remove conflict

.gitignore

@@ -28,3 +28,4 @@ cmake_install.cmake
 paddle/.timestamp
 python/paddlepaddle.egg-info/
 paddle/pybind/pybind.h
+python/paddle/v2/framework/tests/tmp/*

paddle/framework/CMakeLists.txt

@@ -15,7 +15,7 @@ nv_test(lod_tensor_gpu_test SRCS lod_tensor_test.cu DEPS lod_tensor)
 
 cc_test(variable_test SRCS variable_test.cc)
 
-cc_library(scope SRCS scope.cc)
+cc_library(scope SRCS scope.cc DEPS glog)
 cc_test(scope_test SRCS scope_test.cc DEPS scope)
 
 
@@ -24,9 +24,10 @@ cc_test(program_desc_test SRCS program_desc_test.cc DEPS proto_desc)
 cc_library(op_proto_maker SRCS op_proto_maker.cc DEPS framework_proto attribute)
 cc_test(op_proto_maker_test SRCS op_proto_maker_test.cc DEPS op_proto_maker)
 cc_library(op_info SRCS op_info.cc DEPS attribute framework_proto)
-cc_library(operator SRCS operator.cc DEPS op_info device_context tensor scope glog)
+cc_library(shape_inference SRCS shape_inference.cc DEPS ddim attribute)
+cc_library(operator SRCS operator.cc DEPS op_info device_context tensor scope glog shape_inference)
 cc_test(operator_test SRCS operator_test.cc DEPS operator op_registry)
-cc_library(proto_desc SRCS var_desc.cc op_desc.cc block_desc.cc program_desc.cc DEPS attribute ddim op_info operator glog)
+cc_library(proto_desc SRCS var_desc.cc op_desc.cc block_desc.cc program_desc.cc DEPS shape_inference op_info operator glog)
 
 cc_library(op_registry SRCS op_registry.cc DEPS op_proto_maker op_info operator glog proto_desc)
 cc_test(op_registry_test SRCS op_registry_test.cc DEPS op_registry)

paddle/framework/data_type.h

@@ -34,5 +34,25 @@ inline DataType ToDataType(std::type_index type) {
   }
 }
 
+template <typename Visitor>
+inline void VisitDataType(DataType type, Visitor visitor) {
+  switch (type) {
+    case DataType::FP32:
+      visitor.template operator()<float>();
+      break;
+    case DataType::FP64:
+      visitor.template operator()<double>();
+      break;
+    case DataType::INT32:
+      visitor.template operator()<int>();
+      break;
+    case DataType::INT64:
+      visitor.template operator()<int64_t>();
+      break;
+    default:
+      PADDLE_THROW("Not supported");
+  }
+}
+
 }  // namespace framework
 }  // namespace paddle

paddle/framework/op_desc.cc

@@ -16,15 +16,51 @@ limitations under the License. */
 #include <functional>
 #include <mutex>
 #include <unordered_map>
+#include "glog/logging.h"
 #include "paddle/framework/block_desc.h"
 #include "paddle/framework/operator.h"
 #include "paddle/framework/program_desc.h"
-
+#include "paddle/framework/shape_inference.h"
-#include "glog/logging.h"
 
 namespace paddle {
 namespace framework {
 
+class OpDescBind;
+class BlockDescBind;
+class CompileTimeInferShapeContext : public InferShapeContext {
+ public:
+  CompileTimeInferShapeContext(const OpDescBind &op,
+                               const BlockDescBind &block);
+
+  bool HasInput(const std::string &name) const override;
+
+  bool HasOutput(const std::string &name) const override;
+
+  bool HasInputs(const std::string &name) const override;
+
+  bool HasOutputs(const std::string &name) const override;
+
+  DDim GetInputDim(const std::string &name) const override;
+
+  void SetOutputDim(const std::string &name, const DDim &dim) override;
+
+  AttrReader Attrs() const override;
+
+  const std::vector<std::string> &Inputs(
+      const std::string &name) const override;
+
+  const std::vector<std::string> &Outputs(
+      const std::string &name) const override;
+
+ private:
+  DDim GetDim(const std::string &name) const override;
+
+  void SetDim(const std::string &name, const DDim &dim) override;
+
+  const OpDescBind &op_;
+  const BlockDescBind &block_;
+};
+
 OpDescBind::OpDescBind(const std::string &type, const VariableNameMap &inputs,
                        const VariableNameMap &outputs,
                        const AttributeMap &attrs) {
@@ -288,5 +324,97 @@ void OpDescBind::InferVarType(BlockDescBind *block) const {
   }
 }
 
+CompileTimeInferShapeContext::CompileTimeInferShapeContext(
+    const OpDescBind &op, const BlockDescBind &block)
+    : op_(op), block_(block) {}
+
+bool CompileTimeInferShapeContext::HasInput(const std::string &name) const {
+  const std::vector<std::string> &input_names = op_.Input(name);
+  auto length = input_names.size();
+  if (length == 0) {
+    return false;
+  }
+  PADDLE_ENFORCE_EQ(length, 1UL,
+                    "Input(%s) should have only one value, "
+                    "but it have %d now",
+                    name, length);
+  return block_.HasVarRecursive(input_names[0]);
+}
+
+bool CompileTimeInferShapeContext::HasOutput(const std::string &name) const {
+  const std::vector<std::string> &output_names = op_.Output(name);
+  auto length = output_names.size();
+  if (length == 0) {
+    return false;
+  }
+  PADDLE_ENFORCE_EQ(length, 1UL,
+                    "Output(%s) should have only one value, "
+                    "but it have %d now",
+                    name, length);
+  return block_.HasVarRecursive(output_names[0]);
+}
+
+bool CompileTimeInferShapeContext::HasInputs(const std::string &name) const {
+  const std::vector<std::string> &input_names = op_.Input(name);
+  if (input_names.empty()) {
+    return false;
+  }
+  for (auto &input : input_names) {
+    if (!block_.HasVarRecursive(input)) return false;
+  }
+  return true;
+}
+
+bool CompileTimeInferShapeContext::HasOutputs(const std::string &name) const {
+  const std::vector<std::string> &output_names = op_.Output(name);
+  if (output_names.empty()) {
+    return false;
+  }
+  for (auto &output : output_names) {
+    if (!block_.HasVarRecursive(output)) return false;
+  }
+  return true;
+}
+
+DDim CompileTimeInferShapeContext::GetInputDim(const std::string &name) const {
+  std::vector<DDim> ddims = GetInputsDim(name);
+  auto length = ddims.size();
+  PADDLE_ENFORCE_EQ(length, 1UL,
+                    "Input(%s) should have 1 value, "
+                    "but it has %d now",
+                    name, length);
+  return ddims[0];
+}
+
+void CompileTimeInferShapeContext::SetOutputDim(const std::string &name,
+                                                const DDim &dim) {
+  SetOutputsDim(name, {dim});
+}
+
+AttrReader CompileTimeInferShapeContext::Attrs() const {
+  return AttrReader(op_.GetAttrMap());
+}
+
+const std::vector<std::string> &CompileTimeInferShapeContext::Inputs(
+    const std::string &name) const {
+  return op_.Input(name);
+}
+
+const std::vector<std::string> &CompileTimeInferShapeContext::Outputs(
+    const std::string &name) const {
+  return op_.Output(name);
+}
+
+DDim CompileTimeInferShapeContext::GetDim(const std::string &name) const {
+  auto var = block_.FindVarRecursive(name);
+  PADDLE_ENFORCE(var != nullptr, "Cannot find variable %s", name);
+  return framework::make_ddim(var->Shape());
+}
+
+void CompileTimeInferShapeContext::SetDim(const std::string &name,
+                                          const DDim &dim) {
+  block_.FindVarRecursive(name)->SetShape(framework::vectorize(dim));
+}
+
 }  // namespace framework
 }  // namespace paddle

paddle/framework/op_desc.h

@@ -107,6 +107,8 @@ class OpDescBind {
 
   void InferVarType(BlockDescBind *block) const;
 
+  void MarkAsTarget() { desc_.set_is_target(true); }
+
   void Flush();
 
  private:

paddle/framework/op_registry.h

@@ -29,6 +29,7 @@ limitations under the License. */
 #include "paddle/framework/op_desc.h"
 #include "paddle/framework/operator.h"
 #include "paddle/framework/scope.h"
+#include "paddle/framework/shape_inference.h"
 
 namespace paddle {
 namespace framework {
@@ -161,6 +162,10 @@ class OpKernelRegistrar : public Registrar {
   REGISTER_OPERATOR(op_type, op_class, _GradOpDescMaker_##grad_op_type##_, \
                     op_maker_class);
 
+#define REGISTER_OP_WITH_KERNEL(op_type, ...)                         \
+  REGISTER_OPERATOR(op_type, ::paddle::framework::OperatorWithKernel, \
+                    ##__VA_ARGS__)
+
 #define REGISTER_OP_WITHOUT_GRADIENT(op_type, op_class, op_maker_class) \
   REGISTER_OPERATOR(op_type, op_class, op_maker_class)
 
@@ -223,6 +228,10 @@ class OpKernelRegistrar : public Registrar {
   USE_OP_ITSELF(op_type);        \
   USE_OP_DEVICE_KERNEL(op_type, CPU);
 
+#define USE_GPU_ONLY_OP(op_type) \
+  USE_OP_ITSELF(op_type);        \
+  USE_OP_DEVICE_KERNEL(op_type, GPU)
+
 #define USE_OP(op_type)   \
   USE_OP_ITSELF(op_type); \
   USE_OP_KERNEL(op_type)

paddle/framework/operator.cc

@@ -15,6 +15,7 @@ limitations under the License. */
 #include "paddle/framework/operator.h"
 #include <algorithm>
 #include <atomic>
+#include "paddle/framework/shape_inference.h"
 
 namespace paddle {
 namespace framework {
@@ -273,5 +274,137 @@ bool OpSupportGPU(const std::string& op_type) {
   return false;
 }
 
+class RuntimeInferShapeContext : public InferShapeContext {
+ public:
+  RuntimeInferShapeContext(const OperatorBase& op, const Scope& scope)
+      : op_(op), scope_(scope) {}
+
+  bool HasInput(const std::string& name) const override {
+    auto& ins = Inputs(name);
+    size_t length = ins.size();
+    if (length == 0) {
+      return false;
+    }
+    PADDLE_ENFORCE_EQ(length, 1UL, "Input %s should have more than one inputs",
+                      name);
+    auto ipt = ins[0];
+    auto* var = ipt == kEmptyVarName ? nullptr : scope_.FindVar(ipt);
+    return var != nullptr;
+  }
+
+  bool HasOutput(const std::string& name) const override {
+    auto& outs = Outputs(name);
+    size_t length = outs.size();
+    if (length == 0) {
+      return false;
+    }
+    PADDLE_ENFORCE_EQ(length, 1UL, "Output %s should have more than one inputs",
+                      name);
+    auto ipt = outs[0];
+    auto* var = ipt == kEmptyVarName ? nullptr : scope_.FindVar(ipt);
+    return var != nullptr;
+  }
+
+  bool HasInputs(const std::string& name) const override {
+    auto inputs = op_.Inputs(name);
+    if (inputs.empty()) {
+      return false;
+    }
+    for (auto& input : inputs) {
+      if (scope_.FindVar(input) == nullptr) {
+        return false;
+      }
+    }
+    return true;
+  }
+
+  bool HasOutputs(const std::string& name) const override {
+    auto outputs = op_.Outputs(name);
+    if (outputs.empty()) {
+      return false;
+    }
+    for (auto& output : outputs) {
+      if (scope_.FindVar(output) == nullptr) {
+        return false;
+      }
+    }
+    return true;
+  }
+
+  DDim GetInputDim(const std::string& name) const override {
+    return GetDim(op_.Input(name));
+  }
+
+  void SetOutputDim(const std::string& name, const DDim& dim) override {
+    SetDim(op_.Output(name), dim);
+  }
+
+  AttrReader Attrs() const override { return AttrReader(op_.Attrs()); }
+
+  const std::vector<std::string>& Inputs(
+      const std::string& name) const override {
+    return op_.Inputs(name);
+  }
+
+  const std::vector<std::string>& Outputs(
+      const std::string& name) const override {
+    return op_.Outputs(name);
+  }
+
+ private:
+  DDim GetDim(const std::string& name) const override {
+    Variable* var = scope_.FindVar(name);
+    if (var->IsType<LoDTensor>()) {
+      return var->Get<LoDTensor>().dims();
+    } else if (var->IsType<SelectedRows>()) {
+      return var->Get<SelectedRows>().GetCompleteDims();
+    } else {
+      PADDLE_THROW("Variable type must be LoDTensor/SelectedRows.");
+    }
+  }
+
+  void SetDim(const std::string& name, const DDim& dim) override {
+    Variable* var = scope_.FindVar(name);
+    if (var->IsType<LoDTensor>()) {
+      var->GetMutable<LoDTensor>()->Resize(dim);
+    } else if (var->IsType<SelectedRows>()) {
+      var->GetMutable<SelectedRows>()->set_height(dim[0]);
+    } else {
+      PADDLE_THROW("Variable type must be LoDTensor/SelectedRows.");
+    }
+  }
+
+  const OperatorBase& op_;
+  const Scope& scope_;
+};
+
+void OperatorWithKernel::Run(const Scope& scope,
+                             const platform::DeviceContext& dev_ctx) const {
+  VLOG(3) << "Running operator " << this->Type();
+  RuntimeInferShapeContext infer_shape_ctx(*this, scope);
+  this->InferShape(&infer_shape_ctx);
+
+  ExecutionContext ctx(*this, scope, dev_ctx);
+
+  // check if op[type] has kernel registered.
+  auto& all_op_kernels = AllOpKernels();
+  auto kernels_iter = all_op_kernels.find(type_);
+  if (kernels_iter == all_op_kernels.end()) {
+    PADDLE_THROW(
+        "There are no kernels which are registered in the %s operator.", type_);
+  }
+
+  // check if op[type] have kernel for kernel_key
+  OpKernelMap& kernels = kernels_iter->second;
+  auto kernel_key = OpKernelKey(IndicateDataType(ctx), dev_ctx);
+  auto kernel_iter = kernels.find(kernel_key);
+
+  if (kernel_iter == kernels.end()) {
+    PADDLE_THROW("The operator %s does not support %s", type_, kernel_key);
+  }
+
+  kernel_iter->second->Compute(ctx);
+}
+
 }  // namespace framework
 }  // namespace paddle

paddle/framework/program_desc.cc

@@ -49,6 +49,13 @@ ProgramDescBind::ProgramDescBind(const ProgramDescBind &o) {
   }
 }
 
+ProgramDescBind::ProgramDescBind(const ProgramDesc &desc) {
+  desc_ = desc;
+  for (auto &block_desc : *desc_.mutable_blocks()) {
+    blocks_.emplace_back(new BlockDescBind(this, &block_desc));
+  }
+}
+
 ProgramDescBind::ProgramDescBind(const std::string &binary_str) {
   PADDLE_ENFORCE(desc_.ParseFromString(binary_str),
                  "Fail to parse program_desc from binary string.");

paddle/framework/program_desc.h

@@ -29,6 +29,8 @@ class ProgramDescBind {
  public:
   ProgramDescBind();
 
+  explicit ProgramDescBind(const ProgramDesc &desc);
+
   ProgramDescBind(const ProgramDescBind &o);
 
   explicit ProgramDescBind(const std::string &binary_str);

paddle/framework/prune.cc

@@ -46,7 +46,7 @@ bool IsTarget(const OpDesc& op_desc) {
   return false;
 }
 
-void prune_impl(const ProgramDesc& input, ProgramDesc& output, int block_id) {
+void prune_impl(const ProgramDesc& input, ProgramDesc* output, int block_id) {
   // TODO(tonyyang-svail):
   //    - will change to use multiple blocks for RNN op and Cond Op
 
@@ -91,8 +91,8 @@ void prune_impl(const ProgramDesc& input, ProgramDesc& output, int block_id) {
   // we reverse the should_run vector
   std::reverse(should_run.begin(), should_run.end());
 
-  output = input;
+  *output = input;
-  auto* op_field = output.mutable_blocks(block_id)->mutable_ops();
+  auto* op_field = output->mutable_blocks(block_id)->mutable_ops();
   op_field->Clear();
   for (size_t i = 0; i < should_run.size(); ++i) {
     if (should_run[i]) {
@@ -101,7 +101,8 @@ void prune_impl(const ProgramDesc& input, ProgramDesc& output, int block_id) {
   }
 }
 
-void Prune(const ProgramDesc& input, ProgramDesc& output) {
+// TODO(fengjiayi): Prune() could be inplaced to avoid unnecessary copies
+void Prune(const ProgramDesc& input, ProgramDesc* output) {
   prune_impl(input, output, 0);
 }
 

paddle/framework/prune.h

@@ -20,7 +20,7 @@ limitations under the License. */
 namespace paddle {
 namespace framework {
 
-void Prune(const ProgramDesc& input, ProgramDesc& output);
+void Prune(const ProgramDesc& input, ProgramDesc* output);
 
 }  // namespace framework
 }  // namespace paddle

paddle/framework/prune_test.cc

@@ -59,11 +59,11 @@ TEST(Prune, one_operator) {
   f::ProgramDesc *pdesc = program.Proto();
   f::ProgramDesc pruned;
 
-  Prune(*pdesc, pruned);
+  Prune(*pdesc, &pruned);
   PADDLE_ENFORCE_EQ(pruned.blocks(0).ops_size(), 0);
 
   pdesc->mutable_blocks(0)->mutable_ops(0)->set_is_target(true);
-  Prune(*pdesc, pruned);
+  Prune(*pdesc, &pruned);
   PADDLE_ENFORCE_EQ(pruned.blocks(0).ops_size(), 1);
 }
 
@@ -81,7 +81,7 @@ TEST(Prune, forward) {
   for (int i = 0; i < pdesc->blocks(0).ops_size(); ++i) {
     f::ProgramDesc pruned;
     pdesc->mutable_blocks(0)->mutable_ops(i)->set_is_target(true);
-    Prune(*pdesc, pruned);
+    Prune(*pdesc, &pruned);
     PADDLE_ENFORCE_EQ(pruned.blocks(0).ops_size(), i + 1);
   }
 }
@@ -100,7 +100,7 @@ TEST(Prune, multi_input_op) {
   pdesc->mutable_blocks(0)->mutable_ops(3)->set_is_target(true);
 
   f::ProgramDesc pruned;
-  Prune(*pdesc, pruned);
+  Prune(*pdesc, &pruned);
   PADDLE_ENFORCE_EQ(pruned.blocks(0).ops_size(), 4);
 }
 
@@ -116,7 +116,7 @@ TEST(Prune, multi_output_op) {
   pdesc->mutable_blocks(0)->mutable_ops(2)->set_is_target(true);
 
   f::ProgramDesc pruned;
-  Prune(*pdesc, pruned);
+  Prune(*pdesc, &pruned);
   PADDLE_ENFORCE_EQ(pruned.blocks(0).ops_size(), 2);
 }
 
@@ -133,6 +133,6 @@ TEST(Prune, multi_target) {
   pdesc->mutable_blocks(0)->mutable_ops(2)->set_is_target(true);
 
   f::ProgramDesc pruned;
-  Prune(*pdesc, pruned);
+  Prune(*pdesc, &pruned);
   PADDLE_ENFORCE_EQ(pruned.blocks(0).ops_size(), 3);
 }

paddle/framework/scope.cc

@@ -16,6 +16,7 @@ limitations under the License. */
 
 #include <memory>  // for unique_ptr
 #include <mutex>   // for call_once
+#include "glog/logging.h"
 #include "paddle/string/printf.h"
 
 namespace paddle {
@@ -23,7 +24,10 @@ namespace framework {
 
 Scope::~Scope() {
   DropKids();
-  for (auto& kv : vars_) delete kv.second;
+  for (auto& kv : vars_) {
+    VLOG(3) << "Destroy variable " << kv.first;
+    delete kv.second;
+  }
 }
 
 Scope& Scope::NewScope() const {
@@ -38,6 +42,7 @@ Variable* Scope::Var(const std::string& name) {
   }
   Variable* v = new Variable();
   vars_[name] = v;
+  VLOG(3) << "Create variable " << name << " on scope";
   v->name_ = &(vars_.find(name)->first);
   return v;
 }

paddle/framework/shape_inference.cc

@@ -0,0 +1,54 @@
+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+   http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License. */
+#include "paddle/framework/shape_inference.h"
+
+namespace paddle {
+namespace framework {
+
+std::vector<framework::DDim> InferShapeContext::GetInputsDim(
+    const std::string &name) const {
+  const std::vector<std::string> &names = Inputs(name);
+  return GetDims(names);
+}
+
+void InferShapeContext::SetOutputsDim(
+    const std::string &name, const std::vector<framework::DDim> &dims) {
+  auto &names = Outputs(name);
+  SetDims(names, dims);
+}
+
+void InferShapeContext::ShareLoD(const std::string &in, const std::string &out,
+                                 size_t i, size_t j) const {}
+
+std::vector<framework::DDim> InferShapeContext::GetDims(
+    const std::vector<std::string> &names) const {
+  std::vector<framework::DDim> ret;
+  ret.reserve(names.size());
+  std::transform(
+      names.begin(), names.end(), std::back_inserter(ret),
+      [this](const std::string &name) { return this->GetDim(name); });
+  return ret;
+}
+
+void InferShapeContext::SetDims(const std::vector<std::string> &names,
+                                const std::vector<framework::DDim> &dims) {
+  size_t length = names.size();
+  PADDLE_ENFORCE_EQ(length, dims.size());
+  for (size_t i = 0; i < length; ++i) {
+    SetDim(names[i], dims[i]);
+  }
+}
+
+}  // namespace framework
+}  // namespace paddle

paddle/framework/shape_inference.h

@@ -14,6 +14,7 @@ limitations under the License. */
 
 #pragma once
 
+#include "paddle/framework/attribute.h"
 #include "paddle/framework/ddim.h"
 
 namespace paddle {
@@ -21,7 +22,7 @@ namespace framework {
 
 class InferShapeContext {
  public:
-  virtual ~InferShapeContext() {}
+  virtual ~InferShapeContext() = default;
   virtual bool HasInput(const std::string &name) const = 0;
   virtual bool HasOutput(const std::string &name) const = 0;
 
@@ -29,57 +30,32 @@ class InferShapeContext {
   virtual bool HasOutputs(const std::string &name) const = 0;
 
   virtual framework::DDim GetInputDim(const std::string &name) const = 0;
-  std::vector<framework::DDim> GetInputsDim(const std::string &name) const {
+
-    const std::vector<std::string> &names = Inputs(name);
+  std::vector<framework::DDim> GetInputsDim(const std::string &name) const;
-    return GetDims(names);
+
-  }
-  virtual void SetInputDim(const std::string &name,
-                           const framework::DDim &dim) = 0;
-  void SetInputsDim(const std::string &name,
-                    const std::vector<framework::DDim> &dims) {
-    auto &names = Inputs(name);
-    SetDims(names, dims);
-  }
-  virtual framework::DDim GetOutputDim(const std::string &name) const = 0;
-  std::vector<framework::DDim> GetOutputsDim(const std::string &name) const {
-    const std::vector<std::string> &names = Outputs(name);
-    return GetDims(names);
-  }
   virtual void SetOutputDim(const std::string &name, const DDim &dim) = 0;
   void SetOutputsDim(const std::string &name,
-                     const std::vector<framework::DDim> &dims) {
+                     const std::vector<framework::DDim> &dims);
-    auto &names = Outputs(name);
+
-    SetDims(names, dims);
-  }
   virtual AttrReader Attrs() const = 0;
   virtual const std::vector<std::string> &Inputs(
       const std::string &name) const = 0;
   virtual const std::vector<std::string> &Outputs(
       const std::string &name) const = 0;
+
   // TODO(qiao) implement this function
   void ShareLoD(const std::string &in, const std::string &out, size_t i = 0,
-                size_t j = 0) const {}
+                size_t j = 0) const;
 
  protected:
   virtual framework::DDim GetDim(const std::string &name) const = 0;
   virtual void SetDim(const std::string &name, const framework::DDim &dim) = 0;
+
   std::vector<framework::DDim> GetDims(
-      const std::vector<std::string> &names) const {
+      const std::vector<std::string> &names) const;
-    std::vector<framework::DDim> ret;
+
-    ret.reserve(names.size());
-    std::transform(
-        names.begin(), names.end(), std::back_inserter(ret),
-        [this](const std::string &name) { return this->GetDim(name); });
-    return ret;
-  }
   void SetDims(const std::vector<std::string> &names,
-               const std::vector<framework::DDim> &dims) {
+               const std::vector<framework::DDim> &dims);
-    size_t length = names.size();
-    PADDLE_ENFORCE_EQ(length, dims.size());
-    for (size_t i = 0; i < length; ++i) {
-      SetDim(names[i], dims[i]);
-    }
-  }
 };
 
 }  // namespace framework

paddle/framework/tensor.h

@@ -126,11 +126,16 @@ class Tensor {
   inline Tensor Slice(const int& begin_idx, const int& end_idx) const;
 
   platform::Place place() const {
-    PADDLE_ENFORCE_NOT_NULL(holder_, "Tensor get place() must contains holder");
+    PADDLE_ENFORCE_NOT_NULL(
+        holder_, "Tensor not initialized yet when Tensor::place() is called.");
     return holder_->place();
   }
 
-  std::type_index type() const { return holder_->type(); }
+  std::type_index type() const {
+    PADDLE_ENFORCE_NOT_NULL(
+        holder_, "Tensor not initialized yet when Tensor::type() is called.");
+    return holder_->type();
+  }
 
   size_t memory_size() const;
 

paddle/memory/CMakeLists.txt

@@ -1,6 +1,6 @@
 add_subdirectory(detail)
 
-cc_library(memory SRCS memory.cc)
+cc_library(memory SRCS memory.cc DEPS place)
 cc_library(memcpy SRCS memcpy.cc)
 
 cc_library(paddle_memory

paddle/memory/detail/meta_cache.cc

@@ -13,6 +13,7 @@
    limitations under the License. */
 
 #include "paddle/memory/detail/meta_cache.h"
+#include "glog/logging.h"
 #include "paddle/memory/detail/memory_block.h"
 #include "paddle/platform/assert.h"
 
@@ -28,7 +29,9 @@ Metadata MetadataCache::load(const MemoryBlock* block) {
     PADDLE_ASSERT(existing_metadata->second.check_guards());
     return existing_metadata->second;
   } else {
-    PADDLE_ASSERT(reinterpret_cast<const Metadata*>(block)->check_guards());
+    auto* meta = reinterpret_cast<const Metadata*>(block);
+    VLOG(3) << "Load MetaData type=" << meta->type;
+    PADDLE_ASSERT(meta->check_guards());
     return *reinterpret_cast<const Metadata*>(block);
   }
 }

paddle/memory/memory.cc

@@ -39,11 +39,15 @@ BuddyAllocator* GetCPUBuddyAllocator() {
 
 template <>
 void* Alloc<platform::CPUPlace>(platform::CPUPlace place, size_t size) {
-  return GetCPUBuddyAllocator()->Alloc(size);
+  VLOG(3) << "Allocate " << size << " bytes on " << platform::Place(place);
+  void* p = GetCPUBuddyAllocator()->Alloc(size);
+  VLOG(3) << "  pointer=" << p;
+  return p;
 }
 
 template <>
 void Free<platform::CPUPlace>(platform::CPUPlace place, void* p) {
+  VLOG(3) << "Free pointer=" << p << " on " << platform::Place(place);
   GetCPUBuddyAllocator()->Free(p);
 }
 

paddle/operators/CMakeLists.txt

@@ -97,6 +97,13 @@ function(op_library TARGET)
         file(APPEND ${pybind_file} "USE_OP(sigmoid);\n")
     endif()
 
+    # nccl_op contains several operators
+    if ("${TARGET}" STREQUAL "nccl_op")
+        set(pybind_flag 1)
+        # It's enough to just adding one operator to pybind
+        file(APPEND ${pybind_file} "USE_GPU_ONLY_OP(ncclAllReduce);\n")
+    endif()
+    
     # reduce_op contains several operators
     if ("${TARGET}" STREQUAL "reduce_op")
         set(pybind_flag 1)
@@ -128,6 +135,7 @@ function(op_library TARGET)
 endfunction()
 
 add_subdirectory(math)
+add_subdirectory(nccl)
 
 set(DEPS_OPS
     recurrent_op
@@ -139,6 +147,7 @@ set(DEPS_OPS
     pool_with_index_op
     lstm_op
     conv_transpose_op
+    nccl_op
     sequence_conv_op
     lstm_op)
 
@@ -151,6 +160,9 @@ op_library(softmax_with_cross_entropy_op DEPS cross_entropy softmax)
 op_library(sum_op DEPS net_op selected_rows_functor)
 op_library(pool_op DEPS pooling)
 op_library(pool_with_index_op DEPS pooling)
+if(WITH_GPU)
+op_library(nccl_op DEPS nccl_common)
+endif()
 op_library(sequence_conv_op DEPS context_project)
 op_library(lstm_op DEPS sequence2batch lstm_compute)
 op_library(conv_transpose_op DEPS vol2col)
@@ -167,4 +179,8 @@ cc_test(net_op_test SRCS net_op_test.cc DEPS net_op)
 cc_test(scatter_test SRCS scatter_test.cc DEPS tensor)
 cc_test(strided_memcpy_test SRCS strided_memcpy_test.cc DEPS tensor paddle_memory)
 cc_test(dynamic_recurrent_op_test SRCS dynamic_recurrent_op_test.cc DEPS dynamic_recurrent_op recurrent_op tensor_array)
+
+if(WITH_GPU)
+  nv_test(nccl_op_test SRCS nccl_op_test.cu DEPS nccl_op gpu_info device_context)
+endif()
 cc_test(save_load_op_test SRCS save_load_op_test.cc DEPS save_op load_op)

paddle/operators/accuracy_op.cc

@@ -70,7 +70,5 @@ information, or not. But the output only shares the LoD with input `Inference`.
 namespace ops = paddle::operators;
 REGISTER_OP_WITHOUT_GRADIENT(accuracy, ops::AccuracyOp, ops::AccuracyOpMaker);
 REGISTER_OP_CPU_KERNEL(
-    accuracy, ops::AccuracyKernel<paddle::platform::CPUPlace, float>,
+    accuracy, ops::AccuracyKernel<paddle::platform::CPUPlace, int>,
-    ops::AccuracyKernel<paddle::platform::CPUPlace, int>,
-    ops::AccuracyKernel<paddle::platform::CPUPlace, double>,
     ops::AccuracyKernel<paddle::platform::CPUPlace, int64_t>);

paddle/operators/accuracy_op.cu

@@ -81,7 +81,5 @@ class AccuracyOpCUDAKernel : public framework::OpKernel<T> {
 }  // namespace operators
 }  // namespace paddle
 
-REGISTER_OP_GPU_KERNEL(accuracy, paddle::operators::AccuracyOpCUDAKernel<float>,
+REGISTER_OP_GPU_KERNEL(accuracy, paddle::operators::AccuracyOpCUDAKernel<int>,
-                       paddle::operators::AccuracyOpCUDAKernel<double>,
-                       paddle::operators::AccuracyOpCUDAKernel<int>,
                        paddle::operators::AccuracyOpCUDAKernel<int64_t>);

paddle/operators/batch_norm_op.cc

@@ -18,6 +18,7 @@ namespace paddle {
 namespace operators {
 
 using Tensor = framework::Tensor;
+using LoDTensor = framework::LoDTensor;
 template <typename T, int MajorType = Eigen::RowMajor,
           typename IndexType = Eigen::DenseIndex>
 using EigenMatrix = framework::EigenMatrix<T, MajorType, IndexType>;
@@ -64,6 +65,9 @@ class BatchNormOp : public framework::OperatorWithKernel {
         (tensor_format == TensorFormat::NCHW ? x_dims[1]
                                              : x_dims[x_dims.size() - 1]);
 
+    PADDLE_ENFORCE(x_dims.size() >= 3 && x_dims.size() <= 5,
+                   "Input x must have 3 to 5 dimensions.");
+
     PADDLE_ENFORCE_EQ(ctx->GetInputDim("Scale").size(), 1UL);
     PADDLE_ENFORCE_EQ(ctx->GetInputDim("Scale")[0], C);
     PADDLE_ENFORCE_EQ(ctx->GetInputDim("Bias").size(), 1UL);
@@ -108,10 +112,12 @@ class BatchNormOpMaker : public framework::OpProtoAndCheckerMaker {
               "Store the global Variance when training");
     AddOutput("SavedMean",
               "Mean of the current mini batch, "
-              "will apply to output when training");
+              "will apply to output when training")
+        .AsIntermediate();
     AddOutput("SavedVariance",
               "Variance of the current mini batch, "
-              "will apply to output when training");
+              "will apply to output when training")
+        .AsIntermediate();
     AddComment(R"DOC(
 https://arxiv.org/pdf/1502.03167.pdf
 
@@ -135,7 +141,6 @@ class BatchNormKernel<platform::CPUPlace, T> : public framework::OpKernel<T> {
 
     const auto *x = ctx.Input<Tensor>("X");
     const auto &x_dims = x->dims();
-
     PADDLE_ENFORCE(x_dims.size() >= 3 && x_dims.size() <= 5,
                    "The Input dim size should be between 3 and 5");
     const int N = x_dims[0];
@@ -289,6 +294,25 @@ class BatchNormGradOp : public framework::OperatorWithKernel {
     ctx->SetOutputDim(framework::GradVarName("Scale"), {C});
     ctx->SetOutputDim(framework::GradVarName("Bias"), {C});
   }
+
+  framework::DataType IndicateDataType(
+      const framework::ExecutionContext &ctx) const override {
+    VLOG(3) << "IndicateDataType " << this->Type();
+    const auto *var = ctx.InputVar(framework::GradVarName("Y"));
+    if (var == nullptr) {
+      PADDLE_THROW("can't find Y@GRAD");
+    }
+    const Tensor *t = nullptr;
+    if (var->IsType<Tensor>()) {
+      t = &var->Get<Tensor>();
+    } else if (var->IsType<LoDTensor>()) {
+      t = &var->Get<LoDTensor>();
+    }
+    if (t == nullptr) {
+      PADDLE_THROW("can't find Y@GRAD");
+    }
+    return framework::ToDataType(t->type());
+  }
 };
 
 template <typename T>

paddle/operators/batch_norm_op.cu

@@ -117,9 +117,6 @@ class BatchNormKernel<platform::GPUPlace, T> : public framework::OpKernel<T> {
     math::SetConstant<platform::GPUPlace, T> functor;
     functor(ctx.device_context(), saved_mean, 0);
     functor(ctx.device_context(), saved_variance, 0);
-    // FIXME(qiao) should not set zero self
-    functor(ctx.device_context(), mean_out, 0);
-    functor(ctx.device_context(), variance_out, 0);
 
     auto handle = ctx.cuda_device_context().cudnn_handle();
 
@@ -211,8 +208,15 @@ class BatchNormGradKernel<platform::GPUPlace, T>
     mode_ = CUDNN_BATCHNORM_SPATIAL;
 #endif
 
-    std::vector<int> dims = {N, C, H, W, D};
+    std::vector<int> dims;
-    std::vector<int> strides = {H * W * C * D, 1, W * D * C, D * C, C};
+    std::vector<int> strides;
+    if (tensor_format == TensorFormat::NCHW) {
+      dims = {N, C, H, W, D};
+      strides = {C * H * W * D, H * W * D, W * D, D, 1};
+    } else {
+      dims = {N, C, H, W, D};
+      strides = {H * W * C * D, 1, W * D * C, D * C, C};
+    }
     CUDNN_ENFORCE(platform::dynload::cudnnSetTensorNdDescriptor(
         data_desc_, CudnnDataType<T>::type,
         x_dims.size() > 3 ? x_dims.size() : 4, dims.data(), strides.data()));