convert all tuple output to maketuple

cmake/options.cmake

@@ -45,8 +45,8 @@ endif()
 
 if (DEBUG_MODE)
     set(CMAKE_BUILD_TYPE "Debug")
-else()
     add_compile_definitions(MEM_REUSE_DEBUG)
+else()
     set(CMAKE_BUILD_TYPE "Release")
 endif()
 

mindspore/ccsrc/common/trans.cc

@@ -205,8 +205,8 @@ std::vector<int> GetRuntimePaddingShape(const AnfNodePtr &node, size_t index) {
     if (tensor == nullptr) {
       MS_LOG(EXCEPTION) << " the node[ " << node->DebugString() << "]'s cannot convert ";
     }
-    shape = tensor->shape();
-    (void)std::transform(shape.begin(), shape.end(), std::back_inserter(host_shape), IntToSize);
+    auto shape_temp = tensor->shape();
+    (void)std::transform(shape_temp.begin(), shape_temp.end(), std::back_inserter(host_shape), IntToSize);
     if (host_shape.empty()) {
       host_shape.push_back(1);
     }

mindspore/ccsrc/pre_activate/ascend/ascend_helper.cc

@@ -18,6 +18,7 @@
 #include <set>
 #include "common/trans.h"
 #include "common/utils.h"
+#include "pre_activate/common/helper.h"
 #include "utils/utils.h"
 #include "device/kernel_info.h"
 #include "kernel/oplib/oplib.h"
@@ -346,21 +347,6 @@ CNodePtr InsertCastForInput(const FuncGraphPtr &func_graph, const CNodePtr &cnod
   return new_node;
 }
 
-AnfNodePtr CreatTupleGetItemNode(const FuncGraphPtr &func_graph, const AnfNodePtr &node, size_t output_idx) {
-  auto idx = NewValueNode(SizeToInt(output_idx));
-  MS_EXCEPTION_IF_NULL(idx);
-  auto imm = std::make_shared<Int32Imm>(SizeToInt(output_idx));
-  auto abstract_scalar = std::make_shared<abstract::AbstractScalar>(imm);
-  idx->set_abstract(abstract_scalar);
-  AnfNodePtr tuple_getitem = func_graph->NewCNode({NewValueNode(prim::kPrimTupleGetItem), node, idx});
-  MS_EXCEPTION_IF_NULL(tuple_getitem);
-  tuple_getitem->set_scope(node->scope());
-  std::vector<size_t> origin_shape = AnfAlgo::GetOutputInferShape(node, output_idx);
-  TypeId origin_type = AnfAlgo::GetOutputInferDataType(node, output_idx);
-  AnfAlgo::SetOutputInferTypeAndShape({origin_type}, {origin_shape}, tuple_getitem.get());
-  return tuple_getitem;
-}
-
 AnfNodePtr CreateMemcpyAsyncOp(const FuncGraphPtr &graph, const AnfNodePtr &node) {
   MS_EXCEPTION_IF_NULL(graph);
   MS_EXCEPTION_IF_NULL(node);

mindspore/ccsrc/pre_activate/ascend/ascend_helper.h

@@ -64,8 +64,6 @@ AnfNodePtr InsertTransOpForOutput(const FuncGraphPtr &func_graph, const AnfNodeP
 
 CNodePtr InsertCastForInput(const FuncGraphPtr &func_graph, const CNodePtr &cnode);
 
-AnfNodePtr CreatTupleGetItemNode(const FuncGraphPtr &func_graph, const AnfNodePtr &node, size_t output_idx);
-
 AnfNodePtr CreateMemcpyAsyncOp(const FuncGraphPtr &graph, const AnfNodePtr &node);
 }  // namespace opt
 }  // namespace mindspore

mindspore/ccsrc/pre_activate/ascend/enhancer/insert_memcpy_async_for_getnext.cc

@@ -17,6 +17,7 @@
 #include <vector>
 #include <memory>
 #include "pre_activate/ascend/ascend_helper.h"
+#include "pre_activate/common/helper.h"
 #include "session/anf_runtime_algorithm.h"
 
 namespace mindspore {

mindspore/ccsrc/pre_activate/common/common_backend_optimization.cc

@@ -18,6 +18,7 @@
 #include <string>
 #include "pre_activate/common/optimizer.h"
 #include "pre_activate/pass/convert_const_input_to_attr.h"
+#include "pre_activate/pass/convert_tuple_output_to_maketuple.h"
 #include "pre_activate/pass/convert_const_input_to_tensor_input.h"
 #include "pre_activate/pass/convert_tuple_input_to_dynamic_input.h"
 #include "utils/context/ms_context.h"
@@ -42,6 +43,7 @@ void BackendCommonOptimization(const std::shared_ptr<session::KernelGraph> &kern
   common_pm->AddPass(std::make_shared<ConvertConstInputToAttr>());
   common_pm->AddPass(std::make_shared<ConvertConstInputToTensorInput>());
   common_pm->AddPass(std::make_shared<ConvertTupleInputToDynamicInput>());
+  common_pm->AddPass(std::make_shared<ConvertTupleOutputToMaketuple>());
   optimizer->AddPassManager(common_pm);
   (void)optimizer->Optimize(kernel_graph);
   kernel_graph->SetExecOrderByDefault();

mindspore/ccsrc/pre_activate/common/helper.cc

@@ -407,5 +407,20 @@ bool IsUsedByOthers(const FuncGraphPtr &graph, const AnfNodePtr &node) {
   }
   return manager->node_users()[node].size() > 1;
 }
+
+AnfNodePtr CreatTupleGetItemNode(const FuncGraphPtr &func_graph, const AnfNodePtr &node, size_t output_idx) {
+  auto idx = NewValueNode(SizeToInt(output_idx));
+  MS_EXCEPTION_IF_NULL(idx);
+  auto imm = std::make_shared<Int32Imm>(SizeToInt(output_idx));
+  auto abstract_scalar = std::make_shared<abstract::AbstractScalar>(imm);
+  idx->set_abstract(abstract_scalar);
+  AnfNodePtr tuple_getitem = func_graph->NewCNode({NewValueNode(prim::kPrimTupleGetItem), node, idx});
+  MS_EXCEPTION_IF_NULL(tuple_getitem);
+  tuple_getitem->set_scope(node->scope());
+  std::vector<size_t> origin_shape = AnfAlgo::GetOutputInferShape(node, output_idx);
+  TypeId origin_type = AnfAlgo::GetOutputInferDataType(node, output_idx);
+  AnfAlgo::SetOutputInferTypeAndShape({origin_type}, {origin_shape}, tuple_getitem.get());
+  return tuple_getitem;
+}
 }  // namespace opt
 }  // namespace mindspore

mindspore/ccsrc/pre_activate/common/helper.h

@@ -146,6 +146,8 @@ void HideNopNode(session::KernelGraph *const graph);
 
 void RemoveNopNode(session::KernelGraph *const graph);
 
+AnfNodePtr CreatTupleGetItemNode(const FuncGraphPtr &func_graph, const AnfNodePtr &node, size_t output_idx);
+
 bool IsUsedByOthers(const FuncGraphPtr &graph, const AnfNodePtr &node);
 }  // namespace opt
 }  // namespace mindspore

mindspore/ccsrc/pre_activate/pass/convert_tuple_input_to_dynamic_input.cc

@@ -19,6 +19,7 @@
 #include <memory>
 
 #include "session/anf_runtime_algorithm.h"
+#include "pre_activate/common/helper.h"
 #include "session/kernel_graph.h"
 
 namespace mindspore {
@@ -40,13 +41,7 @@ void ConvertTupleOuputToPlantInputs(const FuncGraphPtr &graph, const AnfNodePtr
     convert_inputs = kernel_graph->SplitTupleValueNodeToNodeList(value_node);
   } else {
     for (size_t index = 0; index < output_size; ++index) {
-      auto idx = NewValueNode(SizeToInt(index));
-      MS_EXCEPTION_IF_NULL(idx);
-      auto imm = std::make_shared<Int32Imm>(SizeToInt(index));
-      auto abstract_scalar = std::make_shared<abstract::AbstractScalar>(imm);
-      idx->set_abstract(abstract_scalar);
-      auto tuple_get_item =
-        graph->NewCNode(std::vector<AnfNodePtr>{NewValueNode(prim::kPrimTupleGetItem), input_node, idx});
+      auto tuple_get_item = CreatTupleGetItemNode(graph, input_node, index);
       AnfAlgo::SetOutputInferTypeAndShape({AnfAlgo::GetOutputInferDataType(input_node, index)},
                                           {AnfAlgo::GetOutputInferShape(input_node, index)}, tuple_get_item.get());
       convert_inputs.emplace_back(tuple_get_item);

mindspore/ccsrc/pre_activate/pass/convert_tuple_output_to_maketuple.cc

@@ -0,0 +1,79 @@
+/**
+ * Copyright 2020 Huawei Technologies Co., Ltd
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#include "pre_activate/pass/convert_tuple_output_to_maketuple.h"
+
+#include <algorithm>
+#include <memory>
+
+#include "session/anf_runtime_algorithm.h"
+#include "pre_activate/common/helper.h"
+#include "session/kernel_graph.h"
+
+namespace mindspore {
+namespace opt {
+namespace {
+CNodePtr ConvertTupleInputToMakeTuple(const FuncGraphPtr &graph, const CNodePtr &cnode_ptr) {
+  MS_EXCEPTION_IF_NULL(cnode_ptr);
+  MS_EXCEPTION_IF_NULL(graph);
+  std::vector<AnfNodePtr> convert_inputs = {cnode_ptr->input(0)};
+  for (size_t index = 0; index < AnfAlgo::GetInputTensorNum(cnode_ptr); ++index) {
+    auto input_node = AnfAlgo::GetInputNode(cnode_ptr, index);
+    if (AnfAlgo::IsTupleOutput(input_node)) {
+      std::vector<TypeId> types;
+      std::vector<std::vector<size_t>> shapes;
+      std::vector<AnfNodePtr> make_tuple_inputs_list = {NewValueNode(prim::kPrimMakeTuple)};
+      for (size_t tuple_out_index = 0; tuple_out_index < AnfAlgo::GetOutputTensorNum(input_node); ++tuple_out_index) {
+        make_tuple_inputs_list.emplace_back(CreatTupleGetItemNode(graph, input_node, tuple_out_index));
+        types.push_back(AnfAlgo::GetOutputInferDataType(input_node, tuple_out_index));
+        shapes.emplace_back(AnfAlgo::GetOutputInferShape(input_node, tuple_out_index));
+      }
+      auto make_tuple = graph->NewCNode(make_tuple_inputs_list);
+      AnfAlgo::SetOutputInferTypeAndShape(types, shapes, make_tuple.get());
+      convert_inputs.emplace_back(make_tuple);
+    } else {
+      convert_inputs.push_back(input_node);
+    }
+  }
+  cnode_ptr->set_inputs(convert_inputs);
+  return cnode_ptr;
+}
+}  // namespace
+
+const BaseRef ConvertTupleOutputToMaketuple::DefinePattern() const {
+  VarPtr V = std::make_shared<Var>();
+  VarPtr Xs = std::make_shared<SeqVar>();
+  return VectorRef({V, Xs});
+}
+
+const AnfNodePtr ConvertTupleOutputToMaketuple::Process(const FuncGraphPtr &func_graph, const AnfNodePtr &node,
+                                                        const EquivPtr &) const {
+  if (node == nullptr || !node->isa<CNode>()) {
+    return nullptr;
+  }
+  auto cnode = node->cast<CNodePtr>();
+  MS_EXCEPTION_IF_NULL(cnode);
+  if (AnfAlgo::GetCNodeName(cnode) == prim::kPrimTupleGetItem->name()) {
+    return nullptr;
+  }
+  if (std::any_of(cnode->inputs().begin() + 1, cnode->inputs().end(), [](const AnfNodePtr &node) {
+        return AnfAlgo::IsTupleOutput(node) && AnfAlgo::GetCNodeName(node) != prim::kPrimMakeTuple->name();
+      })) {
+    return ConvertTupleInputToMakeTuple(func_graph, cnode);
+  }
+  return nullptr;
+}
+}  // namespace opt
+}  // namespace mindspore

mindspore/ccsrc/pre_activate/pass/convert_tuple_output_to_maketuple.h

@@ -0,0 +1,40 @@
+/**
+ * Copyright 2020 Huawei Technologies Co., Ltd
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef MINDSPORE_CONVERT_TUPLE_OUTPUT_TO_MAKETUPLE_H
+#define MINDSPORE_CONVERT_TUPLE_OUTPUT_TO_MAKETUPLE_H
+#include <string>
+#include <vector>
+
+#include "ir/anf.h"
+#include "pre_activate/common/optimizer.h"
+
+namespace mindspore {
+namespace opt {
+class ConvertTupleOutputToMaketuple : public PatternProcessPass {
+ public:
+  explicit ConvertTupleOutputToMaketuple(bool multigraph = true)
+      : PatternProcessPass("convert_tuple_output_to_maketuple", multigraph) {}
+
+  ~ConvertTupleOutputToMaketuple() override = default;
+
+  const BaseRef DefinePattern() const override;
+
+  const AnfNodePtr Process(const FuncGraphPtr &, const AnfNodePtr &, const EquivPtr &) const override;
+};
+}  // namespace opt
+}  // namespace mindspore
+#endif  // MINDSPORE_CONVERT_TUPLE_OUTPUT_TO_MAKETUPLE_H

mindspore/ccsrc/session/kernel_graph.cc

@@ -239,7 +239,7 @@ std::vector<AnfNodePtr> KernelGraph::SplitTupleValueNodeToNodeList(const ValueNo
     AddValueNodeToGraph(new_value_node);
     convert_inputs.emplace_back(new_value_node);
   }
-  if (RemoveValueNodeFromGraph(value_node)) {
+  if (!RemoveValueNodeFromGraph(value_node)) {
     MS_LOG(WARNING) << "failed to remove the value_node " << value_node->DebugString();
   }
   return convert_inputs;

tests/ut/cpp/pre_activate/pass/convert_tuple_output_to_maketuple_test.cc

@@ -0,0 +1,65 @@
+/**
+ * Copyright 2020 Huawei Technologies Co., Ltd
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#include "common/backend_common_test.h"
+#include "ir/anf.h"
+#include "ir/meta_tensor.h"
+#include "debug/anf_ir_dump.h"
+#include "common/py_func_graph_fetcher.h"
+#include "session/anf_runtime_algorithm.h"
+#include "pre_activate/common/optimizer.h"
+#include "pre_activate/common/pass_manager.h"
+#include "pre_activate/pass/convert_tuple_output_to_maketuple.h"
+#include "utils/utils.h"
+
+namespace mindspore {
+namespace opt {
+class TestHWTupleOutputToMakeTuple : public BackendCommon {
+ public:
+  TestHWTupleOutputToMakeTuple()
+      : getPyFun_("gtest_input.pre_activate.convert_tuple_output_to_maketuple_test", true) {}
+  ~TestHWTupleOutputToMakeTuple() override = default;
+
+ public:
+  UT::PyFuncGraphFetcher getPyFun_;
+};
+
+TEST_F(TestHWTupleOutputToMakeTuple, test_convert_tuple_output_to_maketuple) {
+  FuncGraphPtr g = getPyFun_.CallAndParseRet("test_convert_tuple_output_to_maketuple", "before");
+  ASSERT_TRUE(g != nullptr);
+  std::vector<int> shp_x{5, 2, 10};
+  std::vector<int> shp_h{1, 2, 2};
+  std::vector<int> shp_c{1, 2, 2};
+  std::vector<int> shp_w{112, 1, 1};
+  auto x_abstract = std::make_shared<abstract::AbstractTensor>(kFloat32, shp_x);
+  auto h_abstract = std::make_shared<abstract::AbstractTensor>(kFloat32, shp_h);
+  auto c_abstract = std::make_shared<abstract::AbstractTensor>(kFloat32, shp_c);
+  auto w_abstract = std::make_shared<abstract::AbstractTensor>(kFloat32, shp_w);
+  AbstractBasePtrList args_spec_list{x_abstract, h_abstract, c_abstract, w_abstract};
+  auto func_graph = GetKernelGraph(g, args_spec_list);
+  ASSERT_TRUE(func_graph != nullptr);
+
+  auto optimizer = std::make_shared<opt::GraphOptimizer>();
+  auto pm = std::make_shared<opt::PassManager>();
+  pm->AddPass(std::make_shared<opt::ConvertTupleOutputToMaketuple>());
+  optimizer->AddPassManager(pm);
+  optimizer->Optimize(func_graph);
+
+  FuncGraphPtr g_after = getPyFun_.CallAndParseRet("test_convert_tuple_output_to_maketuple", "after");
+  ASSERT_TRUE(g_after != nullptr);
+  EXPECT_TRUE(CheckEqualGraph(func_graph, g_after));
+}
+}  // namespace opt
+}  // namespace mindspore

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

@@ -0,0 +1,54 @@
+# Copyright 2020 Huawei Technologies Co., Ltd
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ============================================================================
+from mindspore.ops import operations as P
+from mindspore.ops import Primitive
+import mindspore as ms
+import mindspore.common.dtype as mstype
+from mindspore.common.tensor import Tensor
+import numpy as np
+
+make_tuple = Primitive('make_tuple')
+tuple_get_item = Primitive("tuple_getitem");
+LSTM = P.LSTM(input_size=10,hidden_size=2,num_layers=1,has_bias=True,bidirectional=False,dropout=0.0)
+add = P.TensorAdd()
+
+class FnDict:
+    def __init__(self):
+        self.fnDict = {}
+
+    def __call__(self, fn):
+        self.fnDict[fn.__name__] = fn
+
+    def __getitem__(self, name):
+        return self.fnDict[name]
+
+
+def test_convert_tuple_output_to_maketuple(tag):
+    fns = FnDict()
+
+    @fns
+    def before(x, h, c, w):
+        res = LSTM(x, h, c, w)
+        return res
+
+    @fns
+    def after(x, h, c, w):
+        res = LSTM(x, h, c, w)
+        res = make_tuple(
+            make_tuple(tuple_get_item(res, 0), tuple_get_item(res, 1), tuple_get_item(res, 2), tuple_get_item(res, 3),
+                       tuple_get_item(res, 4)));
+        return res
+
+    return fns[tag]

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

@@ -49,7 +49,10 @@ def test_insert_memcpy_async_for_getnext(tag):
         label = tuple_getitem(res, 1)
         memcpy_async_data = memcpy_async(data)
         memcpy_async_label = memcpy_async(label)
-        tuple = make_tuple(make_tuple(memcpy_async_data, memcpy_async_label))
-        return tuple
+        bind_tuple = make_tuple(memcpy_async_data, memcpy_async_label)
+        get_item0 = tuple_getitem(bind_tuple, 0)
+        get_item1 = tuple_getitem(bind_tuple, 1)
+        bind_tuple = make_tuple(make_tuple(get_item0, get_item1))
+        return bind_tuple
 
     return fns[tag]