
 
 using DependStreamLife = std::map<int64_t, std::map<int64_t, size_t>>;
 
 enum OpMemoryType { kOutput, kWorkspace };
 
 struct NodeTypeIndex {
-  NodeTypeIndex(ge::NodePtr node, OpMemoryType mem_type, uint32_t index, bool ref_input = false, size_t begin = 0)
-      : node(std::move(node)), mem_type(mem_type), index(index), ref_input(ref_input), life_time_begin(begin) {}
+  NodeTypeIndex(ge::NodePtr node, OpMemoryType mem_type, uint32_t index, bool ref_input = false, size_t begin = 0,
+                int32_t thread_scope_id = kInvalidThreadScopeId)
+      : node(std::move(node)), mem_type(mem_type), index(index), ref_input(ref_input), life_time_begin(begin),
+        thread_scope_id(thread_scope_id) {}
 
   ge::NodePtr node = nullptr;
   OpMemoryType mem_type = kOutput;
@@ -48,6 +51,7 @@ struct NodeTypeIndex {
   bool ref_input = false;
   size_t life_time_begin = 0;
   size_t life_time_end = kMaxLifeTime;
+  int32_t thread_scope_id = kInvalidThreadScopeId;
   const string GetMemType() const {
     if (mem_type == kOutput) {
       return "output";
@@ -143,6 +147,9 @@ class MemoryBlock {
         same_stream_ = false;
       }
     }
+    if (node_type_index.thread_scope_id != kInvalidThreadScopeId) {
+      thread_scope_id_.insert(node_type_index.thread_scope_id);
+    }
   }
 
   void AddSymbol(const std::string &symbol) {
@@ -154,6 +161,7 @@ class MemoryBlock {
   const std::vector<size_t> &RealSizeList() const { return real_size_list_; }
   const std::vector<MemoryBlock *> &ChildBlockList() const { return child_blocks_; }
   const std::vector<size_t> &NoAlignSizeList() const { return no_align_size_list_; }
+  const std::set<int32_t> &ThreadScopeId() const { return thread_scope_id_; }
 
   void Resize();
 
@@ -175,6 +183,8 @@ class MemoryBlock {
 
   size_t GetDependLifeBegin(int64_t stream_id, DependStreamLife &node_depend_stream_life);
 
+  bool CanReuse(int32_t thread_scope_id) const;
+
   int ref_count_;
   int64_t stream_id_;
   bool deleted_block_;
@@ -198,6 +208,7 @@ class MemoryBlock {
   std::vector<NodeTypeIndex> node_type_index_list_;
   std::vector<std::string> symbol_list_;
   std::vector<MemoryBlock *> child_blocks_;
+  std::set<int32_t> thread_scope_id_;
 };
 
 class BlockMemAssigner : public MemAssigner {

tests/ut/ge/graph/build/mem_assigner_unittest.cc

@@ -44,12 +44,12 @@ using domi::GetContext;
 
 class UtestMemoryAssignerTest : public testing::Test {
  public:
-  ge::OpDescPtr CreateOpWithWsSize(const string &name, int64_t wsByte, const string &type = "some") {
+  ge::OpDescPtr CreateOpWithWsSize(const string &name, int64_t wsByte, const string &type = "some", int64_t size = 1024) {
     ge::OpDescPtr op_def = make_shared<ge::OpDesc>(name, type);
     auto desc_temp_ptr = make_shared<ge::GeTensorDesc>();
     auto desc_temp = *desc_temp_ptr;
 
-    TensorUtils::SetSize(desc_temp, 1024);
+    TensorUtils::SetSize(desc_temp, size);
     op_def->AddInputDesc(desc_temp);
     op_def->AddOutputDesc(desc_temp);
 
@@ -214,6 +214,44 @@ class UtestMemoryAssignerTest : public testing::Test {
 
     return builder.GetGraph();
   }
+  void make_ffts_reuse_graph(ge::ComputeGraphPtr graph, int32_t thread_scope_id_1 = kInvalidThreadScopeId,
+                            int32_t thread_scope_id_2 = kInvalidThreadScopeId) {
+    ge::OpDescPtr op_def_a = CreateOpWithWsSize("A", 512);
+    ge::OpDescPtr op_def_b = CreateOpWithWsSize("B", 0);
+    ge::OpDescPtr op_def_c = CreateOpWithWsSize("C", 512);
+    ge::OpDescPtr op_def_d = CreateOpWithWsSize("D", 512);
+    ge::OpDescPtr op_def_e = CreateOpWithWsSize("E", 0);
+    ge::OpDescPtr op_def_f = CreateOpWithWsSize("F", 512, "some", 2048UL);
+    ge::OpDescPtr op_def_g = CreateOpWithWsSize("G", 0);
+
+    if (thread_scope_id_1 != kInvalidThreadScopeId) {
+      (void)ge::AttrUtils::SetInt(op_def_a, ATTR_NAME_THREAD_SCOPE_ID, thread_scope_id_1);
+      (void)ge::AttrUtils::SetInt(op_def_b, ATTR_NAME_THREAD_SCOPE_ID, thread_scope_id_1);
+      (void)ge::AttrUtils::SetInt(op_def_c, ATTR_NAME_THREAD_SCOPE_ID, thread_scope_id_1);
+    }
+
+    if (thread_scope_id_2 != kInvalidThreadScopeId) {
+      (void)ge::AttrUtils::SetInt(op_def_d, ATTR_NAME_THREAD_SCOPE_ID, thread_scope_id_2);
+      (void)ge::AttrUtils::SetInt(op_def_e, ATTR_NAME_THREAD_SCOPE_ID, thread_scope_id_2);
+      (void)ge::AttrUtils::SetInt(op_def_f, ATTR_NAME_THREAD_SCOPE_ID, thread_scope_id_2);
+    }
+
+    ge::NodePtr node_a = graph->AddNode(op_def_a);
+    ge::NodePtr node_b = graph->AddNode(op_def_b);
+    ge::NodePtr node_c = graph->AddNode(op_def_c);
+    ge::NodePtr node_d = graph->AddNode(op_def_d);
+    ge::NodePtr node_e = graph->AddNode(op_def_e);
+    ge::NodePtr node_f = graph->AddNode(op_def_f);
+    ge::NodePtr node_g = graph->AddNode(op_def_g);
+
+    ge::GraphUtils::AddEdge(node_a->GetOutDataAnchor(0), node_b->GetInDataAnchor(0));
+    ge::GraphUtils::AddEdge(node_b->GetOutDataAnchor(0), node_c->GetInDataAnchor(0));
+    ge::GraphUtils::AddEdge(node_c->GetOutDataAnchor(0), node_d->GetInDataAnchor(0));
+    ge::GraphUtils::AddEdge(node_d->GetOutDataAnchor(0), node_e->GetInDataAnchor(0));
+    ge::GraphUtils::AddEdge(node_e->GetOutDataAnchor(0), node_f->GetInDataAnchor(0));
+    ge::GraphUtils::AddEdge(node_f->GetOutDataAnchor(0), node_g->GetInDataAnchor(0));
+    graph->TopologicalSorting();
+  }
 
  protected:
   void SetUp() {}
@@ -419,3 +457,33 @@ TEST_F(UtestMemoryAssignerTest, graph_memory_assign_atomic_output_and_workspace)
     EXPECT_EQ(mem_offset.mem_offset_, 1024);
   }
 }
+
+TEST_F(UtestMemoryAssignerTest, Mock_ffts_reuse_no_functinon_op) {
+  ge::ComputeGraphPtr graph = make_shared<ge::ComputeGraph>("");
+  make_ffts_reuse_graph(graph, kInvalidThreadScopeId, kInvalidThreadScopeId);
+  HybridMemAssigner hybridMemAssigner(graph);
+  ge::Status ret = hybridMemAssigner.Assign();
+  size_t offset = hybridMemAssigner.GetMemOffset();
+  EXPECT_EQ(offset, 5120);
+  EXPECT_EQ(ret, SUCCESS);
+}
+
+TEST_F(UtestMemoryAssignerTest, Mock_ffts_reuse_two_functinon_op) {
+  ge::ComputeGraphPtr graph = make_shared<ge::ComputeGraph>("");
+  make_ffts_reuse_graph(graph, 0, 1);
+  HybridMemAssigner hybridMemAssigner(graph);
+  ge::Status ret = hybridMemAssigner.Assign();
+  size_t offset = hybridMemAssigner.GetMemOffset();
+  EXPECT_EQ(offset, 6656);
+  EXPECT_EQ(ret, SUCCESS);
+}
+
+TEST_F(UtestMemoryAssignerTest, Mock_ffts_reuse_one_functinon_op) {
+  ge::ComputeGraphPtr graph = make_shared<ge::ComputeGraph>("");
+  make_ffts_reuse_graph(graph, 0, kInvalidThreadScopeId);
+  HybridMemAssigner hybridMemAssigner(graph);
+  ge::Status ret = hybridMemAssigner.Assign();
+  size_t offset = hybridMemAssigner.GetMemOffset();
+  EXPECT_EQ(offset, 5632);
+  EXPECT_EQ(ret, SUCCESS);
+}