!993 [IRFusion pass enhancement] mul_add_fusion pass supports when add's 2nd is mul

Merge pull request !993 from huanghui/mul-add-fusion-pass-enhancement

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

@@ -24,40 +24,57 @@
 #include "pre_activate/common/helper.h"
 
 namespace mindspore {
-namespace opt {
-const BaseRef MulAddFusion::DefinePattern() const {
-  VarPtr mul_x_ = std::make_shared<Var>();
-  VarPtr mul_y_ = std::make_shared<Var>();
-  VarPtr add_y_ = std::make_shared<Var>();
+bool GetMul(const FuncGraphPtr &graph, const CNodePtr &add, CNodePtr *mul, size_t *mul_index) {
+  MS_EXCEPTION_IF_NULL(graph);
+  MS_EXCEPTION_IF_NULL(add);
 
-  VectorRef mul({prim::kPrimMul, mul_x_, mul_y_});
-  VectorRef add({prim::kPrimTensorAdd, mul, add_y_});
-  return add;
+  for (size_t index = 1; index < add->size(); ++index) {
+    auto input = add->input(index);
+    MS_EXCEPTION_IF_NULL(input);
+    if (input->isa<CNode>()) {
+      auto cnode = input->cast<CNodePtr>();
+      MS_EXCEPTION_IF_NULL(cnode);
+      if (AnfAlgo::GetCNodeName(cnode) == prim::kPrimMul->name()) {
+        if (!opt::IsUsedByOthers(graph, cnode)) {
+          *mul = cnode;
+          *mul_index = index;
+          return true;
+        }
+      }
+    }
+  }
+  return false;
 }
 
-const AnfNodePtr MulAddFusion::Process(const FuncGraphPtr &graph, const AnfNodePtr &node, const EquivPtr &equiv) const {
-  if (graph == nullptr || node == nullptr || equiv == nullptr) {
+namespace opt {
+const BaseRef MulAddFusion::DefinePattern() const {
+  VarPtr x = std::make_shared<Var>();
+  VarPtr y = std::make_shared<Var>();
+  VectorRef pattern({prim::kPrimTensorAdd, x, y});
+  return pattern;
+}
+
+const AnfNodePtr MulAddFusion::Process(const FuncGraphPtr &graph, const AnfNodePtr &node, const EquivPtr &) const {
+  if (graph == nullptr || node == nullptr) {
     return nullptr;
   }
   auto add = node->cast<CNodePtr>();
   if (add == nullptr || add->inputs().size() != kAddInputNum) {
     return nullptr;
   }
-  auto mul_anf = add->input(1);
-  if (mul_anf == nullptr) {
-    return nullptr;
-  }
-  auto mul = mul_anf->cast<CNodePtr>();
-  if (mul == nullptr || mul->inputs().size() != kMulInputNum) {
-    return nullptr;
-  }
-  if (IsUsedByOthers(graph, mul)) {
-    MS_LOG(DEBUG) << "Mul is used by more then two nodes, cannot fuse";
+  CNodePtr mul = nullptr;
+  size_t mul_index = 0;
+  if (!GetMul(graph, add, &mul, &mul_index) || mul == nullptr || mul_index == 0) {
+    MS_LOG(DEBUG) << "Cannot find used-by-only-one-op Mul in Add's inputs";
     return nullptr;
   }
 
   auto prim = std::make_shared<Primitive>(kFusedMulAddOpName);
-  std::vector<AnfNodePtr> inputs = {NewValueNode(prim), mul->input(1), mul->input(2), add->input(2)};
+  std::vector<AnfNodePtr> inputs = {NewValueNode(prim)};
+  for (size_t index = 1; index < mul->size(); ++index) {
+    inputs.push_back(mul->input(index));
+  }
+  inputs.push_back(add->input(add->size() - mul_index));
   auto fusion_node = graph->NewCNode(inputs);
   fusion_node->set_scope(add->scope());
   fusion_node->set_abstract(add->abstract());

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

@@ -28,8 +28,28 @@ class TestHWMulAddFusion : public BackendCommon {
   UT::PyFuncGraphFetcher get_py_fun_;
 };
 
-TEST_F(TestHWMulAddFusion, test_mul_add_fusion) {
-  FuncGraphPtr g = get_py_fun_.CallAndParseRet("test_mul_add_fusion", "before");
+TEST_F(TestHWMulAddFusion, test_mul_add_fusion1) {
+  FuncGraphPtr g = get_py_fun_.CallAndParseRet("test_mul_add_fusion", "before1");
+  std::vector<int> shp{2, 32, 224, 224};
+  auto x_abstract = std::make_shared<abstract::AbstractTensor>(kFloat32, shp);
+  AbstractBasePtrList args_spec_list;
+  for (size_t i = 0; i < 3; ++i) {
+    args_spec_list.push_back(x_abstract);
+  }
+  auto fg = 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::MulAddFusion>());
+  optimizer->AddPassManager(pm);
+  FuncGraphPtr new_graph = optimizer->Optimize(fg);
+
+  FuncGraphPtr g_after = get_py_fun_.CallAndParseRet("test_mul_add_fusion", "after");
+  EXPECT_TRUE(CheckEqualGraph(g_after, new_graph));
+}
+
+TEST_F(TestHWMulAddFusion, test_mul_add_fusion2) {
+  FuncGraphPtr g = get_py_fun_.CallAndParseRet("test_mul_add_fusion", "before2");
   std::vector<int> shp{2, 32, 224, 224};
   auto x_abstract = std::make_shared<abstract::AbstractTensor>(kFloat32, shp);
   AbstractBasePtrList args_spec_list;

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

@@ -21,7 +21,6 @@ fused_mul_add = Primitive('FusedMulAdd')
 make_tuple = Primitive('make_tuple')
 tuple_getitem = Primitive('tuple_getitem')
 
-
 class FnDict:
     def __init__(self):
         self.fnDict = {}
@@ -32,16 +31,21 @@ class FnDict:
     def __getitem__(self, name):
         return self.fnDict[name]
 
-
 def test_mul_add_fusion(tag):
     fns = FnDict()
 
     @fns
-    def before(x, y, z):
+    def before1(x, y, z):
         res = mul(x, y)
         res = add(res, z)
         return res
 
+    @fns
+    def before2(x, y, z):
+        res = mul(x, y)
+        res = add(z, res)
+        return res
+
     @fns
     def after(x, y, z):
         res = fused_mul_add(x, y, z)