add activation attr for gelu

mindspore/ccsrc/backend/kernel_compiler/cpu/nnacl/infer/slice_infer.c

@@ -34,7 +34,7 @@ static bool CheckInputsDataType(const TensorC *const *inputs, size_t inputs_size
 
 int SliceInferShape(const TensorC *const *inputs, size_t inputs_size, TensorC **outputs, size_t outputs_size,
                     OpParameter *parameter) {
-  int ret = CheckAugmentWithMinSize(inputs, inputs_size, outputs, outputs_size, parameter, 1, 1);
+  int ret = CheckAugmentWithMinSize(inputs, inputs_size, outputs, outputs_size, parameter, 3, 1);
   if (ret != NNACL_OK) {
     return ret;
   }

mindspore/core/ops/fusion/activation.cc

@@ -54,6 +54,14 @@ ActivationType Activation::get_activation_type() const {
   auto value_ptr = GetAttr(kActivationType);
   return ActivationType(GetValue<int64_t>(value_ptr));
 }
+
+void Activation::set_approximate(bool approximate) { this->AddAttr(kApproximate, MakeValue(approximate)); }
+
+bool Activation::get_approximate() const {
+  auto value_ptr = this->GetAttr(kApproximate);
+  return GetValue<bool>(value_ptr);
+}
+
 void Activation::Init(const float alpha, const float min_val, const float max_val,
                       const ActivationType &activation_type) {
   this->set_alpha(alpha);

mindspore/core/ops/fusion/activation.h

@@ -38,6 +38,8 @@ class Activation : public PrimitiveC {
   float get_min_val() const;
   float get_max_val() const;
   ActivationType get_activation_type() const;
+  void set_approximate(bool approximate);
+  bool get_approximate() const;
 };
 }  // namespace ops
 }  // namespace mindspore

mindspore/core/ops/op_utils.h

@@ -251,6 +251,7 @@ constexpr auto kNumberSplit = "number_split";
 constexpr auto kSplitDim = "split_dim";
 constexpr auto kPadTop = "pad_top";
 constexpr auto kTransFormat = "trans_format";
+constexpr auto kApproximate = "approximate";
 const std::set<TypePtr> common_valid_types = {kInt8,   kInt16,  kInt32,   kInt64,   kUInt8,  kUInt16,
                                               kUInt32, kUInt64, kFloat16, kFloat32, kFloat64};
 

mindspore/lite/schema/ops.fbs

@@ -229,6 +229,7 @@ table Activation {
     alpha: float;
     min_val: float;
     max_val: float;
+    approximate: bool = false;
 }
 
 table ActivationGrad {

mindspore/lite/src/ops/ops_def.cc

@@ -229,6 +229,7 @@ OP_ATTR_ENUM_WITH_VALUE(activation_type, ActivationType, 0)
 OP_ATTR(alpha, float)
 OP_ATTR(min_val, float)
 OP_ATTR(max_val, float)
+OP_ATTR_WITH_VALUE(approximate, bool, false)
 OP_SCHEMA_DEF_END(Activation)
 
 OP_SCHEMA_DEF(ActivationGrad)

mindspore/lite/tools/optimizer/fusion/gelu_fusion.cc

@@ -29,6 +29,7 @@ CNodePtr GeLUFusion::CreateGeLUNode(const FuncGraphPtr &func_graph, const AnfNod
   MS_ASSERT(node != nullptr);
   auto gelu_prim = std::make_shared<ops::Activation>();
   gelu_prim->set_activation_type(mindspore::GELU);
+  gelu_prim->set_approximate(approximate_);
   auto input_node = utils::cast<AnfNodePtr>((*equiv)[input_]);
   MS_ASSERT(input_node != nullptr);
   auto gelu_cnode = func_graph->NewCNode(gelu_prim, {input_node});

mindspore/lite/tools/optimizer/fusion/gelu_fusion.h

@@ -41,6 +41,9 @@ class GeLUFusion : public PatternProcessPass {
 
  private:
   CNodePtr CreateGeLUNode(const FuncGraphPtr &func_graph, const AnfNodePtr &node, const EquivPtr &equiv) const;
+
+ protected:
+  mutable bool approximate_{false};
 };
 }  // namespace opt
 }  // namespace mindspore

mindspore/lite/tools/optimizer/fusion/tf_gelu_fusion.cc

@@ -82,6 +82,7 @@ bool TfGeLUFusion::CheckPattern(const EquivPtr &equiv) const {
   if (mul3_x < 0 || fabs(mul3_x - MUL3_X) > DIFF_THRESHOLD) {
     return false;
   }
+  approximate_ = true;
   return true;
 }
 }  // namespace opt