!1163 Add BatchNorm fusion pattern with mix precision

Merge pull request !1163 from YuJianfeng/bn_fusion_with_cast

mindspore/ccsrc/pre_activate/ascend/ascend_backend_optimization.cc

@@ -201,6 +201,7 @@ void AscendBackendIRFusionOptimization(const std::shared_ptr<session::KernelGrap
   } else {
     ir_fusion_pm->AddPass(std::make_shared<BatchNormGradSplit>());
     ir_fusion_pm->AddPass(std::make_shared<FusedBatchNormFusion>());
+    ir_fusion_pm->AddPass(std::make_shared<FusedBatchNormMixPrecisionFusion>());
   }
   ir_fusion_pm->AddPass(std::make_shared<AddMemcpyAsync>());
   if (context_ptr->ir_fusion_flag()) {

mindspore/ccsrc/pre_activate/ascend/ir_fusion/fused_batch_norm_fusion.cc

@@ -277,5 +277,28 @@ const AnfNodePtr FusedBatchNormFusion::Process(const FuncGraphPtr &func_graph, c
   }
   return bn_training_update_outputs[0];
 }
+
+const BaseRef FusedBatchNormMixPrecisionFusion::DefinePattern() const {
+  std::shared_ptr<Var> Xs = std::make_shared<SeqVar>();
+  VarPtr index0 = std::make_shared<CondVar>(IsC);
+  VarPtr index1 = std::make_shared<CondVar>(IsC);
+  VarPtr index2 = std::make_shared<CondVar>(IsC);
+  VectorRef batch_norm = VectorRef({batch_norm_var_, data_input0_var_, data_input1_var_, data_input2_var_, Xs});
+  VectorRef tuple_getitem0 = VectorRef({prim::kPrimTupleGetItem, batch_norm, index0});
+  VectorRef tuple_getitem1 = VectorRef({prim::kPrimTupleGetItem, batch_norm, index1});
+  VectorRef tuple_getitem2 = VectorRef({prim::kPrimTupleGetItem, batch_norm, index2});
+  VectorRef cast_variable_input0 = VectorRef({prim::kPrimCast, variable_input0_var_});
+  VectorRef cast_variable_input1 = VectorRef({prim::kPrimCast, variable_input1_var_});
+  VectorRef sub0 = VectorRef({prim::kPrimSub, cast_variable_input0, tuple_getitem1});
+  VectorRef sub1 = VectorRef({prim::kPrimSub, cast_variable_input1, tuple_getitem2});
+  VectorRef mul0 = VectorRef({prim::kPrimMul, sub0, constant_input0_var_});
+  VectorRef mul1 = VectorRef({prim::kPrimMul, sub1, constant_input1_var_});
+  VectorRef cast2 = VectorRef({prim::kPrimCast, mul0});
+  VectorRef cast3 = VectorRef({prim::kPrimCast, mul1});
+  VectorRef assign_sub0 = VectorRef({prim::kPrimAssignSub, variable_input0_var_, cast2});
+  VectorRef assign_sub1 = VectorRef({prim::kPrimAssignSub, variable_input1_var_, cast3});
+  VectorRef depend0 = VectorRef({prim::kPrimDepend, tuple_getitem0, assign_sub0});
+  return VectorRef({prim::kPrimDepend, depend0, assign_sub1});
+}
 }  // namespace opt
 }  // namespace mindspore

mindspore/ccsrc/pre_activate/ascend/ir_fusion/fused_batch_norm_fusion.h

@@ -18,6 +18,7 @@
 
 #include <vector>
 #include <memory>
+#include <string>
 #include "pre_activate/common/optimizer.h"
 #include "utils/utils.h"
 
@@ -25,8 +26,8 @@ namespace mindspore {
 namespace opt {
 class FusedBatchNormFusion : public PatternProcessPass {
  public:
-  explicit FusedBatchNormFusion(bool multigraph = true)
-      : PatternProcessPass("fused_batch_norm_fusion", multigraph),
+  explicit FusedBatchNormFusion(const std::string &name = "fused_batch_norm_fusion", bool multigraph = true)
+      : PatternProcessPass(name, multigraph),
         data_input0_var_(std::make_shared<Var>()),
         data_input1_var_(std::make_shared<Var>()),
         data_input2_var_(std::make_shared<Var>()),
@@ -39,7 +40,7 @@ class FusedBatchNormFusion : public PatternProcessPass {
   const BaseRef DefinePattern() const override;
   const AnfNodePtr Process(const FuncGraphPtr &, const AnfNodePtr &, const EquivPtr &) const override;
 
- private:
+ protected:
   AnfNodePtr CreateBNTrainingReduce(const FuncGraphPtr &func_graph, const AnfNodePtr &node,
                                     const EquivPtr &equiv) const;
   void GetBNTrainingUpdateInputs(const EquivPtr &equiv, const std::vector<AnfNodePtr> &bn_training_reduce_outputs,
@@ -59,6 +60,15 @@ class FusedBatchNormFusion : public PatternProcessPass {
   VarPtr constant_input1_var_;
   VarPtr batch_norm_var_;
 };
+
+class FusedBatchNormMixPrecisionFusion : public FusedBatchNormFusion {
+ public:
+  explicit FusedBatchNormMixPrecisionFusion(bool multigraph = true)
+      : FusedBatchNormFusion("fused_batch_norm_mix_precision_fusion", multigraph) {}
+
+  ~FusedBatchNormMixPrecisionFusion() override = default;
+  const BaseRef DefinePattern() const override;
+};
 }  // namespace opt
 }  // namespace mindspore
 #endif  // MINDSPORE_CCSRC_PRE_ACTIVATE_ASCEND_IR_FUSION_FUSED_BATCH_NORM_FUSION_H_

tests/ut/cpp/pre_activate/ascend/ir_fusion/fused_batch_norm_fusion_test.cc

@@ -50,5 +50,28 @@ TEST_F(TestHWFusedBatchNormFusion, test_fused_batch_norm_fusion) {
   FuncGraphPtr g_after = get_py_fun_.CallAndParseRet("test_fused_batch_norm_fusion", "after");
   EXPECT_TRUE(CheckEqualGraph(g_after, new_graph));
 }
+
+TEST_F(TestHWFusedBatchNormFusion, test_fused_batch_norm_mix_precision_fusion) {
+  FuncGraphPtr g = get_py_fun_.CallAndParseRet("test_fused_batch_norm_fusion", "before_mix_precision");
+  EXPECT_NE(g, nullptr);
+  std::vector<int> shp_x{32, 64, 112, 112};
+  auto x_abstract = std::make_shared<abstract::AbstractTensor>(kFloat32, shp_x);
+  std::vector<int> shp_y{64};
+  auto y_abstract = std::make_shared<abstract::AbstractTensor>(kFloat32, shp_y);
+  AbstractBasePtrList args_spec_list{x_abstract};
+  for (size_t i = 0; i < 6; ++i) {
+    args_spec_list.push_back(y_abstract);
+  }
+  auto kg = GetKernelGraph(g, args_spec_list);
+
+  auto optimizer = std::make_shared<opt::GraphOptimizer>();
+  auto pm = std::make_shared<opt::PassManager>();
+  pm->AddPass(std::make_shared<opt::FusedBatchNormMixPrecisionFusion>());
+  optimizer->AddPassManager(pm);
+  FuncGraphPtr new_graph = optimizer->Optimize(kg);
+
+  FuncGraphPtr g_after = get_py_fun_.CallAndParseRet("test_fused_batch_norm_fusion", "after");
+  EXPECT_TRUE(CheckEqualGraph(g_after, new_graph));
+}
 }  // namespace opt
 }  // namespace mindspore

tests/ut/cpp/python_input/gtest_input/pre_activate/fused_batch_norm_fusion_test.py

@@ -24,6 +24,7 @@ make_tuple = Primitive('make_tuple')
 tuple_getitem = Primitive('tuple_getitem')
 depend = Primitive('depend')
 BatchNorm = P.BatchNorm()
+Cast = P.Cast()
 BNTrainingReduce = Primitive('BNTrainingReduce')
 BNTrainingUpdate = Primitive('BNTrainingUpdate')
 constant0 = Tensor(0.1, mstype.float32)
@@ -59,6 +60,21 @@ def test_fused_batch_norm_fusion(tag):
         output = tuple_getitem(outputs, 0)
         return output
 
+    @fns
+    def before_mix_precision(input0, input1, input2, input3, input4, var0, var1):
+        batch_norm = BatchNorm(input0, input1, input2, input3, input4)
+        sub0 = Sub(Cast(var0, mstype.float32), tuple_getitem(batch_norm, 1))
+        sub1 = Sub(Cast(var1, mstype.float32), tuple_getitem(batch_norm, 2))
+        mul0 = Mul(sub0, constant0)
+        mul1 = Mul(sub1, constant1)
+        assign_sub0 = AssignSub(var0, Cast(mul0, mstype.float32))
+        assign_sub1 = AssignSub(var1, Cast(mul1, mstype.float32))
+        depend0 = depend(tuple_getitem(batch_norm, 0), assign_sub0)
+        depend1 = depend(depend0, assign_sub1)
+        outputs = make_tuple(depend1, tuple_getitem(batch_norm, 3), tuple_getitem(batch_norm, 4))
+        output = tuple_getitem(outputs, 0)
+        return output
+
     @fns
     def after(input0, input1, input2, input3, input4, var0, var1):
         bn_training_reduce = BNTrainingReduce(input0)