Fix CSE bug for some operations like `DropoutGenMask` which should not

be optimized as it will generate different values each time.

mindspore/ccsrc/optimizer/cse.cc

@@ -90,6 +90,22 @@ bool CSE::BuildOrderGroupAndDoReplace(const FuncGraphManagerPtr manager) const {
   return changed;
 }
 
+bool CSE::CheckRandomEffect(const AnfNodePtr &main, const AnfNodePtr &node) const {
+  bool has_random_effect = false;
+  auto prim_main = GetCNodePrimitive(main);
+  auto prim_node = GetCNodePrimitive(node);
+  if (prim_main == prim_node) {
+    return false;
+  }
+  if (prim_main != nullptr) {
+    auto effect_val = prim_main->GetAttr(GRAPH_FLAG_RANDOM_EFFECT);
+    if (effect_val != nullptr && effect_val->isa<BoolImm>()) {
+      has_random_effect = GetValue<bool>(effect_val);
+    }
+  }
+  return has_random_effect;
+}
+
 bool CSE::CheckReplace(const AnfNodePtr &main, const AnfNodePtr &node) const {
   MS_EXCEPTION_IF_NULL(main);
   MS_EXCEPTION_IF_NULL(node);
@@ -122,7 +138,7 @@ bool CSE::CheckReplace(const AnfNodePtr &main, const AnfNodePtr &node) const {
           break;
         }
       }
-      if (IsPrimitiveCNode(c_main, prim::kPrimDropoutGenMask)) {
+      if (CheckRandomEffect(c_main, c_node)) {
         appsame = false;
       }
       replace = appsame;

mindspore/ccsrc/optimizer/cse.h

@@ -43,6 +43,8 @@ class CSE {
 
   virtual bool CheckReplace(const AnfNodePtr &main, const AnfNodePtr &node) const;
 
+  virtual bool CheckRandomEffect(const AnfNodePtr &main, const AnfNodePtr &node) const;
+
   bool Cse(const FuncGraphPtr root, const FuncGraphManagerPtr manager) const;
 
  private:

mindspore/ccsrc/pybind_api/export_flags.cc

@@ -32,5 +32,6 @@ const char GRAPH_FLAG_MIX_PRECISION_FP32[] = "fp32";
 const char GRAPH_FLAG_LOOP_CAN_UNROLL[] = "loop_can_unroll";
 const char GRAPH_FLAG_HAS_EFFECT[] = "has_effect";
 const char GRAPH_FLAG_EFFECT_PATIAL_ORDER[] = "_effect_patial_order";
+const char GRAPH_FLAG_RANDOM_EFFECT[] = "_random_effect";
 
 }  // namespace mindspore

mindspore/ccsrc/pybind_api/export_flags.h

@@ -33,6 +33,7 @@ extern const char GRAPH_FLAG_MIX_PRECISION_FP32[];
 extern const char GRAPH_FLAG_LOOP_CAN_UNROLL[];
 extern const char GRAPH_FLAG_HAS_EFFECT[];
 extern const char GRAPH_FLAG_EFFECT_PATIAL_ORDER[];
+extern const char GRAPH_FLAG_RANDOM_EFFECT[];
 
 }  // namespace mindspore
 

mindspore/ops/operations/nn_ops.py

@@ -1877,6 +1877,7 @@ class DropoutGenMask(Primitive):
         self.init_prim_io_names(inputs=['shape', 'keep_prob'], outputs=['output'])
         validator.check_value_type("Seed0", Seed0, [int], self.name)
         validator.check_value_type("Seed1", Seed1, [int], self.name)
+        self.add_prim_attr("_random_effect", True)
 
 
 class DropoutDoMask(PrimitiveWithInfer):

tests/st/networks/models/bert/bert_tdt_lossscale.py

@@ -162,7 +162,7 @@ def test_bert_tdt():
 
     # assertion occurs while the loss value, overflow state or loss_scale value is wrong
     loss_value = np.array(callback.loss_list)
-    expect_loss_value = [12.1918125, 11.966035, 11.972114, 11.982188, 11.974092, 12.610916, 12.17565, 12.840416, 12.40291, 12.621661]
+    expect_loss_value = [12.1918125, 11.966035, 11.972114, 11.982189, 11.973948, 12.610932, 12.17564, 12.840248, 12.40294, 12.621653]
     print("loss value: {}".format(loss_value))
     assert np.allclose(loss_value, expect_loss_value, 0.00001, 0.00001)