!44526 support dynamic input for GE

Merge pull request !44526 from xulei/ge_dynamic_shape

mindspore/ccsrc/plugin/device/ascend/hal/common/ascend_utils.h

@@ -25,6 +25,20 @@
 namespace mindspore {
 namespace device {
 namespace ascend {
+template <typename Map, typename K = typename Map::key_type, typename V = typename Map::mapped_type>
+std::string MapToString(const Map &value) {
+  std::stringstream buffer;
+  buffer << "{";
+  for (auto it = value.begin(); it != value.end(); it++) {
+    if (it != value.begin()) {
+      buffer << ", ";
+    }
+    buffer << it->first << ": " << it->second;
+  }
+  buffer << "}";
+  return buffer.str();
+}
+
 std::string GetErrorMessage(bool add_title = false);
 std::string GetWarningMessage();
 void SetErrorManagerContext();

mindspore/ccsrc/plugin/device/ascend/hal/hardware/ge_device_context.cc

@@ -35,6 +35,7 @@
 #include "runtime/hardware/device_context_manager.h"
 #include "plugin/device/ascend/hal/hccl_adapter/hccl_adapter.h"
 #include "plugin/device/ascend/optimizer/ge_optimization.h"
+#include "plugin/device/ascend/hal/common/ascend_utils.h"
 #include "runtime/config.h"
 #include "runtime/dev.h"
 #include "distributed/init.h"
@@ -43,6 +44,8 @@ namespace mindspore {
 namespace device {
 namespace ascend {
 namespace {
+using mindspore::transform::OptionMap;
+
 constexpr auto kMindsporeDumpConfig = "MINDSPORE_DUMP_CONFIG";
 constexpr char kGeDumpMode[3][7] = {"all", "input", "output"};
 
@@ -101,6 +104,38 @@ transform::TensorOrderMap GetParams(const FuncGraphPtr &anf_graph) {
   return res;
 }
 
+std::string ShapesToString(const ShapeArray &shapes) {
+  std::stringstream buffer;
+  for (size_t i = 0; i < shapes.size(); ++i) {
+    if (i != 0) {
+      buffer << ",";
+    }
+    buffer << "[";
+    const auto &shape = shapes[i];
+    for (size_t j = 0; j < shape.size(); ++j) {
+      if (j != 0) {
+        buffer << ",";
+      }
+      buffer << shape[j];
+    }
+    buffer << "]";
+  }
+  return buffer.str();
+}
+
+OptionMap GetComputeGraphOptions(const ShapeArray &input_shapes, bool is_dynamic_shape) {
+  OptionMap options{};
+  if (common::GetEnv("GE_TRAIN") == "1") {
+    (void)options.emplace("ge.exec.variable_acc", "1");
+  }
+  if (!is_dynamic_shape) {
+    return options;
+  }
+  (void)options.emplace("ge.exec.dynamicGraphExecuteMode", "dynamic_execute");
+  (void)options.emplace("ge.exec.dataInputsShapeRange", ShapesToString(input_shapes));
+  return options;
+}
+
 bool AddDFGraph(const FuncGraphPtr &anf_graph, const transform::TensorOrderMap &init_inputs_map, bool export_air) {
   MS_EXCEPTION_IF_NULL(anf_graph);
   auto converter = transform::NewConverter(anf_graph);
@@ -121,11 +156,9 @@ bool AddDFGraph(const FuncGraphPtr &anf_graph, const transform::TensorOrderMap &
   std::string graph_name = anf_graph->ToString();
   std::string init_graph = "init_subgraph." + graph_name;
   std::string checkpoint_name = "save." + GetGraphName(anf_graph);
-  if (common::GetEnv("GE_TRAIN") == "1") {
+  const auto options = GetComputeGraphOptions(converter->input_shapes(), converter->dynamic_shape_inputs());
-    (void)transform::AddGraph(graph_name, transform::GetComputeGraph(converter), {{"ge.exec.variable_acc", "1"}});
+  MS_LOG(INFO) << "Set options of compute graph: " << graph_name << " to " << MapToString(options);
-  } else {
+  (void)transform::AddGraph(graph_name, transform::GetComputeGraph(converter), options);
-    (void)transform::AddGraph(graph_name, transform::GetComputeGraph(converter));
-  }
   (void)transform::AddGraph(init_graph, transform::GetInitGraph(converter));
   (void)transform::AddGraph(BROADCAST_GRAPH_NAME, transform::GetBroadcastGraph(converter));
 
@@ -488,7 +521,7 @@ void GeDeviceContext::InitGe(const std::shared_ptr<MsContext> &inst_context) {
   {
     // Release GIL before calling into (potentially long-running) C++ code
     mindspore::ScopedLongRunning long_running;
-    if (ge::GEInitialize(ge_options) != ge::GRAPH_SUCCESS) {
+    if (::ge::GEInitialize(ge_options) != ::ge::GRAPH_SUCCESS) {
       MS_LOG(EXCEPTION) << "Initialize GE failed!";
     }
   }
@@ -670,7 +703,7 @@ bool GeDeviceContext::FinalizeGe(const std::shared_ptr<MsContext> &inst_context)
       std::string exName(abi::__cxa_current_exception_type()->name());
       MS_LOG(ERROR) << "Error occurred when deleting GE graph runner and session fail. Exception name: " << exName;
     }
-    if (ge::GEFinalize() != ge::GRAPH_SUCCESS) {
+    if (::ge::GEFinalize() != ::ge::GRAPH_SUCCESS) {
       MS_LOG(WARNING) << "Finalize GE failed!";
     }
     inst_context->set_param<bool>(MS_CTX_IS_PYNATIVE_GE_INIT, false);

mindspore/ccsrc/transform/graph_ir/convert.cc

@@ -20,6 +20,7 @@
 #include <algorithm>
 #include <queue>
 #include <stack>
+#include <unordered_set>
 #include "include/common/utils/utils.h"
 #include "mindspore/core/ops/core_ops.h"
 #include "utils/anf_utils.h"
@@ -392,14 +393,16 @@ DfGraphConvertor &DfGraphConvertor::InitParam(const TensorOrderMap &tensors) {
     return *this;
   }
 
-  // Processing input with MakeDatasetHandler
+  // Processing input with MakeDatasetHandler and check whether input is dynamic
   for (auto &it : anf_graph_->inputs()) {
     auto op_itor = op_cache_.find(it.get());  // converted node
     if (it->isa<Parameter>() && op_itor != op_cache_.end()) {
-      string name = std::static_pointer_cast<Parameter>(it)->name();
+      const auto &param = std::static_pointer_cast<Parameter>(it);
+      string name = param->name();
       auto tensor_itor = tensors.find(name);  // in init value map
       if (tensor_itor == tensors.end()) {
-        DfGraphConvertor::MakeDatasetHandler(name, input_idx, it);
+        MakeDatasetHandler(name, input_idx, it);
+        AddGraphDynamicInput(param);
         input_idx++;
       }
     }
@@ -1941,6 +1944,19 @@ void DfGraphConvertor::SetNodeInput(const AnfNodePtr node) {
   DfGraphConvertor::SetOpInput(adpt, cnode);
 }
 
+void DfGraphConvertor::AddGraphDynamicInput(const ParameterPtr &param) {
+  MS_EXCEPTION_IF_NULL(param);
+  const auto &base_shape = param->Shape();
+  MS_EXCEPTION_IF_NULL(base_shape);
+  const auto &shape = base_shape->cast<abstract::ShapePtr>();
+  MS_EXCEPTION_IF_NULL(shape);
+  const auto &sv = shape->shape();
+  if (std::any_of(sv.cbegin(), sv.cend(), [](const auto e) { return e == -1; })) {
+    dynamic_shape_inputs_ = true;
+  }
+  (void)input_shapes_.emplace_back(sv);
+}
+
 std::string DfGraphConvertor::GetGNodeName(const ::ge::GNode &node) const {
   ::ge::AscendString name;
   auto ret = node.GetName(name);
@@ -2224,7 +2240,7 @@ OperatorPtr DfGraphConvertor::Convert(const AnfNodePtr node) {
   return nullptr;
 }
 
-void DfGraphConvertor::ConvertTopK(const CNodePtr node) {
+void DfGraphConvertor::ConvertTopK(const CNodePtr &node) {
   MS_EXCEPTION_IF_NULL(node);
   MS_LOG(INFO) << "Convert TopK second input's type from int64 to int32.";
   auto value_ptr = node->input(2)->cast<ValueNodePtr>();
@@ -2372,7 +2388,7 @@ std::vector<int64_t> DfGraphConvertor::CastToInt(const ValuePtr &value) const {
   return cur_value;
 }
 
-void DfGraphConvertor::ConvertReshape(const CNodePtr node) {
+void DfGraphConvertor::ConvertReshape(const CNodePtr &node) {
   MS_LOG(INFO) << "Convert the second input of reshape to op attr.";
   const auto kInputNum = 3;
   if (node->size() < kInputNum) {
@@ -2386,10 +2402,7 @@ void DfGraphConvertor::ConvertReshape(const CNodePtr node) {
   auto op = adpt->generate(node);
   MS_EXCEPTION_IF_NULL(op);
   // get shape form attr
-  auto value_node = node->input(0)->cast<ValueNodePtr>();
+  auto primitive = GetCNodePrimitive(node);
-  MS_EXCEPTION_IF_NULL(value_node);
-  MS_EXCEPTION_IF_NULL(value_node->value());
-  auto primitive = value_node->value()->cast<PrimitivePtr>();
   MS_EXCEPTION_IF_NULL(primitive);
   auto value = primitive->GetAttr("shape");
   std::vector<int64_t> list;
@@ -2399,7 +2412,34 @@ void DfGraphConvertor::ConvertReshape(const CNodePtr node) {
   op_cache_[node.get()] = op;
 }
 
-void DfGraphConvertor::ConvertAllReduce(const CNodePtr node) {
+void DfGraphConvertor::ConvertDynamicStitch(const CNodePtr &node) {
+  MS_LOG(INFO) << "Convert and set 'N' attr of DynamicStitch.";
+  OpAdapterPtr adpt = FindAdapter(node, training_);
+  if (adpt == nullptr) {
+    return;
+  }
+  auto op = adpt->generate(node);
+  MS_EXCEPTION_IF_NULL(op);
+  int64_t input_length = 0;
+  auto indices = node->input(1);
+  MS_EXCEPTION_IF_NULL(indices);
+  if (indices->isa<CNode>()) {
+    input_length = SizeToLong(indices->cast<CNodePtr>()->size()) - 1;
+  } else if (IsValueNode<ValueSequence>(indices)) {
+    const auto tuple = GetValueNode<ValueSequencePtr>(indices);
+    MS_EXCEPTION_IF_NULL(tuple);
+    input_length = SizeToLong(tuple->size());
+  } else {
+    MS_LOG(EXCEPTION) << "Input 1 of DynamicStitch is neither CNode nor ValueNode contains ValueSequence, but "
+                      << indices->ToString() << ", can not set 'N' attr.";
+  }
+
+  (void)op->SetAttr("N", input_length);
+  MS_LOG(INFO) << "Set 'N' attr of DynamicStitch to " << input_length;
+  op_cache_[node.get()] = op;
+}
+
+void DfGraphConvertor::ConvertAllReduce(const CNodePtr &node) {
   MS_LOG(INFO) << "Add AllReduce fusion_id";
   OpAdapterPtr adpt = FindAdapter(node, training_);
   if (adpt == nullptr) {
@@ -2408,10 +2448,7 @@ void DfGraphConvertor::ConvertAllReduce(const CNodePtr node) {
   auto op = adpt->generate(node);
   MS_EXCEPTION_IF_NULL(op);
   // get shape form attr
-  auto value_node = node->input(0)->cast<ValueNodePtr>();
+  auto primitive = GetCNodePrimitive(node);
-  MS_EXCEPTION_IF_NULL(value_node);
-  MS_EXCEPTION_IF_NULL(value_node->value());
-  auto primitive = value_node->value()->cast<PrimitivePtr>();
   MS_EXCEPTION_IF_NULL(primitive);
   auto fusion_value = primitive->GetAttr("fusion");
   auto fusion = GetValue<int64_t>(fusion_value);
@@ -2425,7 +2462,8 @@ void DfGraphConvertor::ConvertAllReduce(const CNodePtr node) {
   op_cache_[node.get()] = op;
 }
 
-void DfGraphConvertor::ConvertConv2D(const CNodePtr node) {
+void DfGraphConvertor::ConvertConv2D(const CNodePtr &node) {
+  MS_LOG(INFO) << "Convert and set 'padding' attr for Conv2D-like op.";
   MS_EXCEPTION_IF_NULL(node);
   OpAdapterPtr adpt = FindAdapter(node, training_);
   if (adpt == nullptr) {
@@ -2433,20 +2471,33 @@ void DfGraphConvertor::ConvertConv2D(const CNodePtr node) {
   }
   auto op = adpt->generate(node);
   MS_EXCEPTION_IF_NULL(op);
-  auto value_node = node->input(0)->cast<ValueNodePtr>();
-  MS_EXCEPTION_IF_NULL(value_node);
-  MS_EXCEPTION_IF_NULL(value_node->value());
-  auto primitive = value_node->value()->cast<PrimitivePtr>();
-  MS_EXCEPTION_IF_NULL(primitive);
-  auto value = primitive->GetAttr("padding");
-  if (value != nullptr) {
-    std::string pad_mode = GetValue<std::string>(value);
-    (void)op->SetAttr("padding", pad_mode);
-  }
   op_cache_[node.get()] = op;
+  auto primitive = GetCNodePrimitive(node);
+  MS_EXCEPTION_IF_NULL(primitive);
+  std::string pad_mode;
+  if (auto value = primitive->GetAttr("padding"); value != nullptr) {
+    pad_mode = GetValue<std::string>(value);
+  } else if (auto value = primitive->GetAttr("pad_mode"); value != nullptr) {
+    // Get 'pad_mode' attr and set it to 'padding' attr for ge
+    const mindspore::HashMap<int64_t, std::string> pad_mode_map{{1, "SAME"}, {2, "VALID"}};
+    if (value->isa<StringImm>()) {
+      pad_mode = GetValue<std::string>(value);
+      (void)std::transform(pad_mode.cbegin(), pad_mode.cend(), pad_mode.begin(), toupper);
+    } else if (auto it = pad_mode_map.find(GetValue<int64_t>(value)); it != pad_mode_map.cend()) {
+      // 'pad_mode' attr could be an enumeration
+      pad_mode = it->second;
+    } else {
+      return;
+    }
+  } else {
+    MS_LOG(INFO) << "Node: " << node->fullname_with_scope() << " has no 'padding' or 'pad_mode' attr";
+    return;
+  }
+  MS_LOG(INFO) << "Set 'padding' attr of node: " << node->fullname_with_scope() << " to " << pad_mode;
+  (void)op->SetAttr("padding", pad_mode);
 }
 
-void DfGraphConvertor::ConvertOCRRecPreHandle(const CNodePtr node) {
+void DfGraphConvertor::ConvertOCRRecPreHandle(const CNodePtr &node) {
   MS_LOG(INFO) << "Add OCRRecognitionPreHandle _op_max_shape attr";
   OpAdapterPtr adpt = FindAdapter(node, training_);
   if (adpt == nullptr) {
@@ -2455,10 +2506,7 @@ void DfGraphConvertor::ConvertOCRRecPreHandle(const CNodePtr node) {
   auto op = adpt->generate(node);
   MS_EXCEPTION_IF_NULL(op);
   // get shape form attr
-  auto value_node = node->input(0)->cast<ValueNodePtr>();
+  auto primitive = GetCNodePrimitive(node);
-  MS_EXCEPTION_IF_NULL(value_node);
-  MS_EXCEPTION_IF_NULL(value_node->value());
-  auto primitive = value_node->value()->cast<PrimitivePtr>();
   MS_EXCEPTION_IF_NULL(primitive);
   auto value = primitive->GetAttr("_op_max_shape");
   if (value == nullptr) {
@@ -2517,33 +2565,26 @@ bool DfGraphConvertor::CheckCNode(const std::string &name, const CNodePtr node)
     return false;
   }
 
-  // Convert TopK second input from int64 to int32.
+  const mindspore::HashMap<std::string, std::function<void(decltype(this), const CNodePtr &)>>
-  if (name == prim::kPrimTopK->name()) {
+    auxiliary_node_converters{
-    ConvertTopK(node);
+      // Convert TopK second input from int64 to int32.
-    return true;
+      {prim::kPrimTopK->name(), &DfGraphConvertor::ConvertTopK},
-  }
+      // Convert Reshape add const input to attr(shape)
+      {prim::kPrimReshape->name(), &DfGraphConvertor::ConvertReshape},
+      {prim::kPrimOCRRecognitionPreHandle->name(), &DfGraphConvertor::ConvertOCRRecPreHandle},
+      // Add attr 'pad_mode' to Conv2D-like op
+      {prim::kPrimConv2D->name(), &DfGraphConvertor::ConvertConv2D},
+      {prim::kPrimDepthwiseConv2dNative->name(), &DfGraphConvertor::ConvertConv2D},
+      {kNameConv2DBackpropInputV2, &DfGraphConvertor::ConvertConv2D},
+      {prim::kPrimConv2DBackpropInput->name(), &DfGraphConvertor::ConvertConv2D},
+      {prim::kPrimConv2DBackpropFilter->name(), &DfGraphConvertor::ConvertConv2D},
+      // Add attr 'N' to DynamicStitch
+      {prim::kPrimDynamicStitch->name(), &DfGraphConvertor::ConvertDynamicStitch},
+      {prim::kPrimAllReduce->name(), &DfGraphConvertor::ConvertAllReduce},
+    };
 
-  // Convert Reshape add const input to attr(shape)
+  if (const auto it = auxiliary_node_converters.find(name); it != auxiliary_node_converters.cend()) {
-  if (name == prim::kPrimReshape->name()) {
+    it->second(this, node);
-    ConvertReshape(node);
-    return true;
-  }
-
-  if (name == prim::kPrimOCRRecognitionPreHandle->name()) {
-    ConvertOCRRecPreHandle(node);
-    return true;
-  }
-
-  // Add attr pad mode to Conv2D
-  if (name == prim::kPrimConv2D->name() || name == prim::kPrimDepthwiseConv2dNative->name() ||
-      name == kNameConv2DBackpropInputV2) {
-    ConvertConv2D(node);
-    return true;
-  }
-
-  if (name == prim::kPrimAllReduce->name()) {
-    ConvertAllReduce(node);
-    return true;
   }
 
   return true;

mindspore/ccsrc/transform/graph_ir/convert.h

@@ -156,6 +156,8 @@ class DfGraphConvertor {
 
   bool is_training() const { return training_; }
   void set_training(bool is_training) { training_ = is_training; }
+  bool dynamic_shape_inputs() const { return dynamic_shape_inputs_; }
+  std::vector<ShapeVector> input_shapes() { return input_shapes_; }
 
  protected:
   void InitLoopVar(std::vector<::ge::Operator> *init_input);
@@ -178,15 +180,16 @@ class DfGraphConvertor {
   OperatorPtr ConvertValueNode(ValueNodePtr node);
   void SaveParamFormat(CNodePtr node);
   void GetBranchNodeInput(const CNodePtr node);
-  void ConvertTopK(const CNodePtr node);
+  void ConvertTopK(const CNodePtr &node);
   void ConvertResizeBilinear(const FuncGraphPtr anf_graph) const;
   void ConvertSpaceBatchNd(const FuncGraphPtr anf_graph) const;
   void ConvertTile(const FuncGraphPtr anf_graph) const;
   AnfNodePtr CreateCast(const AnfNodePtr &input, const TypePtr &dst_type) const;
-  void ConvertReshape(const CNodePtr node);
+  void ConvertReshape(const CNodePtr &node);
-  void ConvertAllReduce(const CNodePtr node);
+  void ConvertAllReduce(const CNodePtr &node);
-  void ConvertOCRRecPreHandle(const CNodePtr node);
+  void ConvertOCRRecPreHandle(const CNodePtr &node);
-  void ConvertConv2D(const CNodePtr node);
+  void ConvertConv2D(const CNodePtr &node);
+  void ConvertDynamicStitch(const CNodePtr &node);
   std::vector<int64_t> CastToInt(const ValuePtr &value) const;
   bool CheckCNode(const std::string &name, const CNodePtr node);
   void SetNodeInput(AnfNodePtr node);
@@ -198,6 +201,7 @@ class DfGraphConvertor {
   void UpdateDataOpDesc(const AnfNodePtr &it, const OperatorPtr &op) const;
   void UpdateConstOpDesc(const AnfNodePtr &it, const OperatorPtr &op) const;
   void AddGraphConstInput(const OperatorPtr &op);
+  void AddGraphDynamicInput(const ParameterPtr &param);
   AnfNodePtr ParseLoadInput(const CNodePtr &cnode) const;
   void SetGraphInputs(std::vector<Operator> *inputs);
   void TransformConstOp(const CNodePtr &node, const AnfNodePtr &pred);
@@ -264,6 +268,7 @@ class DfGraphConvertor {
   std::vector<OperatorPtr> init_ops_;
   std::vector<OperatorPtr> broadcast_ops_;
   std::vector<AnfNodePtr> inputs_;
+  ShapeArray input_shapes_;
   OperatorPtr dataset_iter_getnext_;
   OperatorPtr queue_data_;
   OperatorPtr get_next_from_queue_;
@@ -271,6 +276,7 @@ class DfGraphConvertor {
   bool training_ = false;
   bool distribute_ = false;
   bool use_inputs_ = false;
+  bool dynamic_shape_inputs_ = false;
 
   AnfNodePtr while_cond_node_ = nullptr;
   mindspore::HashMap<AnfNodePtr, std::shared_ptr<std::vector<DfGraph>>> while_dfgraph_cache_;

mindspore/ccsrc/transform/graph_ir/op_adapter_map.h

@@ -396,6 +396,9 @@ constexpr const char kNameTransData[] = "TransData";
 constexpr const char kNameSend[] = "Send";
 constexpr const char kNameReceive[] = "Receive";
 constexpr const char kNameIndexAdd[] = "IndexAdd";
+constexpr const char kNameUnique[] = "Unique";
+constexpr const char kNameDynamicBroadcastGradientArgs[] = "DynamicBroadcastGradientArgs";
+constexpr const char kNameDynamicStitch[] = "DynamicStitch";
 
 class OpAdapterDesc;
 

mindspore/ccsrc/transform/graph_ir/op_declare/array_ops_declare.cc

@@ -160,5 +160,15 @@ REG_ADPT_DESC(IdentityNMakeTuple, kNameMakeTuple, ADPT_DESC(IdentityN))
 REG_ADPT_DESC(IdentityNDepend, kNameDepend, ADPT_DESC(IdentityN))
 REG_ADPT_DESC(IdentityNReturn, kNameReturn, ADPT_DESC(IdentityN))
 
-// IdentityN
+// Unique
+INPUT_MAP(Unique) = {{1, INPUT_DESC(x)}};
+ATTR_MAP(Unique) = EMPTY_ATTR_MAP;
+OUTPUT_MAP(Unique) = {{0, OUTPUT_DESC(y)}, {1, OUTPUT_DESC(idx)}};
+REG_ADPT_DESC(Unique, kNameUnique, ADPT_DESC(Unique))
+
+// BroadcastGradientArgs
+INPUT_MAP(BroadcastGradientArgs) = {{1, INPUT_DESC(x1)}, {2, INPUT_DESC(x2)}};
+ATTR_MAP(BroadcastGradientArgs) = EMPTY_ATTR_MAP;
+OUTPUT_MAP(BroadcastGradientArgs) = {{0, OUTPUT_DESC(y1)}, {1, OUTPUT_DESC(y2)}};
+REG_ADPT_DESC(BroadcastGradientArgs, kNameDynamicBroadcastGradientArgs, ADPT_DESC(BroadcastGradientArgs))
 }  // namespace mindspore::transform

mindspore/ccsrc/transform/graph_ir/op_declare/array_ops_declare.h

@@ -84,5 +84,11 @@ DECLARE_OP_USE_OUTPUT(Identity)
 
 DECLARE_OP_ADAPTER(IdentityN)
 DECLARE_OP_USE_DYN_OUTPUT(IdentityN)
+
+DECLARE_OP_ADAPTER(Unique)
+DECLARE_OP_USE_OUTPUT(Unique)
+
+DECLARE_OP_ADAPTER(BroadcastGradientArgs)
+DECLARE_OP_USE_OUTPUT(BroadcastGradientArgs)
 }  // namespace mindspore::transform
 #endif  // MINDSPORE_CCSRC_TRANSFORM_GRAPH_IR_OP_DECLARE_ARRAY_OPS_DECLARE_H_

mindspore/ccsrc/transform/graph_ir/op_declare/data_flow_ops_declare.cc

@@ -40,4 +40,11 @@ ATTR_MAP(TensorArrayGather) = {{"dtype", ATTR_DESC(dtype, AnyTraits<GEType>())},
                                {"element_shape", ATTR_DESC(element_shape, AnyTraits<std::vector<int64_t>>())}};
 OUTPUT_MAP(TensorArrayGather) = {{0, OUTPUT_DESC(value)}};
 REG_ADPT_DESC(TensorArrayGather, kNameTensorArrayGather, ADPT_DESC(TensorArrayGather))
+
+// DynamicStitch
+INPUT_MAP(DynamicStitch) = EMPTY_INPUT_MAP;
+DYN_INPUT_MAP(DynamicStitch) = {{1, DYN_INPUT_DESC(indices)}, {2, DYN_INPUT_DESC(x)}};
+ATTR_MAP(DynamicStitch) = EMPTY_ATTR_MAP;
+OUTPUT_MAP(DynamicStitch) = {{0, OUTPUT_DESC(y)}};
+REG_ADPT_DESC(DynamicStitch, kNameDynamicStitch, ADPT_DESC(DynamicStitch))
 }  // namespace mindspore::transform

mindspore/ccsrc/transform/graph_ir/op_declare/data_flow_ops_declare.h

@@ -29,5 +29,8 @@ DECLARE_OP_USE_OUTPUT(TensorArrayWrite)
 
 DECLARE_OP_ADAPTER(TensorArrayGather)
 DECLARE_OP_USE_OUTPUT(TensorArrayGather)
+
+DECLARE_OP_ADAPTER(DynamicStitch)
+DECLARE_OP_USE_OUTPUT(DynamicStitch)
 }  // namespace mindspore::transform
 #endif  // MINDSPORE_CCSRC_TRANSFORM_GRAPH_IR_OP_DECLARE_DATA_FLOW_OPS_DECLARE_H_

tests/st/ge/ge_infer_env.py

@@ -14,5 +14,6 @@
 # ============================================================================
 import os
 
+
 os.environ['MS_ENABLE_GE'] = '1'
 os.environ['MS_GE_TRAIN'] = '0'

tests/st/ge/ge_pass/ge_convert.py

@@ -13,7 +13,7 @@
 # limitations under the License.
 # ============================================================================
 import numpy as np
-import ge_train_env  # pylint: disable=unused-import
+from tests.st.ge import ge_train_env  # pylint: disable=unused-import
 import mindspore as ms
 import mindspore.context as context
 import mindspore.nn as nn

tests/st/ge/ge_pass/pass_avg_pool_grad.py

@@ -15,7 +15,7 @@
 """ test ge frontend pass `AvgPoolGradForGE` """
 import numpy as np
 
-import ge_infer_env  # pylint: disable=unused-import
+from tests.st.ge import ge_infer_env  # pylint: disable=unused-import
 import mindspore.context as context
 import mindspore.nn as nn
 from mindspore.ops import operations as op

tests/st/ge/ge_pass/pass_convert_attr_to_input.py

@@ -14,7 +14,7 @@
 # ============================================================================
 import numpy as np
 
-import ge_infer_env  # pylint: disable=unused-import
+from tests.st.ge import ge_infer_env  # pylint: disable=unused-import
 import mindspore
 from mindspore import context, nn, Tensor, ops
 

tests/st/ge/ge_pass/pass_convert_resize_nearest_neighbor_x_dtype.py

@@ -14,7 +14,7 @@
 # ============================================================================
 import numpy as np
 
-import ge_infer_env  # pylint: disable=unused-import
+from tests.st.ge import ge_infer_env  # pylint: disable=unused-import
 import mindspore
 from mindspore import context, nn, Tensor, ops
 

tests/st/ge/ge_pass/pass_dropout.py

@@ -15,7 +15,7 @@
 """ test ge frontend pass `DropoutForGE` `DropoutGradForGE` """
 import numpy as np
 
-import ge_infer_env  # pylint: disable=unused-import
+from tests.st.ge import ge_infer_env  # pylint: disable=unused-import
 from mindspore import ops, nn, context, Tensor
 from mindspore.ops.composite import GradOperation
 

tests/st/ge/ge_pass/pass_reduce_axis_update.py

@@ -14,7 +14,7 @@
 # ============================================================================
 import numpy as np
 
-import ge_infer_env  # pylint: disable=unused-import
+from tests.st.ge import ge_infer_env  # pylint: disable=unused-import
 import mindspore
 from mindspore import context, nn, Tensor, ops
 

tests/st/ge/ge_pass/pass_tensor_array.py

@@ -15,7 +15,7 @@
 """ test ge frontend pass and op `TensorArray`"""
 import numpy as np
 
-import ge_infer_env  # pylint: disable=unused-import
+from tests.st.ge import ge_infer_env  # pylint: disable=unused-import
 import mindspore.context as context
 from mindspore import nn
 from mindspore import Tensor

tests/st/ge/ge_pass/test_ge_convert.py

@@ -13,7 +13,7 @@
 # limitations under the License.
 # ============================================================================
 import pytest
-import ge_test_utils as utils
+import tests.st.ge.ge_test_utils as utils
 
 
 @pytest.mark.level0

tests/st/ge/ge_pass/test_ge_passes.py

@@ -13,7 +13,7 @@
 # limitations under the License.
 # ============================================================================
 import pytest
-import ge_test_utils as utils
+import tests.st.ge.ge_test_utils as utils
 
 
 @pytest.mark.level1

tests/st/ge/ge_test_utils.py

@@ -12,11 +12,14 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 # ============================================================================
+import inspect
 import os
 import sys
 
 
 def run_testcase(file_name, case_name=""):
+    caller_working_dir = os.path.dirname(inspect.stack()[1][1])
+    os.chdir(caller_working_dir)
     log_file = file_name + "_" + case_name + '.log'
     if case_name == "":
         ret = os.system(f'{sys.executable} {file_name}.py &> {log_file}')

tests/st/ge/ge_train_env.py

@@ -14,5 +14,6 @@
 # ============================================================================
 import os
 
+
 os.environ['MS_ENABLE_GE'] = '1'
 os.environ['MS_GE_TRAIN'] = '1'

tests/st/ge/ops/broadcast_gradient_args.py

@@ -0,0 +1,53 @@
+# Copyright 2022 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 tests.st.ge import ge_infer_env  # pylint: disable=unused-import
+import numpy as np
+import mindspore.nn as nn
+import mindspore.context as context
+from mindspore.ops.operations import _inner_ops
+
+
+context.set_context(mode=context.GRAPH_MODE, device_target="Ascend")
+
+
+class Net(nn.Cell):
+    def __init__(self):
+        super(Net, self).__init__()
+        self.args = _inner_ops.DynamicBroadcastGradientArgs()
+
+    def construct(self, s0, s1):
+        r = self.args(s0, s1)
+        return (r[0], r[1])
+
+
+def test_broadcast_gradient_args():
+    """
+    Feature: for DynamicBroadcastGradientArgs op
+    Description: inputs are two shapes
+    Expectation: the result is correct
+    """
+    shape0 = (4, 2, 1)
+    shape1 = (2, 7)
+    net = Net()
+    r0, r1 = net(shape0, shape1)
+    r0_expected = [2]
+    r1_expected = [0]
+
+    assert np.array_equal(r0_expected, r0.asnumpy())
+    assert np.array_equal(r1_expected, r1.asnumpy())
+
+
+if __name__ == "__main__":
+    test_broadcast_gradient_args()

tests/st/ge/ops/conv2d_backprop_filter.py

@@ -0,0 +1,81 @@
+# Copyright 2022 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 tests.st.ge import ge_infer_env  # pylint: disable=unused-import
+import numpy as np
+import mindspore.context as context
+import mindspore.nn as nn
+from mindspore import Tensor
+from mindspore.common.api import ms_function
+from mindspore.common.initializer import initializer
+from mindspore.common.parameter import Parameter
+from mindspore.ops import operations as P
+from mindspore.ops.operations import _grad_ops as G
+
+
+context.set_context(mode=context.GRAPH_MODE, device_target="Ascend")
+
+
+class Net(nn.Cell):
+    def __init__(self):
+        super(Net, self).__init__()
+        out_channel = 4
+        kernel_size = 1
+        self.conv_filter = G.Conv2DBackpropFilter(out_channel,
+                                                  kernel_size,
+                                                  pad_mode="valid",
+                                                  pad=0,
+                                                  mode=1,
+                                                  stride=1,
+                                                  dilation=1,
+                                                  group=1)
+        self.w = Parameter(
+            initializer(Tensor(np.array(
+                [[[[1, 0, -1], [1, 0, -1], [1, 0, -1]]]]).astype(np.float32)), [1, 1, 3, 3]),
+            name='w')
+        self.x = Parameter(initializer(Tensor(np.array([[[
+            [3, 0, 1, 2, 7, 4],
+            [1, 5, 8, 9, 3, 1],
+            [2, 7, 2, 5, 1, 3],
+            [0, 1, 3, 1, 7, 8],
+            [4, 2, 1, 6, 2, 8],
+            [2, 4, 5, 2, 3, 9]]]]).astype(np.float32)), [1, 1, 6, 6]), name='x')
+        self.out = Parameter(initializer(Tensor(np.array([[[
+            [-5, -4, 0, 8],
+            [-10, -2, 2, 3],
+            [0, -2, -4, -7],
+            [-3, -2, -3, -16]]]]).astype(np.float32)), [1, 1, 4, 4]), name='y')
+        self.get_shape = P.Shape()
+
+    @ms_function
+    def construct(self):
+        return self.conv_filter(self.out, self.x, self.get_shape(self.w))
+
+
+def test_conv2d_backprop_filter():
+    """
+    Feature: for Conv2DBackpropFilter op
+    Description: inputs are integers
+    Expectation: the result is correct
+    """
+    conv2d_filter = Net()
+    output = conv2d_filter()
+    expect = np.array([[[[-60, -142, -265],
+                         [-104, -211, -322],
+                         [-102, -144, -248]]]]).astype(np.float32)
+    assert (output.asnumpy() == expect).all()
+
+
+if __name__ == "__main__":
+    test_conv2d_backprop_filter()

tests/st/ge/ops/conv2d_backprop_input.py

@@ -0,0 +1,83 @@
+# Copyright 2022 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 tests.st.ge import ge_infer_env  # pylint: disable=unused-import
+import numpy as np
+import mindspore.context as context
+import mindspore.nn as nn
+from mindspore import Tensor
+from mindspore.common.api import ms_function
+from mindspore.common.initializer import initializer
+from mindspore.common.parameter import Parameter
+from mindspore.ops import operations as P
+
+
+context.set_context(mode=context.GRAPH_MODE, device_target="Ascend")
+
+
+class Net(nn.Cell):
+    def __init__(self):
+        super(Net, self).__init__()
+        out_channel = 4
+        kernel_size = 1
+        self.conv_input = P.Conv2DBackpropInput(out_channel,
+                                                kernel_size,
+                                                pad_mode="valid",
+                                                pad=0,
+                                                mode=1,
+                                                stride=1,
+                                                dilation=1,
+                                                group=1)
+        self.w = Parameter(
+            initializer(Tensor(np.array(
+                [[[[1, 0, -1], [1, 0, -1], [1, 0, -1]]]]).astype(np.float32)), [1, 1, 3, 3]),
+            name='w')
+        self.x = Parameter(initializer(Tensor(np.array([[[
+            [3, 0, 1, 2, 7, 4],
+            [1, 5, 8, 9, 3, 1],
+            [2, 7, 2, 5, 1, 3],
+            [0, 1, 3, 1, 7, 8],
+            [4, 2, 1, 6, 2, 8],
+            [2, 4, 5, 2, 3, 9]]]]).astype(np.float32)), [1, 1, 6, 6]), name='x')
+        self.out = Parameter(initializer(Tensor(np.array([[[
+            [-5, -4, 0, 8],
+            [-10, -2, 2, 3],
+            [0, -2, -4, -7],
+            [-3, -2, -3, -16]]]]).astype(np.float32)), [1, 1, 4, 4]), name='y')
+        self.get_shape = P.Shape()
+
+    @ms_function
+    def construct(self):
+        return self.conv_input(self.out, self.w, self.get_shape(self.x))
+
+
+def test_conv2d_backprop_input():
+    """
+    Feature: for Conv2DBackpropInput op
+    Description: inputs are integers
+    Expectation: the result is correct
+    """
+    conv2d_input = Net()
+    output = conv2d_input()
+    expect = np.array([[[[-5, -4, 5, 12, 0, -8],
+                         [-15, -6, 17, 17, -2, -11],
+                         [-15, -8, 13, 12, 2, -4],
+                         [-13, -6, 8, -14, 5, 20],
+                         [-3, -4, -4, -19, 7, 23],
+                         [-3, -2, 0, -14, 3, 16]]]]).astype(np.float32)
+    assert (output.asnumpy() == expect).all()
+
+
+if __name__ == "__main__":
+    test_conv2d_backprop_input()

tests/st/ge/ops/conv2d_transpose.py

@@ -0,0 +1,83 @@
+# Copyright 2022 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 tests.st.ge import ge_infer_env  # pylint: disable=unused-import
+import numpy as np
+import mindspore.context as context
+import mindspore.nn as nn
+from mindspore import Tensor
+from mindspore.common.api import ms_function
+from mindspore.common.initializer import initializer
+from mindspore.common.parameter import Parameter
+from mindspore.ops import operations as P
+
+
+context.set_context(mode=context.GRAPH_MODE, device_target="Ascend")
+
+
+class Net(nn.Cell):
+    def __init__(self):
+        super(Net, self).__init__()
+        out_channel = 4
+        kernel_size = 1
+        self.conv_input = P.Conv2DTranspose(out_channel,
+                                            kernel_size,
+                                            pad_mode="valid",
+                                            pad=0,
+                                            mode=1,
+                                            stride=1,
+                                            dilation=1,
+                                            group=1)
+        self.w = Parameter(
+            initializer(Tensor(np.array(
+                [[[[1, 0, -1], [1, 0, -1], [1, 0, -1]]]]).astype(np.float32)), [1, 1, 3, 3]),
+            name='w')
+        self.x = Parameter(initializer(Tensor(np.array([[[
+            [3, 0, 1, 2, 7, 4],
+            [1, 5, 8, 9, 3, 1],
+            [2, 7, 2, 5, 1, 3],
+            [0, 1, 3, 1, 7, 8],
+            [4, 2, 1, 6, 2, 8],
+            [2, 4, 5, 2, 3, 9]]]]).astype(np.float32)), [1, 1, 6, 6]), name='x')
+        self.out = Parameter(initializer(Tensor(np.array([[[
+            [-5, -4, 0, 8],
+            [-10, -2, 2, 3],
+            [0, -2, -4, -7],
+            [-3, -2, -3, -16]]]]).astype(np.float32)), [1, 1, 4, 4]), name='y')
+        self.get_shape = P.Shape()
+
+    @ms_function
+    def construct(self):
+        return self.conv_input(self.out, self.w, self.get_shape(self.x))
+
+
+def test_conv2d_transpose():
+    """
+    Feature: for Conv2DTranspose op
+    Description: inputs are integers
+    Expectation: the result is correct
+    """
+    conv2d_input = Net()
+    output = conv2d_input()
+    expect = np.array([[[[-5, -4, 5, 12, 0, -8],
+                         [-15, -6, 17, 17, -2, -11],
+                         [-15, -8, 13, 12, 2, -4],
+                         [-13, -6, 8, -14, 5, 20],
+                         [-3, -4, -4, -19, 7, 23],
+                         [-3, -2, 0, -14, 3, 16]]]]).astype(np.float32)
+    assert (output.asnumpy() == expect).all()
+
+
+if __name__ == "__main__":
+    test_conv2d_transpose()

tests/st/ge/ops/dynamic_broadcast_to.py

@@ -0,0 +1,54 @@
+# Copyright 2022 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 tests.st.ge import ge_infer_env  # pylint: disable=unused-import
+import numpy as np
+import mindspore.context as context
+import mindspore.common.dtype as mstype
+import mindspore.nn as nn
+import mindspore.ops.operations as ops
+from mindspore import Tensor
+from mindspore.ops.operations import _inner_ops as inner
+
+
+context.set_context(mode=context.GRAPH_MODE, device_target="Ascend")
+
+
+class Net(nn.Cell):
+    def __init__(self):
+        super(Net, self).__init__()
+        self.d_shape = ops.TensorShape()
+        self.d_broadcastto = inner.DynamicBroadcastTo()
+
+    def construct(self, data, shape):
+        shape = self.d_shape(shape)
+        return self.d_broadcastto(data, shape)
+
+
+def test_dynamic_broadcast_to():
+    """
+    Feature: for DynamicBroadcastTo op
+    Description: inputs are data and shape
+    Expectation: the result is correct
+    """
+    data = Tensor(np.array([1, 2, 3]), mstype.float32)
+    shape = Tensor(np.zeros((2, 3)), mstype.int64)
+    expect_data = np.array([[1, 2, 3], [1, 2, 3]]).astype(np.float32)
+    net = Net()
+    output = net(data, shape)
+    assert np.array_equal(output.asnumpy(), expect_data)
+
+
+if __name__ == "__main__":
+    test_dynamic_broadcast_to()

tests/st/ge/ops/test_ge_ops.py

@@ -0,0 +1,94 @@
+# Copyright 2022 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.
+# ============================================================================
+import pytest
+import tests.st.ge.ge_test_utils as utils
+
+
+@pytest.mark.level0
+@pytest.mark.platform_arm_ascend_training
+@pytest.mark.platform_x86_ascend_training
+@pytest.mark.env_onecard
+def test_broadcast_gradient_args():
+    """
+    Feature: for DynamicBroadcastGradientArgs op
+    Description: inputs are two shapes
+    Expectation: the result is correct
+    """
+    utils.run_testcase('broadcast_gradient_args')
+
+
+@pytest.mark.level0
+@pytest.mark.platform_arm_ascend_training
+@pytest.mark.platform_x86_ascend_training
+@pytest.mark.env_onecard
+def test_conv2d_backprop_filter():
+    """
+    Feature: for Conv2DBackpropFilter op
+    Description: inputs are integers
+    Expectation: the result is correct
+    """
+    utils.run_testcase('conv2d_backprop_filter')
+
+
+@pytest.mark.level0
+@pytest.mark.platform_arm_ascend_training
+@pytest.mark.platform_x86_ascend_training
+@pytest.mark.env_onecard
+def test_conv2d_backprop_input():
+    """
+    Feature: for Conv2DBackpropInput op
+    Description: inputs are integers
+    Expectation: the result is correct
+    """
+    utils.run_testcase('conv2d_backprop_input')
+
+
+@pytest.mark.level0
+@pytest.mark.platform_arm_ascend_training
+@pytest.mark.platform_x86_ascend_training
+@pytest.mark.env_onecard
+def test_conv2d_transpose():
+    """
+    Feature: for Conv2DTranspose op
+    Description: inputs are integers
+    Expectation: the result is correct
+    """
+    utils.run_testcase('conv2d_transpose')
+
+
+@pytest.mark.level0
+@pytest.mark.platform_arm_ascend_training
+@pytest.mark.platform_x86_ascend_training
+@pytest.mark.env_onecard
+def test_dynamic_broadcast_to():
+    """
+    Feature: for DynamicBroadcastTo op
+    Description: inputs are data and shape
+    Expectation: the result is correct
+    """
+    utils.run_testcase('dynamic_broadcast_to')
+
+
+@pytest.mark.level0
+@pytest.mark.platform_arm_ascend_training
+@pytest.mark.platform_x86_ascend_training
+@pytest.mark.env_onecard
+def test_unique():
+    """
+    Feature: for Unique op
+    Description: inputs are integers
+    Expectation: the result is correct
+    """
+    utils.run_testcase('unique')

tests/st/ge/ops/unique.py

@@ -0,0 +1,53 @@
+# Copyright 2022 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 tests.st.ge import ge_infer_env  # pylint: disable=unused-import
+import numpy as np
+import mindspore.context as context
+import mindspore.nn as nn
+from mindspore import Tensor
+import mindspore.common.dtype as mstype
+from mindspore.ops import operations as P
+
+
+context.set_context(mode=context.GRAPH_MODE, device_target="Ascend")
+
+
+class Net(nn.Cell):
+    def __init__(self):
+        super(Net, self).__init__()
+        self.unique = P.Unique()
+
+    def construct(self, x):
+        x = self.unique(x)
+        return (x[0], x[1])
+
+
+def test_unique():
+    """
+    Feature: for Unique op
+    Description: inputs are integers
+    Expectation: the result is correct
+    """
+    x = Tensor(np.array([1, 1, 2, 3, 3, 3]), mstype.int32)
+    unique = Net()
+    output = unique(x)
+    expect1 = np.array([1, 2, 3])
+    expect2 = np.array([0, 0, 1, 2, 2, 2])
+    assert (output[0].asnumpy() == expect1).all()
+    assert (output[1].asnumpy() == expect2).all()
+
+
+if __name__ == "__main__":
+    test_unique()