!1119 [bug][auto_mixed_precision]add pass for allreduce addn mul

Merge pull request !1119 from vlne-v1/I1F7ZR-reid-can-not-train-and-run-error

mindspore/ccsrc/operator/ops.cc

@@ -241,6 +241,7 @@ const PrimitivePtr kPrimNotInDict = std::make_shared<Primitive>("not_in_dict");
 const PrimitivePtr kPrimMirror = std::make_shared<Primitive>("_MirrorOperator");
 const PrimitivePtr kPrimVirtualDiv = std::make_shared<Primitive>("_VirtualDiv");
 const PrimitivePtr kPrimVirtualDataset = std::make_shared<Primitive>("_VirtualDataset");
+const PrimitivePtr kPrimAllReduce = std::make_shared<Primitive>("AllReduce");
 
 // Debug ops
 const PrimitivePtr kPrimScalarSummary = std::make_shared<Primitive>("ScalarSummary");

mindspore/ccsrc/operator/ops.h

@@ -245,6 +245,7 @@ extern const PrimitivePtr kPrimInDict;
 extern const PrimitivePtr kPrimNotInDict;
 
 // Comm ops
+extern const PrimitivePtr kPrimAllReduce;
 extern const PrimitivePtr kPrimMirror;
 extern const PrimitivePtr kPrimVirtualDiv;
 extern const PrimitivePtr kPrimVirtualDataset;

mindspore/ccsrc/optimizer/irpass.cc

@@ -53,6 +53,7 @@ OptimizeIRPassLib::OptimizeIRPassLib() {
                                            {prim::kPrimInsertGradientOf, prim::kPrimPrintShapeType,
                                             prim::kPrimGetRefKey, prim::kPrimMirror, prim::kPrimVirtualDiv});
   zero_like_fill_zero_ = MakeSubstitution(ZeroLikeFillZero(), "zero_like_fill_zero", prim::kPrimZerosLikeTensor);
+  adjust_all_reduce_mul_add_ = MakeSubstitution(AdjustAllReduceMulAdd(), "adjust_all_reduce_mul_add", prim::kPrimAddN);
 
   // ops eliminate
   item_tuple_eliminate_ =

mindspore/ccsrc/optimizer/irpass.h

@@ -35,6 +35,7 @@ class OptimizeIRPassLib {
   SubstitutionPtr arithmetic_simplify_;
   SubstitutionPtr special_op_eliminate_;
   SubstitutionPtr zero_like_fill_zero_;
+  SubstitutionPtr adjust_all_reduce_mul_add_;
 
   //  ops eliminate
   SubstitutionPtr item_tuple_eliminate_;

mindspore/ccsrc/optimizer/irpass/arithmetic_simplify.h

@@ -228,6 +228,116 @@ class ConstantDuplicateMul : public AnfVisitor {
   CNodePtr cnode_;
 };
 
+// grad = AllReduce(grad) / worker_number
+// grad = grad + weight * decy
+// ->
+// grad = grad + weight * decy
+// grad = AllReduce(grad) / worker_number
+
+// {prim::kPrimAddN, {prim::kPrimMakeTuple, {prim::kPrimMul, {prim::kPrimAllReduce, X}, Y}, Z}} ->
+// {prim::kPrimMul, {prim::kPrimAllReduce, {prim::kPrimAddN,{prim::kPrimMakeTuple, Z, X}}}, Y}
+class AdjustAllReduceMulAdd : public AnfVisitor {
+ public:
+  AnfNodePtr operator()(const OptimizerPtr &, const AnfNodePtr &node) override {
+    Reset();
+    // {prim::kPrimAddN, Zs}
+    if (!IsPrimitiveCNode(node, prim::kPrimAddN)) {
+      return nullptr;
+    }
+    auto addn = node->cast<CNodePtr>();
+    if (addn->size() != 2) {
+      return nullptr;
+    }
+    AnfVisitor::Match(prim::kPrimMakeTuple, {IsNode, IsNode})(addn->input(1));
+    if (x_ == nullptr || y_ == nullptr || z_ == nullptr || all_reduce_fg_ == nullptr) {
+      return nullptr;
+    }
+    auto addn_maketuple = addn->input(1);
+
+    auto fg = all_reduce_fg_;
+    // addn inputs cross the graph, make the inputs same as allreduce node.
+    if (z_->isa<CNode>() && fg != z_->func_graph()) {
+      auto cnode_z = z_->cast<CNodePtr>();
+      z_ = NewCNode(cnode_z->inputs(), fg);
+    }
+
+    auto addn_op_node = addn->input(0);
+    auto make_tuple_op_node = addn->input(1)->cast<CNodePtr>()->input(0);
+
+    AnfNodePtr tuple = NewCNode({make_tuple_op_node, z_, x_}, fg);
+    AnfNodePtr add = NewCNode({addn_op_node, tuple}, fg);
+    AnfNodePtr all_reduce = NewCNode({all_reduce_, add}, fg);
+    AnfNodePtr mul = NewCNode({mul_, all_reduce, y_}, fg);
+    ProcessDependEdge(fg, addn_maketuple, all_reduce);
+    return mul;
+  }
+  void ProcessDependEdge(const FuncGraphPtr &fg, const AnfNodePtr &addn_maketuple, const AnfNodePtr &new_node) {
+    // If has dynamic loss scale.
+    auto &users_map = fg->manager()->node_users();
+    auto it = users_map.find(mul_cnode_);
+
+    if (it != users_map.end()) {
+      auto users = it->second;
+      for (auto &user_pair : users) {
+        auto node = user_pair.first;
+        if (node != addn_maketuple) {
+          if (IsPrimitiveCNode(node, prim::kPrimMakeTuple)) {
+            fg->manager()->SetEdge(node, user_pair.second, new_node);
+          }
+        }
+      }
+    }
+  }
+  void Visit(const AnfNodePtr &node) override {
+    if (level_ == 0) {
+      level_ = 1;
+      is_reduce_match_ = false;
+      // {prim::kPrimMul, {prim::kPrimAllReduce, X}, Y}
+      AnfVisitor::Match(prim::kPrimMul)(node);
+      level_ = 0;
+      if (is_reduce_match_) {
+        mul_ = node->cast<CNodePtr>()->input(0);
+        mul_cnode_ = node->cast<CNodePtr>();
+        y_ = tmp_;
+      } else {
+        z_ = node;
+      }
+    }
+
+    if (level_ == 1) {
+      // {prim::kPrimAllReduce, X}
+      if (IsPrimitiveCNode(node, prim::kPrimAllReduce)) {
+        auto cnode = node->cast<CNodePtr>();
+        if (cnode->size() > 1) {
+          all_reduce_ = cnode->input(0);
+          x_ = cnode->input(1);
+          is_reduce_match_ = true;
+          all_reduce_fg_ = cnode->func_graph();
+        }
+      } else {
+        tmp_ = node;
+      }
+    }
+  }
+
+  void Reset() {
+    level_ = 0;
+    is_reduce_match_ = false;
+    x_ = nullptr;
+    y_ = nullptr;
+    z_ = nullptr;
+    tmp_ = nullptr;
+    all_reduce_fg_ = nullptr;
+  }
+
+ private:
+  int level_{0};
+  bool is_reduce_match_{false};
+  AnfNodePtr x_{nullptr}, y_{nullptr}, z_{nullptr}, tmp_{nullptr};
+  AnfNodePtr all_reduce_{nullptr}, mul_{nullptr}, mul_cnode_{nullptr};
+  FuncGraphPtr all_reduce_fg_{nullptr};
+};
+
 class ArithmeticSimplify {
  public:
   ArithmeticSimplify()

mindspore/ccsrc/pipeline/parse/function_block.h

@@ -28,6 +28,7 @@
 #include <utility>
 #include "pipeline/parse/parse_base.h"
 #include "utils/log_adapter.h"
+#include "utils/ordered_map.h"
 
 namespace mindspore {
 namespace parse {
@@ -99,7 +100,7 @@ class FunctionBlock : public std::enable_shared_from_this<FunctionBlock> {
   std::unordered_map<ParameterPtr, AnfNodePtr> removable_phis_;
 
   // set state nodes need to insert before function return nodes.
-  std::unordered_map<AnfNodePtr, std::string> state_assign_;
+  OrderedMap<AnfNodePtr, std::string> state_assign_;
 
   // hold declared global variables in function
   std::set<std::string> global_vars_;

mindspore/ccsrc/pipeline/pass.cc

@@ -82,6 +82,7 @@ OptPassGroupMap GetOptPassesA(const opt::irpass::OptimizeIRPassLib &irpass) {
     // Arithmetic simplifications
     irpass.arithmetic_simplify_,
     irpass.addn_zero_filter_,
+    irpass.adjust_all_reduce_mul_add_,
 
     // Miscellaneous
     irpass.item_tuple_eliminate_,

mindspore/ops/operations/array_ops.py

@@ -1213,7 +1213,7 @@ class UnsortedSegmentSum(PrimitiveWithInfer):
         Tensor, the shape is :math:`(z, x_{N+1}, ..., x_R)`.
 
     Examples:
-        >>> input_x = Tensor([1, 2, 3, 4], mindspore.float)
+        >>> input_x = Tensor([1, 2, 3, 4], mindspore.float32)
         >>> segment_ids = Tensor([0, 0, 1, 2], mindspore.int32)
         >>> num_segments = 4
         >>> P.UnsortedSegmentSum()(input_x, segment_ids, num_segments)

mindspore/ops/operations/nn_ops.py

@@ -1765,7 +1765,7 @@ class LayerNorm(Primitive):
     `Layer Normalization <https://arxiv.org/abs/1607.06450>`_.
 
     .. math::
-        y = \frac{x - mean]}{\sqrt{variance + \epsilon}} * \gamma + \beta
+        y = \frac{x - mean}{\sqrt{variance + \epsilon}} * \gamma + \beta
 
     where :math:`\gamma` is scale, :math:`\beta` is bias, :math:`\epsilon` is epsilon.
 

mindspore/ops/primitive.py

@@ -284,7 +284,8 @@ def prim_attr_register(fn):
 
 def constexpr(fn=None, get_instance=True, name=None):
     """
-    Makes a PrimitiveWithInfer operator, which infer the value while compiling.
+    Makes a PrimitiveWithInfer operator, which infer the value while compiling. We can define a function
+    to compute between constant variable and used in constructß.
 
     Args:
         fn (function): A `fn` use as the infer_value of the output operator.

tests/ut/cpp/optimizer/lib_test.cc

@@ -556,5 +556,24 @@ TEST_F(TestOptLib, test_constant_duplicate_mul) {
   ASSERT_TRUE(CheckOpt(beforerl, after, patterns));
   ASSERT_TRUE(CheckOpt(beforerr, after, patterns));
 }
+
+TEST_F(TestOptLib, test_adjust_allreduce_mul_add) {
+  FuncGraphPtr beforell = getPyFun.CallAndParseRet("test_adjust_allreduce_mul_add", "beforell");
+  FuncGraphPtr beforelr = getPyFun.CallAndParseRet("test_adjust_allreduce_mul_add", "beforelr");
+  FuncGraphPtr beforerl = getPyFun.CallAndParseRet("test_adjust_allreduce_mul_add", "beforerl");
+  FuncGraphPtr beforerr = getPyFun.CallAndParseRet("test_adjust_allreduce_mul_add", "beforerr");
+  FuncGraphPtr after1 = getPyFun.CallAndParseRet("test_adjust_allreduce_mul_add", "after1");
+  FuncGraphPtr before2r = getPyFun.CallAndParseRet("test_adjust_allreduce_mul_add", "before2r");
+  FuncGraphPtr before2l = getPyFun.CallAndParseRet("test_adjust_allreduce_mul_add", "before2l");
+  FuncGraphPtr after2 = getPyFun.CallAndParseRet("test_adjust_allreduce_mul_add", "after2");
+  auto patterns = std::vector<SubstitutionPtr>({irpass.adjust_all_reduce_mul_add_});
+  ASSERT_TRUE(CheckOpt(beforell, after1, patterns));
+  ASSERT_TRUE(CheckOpt(beforelr, after1, patterns));
+  ASSERT_TRUE(CheckOpt(beforerl, after1, patterns));
+  ASSERT_TRUE(CheckOpt(beforerr, after1, patterns));
+  ASSERT_TRUE(CheckOpt(before2l, after2, patterns));
+  ASSERT_TRUE(CheckOpt(before2r, after2, patterns));
+}
+
 }  // namespace opt
 }  // namespace mindspore

tests/ut/cpp/python_input/gtest_input/optimizer/opt_test.py

@@ -1045,8 +1045,8 @@ def test_print_tuple_wrapper(tag):
 
 def test_constant_duplicate_mul(tag):
     fns = FnDict()
-    Mul = Primitive('Mul');
-    Sqrt = Primitive('Sqrt');
+    Mul = Primitive('Mul')
+    Sqrt = Primitive('Sqrt')
 
     x = Tensor(np.array([[2, 2], [2, 3]]).astype('float32'))
     tensor1 = Tensor(np.array([[1.2, 2.1], [2.2, 3.2]]).astype('float32'))
@@ -1073,3 +1073,44 @@ def test_constant_duplicate_mul(tag):
         return Mul(Sqrt(x), Mul(tensor1, tensor2))
 
     return fns[tag]
+
+
+def test_adjust_allreduce_mul_add(tag):
+    fns = FnDict()
+    Mul = Primitive('Mul')
+    AddN = Primitive('AddN')
+    AllReduce = Primitive('AllReduce')
+
+    @fns
+    def beforell(x, y, z):
+        return AddN((z, Mul(y, AllReduce(x))))
+
+    @fns
+    def beforelr(x, y, z):
+        return AddN((z, Mul(AllReduce(x), y)))
+
+    @fns
+    def beforerl(x, y, z):
+        return AddN((Mul(y, AllReduce(x)), z))
+
+    @fns
+    def beforerr(x, y, z):
+        return AddN((Mul(AllReduce(x), y), z))
+
+    @fns
+    def after1(x, y, z):
+        return Mul(AllReduce(AddN((z, x))), y)
+
+    @fns
+    def before2r(x, y, z):
+        return AddN((Mul(AllReduce(x), y), Mul(z, z)))
+
+    @fns
+    def before2l(x, y, z):
+        return AddN((Mul(z, z), Mul(AllReduce(x), y)))
+
+    @fns
+    def after2(x, y, z):
+        return Mul(AllReduce(AddN((Mul(z, z), x))), y)
+
+    return fns[tag]

tests/ut/python/train/test_amp.py

@@ -20,9 +20,14 @@ import mindspore.context as context
 from mindspore import Tensor
 from mindspore import amp
 from mindspore import nn
-from mindspore.train import Model
+from mindspore.train import Model, ParallelMode
+from mindspore import Tensor
+from mindspore.common import dtype as mstype
+import mindspore.context as context
+from mindspore.model_zoo.resnet import resnet50
 from ....dataset_mock import MindData
-
+from mindspore.parallel._auto_parallel_context import auto_parallel_context
+from mindspore.communication.management import init
 
 def setup_module(module):
     context.set_context(mode=context.GRAPH_MODE)
@@ -138,3 +143,22 @@ def test_compile_model_train_O2():
     with pytest.raises(ValueError):
         # not actual run, the metrics step will fail, check if compile ok.
         model.eval(dataset)
+
+def test_compile_model_train_O2_parallel():
+    dataset_types = (np.float32, np.float32)
+    dataset_shapes = ((16, 16), (16, 16))
+
+    dataset = MindDataSet(dataset_types, dataset_shapes)
+
+    net = NetNoLoss(16, 16)
+    loss = nn.MSELoss()
+    optimizer = nn.Momentum(net.trainable_params(), 0.1, 0.9, 0.00004, 1024.0)
+
+    context.set_auto_parallel_context(
+        global_rank=0, device_num=8,
+        mirror_mean=True, parameter_broadcast=True,
+        parallel_mode=ParallelMode.DATA_PARALLEL)
+    init()
+
+    model = Model(net, loss_fn=loss, optimizer=optimizer, metrics={"acc"}, amp_level="O2")
+    model.train(2, dataset, dataset_sink_mode=False)