!7925 fix gpu momentum fusion

Merge pull request !7925 from chenweifeng/momentum-fusion-fix

mindspore/ccsrc/backend/optimizer/gpu/apply_momentum_scale_fusion.cc

@@ -26,6 +26,28 @@
 
 namespace mindspore {
 namespace opt {
+bool ApplyMomentumScaleFusion::IsScalar(const BaseRef &n) {
+  if (utils::isa<AnfNodePtr>(n)) {
+    AnfNodePtr in = utils::cast<AnfNodePtr>(n);
+    MS_EXCEPTION_IF_NULL(in);
+    auto shape = in->Shape()->cast<abstract::ShapePtr>();
+    MS_EXCEPTION_IF_NULL(shape);
+    if (shape->shape().size() != 0) {
+      return false;
+    }
+    auto dtype = in->Type();
+    if (dtype->type_id() != kObjectTypeTensorType) {
+      return false;
+    }
+    auto element_type = dyn_cast<TensorType>(dtype)->element()->type_id();
+    if (element_type != kNumberTypeFloat32) {
+      return false;
+    }
+    return true;
+  }
+  return false;
+}
+
 const BaseRef ApplyMomentumScaleFusion::DefinePattern() const {
   VectorRef scale = VectorRef({prim::kPrimMul, gradient_, scale_});
   VectorRef apply_momentum =

mindspore/ccsrc/backend/optimizer/gpu/apply_momentum_scale_fusion.h

@@ -18,13 +18,14 @@
 
 #include <memory>
 #include "backend/optimizer/common/optimizer.h"
+#include "backend/session/anf_runtime_algorithm.h"
 
 namespace mindspore {
 namespace opt {
 class ApplyMomentumScaleFusion : public PatternProcessPass {
  public:
   explicit ApplyMomentumScaleFusion(bool multigraph = true) : PatternProcessPass("momentum_scale_fusion", multigraph) {
-    scale_ = std::make_shared<Var>();
+    scale_ = std::make_shared<CondVar>(IsScalar);
     variable_ = std::make_shared<Var>();
     accumulation_ = std::make_shared<Var>();
     learning_rate_ = std::make_shared<Var>();
@@ -36,6 +37,8 @@ class ApplyMomentumScaleFusion : public PatternProcessPass {
   const AnfNodePtr Process(const FuncGraphPtr &, const AnfNodePtr &, const EquivPtr &) const override;
 
  private:
+  static bool IsScalar(const BaseRef &n);
+
   VarPtr scale_;
   VarPtr variable_;
   VarPtr accumulation_;

mindspore/ccsrc/backend/optimizer/gpu/apply_momentum_weight_scale_fusion.cc

@@ -26,6 +26,28 @@
 
 namespace mindspore {
 namespace opt {
+bool ApplyMomentumWeightDecayScaleFusion::IsScalar(const BaseRef &n) {
+  if (utils::isa<AnfNodePtr>(n)) {
+    AnfNodePtr in = utils::cast<AnfNodePtr>(n);
+    MS_EXCEPTION_IF_NULL(in);
+    auto shape = in->Shape()->cast<abstract::ShapePtr>();
+    MS_EXCEPTION_IF_NULL(shape);
+    if (shape->shape().size() != 0) {
+      return false;
+    }
+    auto dtype = in->Type();
+    if (dtype->type_id() != kObjectTypeTensorType) {
+      return false;
+    }
+    auto element_type = dyn_cast<TensorType>(dtype)->element()->type_id();
+    if (element_type != kNumberTypeFloat32) {
+      return false;
+    }
+    return true;
+  }
+  return false;
+}
+
 const BaseRef ApplyMomentumWeightDecayScaleFusion::DefinePattern() const {
   VectorRef weight = VectorRef(
     {prim::kPrimAddN, VectorRef({prim::kPrimMul, variable_, weight_decay_}), VectorRef({prim::kPrimCast, gradient_})});

mindspore/ccsrc/backend/optimizer/gpu/apply_momentum_weight_scale_fusion.h

@@ -26,7 +26,7 @@ class ApplyMomentumWeightDecayScaleFusion : public PatternProcessPass {
   explicit ApplyMomentumWeightDecayScaleFusion(bool multigraph = true)
       : PatternProcessPass("momentum_weightdecay_scale_fusion", multigraph) {
     weight_decay_ = std::make_shared<Var>();
-    scale_ = std::make_shared<Var>();
+    scale_ = std::make_shared<CondVar>(IsScalar);
     variable_ = std::make_shared<Var>();
     accumulation_ = std::make_shared<Var>();
     learning_rate_ = std::make_shared<Var>();
@@ -38,9 +38,10 @@ class ApplyMomentumWeightDecayScaleFusion : public PatternProcessPass {
   const AnfNodePtr Process(const FuncGraphPtr &, const AnfNodePtr &, const EquivPtr &) const override;
 
  private:
+  static bool IsScalar(const BaseRef &n);
+
   VarPtr weight_decay_;
   VarPtr scale_;
-
   VarPtr variable_;
   VarPtr accumulation_;
   VarPtr learning_rate_;