!14499 [GraphKernel]split UMonad in inputs of op

From: @wenfangpei
Reviewed-by: @dayschan,@ckey_dou,@gaoxiong1
Signed-off-by: @gaoxiong1

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

@@ -26,6 +26,7 @@
 #include "backend/kernel_compiler/common_utils.h"
 #include "backend/kernel_compiler/kernel_build_info.h"
 #include "backend/optimizer/graph_kernel/graph_kernel_helper.h"
+#include "backend/optimizer/graph_kernel/split_umonad.h"
 #include "backend/optimizer/graph_kernel/substitute_dropout.h"
 #include "backend/session/anf_runtime_algorithm.h"
 #include "mindspore/core/ir/graph_utils.h"
@@ -37,10 +38,14 @@
 namespace mindspore {
 namespace opt {
 namespace {
+constexpr size_t kAssignInputIdx = 1;
+constexpr size_t kLambInputIdx = 12;
+
 std::vector<PrimitivePtr> GetExpandOps() {
   std::vector<PrimitivePtr> expand_ops = {
     prim::kPrimSquare,
     prim::kPrimGeLUGrad,
+    prim::kPrimAssignAdd,
 #if ENABLE_D
     prim::kPrimTile,
     prim::kPrimSqrtGrad,
@@ -69,7 +74,6 @@ std::vector<PrimitivePtr> GetExpandOps() {
     prim::kPrimSigmoidCrossEntropyWithLogits,
     prim::kPrimSigmoidCrossEntropyWithLogitsGrad,
     prim::kPrimSoftmaxCrossEntropyWithLogits,
-    prim::kPrimAssignAdd,
 #endif
   };
   const auto &flags = context::GraphKernelFlags::GetInstance();
@@ -167,6 +171,22 @@ AnfNodePtr DefaultExpander::Run(const AnfNodePtr &node) {
   return graph_kernel_node;
 }
 
+ExpanderPtr GraphKernelExpander::GetExpander(const AnfNodePtr &node) {
+  std::vector<std::pair<PrimitivePtr, ExpanderPtr>> expanders = {
+    {prim::kPrimDropout, std::make_shared<DropoutExpander>()},
+    {prim::kPrimAssignAdd, std::make_shared<OpUMonadExpander>(kAssignInputIdx)},
+    {prim::kPrimAssignSub, std::make_shared<OpUMonadExpander>(kAssignInputIdx)},
+    {prim::kLambApplyOptimizerAssign, std::make_shared<OpUMonadExpander>(kLambInputIdx)},
+  };
+
+  for (auto &e : expanders) {
+    if (IsPrimitiveCNode(node, e.first)) {
+      return e.second;
+    }
+  }
+  return std::make_shared<DefaultExpander>();
+}
+
 bool GraphKernelExpander::DoExpand(const FuncGraphPtr &func_graph) {
   bool changed = false;
   auto todos = TopoSort(func_graph->get_return());
@@ -192,18 +212,6 @@ bool GraphKernelExpander::DoExpand(const FuncGraphPtr &func_graph) {
   return changed;
 }
 
-ExpanderPtr GraphKernelExpander::GetExpander(const AnfNodePtr &node) {
-  std::vector<std::pair<PrimitivePtr, ExpanderPtr>> expanders = {
-    {prim::kPrimDropout, std::make_shared<DropoutExpander>()},
-  };
-  for (auto &e : expanders) {
-    if (IsPrimitiveCNode(node, e.first)) {
-      return e.second;
-    }
-  }
-  return std::make_shared<DefaultExpander>();
-}
-
 bool GraphKernelExpander::Run(const FuncGraphPtr &func_graph) {
   expand_ops_ = GetExpandOps();
   return DoExpand(func_graph);

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

@@ -37,7 +37,7 @@
 #include "backend/optimizer/graph_kernel/value_graph_binder.h"
 #include "backend/optimizer/graph_kernel/parallel_fusion.h"
 #include "backend/optimizer/graph_kernel/optimize_assign.h"
-#include "backend/optimizer/graph_kernel/split_assign.h"
+#include "backend/optimizer/graph_kernel/split_umonad.h"
 #include "backend/optimizer/graph_kernel/reorder_ops.h"
 #include "backend/optimizer/graph_kernel/update_state_formatter.h"
 #include "backend/optimizer/graph_kernel/axis_normalizer.h"

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

@@ -13,7 +13,7 @@
  * See the License for the specific language governing permissions and
  * limitations under the License.
  */
-#include "backend/optimizer/graph_kernel/split_assign.h"
+#include "backend/optimizer/graph_kernel/split_umonad.h"
 
 #include <vector>
 #include <string>
@@ -35,31 +35,63 @@ const BaseRef SplitAssign::DefinePattern() const {
   return VectorRef({v, Xs, Us, UMonad});
 }
 
-bool CanSplit(const AnfNodePtr &node) {
+bool CanSplit(const AnfNodePtr &node) { return IsPrimitiveCNode(node, prim::kPrimAssign); }
-  return IsPrimitiveCNode(node, prim::kPrimAssignAdd) || IsPrimitiveCNode(node, prim::kPrimAssign) ||
-         IsPrimitiveCNode(node, prim::kPrimAssignSub);
-}
 
-const AnfNodePtr SplitAssign::Process(const FuncGraphPtr &func_graph, const AnfNodePtr &node, const EquivPtr &) const {
+AnfNodePtr ProcessNode(const FuncGraphPtr &func_graph, const AnfNodePtr &node, int input_idx) {
   MS_EXCEPTION_IF_NULL(node);
-  if (!CanSplit(node)) return node;
   CNodePtr cnode = node->cast<CNodePtr>();
   MS_EXCEPTION_IF_NULL(cnode);
-  CheckCNodeInputSize(cnode, kAssignInputTensorNum);
+
-  // Get original assign op's abstract and inputs
+  // Get original op's abstract and inputs
   AbstractBasePtr original_abstract = cnode->abstract()->Clone();
   auto original_inputs = cnode->inputs();
+
+  int input_node_size = cnode->size() - 1;
   // Create depend node
-  AnfNodePtrList depend_inputs = {NewValueNode(prim::kPrimDepend), original_inputs[1], original_inputs[3]};
+  AnfNodePtrList depend_inputs = {NewValueNode(prim::kPrimDepend), original_inputs[input_idx],
+                                  original_inputs[input_node_size]};
   auto depend_cnode = func_graph->NewCNode(depend_inputs);
-  depend_cnode->set_abstract(original_inputs[1]->abstract());
+  depend_cnode->set_abstract(original_inputs[input_idx]->abstract());
   depend_cnode->set_kernel_info(std::make_shared<device::KernelInfo>());
-  // Create new assign node, delete U from inputs.
+  // Create new node, delete U from inputs.
-  AnfNodePtrList new_assign_inputs = {cnode->input(0), depend_cnode, original_inputs[2]};
+  AnfNodePtrList new_inputs = {cnode->input(0)};
-  auto new_assign_cnode = func_graph->NewCNode(new_assign_inputs);
+  for (int i = 1; i < input_node_size; i++) {
-  new_assign_cnode->set_abstract(original_abstract);
+    if (i == input_idx) {
-  new_assign_cnode->set_kernel_info(cnode->kernel_info_ptr());
+      new_inputs.push_back(depend_cnode);
-  return new_assign_cnode;
+    } else {
+      new_inputs.push_back(cnode->input(i));
+    }
+  }
+  auto new_cnode = func_graph->NewCNode(new_inputs);
+  new_cnode->set_abstract(original_abstract);
+  new_cnode->set_kernel_info(cnode->kernel_info_ptr());
+  return new_cnode;
+}
+
+const AnfNodePtr SplitAssign::Process(const FuncGraphPtr &func_graph, const AnfNodePtr &node, const EquivPtr &) const {
+  MS_EXCEPTION_IF_NULL(node);
+  if (!CanSplit(node)) return node;
+  return ProcessNode(node->func_graph(), node, 1);
+}
+
+AnfNodePtr OpUMonadExpander::Run(const AnfNodePtr &node) {
+  auto cnode = node->cast<CNodePtr>();
+  MS_EXCEPTION_IF_NULL(cnode);
+
+  bool has_umonad = false;
+  for (unsigned int i = 1; i < cnode->size(); i++) {
+    if (HasAbstractUMonad(cnode->input(i))) {
+      has_umonad = true;
+      break;
+    }
+  }
+  if (has_umonad) {
+    auto new_node = ProcessNode(node->func_graph(), node, input_idx_);
+    return DefaultExpander::Run(new_node);
+  }
+
+  return DefaultExpander::Run(node);
 }
+
 }  // namespace opt
 }  // namespace mindspore

mindspore/ccsrc/backend/optimizer/graph_kernel/split_umonad.h

@@ -13,11 +13,11 @@
  * See the License for the specific language governing permissions and
  * limitations under the License.
  */
-#ifndef MINDSPORE_CCSRC_BACKEND_OPTIMIZER_GRAPH_KERNEL_SPLIT_ASSIGN_H_
+#ifndef MINDSPORE_CCSRC_BACKEND_OPTIMIZER_GRAPH_KERNEL_SPLIT_UMONAD_H_
-#define MINDSPORE_CCSRC_BACKEND_OPTIMIZER_GRAPH_KERNEL_SPLIT_ASSIGN_H_
+#define MINDSPORE_CCSRC_BACKEND_OPTIMIZER_GRAPH_KERNEL_SPLIT_UMONAD_H_
 
 #include "backend/optimizer/common/optimizer.h"
-
+#include "backend/optimizer/graph_kernel/graph_kernel_expander.h"
 namespace mindspore {
 namespace opt {
 class SplitAssign : public PatternProcessPass {
@@ -27,6 +27,16 @@ class SplitAssign : public PatternProcessPass {
   const BaseRef DefinePattern() const override;
   const AnfNodePtr Process(const FuncGraphPtr &, const AnfNodePtr &, const EquivPtr &) const override;
 };
+
+class OpUMonadExpander : public DefaultExpander {
+ public:
+  explicit OpUMonadExpander(int input_idx) : input_idx_(input_idx) {}
+  ~OpUMonadExpander() = default;
+  AnfNodePtr Run(const AnfNodePtr &node) override;
+
+ private:
+  int input_idx_;
+};
 }  // namespace opt
 }  // namespace mindspore
-#endif  // MINDSPORE_CCSRC_BACKEND_OPTIMIZER_GRAPH_KERNEL_SPLIT_ASSIGN_H_
+#endif  // MINDSPORE_CCSRC_BACKEND_OPTIMIZER_GRAPH_KERNEL_SPLIT_UMONAD_H_

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

@@ -219,7 +219,9 @@ bool ExtendOutputForUpdateState::ProcessIndex(const FuncGraphPtr &func_graph, co
   auto mng = func_graph->manager();
   MS_EXCEPTION_IF_NULL(mng);
   for (auto user : mng->node_users()[getitems_[index]]) {
-    user.first->cast<CNodePtr>()->set_input(user.second, new_node);
+    if (IsPrimitiveCNode(user.first, prim::kPrimUpdateState)) {
+      user.first->cast<CNodePtr>()->set_input(user.second, new_node);
+    }
   }
   return true;
 }

tests/st/ops/graph_kernel/test_assign_add.py

@@ -32,26 +32,38 @@ class AssignAdd(nn.Cell):
         self.add(self.var, y)
         return self.var
 
-
+def get_output(x2, y2, enable_graph_kernel=False):
-@pytest.mark.level0
+    context.set_context(enable_graph_kernel=enable_graph_kernel)
-@pytest.mark.platform_x86_gpu_training
-@pytest.mark.env_onecard
-def test_assign_add():
-    x2 = Tensor(np.arange(1 * 3 * 3 * 3).reshape(1, 3, 3, 3).astype(np.float32))
-    y2 = Tensor(np.arange(1 * 3 * 3 * 3).reshape(1, 3, 3, 3).astype(np.float32))
-
-    context.set_context(mode=context.GRAPH_MODE,
-                        enable_graph_kernel=True, device_target="GPU")
     add = AssignAdd(x2)
     result_gk_on_1 = add(y2)
     add_2 = AssignAdd(result_gk_on_1)
     result_gk_on_2 = add_2(y2)
+    output = [result_gk_on_1, result_gk_on_2]
+    return output
+
+def assign_add():
+    x2 = Tensor(np.arange(1 * 3 * 3 * 3).reshape(1, 3, 3, 3).astype(np.float32))
+    y2 = Tensor(np.arange(1 * 3 * 3 * 3).reshape(1, 3, 3, 3).astype(np.float32))
+
+    expect = get_output(x2, y2, False)
+    output = get_output(x2, y2, True)
+    e1, e2 = list(expect)
+    o1, o2 = list(output)
+
+    assert np.allclose(o1.asnumpy(), e1.asnumpy())
+    assert np.allclose(o2.asnumpy(), e2.asnumpy())
 
-    context.set_context(mode=context.GRAPH_MODE,
+@pytest.mark.level0
-                        enable_graph_kernel=False, device_target="GPU")
+@pytest.mark.platform_x86_gpu_training
-    add_beta = AssignAdd(x2)
+@pytest.mark.env_onecard
-    result_gk_off_1 = add_beta(y2)
+def test_assign_add_gpu():
-    add_beta_2 = AssignAdd(result_gk_off_1)
+    context.set_context(mode=context.GRAPH_MODE, device_target="GPU")
-    result_gk_off_2 = add_beta_2(y2)
+    assign_add()
-    assert (result_gk_on_1.asnumpy() == result_gk_off_1.asnumpy()).all()
+
-    assert (result_gk_on_2.asnumpy() == result_gk_off_2.asnumpy()).all()
+@pytest.mark.level0
+@pytest.mark.platform_arm_ascend_training
+@pytest.mark.platform_x86_ascend_training
+@pytest.mark.env_onecard
+def test_assign_add_ascend():
+    context.set_context(mode=context.GRAPH_MODE, device_target="Ascend")
+    assign_add()

tests/st/ops/graph_kernel/test_lamb_apply_optimizer_assign.py

@@ -14,6 +14,7 @@
 # ============================================================================
 
 import numpy as np
+import pytest
 import mindspore.context as context
 import mindspore.nn as nn
 from mindspore import Tensor
@@ -67,6 +68,10 @@ def lamb_apply_optimizer_assign():
     assert np.allclose(o2.asnumpy(), e2.asnumpy())
     assert np.allclose(o3.asnumpy(), e3.asnumpy())
 
+@pytest.mark.level0
+@pytest.mark.platform_arm_ascend_training
+@pytest.mark.platform_x86_ascend_training
+@pytest.mark.env_onecard
 def test_lamb_apply_optimizer_assign_ascend():
     context.set_context(mode=context.GRAPH_MODE, device_target="Ascend")
     lamb_apply_optimizer_assign()