!16537 Bugfix in arithmetic_simplify of graphkernel

From: @dayschan
Reviewed-by: @gaoxiong1,@ckey_dou
Signed-off-by: @ckey_dou

mindspore/ccsrc/backend/optimizer/graph_kernel/arithmetic_simplify.cc

@@ -656,27 +656,36 @@ AnfNodePtr SimplifyReduce2(const AnfNodePtr &node) {
   }
   PatternNode<AnfNodePtr> x;
   auto reduce_reduce_lambda = [&node, &x](PrimitivePtr &operation) -> AnfNodePtr {
-    auto tmp_node = node->cast<CNodePtr>();
+    auto reduce2 = node->cast<CNodePtr>();
-    auto arg_node = tmp_node->input(1);
+    auto reduce1 = reduce2->input(1);
-    auto arg_dimensions = TransAxisValueToVector(AnfAlgo::GetNodeAttr<ValuePtr>(arg_node, "axis"));
+    // if the "keep_dims" of two nodes are different, the result "keep_dims" could not match the output shape.
-    auto reduce_dimensions = TransAxisValueToVector(AnfAlgo::GetNodeAttr<ValuePtr>(tmp_node, "axis"));
+    // inferring shape and another "Reshape" is needed, or skip this case.
-    ShapeVector new_dimensions;
+    if (AnfAlgo::GetBooleanAttr(reduce1, "keep_dims") != AnfAlgo::GetBooleanAttr(reduce2, "keep_dims")) {
-    for (size_t i = 0; i < arg_dimensions.size(); ++i) {
+      return nullptr;
-      for (size_t j = 0; j < reduce_dimensions.size(); ++j) {
+    }
-        if (reduce_dimensions[j] >= arg_dimensions[i]) {
+    auto reduce1_axis = TransAxisValueToVector(AnfAlgo::GetNodeAttr<ValuePtr>(reduce1, "axis"));
-          ++reduce_dimensions[j];
+    auto reduce2_axis = TransAxisValueToVector(AnfAlgo::GetNodeAttr<ValuePtr>(reduce2, "axis"));
+    ShapeVector new_axis;
+    // offset the second node's reduction axes.
+    if (!AnfAlgo::GetBooleanAttr(reduce1, "keep_dims")) {
+      for (size_t i = 0; i < reduce1_axis.size(); ++i) {
+        for (size_t j = 0; j < reduce2_axis.size(); ++j) {
+          if (reduce2_axis[j] >= reduce1_axis[i]) {
+            ++reduce2_axis[j];
+          }
         }
       }
     }
-    std::merge(arg_dimensions.begin(), arg_dimensions.end(), reduce_dimensions.begin(), reduce_dimensions.end(),
+    std::merge(reduce1_axis.begin(), reduce1_axis.end(), reduce2_axis.begin(), reduce2_axis.end(),
-               std::back_inserter(new_dimensions));
+               std::back_inserter(new_axis));
+    new_axis.erase(std::unique(new_axis.begin(), new_axis.end()), new_axis.end());
     auto new_cnode = NewCNodeWithInfo({NewValueNode(operation), x.GetNode(node)}, node);
-    AnfAlgo::SetNodeAttr("axis", MakeValue(new_dimensions), new_cnode);
+    SetNodeAttrSafely("axis", MakeValue(new_axis), new_cnode);
     AnfAlgo::CopyNodeAttr("keep_dims", node, new_cnode);
     return new_cnode;
   };
   auto neg_reducesum_lambda = [&node, &x]() -> AnfNodePtr {
-    auto arg_node = NewCNodeWithInfo({NewValueNode(prim::kPrimReduceSum), x.GetNode(node)}, node);
+    auto arg_node = NewCNodeWithInfo({NewValueNode(prim::kPrimReduceSum->Clone()), x.GetNode(node)}, node);
     AnfAlgo::CopyNodeAttr("axis", node, arg_node);
     AnfAlgo::CopyNodeAttr("keep_dims", node, arg_node);
     auto new_cnode = NewCNodeWithInfo({NewValueNode(prim::kPrimNeg), arg_node}, node);

mindspore/ccsrc/backend/optimizer/graph_kernel/axis_normalizer.cc

@@ -51,9 +51,9 @@ bool AxisNormalizer::Process(const FuncGraphPtr &func_graph) const {
         auto v1 = GetValue<int64_t>(axis);
         auto v2 = NormAxis(v1, rank);
         axis_vec.push_back(v2);
-        diff = diff || (v1 != v2);
+        diff = true;
-      } else if (axis->isa<ValueList>() || axis->isa<ValueTuple>()) {
+      } else if (axis->isa<ValueSequeue>()) {
-        auto vec = axis->isa<ValueList>() ? axis->cast<ValueListPtr>()->value() : axis->cast<ValueTuplePtr>()->value();
+        auto vec = axis->cast<ValueSequeuePtr>()->value();
         if (vec.empty()) {
           diff = true;
           for (size_t i = 0; i < rank; i++) {

mindspore/ccsrc/backend/optimizer/graph_kernel/graph_kernel_optimization.cc

@@ -47,8 +47,9 @@ namespace mindspore {
 namespace opt {
 PassManagerPtr GraphKernelOptimizer::PreProcess() const {
   auto pm = std::make_shared<PassManager>("graphkernel_stage1_preprocess");
-  // Add cse at beginning, otherwise the SpreadUpdateState ir size may be huge in yolov3 network.
+  // Do cse before all passes of graphkernel
   pm->AddPass(std::make_shared<CommonSubexpressionElimination>());
+
   // Change Assign(p, a, U) to Assign(Depend(p, U), a)
   pm->AddPass(std::make_shared<SplitAssign>());
 
@@ -59,7 +60,7 @@ PassManagerPtr GraphKernelOptimizer::PreProcess() const {
 
   // Spread the MakeTuple input of UpdateState
   pm->AddPass(std::make_shared<SpreadUpdateState>());
-  // Add cse to reduce the SpreadUpdateState ir size, which is huge in yolov3 network.
+  // Eliminate the common nodes that generated in SpreadUpdateState
   pm->AddPass(std::make_shared<CommonSubexpressionElimination>());
   return pm;
 }