!337 optimize execute order sort

Merge pull request !337 from kisnwang/optimize-execute-order-sort

mindspore/ccsrc/device/ascend/ascend_memory_manager.h

@@ -23,7 +23,7 @@ namespace ascend {
 class AscendMemoryManager : public MemoryManager {
  public:
   AscendMemoryManager() = default;
-  virtual ~AscendMemoryManager() = default;
+  ~AscendMemoryManager() override = default;
 
   void MallocDeviceMemory() override;
   void FreeDeviceMemory() override;

mindspore/ccsrc/device/ascend/ascend_memory_pool.h

@@ -26,6 +26,8 @@ namespace ascend {
 class AscendMemoryPool : public DynamicMemPoolBestFit {
  public:
   ~AscendMemoryPool() override = default;
+  AscendMemoryPool(const AscendMemoryPool&) = delete;
+  AscendMemoryPool& operator=(const AscendMemoryPool&) = delete;
 
   size_t AllocDeviceMem(size_t size, DeviceMemPtr* addr) override;
   bool FreeDeviceMem(const DeviceMemPtr& addr) override;
@@ -51,13 +53,11 @@ class AscendMemoryPool : public DynamicMemPoolBestFit {
 
  private:
   AscendMemoryPool() = default;
-  AscendMemoryPool(const AscendMemoryPool&) = delete;
-  AscendMemoryPool& operator=(const AscendMemoryPool&) = delete;
   bool has_malloc_{false};
   uint8_t* device_mem_pool_base_{nullptr};
   uint64_t device_mem_pool_size_{0};
-  size_t free_mem_size_;
-  size_t total_mem_size_;
+  size_t free_mem_size_{0};
+  size_t total_mem_size_{0};
 };
 }  // namespace ascend
 }  // namespace device

mindspore/ccsrc/session/anf_runtime_algorithm.cc

@@ -858,6 +858,14 @@ bool AnfRuntimeAlgorithm::IsCommunicationOp(const AnfNodePtr &node) {
   return false;
 }
 
+bool AnfRuntimeAlgorithm::IsAllReduceOp(const AnfNodePtr &node) {
+  MS_EXCEPTION_IF_NULL(node);
+  if (node->isa<CNode>() && AnfAlgo::GetCNodeName(node) == kAllReduceOpName) {
+    return true;
+  }
+  return false;
+}
+
 bool AnfRuntimeAlgorithm::IsGetNext(const NotNull<AnfNodePtr> &node) {
   auto kernel_name = AnfAlgo::GetCNodeName(node);
   return kernel_name == kGetNextOpName;

mindspore/ccsrc/session/anf_runtime_algorithm.h

@@ -176,6 +176,7 @@ class AnfRuntimeAlgorithm {
   // get real input index for some tbe ops which input order is different between me and tbe impl
   static size_t GetRealInputIndex(const AnfNodePtr &anf_node, const size_t cur_index);
   static bool IsCommunicationOp(const AnfNodePtr &node);
+  static bool IsAllReduceOp(const AnfNodePtr &node);
   static bool IsGetNext(const NotNull<AnfNodePtr> &node);
 };
 }  // namespace session

mindspore/ccsrc/session/kernel_graph.cc

@@ -50,90 +50,127 @@ std::vector<AnfNodePtr> KernelGraph::outputs() const {
 }
 
 void KernelGraph::SetExecOrderByDefault() {
-  BfsToUpdateNodeOutput();
+  std::stack<AnfNodePtr> seed_nodes;
+  UpdateNodeEdgeList(&seed_nodes);
   execution_order_.clear();
-  std::queue<AnfNodePtr> allreduce_nodes;
-  std::queue<AnfNodePtr> zero_output_nodes;
   std::unordered_set<AnfNodePtr> visited_nodes;
-  auto clear_output = [&zero_output_nodes, &allreduce_nodes, &visited_nodes, this](const AnfNodePtr &input) -> void {
-    if (node_output_num_[input] == 0 && visited_nodes.find(input) == visited_nodes.end()) {
-      MS_EXCEPTION_IF_NULL(input);
-      MS_LOG(DEBUG) << "Clear output num:" << input->DebugString();
-      (void)visited_nodes.insert(input);
-      if (input->isa<CNode>() && AnfAlgo::GetCNodeName(input) == kAllReduceOpName) {
-        allreduce_nodes.push(input);
-      } else {
-        zero_output_nodes.push(input);
+  std::queue<AnfNodePtr> zero_input_nodes;
+
+  auto visit_node_descendant = [&visited_nodes, this](const AnfNodePtr &node, std::queue<AnfNodePtr> *visit_queue) {
+    auto it = node_output_edges_.find(node);
+    if (it == node_output_edges_.end()) {
+      // value node and parameter has no input,no need to print log
+      if (node->isa<CNode>()) {
+        MS_LOG(DEBUG) << "Can not find node [" << node->DebugString() << "]";
       }
+      return;
+    }
+
+    // visit all reduce node first, then other nodes
+    std::vector<AnfNodePtr> active_nodes;
+    for (const auto &output_edge : it->second) {
+      auto next_node = output_edge.first;
+      if (node_input_num_.find(next_node) == node_input_num_.end()) {
+        MS_EXCEPTION_IF_NULL(next_node);
+        MS_LOG(EXCEPTION) << "Can't find node[" << next_node->DebugString() << "]";
+      }
+      MS_EXCEPTION_IF_NULL(next_node);
+      MS_LOG(DEBUG) << "Decrease input:" << next_node->DebugString() << ",node:" << node->DebugString()
+                    << ",num: " << node_input_num_[next_node] << ",decrease num:" << output_edge.second;
+      if (node_input_num_[next_node] < output_edge.second) {
+        MS_LOG(EXCEPTION) << "Input node:" << next_node->DebugString() << ",node_output_num"
+                          << node_input_num_[next_node] << ",depend edge:" << output_edge.second;
+      }
+      node_input_num_[next_node] = node_input_num_[next_node] - output_edge.second;
+      // allreduce first
+      if (node_input_num_[next_node] == 0 && visited_nodes.find(next_node) == visited_nodes.end()) {
+        (void)visited_nodes.insert(next_node);
+        if (AnfAlgo::IsAllReduceOp(next_node)) {
+          MS_LOG(DEBUG) << "visit node:" << next_node->DebugString();
+          visit_queue->push(next_node);
+        } else {
+          active_nodes.emplace_back(next_node);
+        }
+      }
+    }
+
+    for (auto &node : active_nodes) {
+      MS_LOG(DEBUG) << "visit node:" << node->DebugString();
+      visit_queue->push(node);
     }
   };
-  zero_output_nodes.emplace(get_return());
-  while (!zero_output_nodes.empty() || !allreduce_nodes.empty()) {
-    AnfNodePtr node;
-    if (!zero_output_nodes.empty()) {
-      node = zero_output_nodes.front();
-      zero_output_nodes.pop();
+
+  AnfNodePtr last_allreduce_node = nullptr;
+  std::queue<AnfNodePtr> allreduce_descendants;
+  while (!seed_nodes.empty() || last_allreduce_node != nullptr) {
+    // seed nodes first, then visit last all reduce node descendant
+    if (seed_nodes.empty()) {
+      visit_node_descendant(last_allreduce_node, &allreduce_descendants);
+      last_allreduce_node = nullptr;
     } else {
-      node = allreduce_nodes.front();
-      allreduce_nodes.pop();
-    }
-    MS_EXCEPTION_IF_NULL(node);
-    if (node->isa<CNode>() && AnfAlgo::IsRealKernel(node)) {
-      execution_order_.push_back(node->cast<CNodePtr>());
+      zero_input_nodes.push(seed_nodes.top());
+      seed_nodes.pop();
     }
-    auto it = node_input_edges_.find(node);
-    if (it == node_input_edges_.end()) {
-      // value node and parameter has no input,no need to print log
-      if (node->isa<CNode>()) {
-        MS_LOG(DEBUG) << "Can not find node [" << node->DebugString() << "]";
+
+    // all reduce node descendant first, then common queue
+    while (!zero_input_nodes.empty() || !allreduce_descendants.empty()) {
+      AnfNodePtr node = nullptr;
+      bool is_allreduce_descendant = false;
+      if (allreduce_descendants.empty()) {
+        node = zero_input_nodes.front();
+        zero_input_nodes.pop();
+      } else {
+        node = allreduce_descendants.front();
+        allreduce_descendants.pop();
+        is_allreduce_descendant = true;
       }
-      continue;
-    }
-    for (const auto &input_edge : it->second) {
-      if (node_output_num_.find(input_edge.first) == node_output_num_.end()) {
-        MS_EXCEPTION_IF_NULL(input_edge.first);
-        MS_LOG(EXCEPTION) << "Can't find node[" << input_edge.first->DebugString() << "]";
+      // add execute node
+      MS_EXCEPTION_IF_NULL(node);
+      if (node->isa<CNode>() && AnfAlgo::IsRealKernel(node)) {
+        execution_order_.push_back(node->cast<CNodePtr>());
       }
-      MS_EXCEPTION_IF_NULL(input_edge.first);
-      MS_LOG(DEBUG) << "Decrease input:" << input_edge.first->DebugString() << ",node:" << node->DebugString()
-                    << ",num: " << node_output_num_[input_edge.first] << ",decrease num:" << input_edge.second;
-      if (node_output_num_[input_edge.first] < input_edge.second) {
-        MS_LOG(EXCEPTION) << "Input node:" << input_edge.first->DebugString() << ",node_output_num"
-                          << node_output_num_[input_edge.first] << "depend edge:" << input_edge.second;
+      // for all reduce node, visit last all reduce node descendant
+      if (AnfAlgo::IsAllReduceOp(node)) {
+        if (last_allreduce_node != nullptr) {
+          visit_node_descendant(last_allreduce_node, &allreduce_descendants);
+        }
+        last_allreduce_node = node;
+      } else if (is_allreduce_descendant) {
+        visit_node_descendant(node, &allreduce_descendants);
+      } else {
+        visit_node_descendant(node, &zero_input_nodes);
       }
-      node_output_num_[input_edge.first] = node_output_num_[input_edge.first] - input_edge.second;
-      clear_output(input_edge.first);
     }
   }
+
   CheckLoop();
-  std::reverse(execution_order_.begin(), execution_order_.end());
 }
 
 void KernelGraph::CheckLoop() {
-  std::map<AnfNodePtr, size_t> none_zero_output;
-  if (node_output_edges_.size() != node_output_num_.size()) {
-    MS_LOG(EXCEPTION) << "node_output_edges_ size :" << node_output_edges_.size()
-                      << "not equal to node_output_num_ size:" << node_output_num_.size();
+  std::map<AnfNodePtr, size_t> none_zero_nodes;
+  if (node_input_edges_.size() != node_input_num_.size()) {
+    MS_LOG(EXCEPTION) << "node_input_edges_ size :" << node_input_edges_.size()
+                      << "not equal to node_input_num_ size:" << node_input_num_.size();
   }
-  for (auto &it : node_output_num_) {
+  for (auto &it : node_input_num_) {
     MS_EXCEPTION_IF_NULL(it.first);
     string str;
-    auto node_output_it = node_output_edges_.find(it.first);
-    if (node_output_it == node_output_edges_.end()) {
+    auto node_input_it = node_input_edges_.find(it.first);
+    if (node_input_it == node_input_edges_.end()) {
       MS_LOG(EXCEPTION) << "Can't find node [" << it.first->DebugString() << "]";
     }
-    for (const auto &output_edge : node_output_edges_[it.first]) {
-      MS_EXCEPTION_IF_NULL(output_edge.first);
-      str = str.append(output_edge.first->DebugString()).append("|");
+    for (const auto &input_edge : node_input_edges_[it.first]) {
+      MS_EXCEPTION_IF_NULL(input_edge.first);
+      str = str.append(input_edge.first->DebugString()).append("|");
     }
     if (it.second != 0) {
-      MS_LOG(WARNING) << "Node:" << it.first->DebugString() << ",outputs:" << str << ",output num:" << it.second;
-      none_zero_output[it.first] = it.second;
+      MS_LOG(WARNING) << "Node:" << it.first->DebugString() << ",inputs:" << str << ",input num:" << it.second;
+      none_zero_nodes[it.first] = it.second;
     }
   }
   // if don't consider control depend and loop exit,a exception will be throw
-  if (!none_zero_output.empty()) {
-    MS_LOG(EXCEPTION) << "Nodes have loop, left node num:" << none_zero_output.size();
+  if (!none_zero_nodes.empty()) {
+    MS_LOG(EXCEPTION) << "Nodes have loop, left node num:" << none_zero_nodes.size();
   }
 }
 
@@ -346,12 +383,13 @@ void KernelGraph::AddDependEdge(const AnfNodePtr &node, const AnfNodePtr &input,
   } else {
     input_it->second.push_back(input_depend_edge);
   }
-  // add the depend sum of node
-  auto depend_it = node_output_num_.find(input);
-  if (depend_it == node_output_num_.end()) {
-    node_output_num_[input] = 0;
+  // add node input depend num
+  auto depend_it = node_input_num_.find(node);
+  if (depend_it == node_input_num_.end()) {
+    node_input_num_[node] = depend_edge_num;
+  } else {
+    depend_it->second += depend_edge_num;
   }
-  node_output_num_[input] += depend_edge_num;
 }
 
 std::vector<AnfNodePtr> KernelGraph::GetOutputNodes(const AnfNodePtr &node) {
@@ -429,9 +467,9 @@ bool KernelGraph::HandleControlDependNode(const AnfNodePtr &node, std::queue<Anf
   return true;
 }
 
-void KernelGraph::BfsToUpdateNodeOutput() {
+void KernelGraph::UpdateNodeEdgeList(std::stack<AnfNodePtr> *seed_nodes) {
   node_output_edges_.clear();
-  node_output_num_.clear();
+  node_input_num_.clear();
   node_input_edges_.clear();
   std::vector<AnfNodePtr> control_depends;
   std::unordered_set<AnfNodePtr> visited_nodes;
@@ -441,6 +479,11 @@ void KernelGraph::BfsToUpdateNodeOutput() {
     auto node = que.front();
     que.pop();
     MS_EXCEPTION_IF_NULL(node);
+    if (node->isa<Parameter>() || node->isa<ValueNode>()) {
+      seed_nodes->push(node);
+      continue;
+    }
+
     if (!node->isa<CNode>()) {
       continue;
     }
@@ -454,10 +497,6 @@ void KernelGraph::BfsToUpdateNodeOutput() {
         control_depends.push_back(input);
         depend_edge_num = 0;
       }
-      // the 2rd input of depend is no depend edge
-      if (AnfAlgo::CheckPrimitiveType(node, prim::kPrimDepend) && input == cnode->input(kDependAttachNodeIndex)) {
-        depend_edge_num = 0;
-      }
       PushNoVisitedNode(input, &que, &visited_nodes);
       AddDependEdge(node, input, depend_edge_num);
     }

mindspore/ccsrc/session/kernel_graph.h

@@ -22,6 +22,7 @@
 #include <utility>
 #include <string>
 #include <queue>
+#include <stack>
 #include <map>
 #include <unordered_set>
 #include "ir/func_graph.h"
@@ -93,8 +94,8 @@ class KernelGraph : public FuncGraph {
  private:
   // remove value node form graph
   bool RemoveValueNodeFromGraph(const ValueNodePtr &value_node);
-  // BFS to update all nodes' output
-  void BfsToUpdateNodeOutput();
+  // update node edge list
+  void UpdateNodeEdgeList(std::stack<AnfNodePtr> *seed_nodes);
   // add node depend edge by data edge or control depend
   void AddDependEdge(const AnfNodePtr &node, const AnfNodePtr &input, size_t depend_edge_num);
   // handle control depend
@@ -114,7 +115,7 @@ class KernelGraph : public FuncGraph {
   std::unordered_map<tensor::TensorPtr, ValueNodePtr> tensor_to_value_node_map_;
   // include all value nodes
   std::unordered_set<ValueNodePtr> graph_value_nodes_;
-  std::unordered_map<AnfNodePtr, size_t> node_output_num_;
+  std::unordered_map<AnfNodePtr, size_t> node_input_num_;
   std::unordered_map<AnfNodePtr, std::vector<std::pair<AnfNodePtr, size_t>>> node_input_edges_;
   // record map between ref final output anf with index and ref origin input with index
   std::map<AnfWithOutIndex, AnfWithOutIndex> ref_out_in_map_;

tests/st/networks/test_gpu_lstm.py

@@ -135,4 +135,5 @@ def test_LSTM():
     for epoch in range(num_epochs):
         loss = train_network(train_features, train_labels)
         losses.append(loss)
+        print("loss:", loss.asnumpy())
     assert(losses[-1].asnumpy() < 0.01)