!4188 fix switch layer sigle prim cell

Merge pull request !4188 from riemann_penn/fix_switch_layer_sigle_prim_cell

mindspore/ccsrc/debug/anf_ir_utils.cc

@@ -607,6 +607,9 @@ void AnfExporter::ExportOneFuncGraph(std::ofstream &ofs, const FuncGraphPtr &fun
   std::vector<AnfNodePtr> parameters = func_graph->parameters();
   OrderedMap<AnfNodePtr, int, ParamPtrHasher, ParamPtrEqual> param_map;
 
+  if (*(func_graph->switch_layer_input())) {
+    ofs << "switch_layer_input: " << *(func_graph->switch_layer_input()) << "\n";
+  }
   ofs << "# [No." << (exported.size() + 1) << "] " << func_graph->DumpText() << "."
       << func_graph->debug_info()->get_id() << "\n";
   if (label_manage::GetGlobalTraceLabelType() == label_manage::TraceLabelType::kWithUniqueId) {

mindspore/ccsrc/frontend/optimizer/ad/dfunctor.cc

@@ -49,6 +49,8 @@ DFunctor::DFunctor(const FuncGraphPtr &primal_graph, const pipeline::ResourceBas
     std::string grad_op_name = GetValue<std::string>(primal_graph->get_attr(FUNC_GRAPH_ATTR_GRAPH_KERNEL));
     k_graph_->set_attr(FUNC_GRAPH_ATTR_GRAPH_KERNEL, MakeValue(grad_op_name));
   }
+  // To keep switch_layer's inputs from being inlined
+  k_graph_->set_switch_layer_input(primal_graph->switch_layer_input());
   TraceManager::EndTrace();
 
   TraceManager::DebugTrace(std::make_shared<TraceGradBprop>(primal_graph->debug_info()));

mindspore/ccsrc/frontend/optimizer/irpass.cc

@@ -45,6 +45,7 @@
 #include "frontend/optimizer/opt.h"
 #include "frontend/optimizer/irpass/row_tensor_eliminate.h"
 #include "frontend/optimizer/irpass/sparse_tensor_eliminate.h"
+#include "frontend/optimizer/irpass/switch_layer_defer_inline.h"
 
 namespace mindspore {
 namespace opt {
@@ -170,6 +171,10 @@ OptimizeIRPassLib::OptimizeIRPassLib() {
   // Value_Based Eliminate
   value_based_eliminate_ = MakeSubstitution(std::make_shared<ValueBasedEliminate>(), "value_based_eliminate",
                                             {prim::kPrimSelect, prim::kPrimMinimum, prim::kPrimMaximum});
+
+  // switch_layer defer inline
+  switch_layer_defer_inline_ =
+    MakeSubstitution(std::make_shared<SwitchLayerDeferInline>(), "switch_layer_defer_inline", prim::kPrimSwitchLayer);
 }
 
 ResolveIRPassLib::ResolveIRPassLib() {

mindspore/ccsrc/frontend/optimizer/irpass.h

@@ -113,6 +113,9 @@ class OptimizeIRPassLib {
 
   // Value_Based Eliminate
   SubstitutionPtr value_based_eliminate_;
+
+  // SwitchLayer defer inline
+  SubstitutionPtr switch_layer_defer_inline_;
 };
 
 // the collection of irpass for resolve action

mindspore/ccsrc/frontend/optimizer/irpass/inline.h

@@ -39,7 +39,7 @@ class ReplaceApplicator : public AnfVisitor {
     }
 
     auto fg = GetValueNode<FuncGraphPtr>(node);
-    if (fg->has_flag(FUNC_GRAPH_FLAG_DEFER_INLINE) || fg->stub()) {
+    if (fg->has_flag(FUNC_GRAPH_FLAG_DEFER_INLINE) || fg->stub() || *(fg->switch_layer_input())) {
       return nullptr;
     }
 

mindspore/ccsrc/frontend/optimizer/irpass/switch_layer_defer_inline.h

@@ -0,0 +1,47 @@
+/**
+ * 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_FRONTEND_OPTIMIZER_IRPASS_SWITCH_LAYER_DEFER_INLINE_H_
+#define MINDSPORE_CCSRC_FRONTEND_OPTIMIZER_IRPASS_SWITCH_LAYER_DEFER_INLINE_H_
+
+#include <vector>
+#include <algorithm>
+
+#include "frontend/optimizer/irpass.h"
+#include "frontend/optimizer/optimizer.h"
+#include "frontend/optimizer/anf_visitor.h"
+#include "frontend/operator/ops.h"
+
+namespace mindspore {
+namespace opt {
+namespace irpass {
+// {prim::kPrimSwitchLayer, {Index, layers}}
+class SwitchLayerDeferInline : public AnfVisitor {
+ public:
+  AnfNodePtr operator()(const OptimizerPtr &, const AnfNodePtr &node) override {
+    auto cnode = node->cast<CNodePtr>();
+    auto tuple = dyn_cast<abstract::AbstractTuple>(cnode->inputs()[2]->abstract());
+    for (auto elem : tuple->elements()) {
+      auto abstract = dyn_cast<abstract::FuncGraphAbstractClosure>(elem);
+      *(abstract->func_graph()->switch_layer_input()) = true;
+    }
+    return nullptr;
+  }
+};
+}  // namespace irpass
+}  // namespace opt
+}  // namespace mindspore
+#endif  // MINDSPORE_CCSRC_FRONTEND_OPTIMIZER_IRPASS_SWITCH_LAYER_DEFER_INLINE_H_

mindspore/ccsrc/pipeline/jit/pass.cc

@@ -90,6 +90,7 @@ bool CleanAfterOptAPass(const ResourcePtr &res) {
 namespace {
 OptPassGroupMap GetOptPassesA(const opt::irpass::OptimizeIRPassLib &irpass) {
   opt::OptPassConfig a_1 = opt::OptPassConfig({
+    irpass.switch_layer_defer_inline_,
     irpass.switch_simplify_,
 
     // Safe inlining

mindspore/core/ir/func_graph.cc

@@ -48,6 +48,7 @@ FuncGraph::FuncGraph()
       manager_(std::weak_ptr<FuncGraphManager>()),
       stub_(false) {
   debug_info_ = std::make_shared<GraphDebugInfo>();
+  switch_layer_input_ = std::make_shared<bool>(false);
 }
 
 abstract::AbstractBasePtr FuncGraph::ToAbstract() {

mindspore/core/ir/func_graph.h

@@ -353,6 +353,8 @@ class FuncGraph : public FuncGraphBase {
   bool stub() const { return stub_; }
   void set_stub(bool stub) { stub_ = stub; }
   static void set_drawer(Drawer drawer) { drawer_ = drawer; }
+  std::shared_ptr<bool> switch_layer_input() const { return switch_layer_input_; }
+  void set_switch_layer_input(std::shared_ptr<bool> switch_layer_input) { switch_layer_input_ = switch_layer_input; }
 
  private:
   // graph is manipulated by manager and others
@@ -414,6 +416,9 @@ class FuncGraph : public FuncGraphBase {
   std::list<CNodePtr> order_;
   bool stub_;
   inline static Drawer drawer_ = nullptr;
+  // Design switch_layer_input as a ptr to
+  // share between derived backpropagator and cloned graphs
+  std::shared_ptr<bool> switch_layer_input_;
 };
 
 inline CNodePtr NewCNode(const std::vector<AnfNodePtr> &inputs, const FuncGraphPtr &fg) {

mindspore/core/ir/func_graph_cloner.cc

@@ -228,6 +228,7 @@ void Cloner::SetFuncGraphInfo(const FuncGraphPtr &func_graph, FuncGraphPtr *cons
   (*target_func_graph)->set_hyper_param_count(func_graph->hyper_param_count());
   (*target_func_graph)->set_is_generate(func_graph->is_generated());
   (*target_func_graph)->set_stub(func_graph->stub());
+  (*target_func_graph)->set_switch_layer_input(func_graph->switch_layer_input());
   TraceManager::EndTrace();
 }
 
@@ -645,6 +646,7 @@ FuncGraphPtr TransformableClone(const FuncGraphPtr &func_graph, const TraceInfoP
   new_func_graph->set_hyper_param_count(func_graph->hyper_param_count());
   new_func_graph->set_is_generate(func_graph->is_generated());
   new_func_graph->set_stub(func_graph->stub());
+  new_func_graph->set_switch_layer_input(func_graph->switch_layer_input());
   for (auto &item : func_graph->parameter_default_value()) {
     new_func_graph->set_param_default_value(item.first, cloner[item.second]);
   }

tests/ut/python/ops/test_control_ops.py

@@ -444,6 +444,26 @@ def test_index_to_switch_layer():
     C.grad_all(net)(index, Tensor(np.full([128, 96], 0.6, dtype=np.float32)))
 
 
+def test_switch_layer_with_single_prim():
+    class SwitchLayerCell(nn.Cell):
+        def __init__(self):
+            super(SwitchLayerCell, self).__init__()
+            self.layers = (nn.ReLU(), nn.ReLU())
+            self.z3 = Parameter(
+                Tensor(np.full([128, 96], 0.6, dtype=np.float32)), name='z3')
+
+        def construct(self, index, x):
+            ret = self.layers[index](x) * self.z3
+            return ret
+
+    index = Tensor(0, dtype=mstype.int32)
+    net = SwitchLayerCell()
+    net(index, Tensor(np.full([128, 96], 0.6, dtype=np.float32)))
+    C.grad_by_list(net, ParameterTuple(net.trainable_params()))(index,
+                                                                Tensor(np.full([128, 96], 0.6, dtype=np.float32)))
+    C.grad_all(net)(index, Tensor(np.full([128, 96], 0.6, dtype=np.float32)))
+
+
 def test_control_depend_check():
     with pytest.raises(TypeError) as e:
         P.ControlDepend(0.0)