From 32022d4df0200284cb4ab2ee98cc9e380f31e62d Mon Sep 17 00:00:00 2001
From: willzhang4a58 <willzhang4a58@gmail.com>
Date: Fri, 18 Aug 2017 16:25:46 +0800
Subject: [PATCH] blob desc

---
 cmake/oneflow.cmake                           |  8 +--
 oneflow/core/graph/boxing_task_node.cpp       | 14 ++---
 oneflow/core/graph/boxing_task_node.h         |  6 +--
 oneflow/core/graph/copy_task_node.cpp         |  4 +-
 oneflow/core/graph/copy_task_node.h           |  2 +-
 oneflow/core/graph/data_comp_task_node.cpp    | 19 ++++---
 oneflow/core/graph/data_comp_task_node.h      |  6 +--
 oneflow/core/graph/exec_graph.cpp             |  6 +--
 oneflow/core/graph/exec_graph.h               |  2 +-
 .../graph/loss_accumulate_comp_task_node.cpp  |  4 +-
 .../graph/loss_accumulate_comp_task_node.h    |  2 +-
 .../core/graph/loss_record_comp_task_node.cpp |  3 +-
 .../core/graph/loss_record_comp_task_node.h   |  2 +-
 .../model_diff_accumulate_comp_task_node.cpp  |  4 +-
 .../model_diff_accumulate_comp_task_node.h    |  2 +-
 .../core/graph/model_save_comp_task_node.cpp  |  2 +-
 .../core/graph/model_save_comp_task_node.h    |  2 +-
 .../graph/model_update_comp_task_node.cpp     | 12 ++---
 .../core/graph/model_update_comp_task_node.h  |  2 +-
 oneflow/core/graph/task_graph.cpp             |  7 ++-
 oneflow/core/graph/task_graph.h               |  2 +-
 oneflow/core/graph/task_node.h                |  2 +-
 oneflow/core/job/compiler.cpp                 |  2 +-
 oneflow/core/operator/boxing_op.cpp           | 17 +++---
 oneflow/core/operator/boxing_op.h             |  4 +-
 oneflow/core/operator/boxing_op_test.cpp      |  6 +--
 oneflow/core/operator/clone_op.cpp            |  8 +--
 oneflow/core/operator/clone_op.h              |  4 +-
 oneflow/core/operator/clone_op_test.cpp       |  2 +-
 oneflow/core/operator/concat_op.cpp           | 23 ++++----
 oneflow/core/operator/concat_op.h             |  4 +-
 oneflow/core/operator/concat_op_test.cpp      |  2 +-
 oneflow/core/operator/convolution_op.cpp      | 43 +++++++--------
 oneflow/core/operator/convolution_op.h        |  4 +-
 oneflow/core/operator/convolution_op_test.cpp |  4 +-
 oneflow/core/operator/data_loader_op.cpp      |  8 +--
 oneflow/core/operator/data_loader_op.h        |  4 +-
 oneflow/core/operator/innerproduct_op.cpp     | 20 ++++---
 oneflow/core/operator/innerproduct_op.h       |  4 +-
 .../core/operator/innerproduct_op_test.cpp    |  4 +-
 oneflow/core/operator/model_update_op.h       |  4 +-
 .../operator/momentum_model_update_op.cpp     |  7 ++-
 .../core/operator/momentum_model_update_op.h  |  4 +-
 .../operator/multinomial_logistic_loss_op.cpp |  8 +--
 .../operator/multinomial_logistic_loss_op.h   |  4 +-
 .../multinomial_logistic_loss_op_test.cpp     |  2 +-
 oneflow/core/operator/operator.h              | 10 ++--
 oneflow/core/operator/pooling_op.cpp          | 26 +++++----
 oneflow/core/operator/pooling_op.h            |  4 +-
 oneflow/core/operator/pooling_op_test.cpp     |  2 +-
 oneflow/core/operator/record_op.h             |  4 +-
 oneflow/core/operator/relu_op.cpp             |  8 ++-
 oneflow/core/operator/relu_op.h               |  4 +-
 oneflow/core/operator/relu_op_test.cpp        |  2 +-
 .../core/operator/rmsprop_model_update_op.cpp |  7 ++-
 .../core/operator/rmsprop_model_update_op.h   |  4 +-
 oneflow/core/operator/softmax_loss_op.cpp     | 14 ++---
 oneflow/core/operator/softmax_loss_op.h       |  4 +-
 .../core/operator/softmax_loss_op_test.cpp    |  2 +-
 oneflow/core/operator/softmax_op.cpp          | 11 ++--
 oneflow/core/operator/softmax_op.h            |  4 +-
 oneflow/core/operator/softmax_op_test.cpp     |  2 +-
 oneflow/core/register/blob.h                  |  2 +-
 oneflow/core/register/blob_desc.h             | 30 +++++++++++
 oneflow/core/register/blob_desc.proto         |  8 +++
 oneflow/core/register/register_desc.cpp       | 53 ++++++++++---------
 oneflow/core/register/register_desc.h         | 21 ++++----
 oneflow/core/register/register_desc.proto     |  4 +-
 oneflow/core/register/register_manager.cpp    | 13 ++---
 .../core/register/runtime_register_desc.cpp   |  5 +-
 oneflow/core/register/runtime_register_desc.h |  8 +--
 71 files changed, 296 insertions(+), 261 deletions(-)
 create mode 100644 oneflow/core/register/blob_desc.h
 create mode 100644 oneflow/core/register/blob_desc.proto

diff --git a/cmake/oneflow.cmake b/cmake/oneflow.cmake
index 5cebccd794..2de16af20f 100644
--- a/cmake/oneflow.cmake
+++ b/cmake/oneflow.cmake
@@ -108,6 +108,8 @@ foreach(cc ${of_main_cc})
 endforeach()
 
 # build test
-cuda_add_executable(oneflow_testexe ${of_all_test_cc})
-target_link_libraries(oneflow_testexe ${of_libs} ${oneflow_third_party_libs})
-add_test(NAME oneflow_test COMMAND oneflow_testexe)
+if(BUILD_TESTING)
+  cuda_add_executable(oneflow_testexe ${of_all_test_cc})
+  target_link_libraries(oneflow_testexe ${of_libs} ${oneflow_third_party_libs})
+  add_test(NAME oneflow_test COMMAND oneflow_testexe)
+endif()
diff --git a/oneflow/core/graph/boxing_task_node.cpp b/oneflow/core/graph/boxing_task_node.cpp
index 2d445165f2..af700d13fe 100644
--- a/oneflow/core/graph/boxing_task_node.cpp
+++ b/oneflow/core/graph/boxing_task_node.cpp
@@ -173,11 +173,11 @@ void BoxingTaskNode::FwBuildChainSortedEdgesPair(
   }
 }
 
-void BoxingTaskNode::FwInferShapeOfBlobsInProducedRegsts(TaskGraph*) {
+void BoxingTaskNode::FwInferBlobDescInProducedRegsts(TaskGraph*) {
   exec_gph().ConstForEachNode([this](const ExecNode* exec_node) {
-    exec_node->op()->InferShape4FwBlobs(exec_node->GetMutShapePtr4BnInOpFunc(),
-                                        chain_node()->parallel_desc()->policy(),
-                                        0, 0);
+    exec_node->op()->InferBlobDesc4FwBlobs(
+        exec_node->GetBlobDesc4BnInOpFunc(),
+        chain_node()->parallel_desc()->policy(), 0, 0);
   });
 }
 
@@ -232,16 +232,16 @@ void BoxingTaskNode::BpBuildExecAndEnrollLbn2Regsts(TaskGraph*) {
   mut_exec_gph().UpdateSourceAndSink();
 }
 
-void BoxingTaskNode::BpInferShapeOfBlobsInProducedRegsts(TaskGraph*) {
+void BoxingTaskNode::BpInferBlobDescInProducedRegsts(TaskGraph*) {
   for (TaskEdge* fw_in_edge : GetFwNode()->in_edges()) {
     auto in_regst = GetRelatedRegst(fw_in_edge);
     if (auto in_diff_regst = GetBpRegstFromFwRegst(in_regst)) {
-      in_diff_regst->CopyShapeFrom(in_regst.get());
+      in_diff_regst->CopyBlobDescFrom(in_regst.get());
     }
   }
   auto fw_middle_regst = GetFwNode()->GetProducedRegstDesc("middle");
   auto bp_middle_regst = GetProducedRegstDesc("middle");
-  bp_middle_regst->CopyShapeFrom(fw_middle_regst.get());
+  bp_middle_regst->CopyBlobDescFrom(fw_middle_regst.get());
 }
 
 }  // namespace oneflow
diff --git a/oneflow/core/graph/boxing_task_node.h b/oneflow/core/graph/boxing_task_node.h
index 5101ee2443..ecb50731cb 100644
--- a/oneflow/core/graph/boxing_task_node.h
+++ b/oneflow/core/graph/boxing_task_node.h
@@ -41,12 +41,12 @@ class BoxingTaskNode : public TaskNode {
 
  private:
   OVERRIDE_IF_FW_BP_FOR_FUNC(BuildExecAndEnrollLbn2Regsts);
-  OVERRIDE_IF_FW_BP_FOR_FUNC(InferShapeOfBlobsInProducedRegsts);
+  OVERRIDE_IF_FW_BP_FOR_FUNC(InferBlobDescInProducedRegsts);
 
   void FwBuildExecAndEnrollLbn2Regsts(TaskGraph*);
-  void FwInferShapeOfBlobsInProducedRegsts(TaskGraph*);
+  void FwInferBlobDescInProducedRegsts(TaskGraph*);
   void BpBuildExecAndEnrollLbn2Regsts(TaskGraph*);
-  void BpInferShapeOfBlobsInProducedRegsts(TaskGraph*);
+  void BpInferBlobDescInProducedRegsts(TaskGraph*);
 
   void EnrollAllRegstAndBindRelatedEdge();
   TaskType task_type() const override { return kBoxingTask; }
diff --git a/oneflow/core/graph/copy_task_node.cpp b/oneflow/core/graph/copy_task_node.cpp
index 0d8d3d8792..39ba616de8 100644
--- a/oneflow/core/graph/copy_task_node.cpp
+++ b/oneflow/core/graph/copy_task_node.cpp
@@ -25,10 +25,10 @@ void CopyTaskNode::BuildExecAndEnrollLbn2Regsts(TaskGraph*) {
   mut_exec_gph().UpdateSourceAndSink();
 }
 
-void CopyTaskNode::InferShapeOfBlobsInProducedRegsts(TaskGraph*) {
+void CopyTaskNode::InferBlobDescInProducedRegsts(TaskGraph*) {
   std::shared_ptr<RegstDesc> in_regst = GetRelatedRegst(SoleInEdge());
   std::shared_ptr<RegstDesc> out_regst = GetRelatedRegst(SoleOutEdge());
-  out_regst->CopyShapeFrom(in_regst.get());
+  out_regst->CopyBlobDescFrom(in_regst.get());
 }
 
 void CopyHDTaskNode::SetFwInCopy() {
diff --git a/oneflow/core/graph/copy_task_node.h b/oneflow/core/graph/copy_task_node.h
index 7c473c2c8f..ff3904a5a5 100644
--- a/oneflow/core/graph/copy_task_node.h
+++ b/oneflow/core/graph/copy_task_node.h
@@ -16,7 +16,7 @@ class CopyTaskNode : public TaskNode {
 
  private:
   void BuildExecAndEnrollLbn2Regsts(TaskGraph*) override;
-  void InferShapeOfBlobsInProducedRegsts(TaskGraph*) override;
+  void InferBlobDescInProducedRegsts(TaskGraph*) override;
 };
 
 class CopyHDTaskNode final : public CopyTaskNode {
diff --git a/oneflow/core/graph/data_comp_task_node.cpp b/oneflow/core/graph/data_comp_task_node.cpp
index bd5485eb3f..f705ed0300 100644
--- a/oneflow/core/graph/data_comp_task_node.cpp
+++ b/oneflow/core/graph/data_comp_task_node.cpp
@@ -24,17 +24,16 @@ void DataCompTaskNode::FwBuildExecAndEnrollLbn2Regsts(TaskGraph*) {
   FwEnrollLbn2ModelAndTmpRegsts();  // model model_tmp data_tmp
 }
 
-void DataCompTaskNode::FwInferShapeOfBlobsInProducedRegsts(TaskGraph*) {
+void DataCompTaskNode::FwInferBlobDescInProducedRegsts(TaskGraph*) {
   exec_gph().ConstTopoForEachNode([this](const ExecNode* node) {
-    node->op()->InferShape4FwBlobs(
-        node->GetMutShapePtr4BnInOpFunc(),
-        chain_node()->parallel_desc()->policy(), parallel_id(),
-        chain_node()->parallel_desc()->parallel_num());
+    node->op()->InferBlobDesc4FwBlobs(
+        node->GetBlobDesc4BnInOpFunc(), chain_node()->parallel_desc()->policy(),
+        parallel_id(), chain_node()->parallel_desc()->parallel_num());
   });
   if (IsLossNode()) {
     auto out_regst = GetRelatedRegst(SoleOutEdge());
     auto in_regst = GetRelatedRegst(SoleInEdge());
-    out_regst->CopyShapeFrom(in_regst.get());
+    out_regst->CopyBlobDescFrom(in_regst.get());
   }
 }
 
@@ -173,20 +172,20 @@ void DataCompTaskNode::BpBuildExecAndEnrollLbn2Regsts(TaskGraph*) {
   BpEnrollLbn2ProducedRegst();
 }
 
-void DataCompTaskNode::BpInferShapeOfBlobsInProducedRegsts(TaskGraph*) {
+void DataCompTaskNode::BpInferBlobDescInProducedRegsts(TaskGraph*) {
   // in_diff_regst
   auto in_diff_regst = GetProducedRegstDesc("in_diff");
   auto in_regst = GetRelatedRegst(GetFwNode()->SoleInEdge());
-  in_diff_regst->CopyShapeFrom(in_regst.get());
+  in_diff_regst->CopyBlobDescFrom(in_regst.get());
   // model_diff_regst
   if (auto md_diff_regst = GetProducedRegstDesc("model_diff")) {
-    md_diff_regst->CopyShapeFrom(
+    md_diff_regst->CopyBlobDescFrom(
         GetFwNode()->GetConsumedRegstDesc("model").get());
   }
   // activation_diff_regst
   if (auto acti_diff_regst = GetProducedRegstDesc("activation_diff")) {
     auto acti_regst = GetFwNode()->GetProducedRegstDesc("activation");
-    acti_diff_regst->CopyShapeFrom(acti_regst.get());
+    acti_diff_regst->CopyBlobDescFrom(acti_regst.get());
   }
 }
 
diff --git a/oneflow/core/graph/data_comp_task_node.h b/oneflow/core/graph/data_comp_task_node.h
index 2dc336bb51..eb5e35055d 100644
--- a/oneflow/core/graph/data_comp_task_node.h
+++ b/oneflow/core/graph/data_comp_task_node.h
@@ -41,12 +41,12 @@ class DataCompTaskNode final : public CompTaskNode {
 
  private:
   OVERRIDE_IF_FW_BP_FOR_FUNC(BuildExecAndEnrollLbn2Regsts);
-  OVERRIDE_IF_FW_BP_FOR_FUNC(InferShapeOfBlobsInProducedRegsts);
+  OVERRIDE_IF_FW_BP_FOR_FUNC(InferBlobDescInProducedRegsts);
 
   using Lbn2NodeBnMap = HashMap<std::string, std::pair<ExecNode*, std::string>>;
 
   void FwBuildExecAndEnrollLbn2Regsts(TaskGraph* gph);
-  void FwInferShapeOfBlobsInProducedRegsts(TaskGraph* gph);
+  void FwInferBlobDescInProducedRegsts(TaskGraph* gph);
   void FwBuildFromUserOps(Lbn2NodeBnMap* lbn2producer,
                           Lbn2NodeBnMap* extern_in_lbn2consumer);
   void FwSetExecNodeFromInRegst(const Lbn2NodeBnMap& extern_in_lbn2consumer);
@@ -56,7 +56,7 @@ class DataCompTaskNode final : public CompTaskNode {
   void FwEnrollLbn2ActivationRegst();
   void FwEnrollLbn2ModelAndTmpRegsts();
   void BpBuildExecAndEnrollLbn2Regsts(TaskGraph*);
-  void BpInferShapeOfBlobsInProducedRegsts(TaskGraph*);
+  void BpInferBlobDescInProducedRegsts(TaskGraph*);
   void BpBuildExecGraph();
   void BpEnrollLbn2ProducedRegst();
   void BpEnrollLbn2ActivationDiffRegst();
diff --git a/oneflow/core/graph/exec_graph.cpp b/oneflow/core/graph/exec_graph.cpp
index 7b1affb9df..b49859119d 100644
--- a/oneflow/core/graph/exec_graph.cpp
+++ b/oneflow/core/graph/exec_graph.cpp
@@ -4,14 +4,14 @@ namespace oneflow {
 
 void ExecEdge::set_lbn(const std::string& lbn) { lbn_ = lbn; }
 
-std::function<Shape*(const std::string&)> ExecNode::GetMutShapePtr4BnInOpFunc()
+std::function<BlobDesc*(const std::string&)> ExecNode::GetBlobDesc4BnInOpFunc()
     const {
-  return [this](const std::string& bn_in_op) -> Shape* {
+  return [this](const std::string& bn_in_op) -> BlobDesc* {
     auto it = this->bn_in_op2regst_.find(bn_in_op);
     if (it == this->bn_in_op2regst_.end()) { return nullptr; }
     std::shared_ptr<RegstDesc> regst = it->second.lock();
     const std::string& lbn = this->op()->Lbn4BnInOp(bn_in_op);
-    return regst->GetMutShapePtr(lbn);
+    return regst->GetMutBlobDesc(lbn);
   };
 }
 
diff --git a/oneflow/core/graph/exec_graph.h b/oneflow/core/graph/exec_graph.h
index 28ec35c062..abaaf79920 100644
--- a/oneflow/core/graph/exec_graph.h
+++ b/oneflow/core/graph/exec_graph.h
@@ -55,7 +55,7 @@ class ExecNode final : public Node<ExecNode, ExecEdge> {
     return bn_in_op2regst_;
   }
 
-  std::function<Shape*(const std::string&)> GetMutShapePtr4BnInOpFunc() const;
+  std::function<BlobDesc*(const std::string&)> GetBlobDesc4BnInOpFunc() const;
 
   std::string VisualStr() const { return op_->op_name(); }
 
diff --git a/oneflow/core/graph/loss_accumulate_comp_task_node.cpp b/oneflow/core/graph/loss_accumulate_comp_task_node.cpp
index 3111000cd7..d4b6a4f09a 100644
--- a/oneflow/core/graph/loss_accumulate_comp_task_node.cpp
+++ b/oneflow/core/graph/loss_accumulate_comp_task_node.cpp
@@ -22,11 +22,11 @@ void LossAccCompTaskNode::BuildExecAndEnrollLbn2Regsts(TaskGraph* gph) {
   mut_exec_gph().UpdateSourceAndSink();
 }
 
-void LossAccCompTaskNode::InferShapeOfBlobsInProducedRegsts(TaskGraph* gph) {
+void LossAccCompTaskNode::InferBlobDescInProducedRegsts(TaskGraph* gph) {
   if (!chain_node()->op_vec().empty()) {
     auto loss_regst = GetConsumedRegstDesc("loss");
     auto loss_acc_regst = GetProducedRegstDesc("loss_acc");
-    loss_acc_regst->CopyShapeFrom(loss_regst.get());
+    loss_acc_regst->CopyBlobDescFrom(loss_regst.get());
   }
 }
 
diff --git a/oneflow/core/graph/loss_accumulate_comp_task_node.h b/oneflow/core/graph/loss_accumulate_comp_task_node.h
index fee4da89b0..edf8151382 100644
--- a/oneflow/core/graph/loss_accumulate_comp_task_node.h
+++ b/oneflow/core/graph/loss_accumulate_comp_task_node.h
@@ -13,7 +13,7 @@ class LossAccCompTaskNode final : public CompTaskNode {
 
  private:
   void BuildExecAndEnrollLbn2Regsts(TaskGraph* gph) override;
-  void InferShapeOfBlobsInProducedRegsts(TaskGraph* gph) override;
+  void InferBlobDescInProducedRegsts(TaskGraph* gph) override;
   TaskType task_type() const override { return kLossAccCompTask; }
   std::unique_ptr<TaskNode> CreateSameTypeNode() const override {
     return of_make_unique<LossAccCompTaskNode>();
diff --git a/oneflow/core/graph/loss_record_comp_task_node.cpp b/oneflow/core/graph/loss_record_comp_task_node.cpp
index 0fd6707303..f84bff7112 100644
--- a/oneflow/core/graph/loss_record_comp_task_node.cpp
+++ b/oneflow/core/graph/loss_record_comp_task_node.cpp
@@ -20,7 +20,6 @@ void LossRecordCompTaskNode::BuildExecAndEnrollLbn2Regsts(TaskGraph* gph) {
   mut_exec_gph().UpdateSourceAndSink();
 }
 
-void LossRecordCompTaskNode::InferShapeOfBlobsInProducedRegsts(TaskGraph* gph) {
-}
+void LossRecordCompTaskNode::InferBlobDescInProducedRegsts(TaskGraph* gph) {}
 
 }  // namespace oneflow
diff --git a/oneflow/core/graph/loss_record_comp_task_node.h b/oneflow/core/graph/loss_record_comp_task_node.h
index fea270d6c9..7387cf97a7 100644
--- a/oneflow/core/graph/loss_record_comp_task_node.h
+++ b/oneflow/core/graph/loss_record_comp_task_node.h
@@ -13,7 +13,7 @@ class LossRecordCompTaskNode final : public CompTaskNode {
 
  private:
   void BuildExecAndEnrollLbn2Regsts(TaskGraph* gph) override;
-  void InferShapeOfBlobsInProducedRegsts(TaskGraph* gph) override;
+  void InferBlobDescInProducedRegsts(TaskGraph* gph) override;
   bool IsMeaningLess() const override {
     return !GetConsumedRegstDesc("loss_acc");
   }
diff --git a/oneflow/core/graph/model_diff_accumulate_comp_task_node.cpp b/oneflow/core/graph/model_diff_accumulate_comp_task_node.cpp
index 53bd780c41..3a00d46360 100644
--- a/oneflow/core/graph/model_diff_accumulate_comp_task_node.cpp
+++ b/oneflow/core/graph/model_diff_accumulate_comp_task_node.cpp
@@ -35,13 +35,13 @@ void MdDiffAccCompTaskNode::BuildExecAndEnrollLbn2Regsts(TaskGraph* gph) {
   mut_exec_gph().UpdateSourceAndSink();
 }
 
-void MdDiffAccCompTaskNode::InferShapeOfBlobsInProducedRegsts(TaskGraph* gph) {
+void MdDiffAccCompTaskNode::InferBlobDescInProducedRegsts(TaskGraph* gph) {
   CHECK(IsFwNode());
   if (!chain_node()->op_vec().empty()) {
     std::shared_ptr<RegstDesc> in_regst = GetConsumedRegstDesc("model_diff");
     std::shared_ptr<RegstDesc> out_regst =
         GetProducedRegstDesc("model_diff_acc");
-    out_regst->CopyShapeFrom(in_regst.get());
+    out_regst->CopyBlobDescFrom(in_regst.get());
   }
 }
 
diff --git a/oneflow/core/graph/model_diff_accumulate_comp_task_node.h b/oneflow/core/graph/model_diff_accumulate_comp_task_node.h
index c5af6371c5..d8c6ef34a4 100644
--- a/oneflow/core/graph/model_diff_accumulate_comp_task_node.h
+++ b/oneflow/core/graph/model_diff_accumulate_comp_task_node.h
@@ -19,7 +19,7 @@ class MdDiffAccCompTaskNode final : public CompTaskNode {
 
  private:
   void BuildExecAndEnrollLbn2Regsts(TaskGraph* gph) override;
-  void InferShapeOfBlobsInProducedRegsts(TaskGraph* gph) override;
+  void InferBlobDescInProducedRegsts(TaskGraph* gph) override;
   TaskType task_type() const override { return kMdDiffAccCompTask; }
   std::unique_ptr<TaskNode> CreateSameTypeNode() const override {
     return of_make_unique<MdDiffAccCompTaskNode>();
diff --git a/oneflow/core/graph/model_save_comp_task_node.cpp b/oneflow/core/graph/model_save_comp_task_node.cpp
index 111583c8db..293cafee43 100644
--- a/oneflow/core/graph/model_save_comp_task_node.cpp
+++ b/oneflow/core/graph/model_save_comp_task_node.cpp
@@ -32,7 +32,7 @@ void MdSaveCompTaskNode::BuildExecAndEnrollLbn2Regsts(TaskGraph* gph) {
   }
 }
 
-void MdSaveCompTaskNode::InferShapeOfBlobsInProducedRegsts(TaskGraph* gph) {
+void MdSaveCompTaskNode::InferBlobDescInProducedRegsts(TaskGraph* gph) {
   CHECK(IsFwNode());
 }
 
diff --git a/oneflow/core/graph/model_save_comp_task_node.h b/oneflow/core/graph/model_save_comp_task_node.h
index 1031d6b374..d6bfd624de 100644
--- a/oneflow/core/graph/model_save_comp_task_node.h
+++ b/oneflow/core/graph/model_save_comp_task_node.h
@@ -22,7 +22,7 @@ class MdSaveCompTaskNode final : public CompTaskNode {
 
  private:
   void BuildExecAndEnrollLbn2Regsts(TaskGraph* gph) override;
-  void InferShapeOfBlobsInProducedRegsts(TaskGraph* gph) override;
+  void InferBlobDescInProducedRegsts(TaskGraph* gph) override;
   bool IsMeaningLess() const override { return !GetConsumedRegstDesc("model"); }
 
   TaskType task_type() const override { return kMdSaveCompTask; }
diff --git a/oneflow/core/graph/model_update_comp_task_node.cpp b/oneflow/core/graph/model_update_comp_task_node.cpp
index 9dbf88915f..0220874253 100644
--- a/oneflow/core/graph/model_update_comp_task_node.cpp
+++ b/oneflow/core/graph/model_update_comp_task_node.cpp
@@ -34,17 +34,17 @@ void MdUpdtCompTaskNode::BuildExecAndEnrollLbn2Regsts(TaskGraph* gph) {
   mut_exec_gph().UpdateSourceAndSink();
 }
 
-void MdUpdtCompTaskNode::InferShapeOfBlobsInProducedRegsts(TaskGraph* gph) {
+void MdUpdtCompTaskNode::InferBlobDescInProducedRegsts(TaskGraph* gph) {
   CHECK(IsFwNode());
   ExecNode* exec_node = exec_gph().SoleNode();
   auto model_diffs_regst = GetConsumedRegstDesc("model_diffs");
-  Shape packed_model_diffs_shape({model_diffs_regst->CompElemCntOfAllBlob()});
-  exec_node->op()->InferShape4FwBlobs(
-      [&](const std::string& bn_in_op) -> Shape* {
+  BlobDesc packed_blob_desc = model_diffs_regst->CompPackedBlobDesc();
+  exec_node->op()->InferBlobDesc4FwBlobs(
+      [&](const std::string& bn_in_op) -> BlobDesc* {
         if (bn_in_op == "model_diffs") {
-          return &packed_model_diffs_shape;
+          return &packed_blob_desc;
         } else {
-          return exec_node->GetMutShapePtr4BnInOpFunc()(bn_in_op);
+          return exec_node->GetBlobDesc4BnInOpFunc()(bn_in_op);
         }
       },
       kDataParallel, 0, 0);
diff --git a/oneflow/core/graph/model_update_comp_task_node.h b/oneflow/core/graph/model_update_comp_task_node.h
index 9f68eec5b1..80667d4577 100644
--- a/oneflow/core/graph/model_update_comp_task_node.h
+++ b/oneflow/core/graph/model_update_comp_task_node.h
@@ -35,7 +35,7 @@ class MdUpdtCompTaskNode final : public CompTaskNode {
 
  private:
   void BuildExecAndEnrollLbn2Regsts(TaskGraph* gph) override;
-  void InferShapeOfBlobsInProducedRegsts(TaskGraph* gph) override;
+  void InferBlobDescInProducedRegsts(TaskGraph* gph) override;
   TaskType task_type() const override { return kMdUpdtCompTask; }
   std::unique_ptr<TaskNode> CreateSameTypeNode() const override {
     return of_make_unique<MdUpdtCompTaskNode>();
diff --git a/oneflow/core/graph/task_graph.cpp b/oneflow/core/graph/task_graph.cpp
index 00fe01ef5d..b2da096f75 100644
--- a/oneflow/core/graph/task_graph.cpp
+++ b/oneflow/core/graph/task_graph.cpp
@@ -33,10 +33,9 @@ void TaskGraph::BuildExecAndEnrollLbn2Regsts() {
       [this](TaskNode* node) { node->BuildExecAndEnrollLbn2Regsts(this); });
 }
 
-void TaskGraph::InferShapeOfBlobsInProducedRegsts() {
-  TopoForEachNode([this](TaskNode* node) {
-    node->InferShapeOfBlobsInProducedRegsts(this);
-  });
+void TaskGraph::InferBlobDescInProducedRegsts() {
+  TopoForEachNode(
+      [this](TaskNode* node) { node->InferBlobDescInProducedRegsts(this); });
 }
 
 std::vector<CompTaskNode*> TaskGraph::CompTasksInChain(const ChainNode* chain) {
diff --git a/oneflow/core/graph/task_graph.h b/oneflow/core/graph/task_graph.h
index ac89465f59..241c18b575 100644
--- a/oneflow/core/graph/task_graph.h
+++ b/oneflow/core/graph/task_graph.h
@@ -22,7 +22,7 @@ class TaskGraph : public Graph<TaskNode, TaskEdge> {
   const ChainGraph* chain_gph() const { return stage_gph_->chain_gph(); }
   std::vector<CompTaskNode*> CompTasksInChain(const ChainNode*);
 
-  void InferShapeOfBlobsInProducedRegsts();
+  void InferBlobDescInProducedRegsts();
 
   const std::string& name() const { return name_; }
 
diff --git a/oneflow/core/graph/task_node.h b/oneflow/core/graph/task_node.h
index e6209ca452..9a428b48b1 100644
--- a/oneflow/core/graph/task_node.h
+++ b/oneflow/core/graph/task_node.h
@@ -42,7 +42,7 @@ class TaskNode : public Node<TaskNode, TaskEdge> {
 
   //
   virtual void BuildExecAndEnrollLbn2Regsts(TaskGraph*) = 0;
-  virtual void InferShapeOfBlobsInProducedRegsts(TaskGraph*) = 0;
+  virtual void InferBlobDescInProducedRegsts(TaskGraph*) = 0;
 
 #define OVERRIDE_IF_FW_BP_FOR_FUNC(func_name) \
   void func_name(TaskGraph* gph) override {   \
diff --git a/oneflow/core/job/compiler.cpp b/oneflow/core/job/compiler.cpp
index 9133b781a0..50729e2c27 100644
--- a/oneflow/core/job/compiler.cpp
+++ b/oneflow/core/job/compiler.cpp
@@ -168,7 +168,7 @@ void Compiler::BuildLossGraph(
 void Compiler::InferShape4Regsts() {
   for (auto& task_gph : ordered_task_gphs_) {
     LOG(INFO) << "InferShape for " << task_gph->name();
-    task_gph->InferShapeOfBlobsInProducedRegsts();
+    task_gph->InferBlobDescInProducedRegsts();
   }
 }
 
diff --git a/oneflow/core/operator/boxing_op.cpp b/oneflow/core/operator/boxing_op.cpp
index e4b296dd26..b0e31864d8 100644
--- a/oneflow/core/operator/boxing_op.cpp
+++ b/oneflow/core/operator/boxing_op.cpp
@@ -31,13 +31,13 @@ std::string BoxingOp::obn2lbn(const std::string& output_bn) const {
   return GetStringFromSpecialConf("lbn");
 }
 
-void BoxingOp::InferShape4FwBlobs(
-    std::function<Shape*(const std::string&)> GetShapePtr4BnInOp,
+void BoxingOp::InferBlobDesc4FwBlobs(
+    std::function<BlobDesc*(const std::string)> GetBlobDesc4BnInOp,
     ParallelPolicy policy, int64_t parallel_id, int64_t parallel_num) const {
   auto boxing_conf = op_conf().boxing_conf();
   auto in_box_case = boxing_conf.in_box_case();
   std::vector<int64_t> data_tmp_blob_shape_vec =
-      GetShapePtr4BnInOp(input_bns().at(0))->dim_vec();
+      GetBlobDesc4BnInOp(input_bns().at(0))->shape().dim_vec();
 
   // if it is a concat-box, accumulate the dimensions on concat-axis.
   // otherwise only check all boxes are in the same shape.
@@ -47,7 +47,8 @@ void BoxingOp::InferShape4FwBlobs(
     CHECK(concat_axis == 0 || concat_axis == 1);
   }
   for (size_t ib_idx = 1; ib_idx < input_bns().size(); ++ib_idx) {
-    auto ib_shape_vec = GetShapePtr4BnInOp(input_bns().at(ib_idx))->dim_vec();
+    auto ib_shape_vec =
+        GetBlobDesc4BnInOp(input_bns().at(ib_idx))->shape().dim_vec();
     for (size_t i = 0; i < ib_shape_vec.size(); ++i) {
       if (in_box_case == BoxingOpConf::kConcatBox && i == concat_axis) {
         data_tmp_blob_shape_vec[i] += ib_shape_vec[i];
@@ -61,7 +62,8 @@ void BoxingOp::InferShape4FwBlobs(
   // it is stored back if and only if this is a concat-clone box
   if (in_box_case == BoxingOpConf::kConcatBox
       && out_box_case == BoxingOpConf::kCloneBox) {
-    *GetShapePtr4BnInOp(SoleDtbn()) = Shape(data_tmp_blob_shape_vec);
+    GetBlobDesc4BnInOp(SoleDtbn())->mut_shape() =
+        Shape(data_tmp_blob_shape_vec);
   }
   CHECK_NE(out_box_case, BoxingOpConf::OUT_BOX_NOT_SET);
   if (out_box_case == BoxingOpConf::kDataSplitBox) {
@@ -70,11 +72,12 @@ void BoxingOp::InferShape4FwBlobs(
     auto output_shape_vec = data_tmp_blob_shape_vec;
     for (size_t i = 0; i < out_num; ++i) {
       output_shape_vec[0] = splitter.At(i).size();
-      *GetShapePtr4BnInOp(output_bns()[i]) = Shape(output_shape_vec);
+      GetBlobDesc4BnInOp(output_bns()[i])->mut_shape() =
+          Shape(output_shape_vec);
     }
   } else if (out_box_case == BoxingOpConf::kCloneBox) {
     for (auto obn : output_bns()) {
-      *GetShapePtr4BnInOp(obn) = Shape(data_tmp_blob_shape_vec);
+      GetBlobDesc4BnInOp(obn)->mut_shape() = Shape(data_tmp_blob_shape_vec);
     }
   } else {
     UNEXPECTED_RUN();
diff --git a/oneflow/core/operator/boxing_op.h b/oneflow/core/operator/boxing_op.h
index 9b39639984..1ffb1cb707 100644
--- a/oneflow/core/operator/boxing_op.h
+++ b/oneflow/core/operator/boxing_op.h
@@ -14,8 +14,8 @@ class BoxingOp final : public SysOperator {
   void InitFromOpConf(const OperatorConf& op_conf) override;
   const PbMessage& GetSpecialConf() const override;
 
-  void InferShape4FwBlobs(
-      std::function<Shape*(const std::string&)> GetShapePtr4BnInOp,
+  void InferBlobDesc4FwBlobs(
+      std::function<BlobDesc*(const std::string)> GetBlobDesc4BnInOp,
       ParallelPolicy policy, int64_t parallel_id,
       int64_t parallel_num) const override;
 
diff --git a/oneflow/core/operator/boxing_op_test.cpp b/oneflow/core/operator/boxing_op_test.cpp
index a3ade1b7b2..1583b88b69 100644
--- a/oneflow/core/operator/boxing_op_test.cpp
+++ b/oneflow/core/operator/boxing_op_test.cpp
@@ -36,7 +36,7 @@ TEST(BoxingOp, box_4_10x5x6x6) {
   };
 
   // do infer shape
-  boxing_op->InferShape4FwBlobs(fp, kModelParallel, 0, 1);
+  boxing_op->InferBlobDesc4FwBlobs(fp, kModelParallel, 0, 1);
 
   // test results
   // output_shape should be:
@@ -58,7 +58,7 @@ TEST(BoxingOp, box_4_10x5x6x6) {
   boxing_op = ConstructOp(op_conf);
 
   // do infer shape
-  boxing_op->InferShape4FwBlobs(fp, kModelParallel, 0, 1);
+  boxing_op->InferBlobDesc4FwBlobs(fp, kModelParallel, 0, 1);
 
   // test results
   // output shape should be the same as input
@@ -75,7 +75,7 @@ TEST(BoxingOp, box_4_10x5x6x6) {
   boxing_op = ConstructOp(op_conf);
 
   // do infer shape
-  boxing_op->InferShape4FwBlobs(fp, kModelParallel, 0, 1);
+  boxing_op->InferBlobDesc4FwBlobs(fp, kModelParallel, 0, 1);
 
   // data_tmp_shape is {10, 17, 6, 6}, and the 17 = 4 + 4 + 4 + 5
   Shape* data_tmp_shape_ptr = bn2shape_ptr.at(boxing_op->SoleDtbn());
diff --git a/oneflow/core/operator/clone_op.cpp b/oneflow/core/operator/clone_op.cpp
index f49f6abf3d..cd3ba125dc 100644
--- a/oneflow/core/operator/clone_op.cpp
+++ b/oneflow/core/operator/clone_op.cpp
@@ -16,12 +16,12 @@ const PbMessage& CloneOp::GetSpecialConf() const {
   return op_conf().clone_conf();
 }
 
-void CloneOp::InferShape4FwBlobs(
-    std::function<Shape*(const std::string&)> GetShapePtr4BnInOp,
+void CloneOp::InferBlobDesc4FwBlobs(
+    std::function<BlobDesc*(const std::string)> GetBlobDesc4BnInOp,
     ParallelPolicy policy, int64_t parallel_id, int64_t parallel_num) const {
-  Shape* input_shape_ptr = GetShapePtr4BnInOp(SoleIbn());
+  const BlobDesc* input_blob_desc = GetBlobDesc4BnInOp(SoleIbn());
   for (std::string obn : output_bns()) {
-    *GetShapePtr4BnInOp(obn) = *input_shape_ptr;
+    *GetBlobDesc4BnInOp(obn) = *input_blob_desc;
   }
 }
 
diff --git a/oneflow/core/operator/clone_op.h b/oneflow/core/operator/clone_op.h
index 9c70a598b5..1c7c1f5b8a 100644
--- a/oneflow/core/operator/clone_op.h
+++ b/oneflow/core/operator/clone_op.h
@@ -15,8 +15,8 @@ class CloneOp final : public SysOperator {
 
   void InitFromOpConf(const OperatorConf& op_conf) override;
   const PbMessage& GetSpecialConf() const override;
-  void InferShape4FwBlobs(
-      std::function<Shape*(const std::string&)> GetShapePtr4BnInOp,
+  void InferBlobDesc4FwBlobs(
+      std::function<BlobDesc*(const std::string)> GetBlobDesc4BnInOp,
       ParallelPolicy policy, int64_t parallel_id,
       int64_t parallel_num) const override;
 
diff --git a/oneflow/core/operator/clone_op_test.cpp b/oneflow/core/operator/clone_op_test.cpp
index 457bcf5fa0..e534bbc811 100644
--- a/oneflow/core/operator/clone_op_test.cpp
+++ b/oneflow/core/operator/clone_op_test.cpp
@@ -18,7 +18,7 @@ TEST(CloneOp, clone_4x3_3_times) {
     return bn2shape_ptr.at(bn);
   };
 
-  clone_op->InferShape4FwBlobs(fp, kDataParallel, 3, 10);
+  clone_op->InferBlobDesc4FwBlobs(fp, kDataParallel, 3, 10);
 
   Shape* input_shape_ptr = bn2shape_ptr.at(clone_op->SoleIbn());
   for (std::string obn : clone_op->output_bns()) {
diff --git a/oneflow/core/operator/concat_op.cpp b/oneflow/core/operator/concat_op.cpp
index 5b790c8684..a8ce7ceba8 100644
--- a/oneflow/core/operator/concat_op.cpp
+++ b/oneflow/core/operator/concat_op.cpp
@@ -18,23 +18,26 @@ const PbMessage& ConcatOp::GetSpecialConf() const {
   return op_conf().concat_conf();
 }
 
-void ConcatOp::InferShape4FwBlobs(
-    std::function<Shape*(const std::string&)> GetShapePtr4BnInOp,
+void ConcatOp::InferBlobDesc4FwBlobs(
+    std::function<BlobDesc*(const std::string)> GetBlobDesc4BnInOp,
     ParallelPolicy policy, int64_t parallel_id, int64_t parallel_num) const {
-  std::vector<int64_t> vec = GetShapePtr4BnInOp(input_bns().at(0))->dim_vec();
+  std::vector<int64_t> vec =
+      GetBlobDesc4BnInOp(input_bns().at(0))->shape().dim_vec();
   for (size_t ibn_idx = 1; ibn_idx < input_bns().size(); ++ibn_idx) {
-    Shape* ib_shape = GetShapePtr4BnInOp(input_bns().at(ibn_idx));
+    const Shape& ib_shape =
+        GetBlobDesc4BnInOp(input_bns().at(ibn_idx))->shape();
     int32_t concat_axis = op_conf().concat_conf().axis();
-    for (int64_t j = 0; j < ib_shape->NumAxes(); ++j) {
-      if (j == concat_axis || j == concat_axis + ib_shape->NumAxes()) {
-        vec[j] += ib_shape->At(j);
+    for (int64_t j = 0; j < ib_shape.NumAxes(); ++j) {
+      if (j == concat_axis || j == concat_axis + ib_shape.NumAxes()) {
+        vec[j] += ib_shape.At(j);
       } else {
-        CHECK_EQ(vec[j], ib_shape->At(j));
+        CHECK_EQ(vec[j], ib_shape.At(j));
       }
     }
   }
-  CHECK_EQ(vec.size(), GetShapePtr4BnInOp(input_bns().at(0))->NumAxes());
-  *GetShapePtr4BnInOp(SoleObn()) = Shape(vec);
+  CHECK_EQ(vec.size(),
+           GetBlobDesc4BnInOp(input_bns().at(0))->shape().NumAxes());
+  GetBlobDesc4BnInOp(SoleObn())->mut_shape() = Shape(vec);
 }
 
 REGISTER_OP(OperatorConf::kConcatConf, ConcatOp);
diff --git a/oneflow/core/operator/concat_op.h b/oneflow/core/operator/concat_op.h
index f8f2b8b6f7..2f848c429f 100644
--- a/oneflow/core/operator/concat_op.h
+++ b/oneflow/core/operator/concat_op.h
@@ -15,8 +15,8 @@ class ConcatOp final : public UserOperator {
 
   const PbMessage& GetSpecialConf() const override;
 
-  void InferShape4FwBlobs(
-      std::function<Shape*(const std::string&)> GetShapePtr4BnInOp,
+  void InferBlobDesc4FwBlobs(
+      std::function<BlobDesc*(const std::string)> GetBlobDesc4BnInOp,
       ParallelPolicy policy, int64_t parallel_id,
       int64_t parallel_num) const override;
 
diff --git a/oneflow/core/operator/concat_op_test.cpp b/oneflow/core/operator/concat_op_test.cpp
index 4f74601f47..35fd2f5ea9 100644
--- a/oneflow/core/operator/concat_op_test.cpp
+++ b/oneflow/core/operator/concat_op_test.cpp
@@ -21,7 +21,7 @@ TEST(ConcatOp, concat_two_3x3) {
     return bn2shape_ptr.at(bn);
   };
   // infershape
-  concat_op->InferShape4FwBlobs(fp, kDataParallel, 0, 1);
+  concat_op->InferBlobDesc4FwBlobs(fp, kDataParallel, 0, 1);
   // test
   Shape* output_shape_ptr = fp(concat_op->SoleObn());
   ASSERT_EQ(*output_shape_ptr, Shape({3, 6}));
diff --git a/oneflow/core/operator/convolution_op.cpp b/oneflow/core/operator/convolution_op.cpp
index cf4fb2bb6c..3b60e4ac9d 100644
--- a/oneflow/core/operator/convolution_op.cpp
+++ b/oneflow/core/operator/convolution_op.cpp
@@ -22,15 +22,13 @@ const PbMessage& ConvolutionOp::GetSpecialConf() const {
   return op_conf().convolution_conf();
 }
 
-void ConvolutionOp::InferShape4FwBlobs(
-    std::function<Shape*(const std::string&)> GetShapePtr4BnInOp,
+void ConvolutionOp::InferBlobDesc4FwBlobs(
+    std::function<BlobDesc*(const std::string)> GetBlobDesc4BnInOp,
     ParallelPolicy policy, int64_t parallel_id, int64_t parallel_num) const {
-  Shape* input_shape_ptr = GetShapePtr4BnInOp(SoleIbn());
-  Shape* output_shape_ptr = GetShapePtr4BnInOp(SoleObn());
-  Shape* colbuf_shape_ptr = GetShapePtr4BnInOp("col_buf");
+  const Shape& input_shape = GetBlobDesc4BnInOp(SoleIbn())->shape();
   auto conv_conf = op_conf().convolution_conf();
-  int64_t batch_size = input_shape_ptr->At(0);
-  int64_t c_i = input_shape_ptr->At(1);
+  int64_t batch_size = input_shape.At(0);
+  int64_t c_i = input_shape.At(1);
 
   int32_t out_num = GetInt32FromSpecialConf("out_num");
   if (policy == kModelParallel) {
@@ -43,32 +41,29 @@ void ConvolutionOp::InferShape4FwBlobs(
   int64_t output_size = 1;
   std::vector<int64_t> output_shape_vec = {batch_size, c_o};
 
-  int64_t h_len = (input_shape_ptr->At(2) + 2 * conv_conf.pad_h()
-                   - conv_conf.kernel_size_h())
-                      / conv_conf.stride_h()
-                  + 1;
+  int64_t h_len =
+      (input_shape.At(2) + 2 * conv_conf.pad_h() - conv_conf.kernel_size_h())
+          / conv_conf.stride_h()
+      + 1;
   output_shape_vec.push_back(h_len);
-  int64_t w_len = (input_shape_ptr->At(3) + 2 * conv_conf.pad_w()
-                   - conv_conf.kernel_size_w())
-                      / conv_conf.stride_w()
-                  + 1;
+  int64_t w_len =
+      (input_shape.At(3) + 2 * conv_conf.pad_w() - conv_conf.kernel_size_w())
+          / conv_conf.stride_w()
+      + 1;
   output_shape_vec.push_back(w_len);
   kernel_size *= conv_conf.kernel_size_h();
   kernel_size *= conv_conf.kernel_size_w();
   output_size *= h_len;
   output_size *= w_len;
 
-  *output_shape_ptr = Shape(output_shape_vec);
-  CHECK_EQ(output_shape_ptr->NumAxes(), input_shape_ptr->NumAxes());
-  *colbuf_shape_ptr = Shape({batch_size, output_size, c_i * kernel_size});
-  Shape* weight = GetShapePtr4BnInOp("weight");
-  *weight = Shape({c_o, c_i * kernel_size});
+  GetBlobDesc4BnInOp(SoleObn())->mut_shape() = Shape(output_shape_vec);
+  GetBlobDesc4BnInOp("col_buf")->mut_shape() =
+      Shape({batch_size, output_size, c_i * kernel_size});
+  GetBlobDesc4BnInOp("weight")->mut_shape() = Shape({c_o, c_i * kernel_size});
 
   if (GetBoolFromSpecialConf("has_bias_term")) {
-    Shape* bias = GetShapePtr4BnInOp("bias");
-    Shape* biasmult_shape_ptr = GetShapePtr4BnInOp("bias_multiplier");
-    *bias = Shape({c_o});
-    *biasmult_shape_ptr = Shape({output_size});
+    GetBlobDesc4BnInOp("bias")->mut_shape() = Shape({c_o});
+    GetBlobDesc4BnInOp("bias_multiplier")->mut_shape() = Shape({output_size});
   }
 }
 
diff --git a/oneflow/core/operator/convolution_op.h b/oneflow/core/operator/convolution_op.h
index 26b1ed1732..bb37bc7730 100644
--- a/oneflow/core/operator/convolution_op.h
+++ b/oneflow/core/operator/convolution_op.h
@@ -13,8 +13,8 @@ class ConvolutionOp final : public UserOperator {
 
   void InitFromOpConf(const OperatorConf& op_conf) override;
   const PbMessage& GetSpecialConf() const override;
-  void InferShape4FwBlobs(
-      std::function<Shape*(const std::string&)> GetShapePtr4BnInOp,
+  void InferBlobDesc4FwBlobs(
+      std::function<BlobDesc*(const std::string)> GetBlobDesc4BnInOp,
       ParallelPolicy policy, int64_t parallel_id,
       int64_t parallel_num) const override;
   void FixParallelDesc(ParallelDesc* pr_desc) const override {
diff --git a/oneflow/core/operator/convolution_op_test.cpp b/oneflow/core/operator/convolution_op_test.cpp
index 31874fbf93..3338a05f47 100644
--- a/oneflow/core/operator/convolution_op_test.cpp
+++ b/oneflow/core/operator/convolution_op_test.cpp
@@ -37,7 +37,7 @@ void TestDataParallelConvolutionOp() {
   };
 
   // infershape
-  convolution_op->InferShape4FwBlobs(fp, kDataParallel, 0, 1);
+  convolution_op->InferBlobDesc4FwBlobs(fp, kDataParallel, 0, 1);
 
   // test
   Shape* output_shape_ptr = fp(convolution_op->SoleObn());
@@ -71,7 +71,7 @@ void TestModelParallelConvolutionOp() {
   };
 
   // infershape
-  convolution_op->InferShape4FwBlobs(fp, kModelParallel, 3, 8);
+  convolution_op->InferBlobDesc4FwBlobs(fp, kModelParallel, 3, 8);
 
   // test
   Shape* output_shape_ptr = fp(convolution_op->SoleObn());
diff --git a/oneflow/core/operator/data_loader_op.cpp b/oneflow/core/operator/data_loader_op.cpp
index 9185c5c15b..89bb75950b 100644
--- a/oneflow/core/operator/data_loader_op.cpp
+++ b/oneflow/core/operator/data_loader_op.cpp
@@ -15,8 +15,8 @@ const PbMessage& DataLoaderOp::GetSpecialConf() const {
   return op_conf().data_loader_conf();
 }
 
-void DataLoaderOp::InferShape4FwBlobs(
-    std::function<Shape*(const std::string&)> GetShapePtr4BnInOp,
+void DataLoaderOp::InferBlobDesc4FwBlobs(
+    std::function<BlobDesc*(const std::string)> GetBlobDesc4BnInOp,
     ParallelPolicy policy, int64_t parallel_id, int64_t parallel_num) const {
   // useful vars
   int32_t piece_size = JobDesc::Singleton()->piece_size();
@@ -27,9 +27,9 @@ void DataLoaderOp::InferShape4FwBlobs(
   feature_shape.insert(feature_shape.end(),
                        feature_shape_of_one_ins.dim_vec().begin(),
                        feature_shape_of_one_ins.dim_vec().end());
-  *GetShapePtr4BnInOp("feature") = Shape(feature_shape);
+  GetBlobDesc4BnInOp("feature")->mut_shape() = Shape(feature_shape);
   // label shape
-  *GetShapePtr4BnInOp("label") = Shape({piece_size});
+  GetBlobDesc4BnInOp("label")->mut_shape() = Shape({piece_size});
 }
 
 REGISTER_OP(OperatorConf::kDataLoaderConf, DataLoaderOp);
diff --git a/oneflow/core/operator/data_loader_op.h b/oneflow/core/operator/data_loader_op.h
index 9aa45a0105..f51a982436 100644
--- a/oneflow/core/operator/data_loader_op.h
+++ b/oneflow/core/operator/data_loader_op.h
@@ -14,8 +14,8 @@ class DataLoaderOp final : public SysOperator {
   void InitFromOpConf(const OperatorConf& op_conf) override;
   const PbMessage& GetSpecialConf() const override;
 
-  void InferShape4FwBlobs(
-      std::function<Shape*(const std::string&)> GetShapePtr4BnInOp,
+  void InferBlobDesc4FwBlobs(
+      std::function<BlobDesc*(const std::string)> GetBlobDesc4BnInOp,
       ParallelPolicy policy, int64_t parallel_id,
       int64_t parallel_num) const override;
 
diff --git a/oneflow/core/operator/innerproduct_op.cpp b/oneflow/core/operator/innerproduct_op.cpp
index f141501d57..575db254b7 100644
--- a/oneflow/core/operator/innerproduct_op.cpp
+++ b/oneflow/core/operator/innerproduct_op.cpp
@@ -21,10 +21,10 @@ const PbMessage& InnerProductOp::GetSpecialConf() const {
   return op_conf().innerproduct_conf();
 }
 
-void InnerProductOp::InferShape4FwBlobs(
-    std::function<Shape*(const std::string&)> GetShapePtr4BnInOp,
+void InnerProductOp::InferBlobDesc4FwBlobs(
+    std::function<BlobDesc*(const std::string)> GetBlobDesc4BnInOp,
     ParallelPolicy policy, int64_t parallel_id, int64_t parallel_num) const {
-  Shape* in_shape_ptr = GetShapePtr4BnInOp(SoleIbn());
+  const Shape& in_shape = GetBlobDesc4BnInOp(SoleIbn())->shape();
   int32_t out_num = GetInt32FromSpecialConf("out_num");
   if (policy == kModelParallel) {
     BalancedSplitter splitter(out_num, parallel_num);
@@ -32,22 +32,20 @@ void InnerProductOp::InferShape4FwBlobs(
   }
 
   // output bn
-  Shape* out_shape_ptr = GetShapePtr4BnInOp(SoleObn());
-  *out_shape_ptr = Shape({in_shape_ptr->At(0), out_num});
+  GetBlobDesc4BnInOp(SoleObn())->mut_shape() = Shape({in_shape.At(0), out_num});
 
   // model bn
-  Shape* weight_shape_ptr = GetShapePtr4BnInOp("weight");
-  *weight_shape_ptr = Shape({out_num, in_shape_ptr->Count(1)});
+  GetBlobDesc4BnInOp("weight")->mut_shape() =
+      Shape({out_num, in_shape.Count(1)});
 
   if (GetBoolFromSpecialConf("has_bias_term")) {
     // model bn
-    Shape* bias_shape_ptr = GetShapePtr4BnInOp("bias");
-    *bias_shape_ptr = Shape({1, out_num});
+    GetBlobDesc4BnInOp("bias")->mut_shape() = Shape({1, out_num});
 
     // model tmp bn
     CHECK_EQ(model_tmp_bns().size(), 1);
-    Shape* bias_multiplier_shape_ptr = GetShapePtr4BnInOp("bias_multiplier");
-    *bias_multiplier_shape_ptr = Shape({in_shape_ptr->At(0), 1});
+    GetBlobDesc4BnInOp("bias_multiplier")->mut_shape() =
+        Shape({in_shape.At(0), 1});
   }
 }
 
diff --git a/oneflow/core/operator/innerproduct_op.h b/oneflow/core/operator/innerproduct_op.h
index 6a0c2a2f52..1076b5e66e 100644
--- a/oneflow/core/operator/innerproduct_op.h
+++ b/oneflow/core/operator/innerproduct_op.h
@@ -13,8 +13,8 @@ class InnerProductOp final : public UserOperator {
 
   void InitFromOpConf(const OperatorConf& op_conf) override;
   const PbMessage& GetSpecialConf() const override;
-  void InferShape4FwBlobs(
-      std::function<Shape*(const std::string&)> GetShapePtr4BnInOp,
+  void InferBlobDesc4FwBlobs(
+      std::function<BlobDesc*(const std::string)> GetBlobDesc4BnInOp,
       ParallelPolicy policy, int64_t parallel_id,
       int64_t parallel_num) const override;
   void FixParallelDesc(ParallelDesc* pr_desc) const override {
diff --git a/oneflow/core/operator/innerproduct_op_test.cpp b/oneflow/core/operator/innerproduct_op_test.cpp
index 139e899601..606f6993dd 100644
--- a/oneflow/core/operator/innerproduct_op_test.cpp
+++ b/oneflow/core/operator/innerproduct_op_test.cpp
@@ -29,7 +29,7 @@ void TestModelParallelInnerProductOp(bool has_bias_term) {
     return bn2shape_ptr.at(bn);
   };
 
-  ip_op->InferShape4FwBlobs(fp, kModelParallel, 3, 10);
+  ip_op->InferBlobDesc4FwBlobs(fp, kModelParallel, 3, 10);
 
   BalancedSplitter splitter(40, 10);
   int out_num = splitter.At(3).size();
@@ -70,7 +70,7 @@ void TestDataParallelInnerProductOp(bool has_bias_term) {
     return bn2shape_ptr.at(bn);
   };
 
-  ip_op->InferShape4FwBlobs(fp, kDataParallel, 3, 10);
+  ip_op->InferBlobDesc4FwBlobs(fp, kDataParallel, 3, 10);
 
   Shape* out_shape_ptr = bn2shape_ptr.at(ip_op->SoleObn());
   CHECK_EQ(*out_shape_ptr, Shape({1000, 40}));
diff --git a/oneflow/core/operator/model_update_op.h b/oneflow/core/operator/model_update_op.h
index a5495d75fc..bf78c0d070 100644
--- a/oneflow/core/operator/model_update_op.h
+++ b/oneflow/core/operator/model_update_op.h
@@ -10,8 +10,8 @@ class ModelUpdtOp : public SysOperator {
   OF_DISALLOW_COPY_AND_MOVE(ModelUpdtOp);
   virtual ~ModelUpdtOp() = default;
 
-  virtual void InferShape4FwBlobs(
-      std::function<Shape*(const std::string&)> GetShapePtr4BnInOp,
+  virtual void InferBlobDesc4FwBlobs(
+      std::function<BlobDesc*(const std::string)> GetBlobDesc4BnInOp,
       ParallelPolicy policy, int64_t parallel_id,
       int64_t parallel_num) const override {
     // do nothing
diff --git a/oneflow/core/operator/momentum_model_update_op.cpp b/oneflow/core/operator/momentum_model_update_op.cpp
index b589c33f3e..778fe991fe 100644
--- a/oneflow/core/operator/momentum_model_update_op.cpp
+++ b/oneflow/core/operator/momentum_model_update_op.cpp
@@ -15,11 +15,10 @@ const PbMessage& MomentumModelUpdateOp::GetSpecialConf() const {
   return op_conf().momentum_mdupdt_conf();
 }
 
-void MomentumModelUpdateOp::InferShape4FwBlobs(
-    std::function<Shape*(const std::string&)> GetShapePtr4BnInOp,
+void MomentumModelUpdateOp::InferBlobDesc4FwBlobs(
+    std::function<BlobDesc*(const std::string)> GetBlobDesc4BnInOp,
     ParallelPolicy policy, int64_t parallel_id, int64_t parallel_num) const {
-  Shape* input_shape_ptr = GetShapePtr4BnInOp("model_diffs");
-  *GetShapePtr4BnInOp("momentum") = *input_shape_ptr;
+  TODO();
 }
 
 REGISTER_OP(OperatorConf::kMomentumMdupdtConf, MomentumModelUpdateOp);
diff --git a/oneflow/core/operator/momentum_model_update_op.h b/oneflow/core/operator/momentum_model_update_op.h
index bfc8da32a1..ba119b20b7 100644
--- a/oneflow/core/operator/momentum_model_update_op.h
+++ b/oneflow/core/operator/momentum_model_update_op.h
@@ -13,8 +13,8 @@ class MomentumModelUpdateOp final : public ModelUpdtOp {
 
   void InitFromOpConf(const OperatorConf& op_conf) override;
   const PbMessage& GetSpecialConf() const override;
-  void InferShape4FwBlobs(
-      std::function<Shape*(const std::string&)> GetShapePtr4BnInOp,
+  void InferBlobDesc4FwBlobs(
+      std::function<BlobDesc*(const std::string)> GetBlobDesc4BnInOp,
       ParallelPolicy policy, int64_t parallel_id,
       int64_t parallel_num) const override;
 
diff --git a/oneflow/core/operator/multinomial_logistic_loss_op.cpp b/oneflow/core/operator/multinomial_logistic_loss_op.cpp
index 06f1e04d6a..0ba8c43c63 100644
--- a/oneflow/core/operator/multinomial_logistic_loss_op.cpp
+++ b/oneflow/core/operator/multinomial_logistic_loss_op.cpp
@@ -16,11 +16,11 @@ const PbMessage& MultinomialLogisticLossOp::GetSpecialConf() const {
   return op_conf().multinomial_logistic_loss_conf();
 }
 
-void MultinomialLogisticLossOp::InferShape4FwBlobs(
-    std::function<Shape*(const std::string&)> GetShapePtr4BnInOp,
+void MultinomialLogisticLossOp::InferBlobDesc4FwBlobs(
+    std::function<BlobDesc*(const std::string)> GetBlobDesc4BnInOp,
     ParallelPolicy policy, int64_t parallel_id, int64_t parallel_num) const {
-  *GetShapePtr4BnInOp("loss") = Shape({1});
-  *GetShapePtr4BnInOp("loss_buffer") = Shape({1});
+  GetBlobDesc4BnInOp("loss")->mut_shape() = Shape({1});
+  GetBlobDesc4BnInOp("loss_buffer")->mut_shape() = Shape({1});
 }
 
 REGISTER_OP(OperatorConf::kMultinomialLogisticLossConf,
diff --git a/oneflow/core/operator/multinomial_logistic_loss_op.h b/oneflow/core/operator/multinomial_logistic_loss_op.h
index 6310da73df..30c5ce8b9b 100644
--- a/oneflow/core/operator/multinomial_logistic_loss_op.h
+++ b/oneflow/core/operator/multinomial_logistic_loss_op.h
@@ -17,8 +17,8 @@ class MultinomialLogisticLossOp final : public UserOperator {
   const PbMessage& GetSpecialConf() const override;
   bool IsLossOp() const override { return true; }
 
-  void InferShape4FwBlobs(
-      std::function<Shape*(const std::string&)> GetShapePtr4BnInOp,
+  void InferBlobDesc4FwBlobs(
+      std::function<BlobDesc*(const std::string)> GetBlobDesc4BnInOp,
       ParallelPolicy policy, int64_t parallel_id,
       int64_t parallel_num) const override;
 
diff --git a/oneflow/core/operator/multinomial_logistic_loss_op_test.cpp b/oneflow/core/operator/multinomial_logistic_loss_op_test.cpp
index a8fa37efeb..b7a6f678c6 100644
--- a/oneflow/core/operator/multinomial_logistic_loss_op_test.cpp
+++ b/oneflow/core/operator/multinomial_logistic_loss_op_test.cpp
@@ -21,7 +21,7 @@ TEST(MultinomialLogisticLossOp, test_loss_op) {
     return bn2shape_ptr.at(bn);
   };
 
-  loss_op->InferShape4FwBlobs(fp, kDataParallel, 2, 10);
+  loss_op->InferBlobDesc4FwBlobs(fp, kDataParallel, 2, 10);
 
   Shape* loss_shape_ptr = bn2shape_ptr.at(loss_op->SoleObn());
   Shape* loss_buffer_shape_ptr = bn2shape_ptr.at(loss_op->SoleDtbn());
diff --git a/oneflow/core/operator/operator.h b/oneflow/core/operator/operator.h
index 2fac09b228..e96cb92fa0 100644
--- a/oneflow/core/operator/operator.h
+++ b/oneflow/core/operator/operator.h
@@ -2,13 +2,13 @@
 #define ONEFLOW_CORE_OPERATOR_OPERATOR_H_
 
 #include "oneflow/core/common/protobuf.h"
-#include "oneflow/core/common/shape.h"
 #include "oneflow/core/common/util.h"
 #include "oneflow/core/job/keyword.h"
 #include "oneflow/core/job/parallel_desc.h"
 #include "oneflow/core/job/placement.pb.h"
 #include "oneflow/core/operator/op_conf.pb.h"
 #include "oneflow/core/operator/operator.pb.h"
+#include "oneflow/core/register/blob_desc.h"
 
 namespace oneflow {
 
@@ -79,8 +79,8 @@ class Operator {
 
   // Read: shape of input_blobs
   // Write: shape of output_blobs, model_blobs, data_tmp_blobs, model_tmp_blobs
-  virtual void InferShape4FwBlobs(
-      std::function<Shape*(const std::string&)> GetShapePtr4BnInOp,
+  virtual void InferBlobDesc4FwBlobs(
+      std::function<BlobDesc*(const std::string)> GetBlobDesc4BnInOp,
       ParallelPolicy policy, int64_t parallel_id,
       int64_t parallel_num) const = 0;
 
@@ -144,8 +144,8 @@ class SysOperator : public Operator {
   SysOperator() = default;
   virtual ~SysOperator() = default;
 
-  virtual void InferShape4FwBlobs(
-      std::function<Shape*(const std::string&)> GetShapePtr4BnInOp,
+  virtual void InferBlobDesc4FwBlobs(
+      std::function<BlobDesc*(const std::string)> GetBlobDesc4BnInOp,
       ParallelPolicy policy, int64_t parallel_id,
       int64_t parallel_num) const override {
     UNEXPECTED_RUN();
diff --git a/oneflow/core/operator/pooling_op.cpp b/oneflow/core/operator/pooling_op.cpp
index a5a2b420dc..0325f814d9 100644
--- a/oneflow/core/operator/pooling_op.cpp
+++ b/oneflow/core/operator/pooling_op.cpp
@@ -15,32 +15,30 @@ const PbMessage& PoolingOp::GetSpecialConf() const {
   return op_conf().pooling_conf();
 }
 
-void PoolingOp::InferShape4FwBlobs(
-    std::function<Shape*(const std::string&)> GetShapePtr4BnInOp,
+void PoolingOp::InferBlobDesc4FwBlobs(
+    std::function<BlobDesc*(const std::string)> GetBlobDesc4BnInOp,
     ParallelPolicy policy, int64_t parallel_id, int64_t parallel_num) const {
-  Shape* input_shape_ptr = GetShapePtr4BnInOp(SoleIbn());
-  CHECK_EQ(input_shape_ptr->NumAxes(), 4);
-  Shape* output_shape_ptr = GetShapePtr4BnInOp(SoleObn());
+  const Shape& input_shape = GetBlobDesc4BnInOp(SoleIbn())->shape();
+  CHECK_EQ(input_shape.NumAxes(), 4);
+  BlobDesc* output_blob_desc = GetBlobDesc4BnInOp(SoleObn());
   const PoolingOpConf& pooling_conf = op_conf().pooling_conf();
 
-  std::vector<int64_t> output_shape_dim_vec = {input_shape_ptr->At(0),
-                                               input_shape_ptr->At(1)};
+  std::vector<int64_t> output_shape_dim_vec = {input_shape.At(0),
+                                               input_shape.At(1)};
 
-  output_shape_dim_vec.push_back((input_shape_ptr->At(2)
-                                  + 2 * pooling_conf.pad_h()
+  output_shape_dim_vec.push_back((input_shape.At(2) + 2 * pooling_conf.pad_h()
                                   - pooling_conf.kernel_size_h())
                                      / pooling_conf.stride_h()
                                  + 1);
 
-  output_shape_dim_vec.push_back((input_shape_ptr->At(3)
-                                  + 2 * pooling_conf.pad_w()
+  output_shape_dim_vec.push_back((input_shape.At(3) + 2 * pooling_conf.pad_w()
                                   - pooling_conf.kernel_size_w())
                                      / pooling_conf.stride_w()
                                  + 1);
 
-  *output_shape_ptr = Shape(output_shape_dim_vec);
-  Shape* data_tmp_shape_ptr = GetShapePtr4BnInOp(SoleDtbn());
-  *data_tmp_shape_ptr = Shape(output_shape_dim_vec);
+  output_blob_desc->mut_shape() = Shape(output_shape_dim_vec);
+  BlobDesc* data_tmp_blob_desc = GetBlobDesc4BnInOp(SoleDtbn());
+  data_tmp_blob_desc->mut_shape() = Shape(output_shape_dim_vec);
 }
 
 REGISTER_OP(OperatorConf::kPoolingConf, PoolingOp);
diff --git a/oneflow/core/operator/pooling_op.h b/oneflow/core/operator/pooling_op.h
index b6dbcfd8b8..ace71984aa 100644
--- a/oneflow/core/operator/pooling_op.h
+++ b/oneflow/core/operator/pooling_op.h
@@ -16,8 +16,8 @@ class PoolingOp final : public UserOperator {
   void InitFromOpConf(const OperatorConf& op_conf) override;
   const PbMessage& GetSpecialConf() const override;
 
-  void InferShape4FwBlobs(
-      std::function<Shape*(const std::string&)> GetShapePtr4BnInOp,
+  void InferBlobDesc4FwBlobs(
+      std::function<BlobDesc*(const std::string)> GetBlobDesc4BnInOp,
       ParallelPolicy policy, int64_t parallel_id,
       int64_t parallel_num) const override;
 
diff --git a/oneflow/core/operator/pooling_op_test.cpp b/oneflow/core/operator/pooling_op_test.cpp
index f525832b8e..9cc9abbf13 100644
--- a/oneflow/core/operator/pooling_op_test.cpp
+++ b/oneflow/core/operator/pooling_op_test.cpp
@@ -30,7 +30,7 @@ TEST(PoolingOp, pool_100x64x11x11) {
     return bn2shape_ptr.at(bn);
   };
   // do infer shape
-  pooling_op->InferShape4FwBlobs(fp, kDataParallel, 0, 1);
+  pooling_op->InferBlobDesc4FwBlobs(fp, kDataParallel, 0, 1);
   // test
   Shape* output_shape_ptr = bn2shape_ptr.at(pooling_op->SoleObn());
   Shape* data_tmp_shape_ptr = bn2shape_ptr.at(pooling_op->SoleDtbn());
diff --git a/oneflow/core/operator/record_op.h b/oneflow/core/operator/record_op.h
index 06e5fe301d..d02a532706 100644
--- a/oneflow/core/operator/record_op.h
+++ b/oneflow/core/operator/record_op.h
@@ -15,8 +15,8 @@ class RecordOp final : public SysOperator {
   const PbMessage& GetSpecialConf() const override;
   bool IsRecordOp() const override { return true; }
 
-  void InferShape4FwBlobs(
-      std::function<Shape*(const std::string&)> GetShapePtr4BnInOp,
+  void InferBlobDesc4FwBlobs(
+      std::function<BlobDesc*(const std::string)> GetBlobDesc4BnInOp,
       ParallelPolicy policy, int64_t parallel_id,
       int64_t parallel_num) const override {}
 
diff --git a/oneflow/core/operator/relu_op.cpp b/oneflow/core/operator/relu_op.cpp
index ec35d5fccb..8fc5b6b200 100644
--- a/oneflow/core/operator/relu_op.cpp
+++ b/oneflow/core/operator/relu_op.cpp
@@ -14,12 +14,10 @@ const PbMessage& ReluOp::GetSpecialConf() const {
   return op_conf().relu_conf();
 }
 
-void ReluOp::InferShape4FwBlobs(
-    std::function<Shape*(const std::string&)> GetShapePtr4BnInOp,
+void ReluOp::InferBlobDesc4FwBlobs(
+    std::function<BlobDesc*(const std::string)> GetBlobDesc4BnInOp,
     ParallelPolicy policy, int64_t parallel_id, int64_t parallel_num) const {
-  Shape* output_shape_ptr = GetShapePtr4BnInOp(SoleObn());
-  Shape* input_shape_ptr = GetShapePtr4BnInOp(SoleIbn());
-  *output_shape_ptr = *input_shape_ptr;
+  *GetBlobDesc4BnInOp(SoleObn()) = *GetBlobDesc4BnInOp(SoleIbn());
 }
 
 REGISTER_OP(OperatorConf::kReluConf, ReluOp);
diff --git a/oneflow/core/operator/relu_op.h b/oneflow/core/operator/relu_op.h
index 631213372f..ee776f0816 100644
--- a/oneflow/core/operator/relu_op.h
+++ b/oneflow/core/operator/relu_op.h
@@ -15,8 +15,8 @@ class ReluOp final : public UserOperator {
   const PbMessage& GetSpecialConf() const override;
   bool IsElemWise() const override { return true; }
 
-  void InferShape4FwBlobs(
-      std::function<Shape*(const std::string&)> GetShapePtr4BnInOp,
+  void InferBlobDesc4FwBlobs(
+      std::function<BlobDesc*(const std::string)> GetBlobDesc4BnInOp,
       ParallelPolicy policy, int64_t parallel_id,
       int64_t parallel_num) const override;
 
diff --git a/oneflow/core/operator/relu_op_test.cpp b/oneflow/core/operator/relu_op_test.cpp
index 1753a0c2e9..74c12bad2a 100644
--- a/oneflow/core/operator/relu_op_test.cpp
+++ b/oneflow/core/operator/relu_op_test.cpp
@@ -17,7 +17,7 @@ TEST(ReluOp, relu_3x5x4) {
     return bn2shape_ptr.at(bn);
   };
   // do infer shape
-  relu_op->InferShape4FwBlobs(fp, kDataParallel, 0, 1);
+  relu_op->InferBlobDesc4FwBlobs(fp, kDataParallel, 0, 1);
   // test
   Shape* input_shape_ptr = bn2shape_ptr.at(relu_op->SoleIbn());
   Shape* output_shape_ptr = bn2shape_ptr.at(relu_op->SoleObn());
diff --git a/oneflow/core/operator/rmsprop_model_update_op.cpp b/oneflow/core/operator/rmsprop_model_update_op.cpp
index c517b113c2..2da4a892d1 100644
--- a/oneflow/core/operator/rmsprop_model_update_op.cpp
+++ b/oneflow/core/operator/rmsprop_model_update_op.cpp
@@ -15,11 +15,10 @@ const PbMessage& RMSPropModelUpdateOp::GetSpecialConf() const {
   return op_conf().rmsprop_mdupdt_conf();
 }
 
-void RMSPropModelUpdateOp::InferShape4FwBlobs(
-    std::function<Shape*(const std::string&)> GetShapePtr4BnInOp,
+void RMSPropModelUpdateOp::InferBlobDesc4FwBlobs(
+    std::function<BlobDesc*(const std::string)> GetBlobDesc4BnInOp,
     ParallelPolicy policy, int64_t parallel_id, int64_t parallel_num) const {
-  Shape* input_shape_ptr = GetShapePtr4BnInOp("model_diffs");
-  *GetShapePtr4BnInOp("mean_square") = *input_shape_ptr;
+  TODO();
 }
 
 REGISTER_OP(OperatorConf::kRmspropMdupdtConf, RMSPropModelUpdateOp);
diff --git a/oneflow/core/operator/rmsprop_model_update_op.h b/oneflow/core/operator/rmsprop_model_update_op.h
index c8ee7dea14..72fbbc32e5 100644
--- a/oneflow/core/operator/rmsprop_model_update_op.h
+++ b/oneflow/core/operator/rmsprop_model_update_op.h
@@ -13,8 +13,8 @@ class RMSPropModelUpdateOp final : public ModelUpdtOp {
 
   void InitFromOpConf(const OperatorConf& op_conf) override;
   const PbMessage& GetSpecialConf() const override;
-  void InferShape4FwBlobs(
-      std::function<Shape*(const std::string&)> GetShapePtr4BnInOp,
+  void InferBlobDesc4FwBlobs(
+      std::function<BlobDesc*(const std::string)> GetBlobDesc4BnInOp,
       ParallelPolicy policy, int64_t parallel_id,
       int64_t parallel_num) const override;
 
diff --git a/oneflow/core/operator/softmax_loss_op.cpp b/oneflow/core/operator/softmax_loss_op.cpp
index 71f47282f7..19369aab9d 100644
--- a/oneflow/core/operator/softmax_loss_op.cpp
+++ b/oneflow/core/operator/softmax_loss_op.cpp
@@ -17,16 +17,16 @@ const PbMessage& SoftmaxLossOp::GetSpecialConf() const {
   return op_conf().softmax_loss_conf();
 }
 
-void SoftmaxLossOp::InferShape4FwBlobs(
-    std::function<Shape*(const std::string&)> GetShapePtr4BnInOp,
+void SoftmaxLossOp::InferBlobDesc4FwBlobs(
+    std::function<BlobDesc*(const std::string)> GetBlobDesc4BnInOp,
     ParallelPolicy policy, int64_t parallel_id, int64_t parallel_num) const {
   const std::vector<int64_t> in_dim_vec =
-      GetShapePtr4BnInOp("prediction")->dim_vec();
+      GetBlobDesc4BnInOp("prediction")->shape().dim_vec();
   CHECK_EQ(in_dim_vec.size(), 2);
-  CHECK_EQ(*GetShapePtr4BnInOp("label"), Shape({in_dim_vec[0]}));
-  *GetShapePtr4BnInOp(SoleObn()) = Shape({1});
-  *GetShapePtr4BnInOp("prob") = Shape(in_dim_vec);
-  *GetShapePtr4BnInOp("tmp_1D") = Shape({in_dim_vec[0]});
+  CHECK_EQ(GetBlobDesc4BnInOp("label")->shape(), Shape({in_dim_vec[0]}));
+  GetBlobDesc4BnInOp(SoleObn())->mut_shape() = Shape({1});
+  GetBlobDesc4BnInOp("prob")->mut_shape() = Shape(in_dim_vec);
+  GetBlobDesc4BnInOp("tmp_1D")->mut_shape() = Shape({in_dim_vec[0]});
 }
 
 REGISTER_OP(OperatorConf::kSoftmaxLossConf, SoftmaxLossOp);
diff --git a/oneflow/core/operator/softmax_loss_op.h b/oneflow/core/operator/softmax_loss_op.h
index f0b2b966ac..cca9bea9c9 100644
--- a/oneflow/core/operator/softmax_loss_op.h
+++ b/oneflow/core/operator/softmax_loss_op.h
@@ -15,8 +15,8 @@ class SoftmaxLossOp final : public UserOperator {
   const PbMessage& GetSpecialConf() const override;
   bool IsLossOp() const override { return true; }
 
-  void InferShape4FwBlobs(
-      std::function<Shape*(const std::string&)> GetShapePtr4BnInOp,
+  void InferBlobDesc4FwBlobs(
+      std::function<BlobDesc*(const std::string)> GetBlobDesc4BnInOp,
       ParallelPolicy policy, int64_t parallel_id,
       int64_t parallel_num) const override;
 
diff --git a/oneflow/core/operator/softmax_loss_op_test.cpp b/oneflow/core/operator/softmax_loss_op_test.cpp
index eb7048024f..558465e8a5 100644
--- a/oneflow/core/operator/softmax_loss_op_test.cpp
+++ b/oneflow/core/operator/softmax_loss_op_test.cpp
@@ -20,7 +20,7 @@ TEST(SoftmaxLossOp, softmax_loss_3x5) {
     return bn2shape_ptr.at(bn);
   };
   // infershape
-  softmax_loss_op->InferShape4FwBlobs(fp, kDataParallel, 0, 1);
+  softmax_loss_op->InferBlobDesc4FwBlobs(fp, kDataParallel, 0, 1);
   // test
   ASSERT_EQ(*fp("loss"), Shape({1}));
   ASSERT_EQ(*fp("prob"), Shape({3, 5}));
diff --git a/oneflow/core/operator/softmax_op.cpp b/oneflow/core/operator/softmax_op.cpp
index 63bceb28a4..e83723bb82 100644
--- a/oneflow/core/operator/softmax_op.cpp
+++ b/oneflow/core/operator/softmax_op.cpp
@@ -15,13 +15,14 @@ const PbMessage& SoftmaxOp::GetSpecialConf() const {
   return op_conf().softmax_conf();
 }
 
-void SoftmaxOp::InferShape4FwBlobs(
-    std::function<Shape*(const std::string&)> GetShapePtr4BnInOp,
+void SoftmaxOp::InferBlobDesc4FwBlobs(
+    std::function<BlobDesc*(const std::string)> GetBlobDesc4BnInOp,
     ParallelPolicy policy, int64_t parallel_id, int64_t parallel_num) const {
-  std::vector<int64_t> vec = GetShapePtr4BnInOp(SoleIbn())->dim_vec();
+  const std::vector<int64_t>& vec =
+      GetBlobDesc4BnInOp(SoleIbn())->shape().dim_vec();
   CHECK_EQ(vec.size(), 2);
-  *GetShapePtr4BnInOp(SoleObn()) = Shape(vec);
-  *GetShapePtr4BnInOp(SoleDtbn()) = Shape({vec[0]});
+  GetBlobDesc4BnInOp(SoleObn())->mut_shape() = Shape(vec);
+  GetBlobDesc4BnInOp(SoleDtbn())->mut_shape() = Shape({vec[0]});
 }
 
 REGISTER_OP(OperatorConf::kSoftmaxConf, SoftmaxOp);
diff --git a/oneflow/core/operator/softmax_op.h b/oneflow/core/operator/softmax_op.h
index 3aba95c97b..ea6d6bf9d0 100644
--- a/oneflow/core/operator/softmax_op.h
+++ b/oneflow/core/operator/softmax_op.h
@@ -14,8 +14,8 @@ class SoftmaxOp final : public UserOperator {
   void InitFromOpConf(const OperatorConf& op_conf) override;
   const PbMessage& GetSpecialConf() const override;
 
-  void InferShape4FwBlobs(
-      std::function<Shape*(const std::string&)> GetShapePtr4BnInOp,
+  void InferBlobDesc4FwBlobs(
+      std::function<BlobDesc*(const std::string)> GetBlobDesc4BnInOp,
       ParallelPolicy policy, int64_t parallel_id,
       int64_t parallel_num) const override;
 
diff --git a/oneflow/core/operator/softmax_op_test.cpp b/oneflow/core/operator/softmax_op_test.cpp
index 7cab6b8a8c..f0b8a3ce06 100644
--- a/oneflow/core/operator/softmax_op_test.cpp
+++ b/oneflow/core/operator/softmax_op_test.cpp
@@ -17,7 +17,7 @@ TEST(SoftmaxOp, softmax_3x5) {
     return bn2shape_ptr.at(bn);
   };
   // infershape
-  softmax_op->InferShape4FwBlobs(fp, kDataParallel, 0, 1);
+  softmax_op->InferBlobDesc4FwBlobs(fp, kDataParallel, 0, 1);
   // test
   Shape* output_shape_ptr = fp(softmax_op->SoleObn());
   Shape* tmp_shape_ptr = fp(softmax_op->SoleDtbn());
diff --git a/oneflow/core/register/blob.h b/oneflow/core/register/blob.h
index 0979098cc3..f1d7b880ce 100644
--- a/oneflow/core/register/blob.h
+++ b/oneflow/core/register/blob.h
@@ -6,7 +6,7 @@
 
 namespace oneflow {
 
-class Blob {
+class Blob final {
  public:
   OF_DISALLOW_COPY_AND_MOVE(Blob);
   Blob(void* dptr, const Shape* shape) : dptr_(dptr), shape_(shape) {}
diff --git a/oneflow/core/register/blob_desc.h b/oneflow/core/register/blob_desc.h
new file mode 100644
index 0000000000..2258ef870a
--- /dev/null
+++ b/oneflow/core/register/blob_desc.h
@@ -0,0 +1,30 @@
+#ifndef ONEFLOW_CORE_REGISTER_BLOB_DESC_H_
+#define ONEFLOW_CORE_REGISTER_BLOB_DESC_H_
+
+#include "oneflow/core/common/shape.h"
+#include "oneflow/core/register/blob_desc.pb.h"
+
+namespace oneflow {
+
+class BlobDesc final {
+ public:
+  // OF_DISALLOW_COPY_AND_MOVE(BlobDesc);
+  BlobDesc() = default;
+  ~BlobDesc() = default;
+
+  BlobDesc(const BlobDescProto& proto) { shape_ = Shape(proto.shape()); }
+
+  const Shape& shape() const { return shape_; }
+  Shape& mut_shape() { return shape_; }
+
+  void ToProto(BlobDescProto* proto) const {
+    shape_.ToProto(proto->mutable_shape());
+  }
+
+ private:
+  Shape shape_;
+};
+
+}  // namespace oneflow
+
+#endif  // ONEFLOW_CORE_REGISTER_BLOB_DESC_H_
diff --git a/oneflow/core/register/blob_desc.proto b/oneflow/core/register/blob_desc.proto
new file mode 100644
index 0000000000..893f18ed47
--- /dev/null
+++ b/oneflow/core/register/blob_desc.proto
@@ -0,0 +1,8 @@
+syntax = "proto3";
+package oneflow;
+
+import "oneflow/core/common/shape.proto";
+
+message BlobDescProto {
+  ShapeProto shape = 1;
+}
diff --git a/oneflow/core/register/register_desc.cpp b/oneflow/core/register/register_desc.cpp
index 4ac2aaa80b..08030d9007 100644
--- a/oneflow/core/register/register_desc.cpp
+++ b/oneflow/core/register/register_desc.cpp
@@ -40,57 +40,57 @@ void RegstDesc::AddConsumer(const TaskNode* new_consumer) {
 }
 
 void RegstDesc::CopyLbnFrom(const RegstDesc* rhs) {
-  lbn2shape_.clear();
-  for (const auto& pair : rhs->lbn2shape_) {
+  CHECK(lbn2blob_desc_.empty());
+  for (const auto& pair : rhs->lbn2blob_desc_) {
     const std::string& lbn = pair.first;
-    auto shape = of_make_unique<Shape>();
-    CHECK(lbn2shape_.emplace(lbn, std::move(shape)).second);
+    CHECK(lbn2blob_desc_.emplace(lbn, of_make_unique<BlobDesc>()).second);
   }
 }
 
-void RegstDesc::CopyShapeFrom(const RegstDesc* rhs) {
-  for (const auto& pair : lbn2shape_) {
+void RegstDesc::CopyBlobDescFrom(const RegstDesc* rhs) {
+  for (const auto& pair : lbn2blob_desc_) {
     const std::string& lbn = pair.first;
-    *(lbn2shape_.at(lbn)) = rhs->GetShape(lbn);
+    *(lbn2blob_desc_.at(lbn)) = rhs->GetBlobDesc(lbn);
   }
 }
 
 void RegstDesc::EnrollLbn(const std::string& lbn) {
-  std::unique_ptr<Shape> ptr(new Shape);
-  CHECK(lbn2shape_.emplace(lbn, std::move(ptr)).second) << lbn;
+  CHECK(lbn2blob_desc_.emplace(lbn, of_make_unique<BlobDesc>()).second) << lbn;
 }
 
-const Shape& RegstDesc::GetShape(const std::string& lbn) const {
-  return *(lbn2shape_.at(lbn));
+const BlobDesc& RegstDesc::GetBlobDesc(const std::string& lbn) const {
+  return *(lbn2blob_desc_.at(lbn));
 }
 
-Shape* RegstDesc::GetMutShapePtr(const std::string& lbn) {
-  return lbn2shape_.at(lbn).get();
+BlobDesc* RegstDesc::GetMutBlobDesc(const std::string& lbn) {
+  return lbn2blob_desc_.at(lbn).get();
 }
 
 void RegstDesc::ForEachLbn(std::function<void(const std::string&)> func) const {
-  for (const auto& p : lbn2shape_) { func(p.first); }
+  for (const auto& p : lbn2blob_desc_) { func(p.first); }
 }
 
 void RegstDesc::EraseZeroSizeBlob() {
-  EraseIf<std::string, std::unique_ptr<Shape>>(
-      &lbn2shape_,
-      [](HashMap<std::string, std::unique_ptr<Shape>>::iterator it) {
-        return it->second->elem_cnt() == 0;
+  EraseIf<std::string, std::unique_ptr<BlobDesc>>(
+      &lbn2blob_desc_,
+      [](HashMap<std::string, std::unique_ptr<BlobDesc>>::iterator it) {
+        return it->second->shape().elem_cnt() == 0;
       });
 }
 
 int64_t RegstDesc::CompElemCntOfAllBlob() const {
   int64_t sum = 0;
-  for (const auto& pair : lbn2shape_) { sum += pair.second->elem_cnt(); }
+  for (const auto& pair : lbn2blob_desc_) {
+    sum += pair.second->shape().elem_cnt();
+  }
   return sum;
 }
 
 std::string RegstDesc::DebugStr() const {
   std::stringstream ss;
   ss << "{";
-  for (const auto& pair : lbn2shape_) {
-    ss << "{" << pair.first << ":" << pair.second->DebugStr() << "}";
+  for (const auto& pair : lbn2blob_desc_) {
+    ss << "{" << pair.first << ":" << pair.second->shape().DebugStr() << "}";
   }
   ss << "}";
   return ss.str();
@@ -104,10 +104,10 @@ void RegstDesc::ToProto(RegstDescProto* ret) const {
       ret->add_consumer_task_id(consumer->task_id());
     }
   }
-  for (const auto& pair : lbn2shape_) {
-    PbMapPair<std::string, ShapeProto> pb_pair(pair.first);
+  for (const auto& pair : lbn2blob_desc_) {
+    PbMapPair<std::string, BlobDescProto> pb_pair(pair.first);
     pair.second->ToProto(&(pb_pair.second));
-    ret->mutable_lbn2shape()->insert(pb_pair);
+    ret->mutable_lbn2blob_desc()->insert(pb_pair);
   }
   ret->set_register_num(register_num_);
   *(ret->mutable_mem_case()) = InferMemCase();
@@ -140,4 +140,9 @@ MemoryCase RegstDesc::InferMemCase() const {
   return mem_case;
 }
 
+BlobDesc RegstDesc::CompPackedBlobDesc() const {
+  BlobDesc packed_blob_desc;
+  packed_blob_desc.mut_shape() = Shape({CompElemCntOfAllBlob()});
+}
+
 }  // namespace oneflow
diff --git a/oneflow/core/register/register_desc.h b/oneflow/core/register/register_desc.h
index c9c9038a35..d69e409485 100644
--- a/oneflow/core/register/register_desc.h
+++ b/oneflow/core/register/register_desc.h
@@ -1,15 +1,11 @@
 #ifndef ONEFLOW_CORE_REGISTER_REGISTER_DESC_H_
 #define ONEFLOW_CORE_REGISTER_REGISTER_DESC_H_
 
-#include "oneflow/core/common/shape.h"
-#include "oneflow/core/common/util.h"
+#include "oneflow/core/register/blob_desc.h"
 #include "oneflow/core/register/register_desc.pb.h"
 
 namespace oneflow {
 
-// Regst  : Register
-// Contig : Contiguous
-
 class TaskNode;
 
 class RegstDesc final {
@@ -27,28 +23,29 @@ class RegstDesc final {
   const HashSet<const TaskNode*>& consumers() const { return consumers_; }
   void AddConsumer(const TaskNode*);
 
-  // Lbn and Shape
+  // Lbn and BlobDesc
   void CopyLbnFrom(const RegstDesc*);
-  void CopyShapeFrom(const RegstDesc*);
+  void CopyBlobDescFrom(const RegstDesc*);
   void EnrollLbn(const std::string& lbn);
-  const Shape& GetShape(const std::string& lbn) const;
-  Shape* GetMutShapePtr(const std::string& lbn);
+  const BlobDesc& GetBlobDesc(const std::string& lbn) const;
+  BlobDesc* GetMutBlobDesc(const std::string& lbn);
   void ForEachLbn(std::function<void(const std::string&)> func) const;
-  size_t NumOfLbn() const { return lbn2shape_.size(); }
+  size_t NumOfLbn() const { return lbn2blob_desc_.size(); }
 
   //
   void EraseZeroSizeBlob();
-  int64_t CompElemCntOfAllBlob() const;
   std::string DebugStr() const;
   void ToProto(RegstDescProto*) const;
   MemoryCase InferMemCase() const;
+  BlobDesc CompPackedBlobDesc() const;
 
  private:
+  int64_t CompElemCntOfAllBlob() const;
   int64_t regst_desc_id_;
   const TaskNode* producer_;
   HashSet<const TaskNode*> consumers_;
 
-  HashMap<std::string, std::unique_ptr<Shape>> lbn2shape_;
+  HashMap<std::string, std::unique_ptr<BlobDesc>> lbn2blob_desc_;
   int64_t register_num_;
 };
 
diff --git a/oneflow/core/register/register_desc.proto b/oneflow/core/register/register_desc.proto
index 58d548e069..54b3d0e814 100644
--- a/oneflow/core/register/register_desc.proto
+++ b/oneflow/core/register/register_desc.proto
@@ -1,14 +1,14 @@
 syntax = "proto3";
 package oneflow;
 
-import "oneflow/core/common/shape.proto";
+import "oneflow/core/register/blob_desc.proto";
 import "oneflow/core/memory/memory_case.proto";
 
 message RegstDescProto {
   int64 regst_desc_id = 1;
   int64 producer_task_id = 2;
   repeated int64 consumer_task_id = 3;
-  map<string, ShapeProto> lbn2shape = 4;
+  map<string, BlobDescProto> lbn2blob_desc = 4;
   int64 register_num = 5;
   MemoryCase mem_case = 6;
 }
diff --git a/oneflow/core/register/register_manager.cpp b/oneflow/core/register/register_manager.cpp
index 881faa62b0..1bfc16d8a7 100644
--- a/oneflow/core/register/register_manager.cpp
+++ b/oneflow/core/register/register_manager.cpp
@@ -19,12 +19,12 @@ void RegstMgr::NewRegsts(const RegstDescProto& regst_desc_proto,
     }
     int64_t elem_cnt = 0;
     std::vector<std::string> lbns;
-    lbns.reserve(regst_desc_proto.lbn2shape().size());
-    for (const auto& pair : regst_desc_proto.lbn2shape()) {
-      const Shape* shape_ptr =
-          runtime_regst_desc->GetShapePtrFromLbn(pair.first);
+    lbns.reserve(regst_desc_proto.lbn2blob_desc().size());
+    for (const auto& pair : regst_desc_proto.lbn2blob_desc()) {
+      const Shape& shape_ptr =
+          runtime_regst_desc->GetBlobDescFromLbn(pair.first)->shape();
       lbns.push_back(pair.first);
-      elem_cnt += shape_ptr->elem_cnt();
+      elem_cnt += shape_ptr.elem_cnt();
     }
     std::sort(lbns.begin(), lbns.end());
     std::pair<char*, std::function<void()>> allocation =
@@ -33,7 +33,8 @@ void RegstMgr::NewRegsts(const RegstDescProto& regst_desc_proto,
 
     int64_t blob_idx = 0;
     for (const std::string& lbn : lbns) {
-      const Shape* shape_ptr = runtime_regst_desc->GetShapePtrFromLbn(lbn);
+      const Shape* shape_ptr =
+          &(runtime_regst_desc->GetBlobDescFromLbn(lbn)->shape());
       auto blob_ptr =
           of_make_unique<Blob>(allocation.first + blob_idx, shape_ptr);
       CHECK(regst->lbn2blob_.emplace(lbn, std::move(blob_ptr)).second);
diff --git a/oneflow/core/register/runtime_register_desc.cpp b/oneflow/core/register/runtime_register_desc.cpp
index e3d78a817b..7d86ddce39 100644
--- a/oneflow/core/register/runtime_register_desc.cpp
+++ b/oneflow/core/register/runtime_register_desc.cpp
@@ -15,8 +15,9 @@ RtRegstDesc::RtRegstDesc(const RegstDescProto& regst_desc_proto) {
     consumers_actor_id_.push_back(IDMgr::Singleton()->ActorId4TaskId(task_id));
   }
 
-  for (const auto& pair : regst_desc_proto.lbn2shape()) {
-    CHECK(lbn2shape_.emplace(pair.first, of_make_unique<Shape>(pair.second))
+  for (const auto& pair : regst_desc_proto.lbn2blob_desc()) {
+    CHECK(lbn2blob_desc_
+              .emplace(pair.first, of_make_unique<BlobDesc>(pair.second))
               .second);
   }
   mem_case_ = regst_desc_proto.mem_case();
diff --git a/oneflow/core/register/runtime_register_desc.h b/oneflow/core/register/runtime_register_desc.h
index 31258039c7..14dc491a26 100644
--- a/oneflow/core/register/runtime_register_desc.h
+++ b/oneflow/core/register/runtime_register_desc.h
@@ -1,8 +1,8 @@
 #ifndef ONEFLOW_CORE_REGISTER_RUNTIME_REGISTER_DESC_H_
 #define ONEFLOW_CORE_REGISTER_RUNTIME_REGISTER_DESC_H_
 
-#include "oneflow/core/common/shape.h"
 #include "oneflow/core/memory/memory_case.pb.h"
+#include "oneflow/core/register/blob_desc.h"
 #include "oneflow/core/register/register_desc.pb.h"
 
 namespace oneflow {
@@ -23,15 +23,15 @@ class RtRegstDesc {
   int64_t register_num() const { return register_num_; }
   const MemoryCase& mem_case() const { return mem_case_; }
 
-  const Shape* GetShapePtrFromLbn(const std::string& lbn) const {
-    return lbn2shape_.at(lbn).get();
+  const BlobDesc* GetBlobDescFromLbn(const std::string& lbn) const {
+    return lbn2blob_desc_.at(lbn).get();
   }
 
  private:
   int64_t regst_desc_id_;
   int64_t producer_actor_id_;
   std::vector<int64_t> consumers_actor_id_;
-  std::unordered_map<std::string, std::unique_ptr<Shape>> lbn2shape_;
+  std::unordered_map<std::string, std::unique_ptr<BlobDesc>> lbn2blob_desc_;
   int64_t register_num_;
   MemoryCase mem_case_;
 };
-- 
GitLab