!2410 Support insert memcpy between two hccl op if the part outputs of prior hccl op linking to next hccl op

Merge pull request !2410 from huanghui/insert-memcpy-async-pass

mindspore/ccsrc/backend/optimizer/ascend/ascend_backend_optimization.cc

@@ -87,6 +87,7 @@
 #include "backend/optimizer/ascend/buffer_fusion/segment_eltwise_fusion_pass.h"
 #include "backend/optimizer/ascend/format_type/deal_ref_trans_and_cast.h"
 #include "backend/optimizer/ascend/enhancer/insert_memcpy_async_for_hccl_op.h"
+#include "backend/optimizer/ascend/enhancer/insert_memcpy_async_for_cascade.h"
 #include "backend/optimizer/ascend/enhancer/insert_pad_for_nms_with_mask.h"
 #include "backend/optimizer/ascend/format_type/insert_transdata_for_runop.h"
 #include "backend/optimizer/ascend/enhancer/getnext_memcpy_elimination.h"
@@ -340,6 +341,7 @@ void AscendBackendOptimization(const std::shared_ptr<session::KernelGraph> &kern
   other_pm->AddPass(std::make_shared<AllGatherFusion>());
   other_pm->AddPass(std::make_shared<ReduceScatterFusion>());
   other_pm->AddPass(std::make_shared<BroadcastFusion>());
+  other_pm->AddPass(std::make_shared<InsertMemcpyAsyncForCascade>());
   other_pm->AddPass(std::make_shared<ParameterTransOpFusion>());
   other_pm->AddPass(std::make_shared<RefreshParameterFormat>());
   optimizer->AddPassManager(other_pm);

mindspore/ccsrc/backend/optimizer/ascend/enhancer/insert_memcpy_async_for_cascade.cc

@@ -0,0 +1,114 @@
+/**
+ * Copyright 2020 Huawei Technologies Co., Ltd
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#include "backend/optimizer/ascend/enhancer/insert_memcpy_async_for_cascade.h"
+#include <vector>
+#include <set>
+#include <string>
+#include "utils/utils.h"
+#include "backend/session/anf_runtime_algorithm.h"
+#include "frontend/optimizer/opt.h"
+#include "backend/optimizer/ascend/ascend_helper.h"
+
+namespace mindspore {
+namespace opt {
+namespace {
+bool IsPartOutputsOfHcclOp(const AnfNodePtr &node, const CNodePtr &cur_hccl, const FuncGraphPtr &graph) {
+  MS_EXCEPTION_IF_NULL(node);
+  MS_EXCEPTION_IF_NULL(cur_hccl);
+  MS_EXCEPTION_IF_NULL(graph);
+  if (!AnfAlgo::CheckPrimitiveType(node, prim::kPrimTupleGetItem)) {
+    return false;
+  }
+  auto cnode = node->cast<CNodePtr>();
+  MS_EXCEPTION_IF_NULL(cnode);
+  auto prev_node = cnode->input(kRealInputNodeIndexInTupleGetItem);
+  MS_EXCEPTION_IF_NULL(prev_node);
+  if (!AnfAlgo::IsCommunicationOp(prev_node)) {
+    return false;
+  }
+  auto prev_hccl_op = prev_node->cast<CNodePtr>();
+  MS_EXCEPTION_IF_NULL(prev_hccl_op);
+
+  auto manager = graph->manager();
+  MS_EXCEPTION_IF_NULL(manager);
+  auto &node_users = manager->node_users();
+  auto iter = node_users.find(prev_hccl_op);
+  if (iter == node_users.end()) {
+    MS_LOG(EXCEPTION) << "node has no output in manager";
+  }
+  for (const auto &node_index : iter->second) {
+    AnfNodePtr output = node_index.first;
+    MS_EXCEPTION_IF_NULL(output);
+    if (IsPrimitiveCNode(output, prim::kPrimTupleGetItem)) {
+      bool is_contain = false;
+      for (size_t i = 1; i < cur_hccl->size(); ++i) {
+        if (cur_hccl->input(i) == output) {
+          is_contain = true;
+          break;
+        }
+      }
+      if (!is_contain) {
+        return true;
+      }
+    }
+  }
+  return false;
+}
+}  // namespace
+
+AnfNodePtr InsertMemcpyAsyncForCascade::InsertMemcpyAsync(const FuncGraphPtr &graph, const CNodePtr &hccl_node) const {
+  MS_EXCEPTION_IF_NULL(graph);
+  MS_EXCEPTION_IF_NULL(hccl_node);
+  std::vector<AnfNodePtr> memcpy_async_list;
+  std::vector<AnfNodePtr> new_inputs = {hccl_node->input(0)};
+  for (size_t i = 1; i < hccl_node->size(); ++i) {
+    auto input = hccl_node->input(i);
+    MS_EXCEPTION_IF_NULL(input);
+    // when input is also a hccl op and just part outputs of it linking with cur_hccl_op
+    if (IsPartOutputsOfHcclOp(input, hccl_node, graph)) {
+      auto memcpy_async = CreateMemcpyAsyncOp(graph, input);
+      auto kernel_info = std::make_shared<device::KernelInfo>();
+      memcpy_async->set_kernel_info(kernel_info);
+      MS_EXCEPTION_IF_NULL(kernel_select_);
+      kernel_select_->SelectKernel(memcpy_async->cast<CNodePtr>());
+      new_inputs.push_back(memcpy_async);
+      memcpy_async_list.push_back(memcpy_async);
+    } else {
+      new_inputs.push_back(input);
+    }
+  }
+
+  if (!memcpy_async_list.empty()) {
+    CNodePtr new_hccl_node = std::make_shared<CNode>(*hccl_node);
+    new_hccl_node->set_inputs(new_inputs);
+    return new_hccl_node;
+  }
+  return nullptr;
+}
+
+const AnfNodePtr InsertMemcpyAsyncForCascade::Process(const FuncGraphPtr &func_graph, const AnfNodePtr &node,
+                                                      const EquivPtr &) const {
+  if (func_graph == nullptr || node == nullptr || !node->isa<CNode>()) {
+    return nullptr;
+  }
+  auto cnode = node->cast<CNodePtr>();
+  if (!AnfAlgo::IsCommunicationOp(node)) {
+    return nullptr;
+  }
+  return InsertMemcpyAsync(func_graph, cnode);
+}
+}  // namespace opt
+}  // namespace mindspore

mindspore/ccsrc/backend/optimizer/ascend/enhancer/insert_memcpy_async_for_cascade.h

@@ -0,0 +1,39 @@
+/**
+ * Copyright 2020 Huawei Technologies Co., Ltd
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#ifndef MINDSPORE_CCSRC_PRE_ACTIVATE_ASCEND_ENHANCER_INSERT_MEMCPY_ASYNC_FOR_CASCADE_H_
+#define MINDSPORE_CCSRC_PRE_ACTIVATE_ASCEND_ENHANCER_INSERT_MEMCPY_ASYNC_FOR_CASCADE_H_
+
+#include <memory>
+#include "backend/optimizer/common/optimizer.h"
+#include "backend/optimizer/ascend/ascend_helper.h"
+
+namespace mindspore {
+namespace opt {
+class InsertMemcpyAsyncForCascade : public PatternProcessPass {
+ public:
+  explicit InsertMemcpyAsyncForCascade(bool multigraph = true)
+      : PatternProcessPass("insert_memcpy_async_for_cascade", multigraph),
+        kernel_select_(std::make_shared<KernelSelect>()) {}
+  ~InsertMemcpyAsyncForCascade() override = default;
+  const AnfNodePtr Process(const FuncGraphPtr &, const AnfNodePtr &, const EquivPtr &) const override;
+
+ private:
+  AnfNodePtr InsertMemcpyAsync(const FuncGraphPtr &graph, const CNodePtr &hccl_node) const;
+  KernelSelectPtr kernel_select_;
+};
+}  // namespace opt
+}  // namespace mindspore
+#endif  // MINDSPORE_CCSRC_PRE_ACTIVATE_ASCEND_ENHANCER_INSERT_MEMCPY_ASYNC_FOR_OP_CASCADE_H_

mindspore/ccsrc/backend/optimizer/ascend/enhancer/insert_memcpy_async_for_hccl_op.cc

@@ -32,12 +32,17 @@ const std::set<std::string> kNeedInsertMemcpyOpSet = {kLambNextMVOpName, kLambNe
 bool IsParameterOrValueNode(const AnfNodePtr &node) {
   MS_EXCEPTION_IF_NULL(node);
   auto kernel_with_index = AnfAlgo::VisitKernelWithReturnType(node, 0, true);
-  return kernel_with_index.first->isa<Parameter>() || kernel_with_index.first->isa<ValueNode>();
+  auto real_node = kernel_with_index.first;
+  MS_EXCEPTION_IF_NULL(real_node);
+  if (real_node->isa<Parameter>()) {
+    return true;
+  }
+  return real_node->isa<ValueNode>();
 }
 
-void TransferControl(const CNodePtr &hccl_node, const AnfNodePtr &memcpy_async, const FuncGraphPtr &graph) {
+void TransferControl(const CNodePtr &hccl_node, const std::vector<AnfNodePtr> &memcpy_async_list,
+                     const FuncGraphPtr &graph) {
   MS_EXCEPTION_IF_NULL(hccl_node);
-  MS_EXCEPTION_IF_NULL(memcpy_async);
   MS_EXCEPTION_IF_NULL(graph);
   auto manager = graph->manager();
   MS_EXCEPTION_IF_NULL(manager);
@@ -48,49 +53,62 @@ void TransferControl(const CNodePtr &hccl_node, const AnfNodePtr &memcpy_async,
   }
   // find hccl_node's output which is a control depend
   for (const auto &node_index : iter->second) {
-    AnfNodePtr output = node_index.first;
+    if (!AnfAlgo::CheckPrimitiveType(node_index.first, prim::kPrimControlDepend)) {
-    int output_index = node_index.second;
+      continue;
-    if (AnfAlgo::CheckPrimitiveType(output, prim::kPrimControlDepend)) {
+    }
-      CNodePtr control_depend = output->cast<CNodePtr>();
+    CNodePtr control_depend = node_index.first->cast<CNodePtr>();
-      MS_EXCEPTION_IF_NULL(control_depend);
+    MS_EXCEPTION_IF_NULL(control_depend);
-      std::vector<AnfNodePtr> new_inputs;
+    std::vector<AnfNodePtr> new_inputs;
-      for (size_t i = 0; i < control_depend->size(); ++i) {
+    for (size_t i = 0; i < control_depend->size(); ++i) {
-        if (i == IntToSize(output_index)) {
+      if (i == IntToSize(node_index.second)) {
-          new_inputs.push_back(memcpy_async);
+        std::vector<AnfNodePtr> make_tuple_inputs = {NewValueNode(prim::kPrimMakeTuple)};
-        } else {
+        make_tuple_inputs.insert(make_tuple_inputs.end(), memcpy_async_list.begin(), memcpy_async_list.end());
-          new_inputs.push_back(control_depend->input(i));
+        make_tuple_inputs.emplace_back(hccl_node);
-        }
+        auto make_tuple = graph->NewCNode(make_tuple_inputs);
+        MS_EXCEPTION_IF_NULL(make_tuple);
+        new_inputs.push_back(make_tuple);
+      } else {
+        new_inputs.push_back(control_depend->input(i));
       }
-      control_depend->set_inputs(new_inputs);
     }
+    control_depend->set_inputs(new_inputs);
   }
 }
 }  // namespace
 
-bool InsertMemcpyAsyncForHcclOp::NeedInsertMemcpy(const FuncGraphPtr &graph, const AnfNodePtr &input) const {
+bool InsertMemcpyAsyncForHcclOp::NeedInsertMemcpy(const FuncGraphPtr &graph, const AnfNodePtr &input,
+                                                  const CNodePtr &cur_node) const {
   MS_EXCEPTION_IF_NULL(graph);
   MS_EXCEPTION_IF_NULL(input);
+  MS_EXCEPTION_IF_NULL(cur_node);
   // when input is a parameter or is a value node
   if (IsParameterOrValueNode(input)) {
     return true;
   }
 
-  // when input is a Ref or some special cnodes
+  if (input->isa<CNode>()) {
-  if (kernel_query_->IsTbeRef(input) ||
+    auto manager = graph->manager();
-      kNeedInsertMemcpyOpSet.find(AnfAlgo::GetCNodeName(input)) != kNeedInsertMemcpyOpSet.end()) {
+    MS_EXCEPTION_IF_NULL(manager);
-    return true;
+    auto &node_users = manager->node_users();
-  }
 
-  auto manager = graph->manager();
+    // when input is a Ref cnode
-  MS_EXCEPTION_IF_NULL(manager);
+    if (kernel_query_->IsTbeRef(input)) {
-  auto &node_users = manager->node_users();
+      return true;
-  auto iter = node_users.find(input);
+    }
-  if (iter == node_users.end()) {
+
-    MS_LOG(EXCEPTION) << "node has no output in manager";
+    // when input is some special cnodes
-  }
+    if (kNeedInsertMemcpyOpSet.find(AnfAlgo::GetCNodeName(input)) != kNeedInsertMemcpyOpSet.end()) {
-  // when input is used by others
+      return true;
-  if (iter->second.size() > 1) {
+    }
-    return true;
+
+    // when input is used by others
+    auto iter = node_users.find(input);
+    if (iter == node_users.end()) {
+      MS_LOG(EXCEPTION) << "node has no output in manager";
+    }
+    if (iter->second.size() > 1) {
+      return true;
+    }
   }
   return false;
 }
@@ -98,21 +116,20 @@ bool InsertMemcpyAsyncForHcclOp::NeedInsertMemcpy(const FuncGraphPtr &graph, con
 void InsertMemcpyAsyncForHcclOp::InsertMemcpyAsync(const FuncGraphPtr &graph, const CNodePtr &hccl_node) const {
   MS_EXCEPTION_IF_NULL(graph);
   MS_EXCEPTION_IF_NULL(hccl_node);
-  bool has_insert_memcpy = false;
+  std::vector<AnfNodePtr> memcpy_async_list;
-  AnfNodePtr memcpy_async = nullptr;
   std::vector<AnfNodePtr> new_inputs = {hccl_node->input(0)};
   for (size_t i = 1; i < hccl_node->size(); ++i) {
     auto input = hccl_node->input(i);
-    if (NeedInsertMemcpy(graph, input)) {
+    if (NeedInsertMemcpy(graph, input, hccl_node)) {
-      memcpy_async = CreateMemcpyAsyncOp(graph, input);
+      auto memcpy_async = CreateMemcpyAsyncOp(graph, input);
-      has_insert_memcpy = true;
       new_inputs.push_back(memcpy_async);
+      memcpy_async_list.push_back(memcpy_async);
     } else {
       new_inputs.push_back(input);
     }
   }
 
-  if (has_insert_memcpy) {
+  if (!memcpy_async_list.empty()) {
     CNodePtr new_hccl_node = std::make_shared<CNode>(*hccl_node);
     new_hccl_node->set_inputs(new_inputs);
     auto manager = graph->manager();
@@ -122,9 +139,7 @@ void InsertMemcpyAsyncForHcclOp::InsertMemcpyAsync(const FuncGraphPtr &graph, co
     MS_LOG(DEBUG) << "end replace";
 
     // transer hccl op's control to the memcpy_async
-    if (hccl_node->size() == 2) {
+    TransferControl(new_hccl_node, memcpy_async_list, graph);
-      TransferControl(new_hccl_node, memcpy_async, graph);
-    }
   }
 }
 

mindspore/ccsrc/backend/optimizer/ascend/enhancer/insert_memcpy_async_for_hccl_op.h

@@ -32,7 +32,7 @@ class InsertMemcpyAsyncForHcclOp : public PatternProcessPass {
 
  private:
   void InsertMemcpyAsync(const FuncGraphPtr &graph, const CNodePtr &hccl_node) const;
-  bool NeedInsertMemcpy(const FuncGraphPtr &graph, const AnfNodePtr &input) const;
+  bool NeedInsertMemcpy(const FuncGraphPtr &graph, const AnfNodePtr &input, const CNodePtr &cur_node) const;
   KernelQueryPtr kernel_query_;
 };
 }  // namespace opt

tests/ut/cpp/pre_activate/ascend/enhancer/insert_memcpy_async_for_hccl_op_test.cc

@@ -22,6 +22,7 @@
 #include "utils/utils.h"
 #include "backend/kernel_compiler/kernel_build_info.h"
 #include "backend/optimizer/common/optimizer.h"
+#include "ir/param_value.h"
 #define private public
 #define protected public
 #include "backend/optimizer/ascend/enhancer/insert_memcpy_async_for_hccl_op.h"
@@ -44,12 +45,10 @@ class MockInsertMemcpyForHcclKernelQuery : public KernelQuery {
   ~MockInsertMemcpyForHcclKernelQuery() override = default;
   bool IsTbeRef(const AnfNodePtr &node) override {
     MS_EXCEPTION_IF_NULL(node);
-    auto cnode = node->cast<CNodePtr>();
+    if (!node->isa<CNode>()) {
-    if (cnode == nullptr) {
       return false;
     }
-    auto name = AnfAlgo::GetCNodeName(cnode);
+    return AnfAlgo::GetCNodeName(node->cast<CNodePtr>()) == "ApplyMomentum";
-    return name == "ApplyMomentum";
   }
 };
 
@@ -105,6 +104,11 @@ TEST_F(TestHWInsertMemcpyForHccl, test_cond2) {
   AbstractBasePtrList args_spec_list{x_abstract};
   auto kg = GetKernelGraph(g, args_spec_list);
   EXPECT_NE(kg, nullptr);
+  for (auto p : kg->parameters()) {
+    auto param = p->cast<ParameterPtr>();
+    EXPECT_NE(param, nullptr);
+    param->set_default_param(std::make_shared<ParamValue>());
+  }
 
   auto optimizer = std::make_shared<opt::GraphOptimizer>();
   auto pm = std::make_shared<opt::PassManager>();
@@ -146,10 +150,16 @@ TEST_F(TestHWInsertMemcpyForHccl, test_cond4) {
   ASSERT_TRUE(g != nullptr);
   std::vector<int> shp_x{1, 64, 112, 112};
   auto x_abstract = std::make_shared<abstract::AbstractTensor>(kFloat32, shp_x);
-  AbstractBasePtrList args_spec_list{x_abstract, x_abstract, x_abstract, x_abstract, x_abstract};
+  AbstractBasePtrList args_spec_list{x_abstract, x_abstract};
   auto kg = GetKernelGraph(g, args_spec_list);
   EXPECT_NE(kg, nullptr);
 
+  for (auto p : kg->parameters()) {
+    auto param = p->cast<ParameterPtr>();
+    EXPECT_NE(param, nullptr);
+    param->set_default_param(std::make_shared<ParamValue>());
+  }
+
   auto optimizer = std::make_shared<opt::GraphOptimizer>();
   auto pm = std::make_shared<opt::PassManager>();
   auto pass = std::make_shared<opt::InsertMemcpyAsyncForHcclOp>();
@@ -161,5 +171,34 @@ TEST_F(TestHWInsertMemcpyForHccl, test_cond4) {
   FuncGraphPtr g_after = get_py_fun_.CallAndParseRet("test_insert_memcpy_async_for_hccl_op_cond4", "after");
   EXPECT_TRUE(CheckEqualGraph(g_after, new_graph));
 }
+
+TEST_F(TestHWInsertMemcpyForHccl, test_cond5) {
+  get_py_fun_.SetDoResolve(true);
+  FuncGraphPtr g = get_py_fun_.CallAndParseRet("test_insert_memcpy_async_for_hccl_op_cond5", "before");
+  ASSERT_TRUE(g != nullptr);
+  std::vector<int> shp_x{1, 64, 112, 112};
+  auto x_abstract = std::make_shared<abstract::AbstractTensor>(kFloat32, shp_x);
+  AbstractBasePtrList args_spec_list{x_abstract, x_abstract, x_abstract};
+  auto kg = GetKernelGraph(g, args_spec_list);
+  EXPECT_NE(kg, nullptr);
+
+  for (auto p : kg->parameters()) {
+    auto param = p->cast<ParameterPtr>();
+    EXPECT_NE(param, nullptr);
+    param->set_default_param(std::make_shared<ParamValue>());
+  }
+
+  auto optimizer = std::make_shared<opt::GraphOptimizer>();
+  auto pm = std::make_shared<opt::PassManager>();
+  auto pass = std::make_shared<opt::InsertMemcpyAsyncForHcclOp>();
+  pass->kernel_query_ = std::make_shared<MockInsertMemcpyForHcclKernelQuery>();
+  pm->AddPass(pass);
+  optimizer->AddPassManager(pm);
+  auto new_graph = optimizer->Optimize(kg);
+  kg->SetExecOrderByDefault();
+
+  FuncGraphPtr g_after = get_py_fun_.CallAndParseRet("test_insert_memcpy_async_for_hccl_op_cond5", "after");
+  EXPECT_TRUE(CheckEqualGraph(g_after, new_graph));
+}
 }  // namespace opt
 }  // namespace mindspore

tests/ut/cpp/python_input/gtest_input/pre_activate/insert_memcpy_async_for_hccl_op.py

@@ -17,6 +17,7 @@ from mindspore.ops import Primitive
 from mindspore.ops import operations as P
 
 all_reduce = P.AllReduce()
+broadcast = P.Broadcast(1)
 memcpy_async = Primitive('memcpy_async')
 make_tuple = Primitive('make_tuple')
 tuple_getitem = Primitive('tuple_getitem')
@@ -101,20 +102,40 @@ def test_insert_memcpy_async_for_hccl_op_cond4(tag):
     fns = FnDict()
 
     @fns
-    def before(a, b, c, d, e):
+    def before(a, b):
-        res1 = apply_momentun(a, b, c, d, e)
+        x = relu(a)
-        res2 = all_reduce(a)
+        y = all_reduce(b)
-        res = control_depend(res1, res2)
+        res = control_depend(x, y)
-        res = make_tuple(res, res2)
         return res
 
     @fns
-    def after(a, b, c, d, e):
+    def after(a, b):
-        res1 = apply_momentun(a, b, c, d, e)
+        x = relu(a)
-        res2 = memcpy_async(a)
+        y1 = memcpy_async(b)
-        res3 = all_reduce(res2)
+        y2 = all_reduce(y1)
-        res = control_depend(res1, res2)
+        res = control_depend(x, make_tuple(y1, y2))
-        res = make_tuple(res, res3)
+        return make_tuple(res)
+
+    return fns[tag]
+
+
+def test_insert_memcpy_async_for_hccl_op_cond5(tag):
+    fns = FnDict()
+
+    @fns
+    def before(a, b, c):
+        x = relu(a)
+        y = broadcast((b, c))
+        res = control_depend(x, y)
+        return res
+
+    @fns
+    def after(a, b, c):
+        x = relu(a)
+        m1 = memcpy_async(b)
+        m2 = memcpy_async(c)
+        y = broadcast(m1, m2)
+        res = control_depend(x, make_tuple(m1, m2, y))
         return make_tuple(res)
 
     return fns[tag]