simplify transpose matmul reduce

mindspore/ccsrc/backend/optimizer/graph_kernel/arithmetic_simplify.cc

@@ -14,18 +14,18 @@
  * limitations under the License.
  */
 #include "backend/optimizer/graph_kernel/arithmetic_simplify.h"
+
 #include <list>
 #include "backend/optimizer/graph_kernel/graph_kernel_helper.h"
 #include "backend/kernel_compiler/common_utils.h"
 #include "backend/session/anf_runtime_algorithm.h"
-#include "ir/pattern_matcher.h"
 #include "frontend/operator/ops.h"
+#include "ir/pattern_matcher.h"
 #include "utils/convert_utils.h"
 #include "utils/utils.h"
 
 namespace mindspore {
 namespace opt {
-
 AnfNodePtr NewCNodeWithInfo(const AnfNodePtrList &inputs, const AnfNodePtr &ori_node) {
   auto func_graph = ori_node->func_graph();
   MS_EXCEPTION_IF_NULL(func_graph);
@@ -401,10 +401,236 @@ AnfNodePtr SimplifyDiv(const AnfNodePtr &node) {
     (FLAG) = true;                                     \
   }
 
+bool TryTransposeToReshape(const AnfNodePtr &node) {
+  auto perm = AnfAlgo::GetNodeAttr<std::vector<int>>(node, "perm");
+  auto ori_shape = AnfAlgo::GetPrevNodeOutputInferShape(node, 0);
+  std::vector<int> remove_one_perm;
+  for (auto idx : perm) {
+    if (idx < 0 || IntToSize(idx) >= ori_shape.size()) {
+      MS_EXCEPTION(ValueError);
+      return false;
+    }
+    if (ori_shape[idx] != 1) {
+      remove_one_perm.emplace_back(idx);
+    }
+  }
+  if (remove_one_perm.size() < 2) {
+    return true;
+  }
+  for (size_t idx = 1; idx < remove_one_perm.size(); idx++) {
+    if (remove_one_perm[idx] < remove_one_perm[idx - 1]) {
+      return false;
+    }
+  }
+  return true;
+}
+
+AnfNodePtr SimplifyTranspose(const AnfNodePtr &node) {
+  if (!IsPrimitiveCNode(node, prim::kPrimTranspose)) {
+    return nullptr;
+  }
+  if (TryTransposeToReshape(node)) {
+    auto new_cnode = NewCNodeWithInfo({NewValueNode(prim::kPrimReshape), node->cast<CNodePtr>()->input(1)}, node);
+    return new_cnode;
+  }
+  return nullptr;
+}
+
+AnfNodePtr SimplifyMatMul(const AnfNodePtr &node) {
+  if (!IsPrimitiveCNode(node, prim::kPrimMatMul)) {
+    return nullptr;
+  }
+  PatternNode<AnfNodePtr> x, y;
+  auto matmul_transpose_lambda = [&node, &x, &y]() -> AnfNodePtr {
+    auto new_matmul = NewCNodeWithInfo({NewValueNode(prim::kPrimMatMul), y.GetNode(node), x.GetNode(node)}, node);
+    auto new_abstract = node->abstract()->Clone();
+    auto ori_shape = node->abstract()->GetShapeTrack()->cast<abstract::ShapePtr>();
+    auto shape_value = ori_shape->shape();
+    ShapeVector new_shape_value;
+    std::copy(shape_value.rbegin(), shape_value.rend(), std::back_inserter(new_shape_value));
+    auto new_shape = std::make_shared<abstract::Shape>(new_shape_value);
+    new_abstract->set_shape(new_shape);
+    new_matmul->set_abstract(new_abstract);
+    auto new_cnode = NewCNodeWithInfo({NewValueNode(prim::kPrimTranspose), new_matmul}, node);
+    auto transpose_a = AnfAlgo::GetNodeAttr<ValuePtr>(node, "transpose_a");
+    auto transpose_b = AnfAlgo::GetNodeAttr<ValuePtr>(node, "transpose_b");
+    auto transpose_x1 = AnfAlgo::GetNodeAttr<ValuePtr>(node, "transpose_x1");
+    auto transpose_x2 = AnfAlgo::GetNodeAttr<ValuePtr>(node, "transpose_x2");
+    auto perm = AnfAlgo::GetNodeAttr<ValuePtr>(node->cast<CNodePtr>()->input(1), "perm");
+    AnfAlgo::SetNodeAttr("transpose_a", transpose_b, new_matmul);
+    AnfAlgo::SetNodeAttr("transpose_b", transpose_a, new_matmul);
+    AnfAlgo::SetNodeAttr("transpose_x1", transpose_x2, new_matmul);
+    AnfAlgo::SetNodeAttr("transpose_x2", transpose_x1, new_matmul);
+    AnfAlgo::SetNodeAttr("perm", perm, new_cnode);
+    return new_cnode;
+  };
+  // MatMul(Transpose(x), Transpose(y)) ==> Transpose(MatMul(y, x))
+  MATCH_REPLACE_LAMBDA(node,
+                       PBinOperation(prim::kPrimMatMul, PUnaryOperation(prim::kPrimTranspose, x),
+                                     PUnaryOperation(prim::kPrimTranspose, y), false),
+                       matmul_transpose_lambda);
+  return nullptr;
+}
+
+ShapeVector TransAxisValueToVector(const ValuePtr &value) {
+  MS_EXCEPTION_IF_NULL(value);
+  ShapeVector axis_vector;
+  if (value->isa<Int32Imm>()) {
+    axis_vector.emplace_back(GetValue<int>(value));
+  }
+  if (value->isa<ValueTuple>() || value->isa<ValueList>()) {
+    axis_vector = GetValue<std::vector<int>>(value);
+  }
+  return axis_vector;
+}
+
+ShapeVector GetNodeShape(const AnfNodePtr &node) {
+  auto base_shape = node->Shape()->cast<abstract::ShapePtr>();
+  std::vector<int> shape;
+  std::transform(base_shape->shape().begin(), base_shape->shape().end(), std::back_inserter(shape), IntToSize);
+  return shape;
+}
+
+std::vector<std::pair<int, int>> GetUnmodifiedDim(const ShapeVector &a, const ShapeVector &b) {
+  std::vector<std::pair<int, int>> unmodified;
+  for (size_t i = 0, j = 0, patial_a = 1, patial_b = 1;;) {
+    if (i >= a.size() && j >= b.size()) {
+      break;
+    }
+    patial_a *= a[i];
+    patial_b *= b[j];
+    if (patial_a == patial_b && a[i] == b[j]) {
+      unmodified.emplace_back(std::make_pair(i, j));
+      ++i;
+      ++j;
+      continue;
+    }
+    if (patial_a < patial_b && b[j] > a[i]) {
+      ++i;
+      patial_a *= a[i];
+      if (patial_a == patial_b) {
+        ++i;
+        ++j;
+      }
+      continue;
+    }
+    if (patial_a > patial_b && b[j] < a[i]) {
+      ++j;
+      patial_b *= b[j];
+      if (patial_a == patial_b) {
+        ++i;
+        ++j;
+      }
+      continue;
+    }
+  }
+  return unmodified;
+}
+
+AnfNodePtr SimplifyReduce(const AnfNodePtr &node) {
+  if (!IsPrimitiveCNode(node, prim::kPrimReduceMax) && !IsPrimitiveCNode(node, prim::kPrimReduceMin) &&
+      !IsPrimitiveCNode(node, prim::kPrimReduceSum)) {
+    return nullptr;
+  }
+  PatternNode<AnfNodePtr> x;
+  auto trans_reduce_lamda = [&node, &x](PrimitivePtr &operation) -> AnfNodePtr {
+    auto shape = GetNodeShape(node);
+    if (shape.size() != 0 && shape.size() != 1) {
+      return node;
+    } else {
+      auto tmp_node = node->cast<CNodePtr>();
+      auto transpose_node = tmp_node->input(1);
+      auto transpose_dimensions = GetValue<std::vector<int>>(AnfAlgo::GetNodeAttr<ValuePtr>(transpose_node, "perm"));
+      ShapeVector new_dimensions;
+      auto reduce_dimensions = TransAxisValueToVector(AnfAlgo::GetNodeAttr<ValuePtr>(tmp_node, "axis"));
+      std::transform(reduce_dimensions.begin(), reduce_dimensions.end(), std::back_inserter(new_dimensions),
+                     [&transpose_dimensions](const int &dim) { return transpose_dimensions[dim]; });
+      std::sort(new_dimensions.begin(), new_dimensions.end());
+      auto new_cnode = NewCNodeWithInfo({NewValueNode(operation), x.GetNode(node)}, node);
+      AnfAlgo::SetNodeAttr("axis", MakeValue(new_dimensions), new_cnode);
+      AnfAlgo::CopyNodeAttr("keep_dims", node, new_cnode);
+      return new_cnode;
+    }
+  };
+  auto reduce_reduce_lamda = [&node, &x](PrimitivePtr &operation) -> AnfNodePtr {
+    auto tmp_node = node->cast<CNodePtr>();
+    auto arg_node = tmp_node->input(1);
+    auto arg_dimensions = TransAxisValueToVector(AnfAlgo::GetNodeAttr<ValuePtr>(arg_node, "axis"));
+    auto reduce_dimensions = TransAxisValueToVector(AnfAlgo::GetNodeAttr<ValuePtr>(tmp_node, "axis"));
+    ShapeVector new_dimensions;
+    for (size_t i = 0; i < arg_dimensions.size(); ++i) {
+      for (size_t j = 0; j < reduce_dimensions.size(); ++j) {
+        if (reduce_dimensions[j] >= arg_dimensions[i]) {
+          ++reduce_dimensions[j];
+        }
+      }
+    }
+    std::merge(arg_dimensions.begin(), arg_dimensions.end(), reduce_dimensions.begin(), reduce_dimensions.end(),
+               std::back_inserter(new_dimensions));
+    auto new_cnode = NewCNodeWithInfo({NewValueNode(operation), x.GetNode(node)}, node);
+    AnfAlgo::SetNodeAttr("axis", MakeValue(new_dimensions), new_cnode);
+    AnfAlgo::CopyNodeAttr("keep_dims", node, new_cnode);
+    return new_cnode;
+  };
+  auto reshape_reduce_lamda = [&node, &x](PrimitivePtr &operation) -> AnfNodePtr {
+    auto tmp_node = node->cast<CNodePtr>();
+    auto arg_node = tmp_node->input(1);
+    auto input_shape = GetNodeShape(arg_node->cast<CNodePtr>()->input(1));
+    auto re_shape = GetNodeShape(arg_node);
+    auto reduce_dimensions = TransAxisValueToVector(AnfAlgo::GetNodeAttr<ValuePtr>(tmp_node, "axis"));
+    auto unmodified_dim_pair = GetUnmodifiedDim(input_shape, re_shape);
+    std::vector<bool> dim_in_output(re_shape.size(), true);
+    std::vector<bool> dim_unmodified(re_shape.size(), false);
+    for (auto dim : reduce_dimensions) {
+      dim_in_output[dim] = false;
+    }
+    for (auto pair_dim : unmodified_dim_pair) {
+      dim_unmodified[pair_dim.second] = true;
+    }
+    bool replace = true;
+    for (size_t i = 0; i < dim_in_output.size(); ++i) {
+      if (dim_in_output[i] && !dim_unmodified[i]) {
+        replace = false;
+      }
+    }
+    if (replace) {
+      ShapeVector un_dimensions;
+      for (auto pair_dim : unmodified_dim_pair) {
+        if (dim_in_output[pair_dim.second]) {
+          un_dimensions.emplace_back(pair_dim.first);
+        }
+      }
+      ShapeVector new_dimensions;
+      for (size_t i = 0; i < input_shape.size(); ++i) {
+        if (std::find(un_dimensions.begin(), un_dimensions.end(), i) == un_dimensions.end()) {
+          new_dimensions.emplace_back(i);
+        }
+      }
+      auto new_cnode = NewCNodeWithInfo({NewValueNode(operation), x.GetNode(node)}, node);
+      AnfAlgo::SetNodeAttr("axis", MakeValue(new_dimensions), new_cnode);
+      AnfAlgo::CopyNodeAttr("keep_dims", node, new_cnode);
+      return new_cnode;
+    }
+    return node;
+  };
+  std::list<PrimitivePtr> ReduceOperations = {prim::kPrimReduceSum, prim::kPrimReduceMax, prim::kPrimReduceMin};
+  for (auto operation : ReduceOperations) {
+    // Reduce(Transpose(A))  = Reduce(A) if result is a scalar or vector
+    MATCH_REPLACE_LAMBDA_FLAG(node, PPrimitive(operation, PPrimitive(prim::kPrimTranspose, x)), trans_reduce_lamda,
+                              operation);
+    // Reduce(Reduce(A)) = Reduce(A)
+    MATCH_REPLACE_LAMBDA_FLAG(node, PPrimitive(operation, PPrimitive(operation, x)), reduce_reduce_lamda, operation);
+    // Reduce(Reshape(A)) = Reduce(A) if reduce dimensions is not in reshape dimensions
+    MATCH_REPLACE_LAMBDA_FLAG(node, PPrimitive(operation, PPrimitive(prim::kPrimReshape, x)), reshape_reduce_lamda,
+                              operation);
+  }
+  return nullptr;
+}
+
 AnfNodePtr TrySimplify(const AnfNodePtr &node) {
   std::list<std::function<AnfNodePtr(AnfNodePtr)>> SimplifyFuncList = {
-    SimplifyAdd, SimplifyDiv,   SimplifyLog,    SimplifyMul,  SimplifyNeg,
-    SimplifyPow, SimplifyRsqrt, SimplifySelect, SimplifySqrt, SimplifySub};
+    SimplifyAdd,    SimplifyDiv,  SimplifyLog, SimplifyMul,       SimplifyNeg,    SimplifyPow,   SimplifyRsqrt,
+    SimplifySelect, SimplifySqrt, SimplifySub, SimplifyTranspose, SimplifyMatMul, SimplifyReduce};
   for (auto f : SimplifyFuncList) {
     auto ret = f(node);
     if (ret != nullptr) {

mindspore/core/ir/pattern_matcher.h

@@ -22,8 +22,8 @@
 #include <tuple>
 #include <vector>
 
-#include "ir/visitor.h"
 #include "base/core_ops.h"
+#include "ir/visitor.h"
 #include "utils/shape_utils.h"
 
 namespace mindspore {
@@ -750,9 +750,18 @@ class PConstant : public PBase<PConstant<T> > {
     if (value->isa<tensor::Tensor>()) {
       tensor::TensorPtr tensor_ptr = dyn_cast<tensor::Tensor>(value);
       TypeId tensor_type = tensor_ptr->Dtype()->type_id();
+      auto tensor_abstract = node->abstract()->cast<abstract::AbstractTensorPtr>();
+      TypePtr tensor_type_ptr = tensor_abstract->element()->BuildType();
+      ShapeVector tensor_shape = tensor_abstract->shape()->shape();
+      auto new_tensor_ptr = std::make_shared<tensor::Tensor>(tensor_type_ptr->type_id(), tensor_shape);
+      size_t mem_size = GetTypeByte(tensor_type_ptr) * IntToSize(new_tensor_ptr->ElementsNum());
       if ((tensor_type == TypeId::kNumberTypeFloat32) || (tensor_type == TypeId::kNumberTypeFloat) ||
           (tensor_type == TypeId::kNumberTypeFloat64)) {
-        float *data2 = reinterpret_cast<float *>(tensor_ptr->data_c());
+        float *data = reinterpret_cast<float *>(tensor_ptr->data_c());
+        float *data2 = reinterpret_cast<float *>(new_tensor_ptr->data_c());
+        if (memcpy_s(data2, mem_size, data, mem_size) != 0) {
+          return nullptr;
+        }
         for (int i = 0; i < tensor_ptr->DataSize(); i++) {
           if (data2[i] == 0 && calcu_type == prim::kPrimReciprocal) {
             return nullptr;
@@ -761,7 +770,11 @@ class PConstant : public PBase<PConstant<T> > {
         }
       }
       if ((tensor_type == TypeId::kNumberTypeInt32) || (tensor_type == TypeId::kNumberTypeInt)) {
-        int *data2 = reinterpret_cast<int *>(tensor_ptr->data_c());
+        int *data = reinterpret_cast<int *>(tensor_ptr->data_c());
+        int *data2 = reinterpret_cast<int *>(new_tensor_ptr->data_c());
+        if (memcpy_s(data2, mem_size, data, mem_size) != 0) {
+          return nullptr;
+        }
         for (int i = 0; i < tensor_ptr->DataSize(); i++) {
           if (data2[i] == 0 && calcu_type == prim::kPrimReciprocal) {
             return nullptr;
@@ -770,7 +783,11 @@ class PConstant : public PBase<PConstant<T> > {
         }
       }
       if (tensor_type == TypeId::kNumberTypeFloat64) {
-        double *data2 = reinterpret_cast<double *>(tensor_ptr->data_c());
+        double *data = reinterpret_cast<double *>(tensor_ptr->data_c());
+        double *data2 = reinterpret_cast<double *>(new_tensor_ptr->data_c());
+        if (memcpy_s(data2, mem_size, data, mem_size) != 0) {
+          return nullptr;
+        }
         for (int i = 0; i < tensor_ptr->DataSize(); i++) {
           if (data2[i] == 0 && calcu_type == prim::kPrimReciprocal) {
             return nullptr;
@@ -778,7 +795,9 @@ class PConstant : public PBase<PConstant<T> > {
           data2[i] = CalcuConstant(data2[i], calcu_type);
         }
       }
-      return node;
+      auto new_vnode = NewValueNode(new_tensor_ptr);
+      new_vnode->set_abstract(tensor_ptr->ToAbstract());
+      return new_vnode;
     }
     return nullptr;
   }
@@ -1005,6 +1024,14 @@ BIN_OPERATION_PATTERN(operator-, prim::kPrimSub, false);
       return rep;                                                         \
     }                                                                     \
   }
+
+#define MATCH_REPLACE_LAMBDA_FLAG(OrigNode, CaptureNode, Lambda, Flag) \
+  if ((CaptureNode).TryCapture(OrigNode)) {                            \
+    auto rep = (Lambda)(Flag);                                         \
+    if (rep != nullptr) {                                              \
+      return rep;                                                      \
+    }                                                                  \
+  }
 }  // namespace mindspore
 
 #endif  // MINDSPORE_CORE_IR_PATTERN_MATCHER_H_

tests/st/ops/graph_kernel/test_simplify.py

@@ -20,7 +20,8 @@ from mindspore import Tensor
 from mindspore.nn import Cell
 import mindspore.ops.operations as P
 
-context.set_context(mode=context.GRAPH_MODE, enable_graph_kernel=True, device_target="GPU")
+context.set_context(mode=context.GRAPH_MODE,
+                    enable_graph_kernel=True, device_target="GPU")
 
 
 class Net(Cell):
@@ -33,6 +34,8 @@ class Net(Cell):
         self.sqrt = P.Sqrt()
         self.pow = P.Pow()
         self.neg = P.Neg()
+        self.reducemin = P.ReduceMin()
+        self.reshape = P.Reshape()
 
     def construct(self, x, y):
         add_res1 = self.add(x, 4)
@@ -42,7 +45,9 @@ class Net(Cell):
         div_res = self.div(mul_res, self.sqrt(mul_res))
         pow_res = self.pow(y, 2)
         neg_res = self.neg(self.neg(pow_res))
-        return self.add(div_res, neg_res)
+        add_res3 = self.add(neg_res, div_res)
+        resh_res = self.reshape(add_res3, (2, 12, 3))
+        return self.reducemin(resh_res, 1)
 
 
 @pytest.mark.level0
@@ -58,10 +63,12 @@ def test_basic():
     div_res = np.sqrt(mul_res)
     pow_res = input_y * input_y
     neg_res = pow_res
-    expect = div_res + neg_res
+    add_res3 = neg_res + div_res
+    expect = np.min(add_res3, (1, 2))
 
     net = Net()
     result = net(Tensor(input_x), Tensor(input_y))
 
-    res = np.allclose(expect, result.asnumpy(), rtol=1.e-4, atol=1.e-7, equal_nan=True)
+    res = np.allclose(expect, result.asnumpy(), rtol=1.e-4,
+                      atol=1.e-7, equal_nan=True)
     assert res