From 402a9f1f24b0b2044940fc73102d652c2164ec5c Mon Sep 17 00:00:00 2001
From: Yi Wang <yi.wang.2005@gmail.com>
Date: Sat, 7 Apr 2018 17:34:33 -0700
Subject: [PATCH 01/16] Rewrite the interface of memroy/detail

---
 paddle/fluid/memory/CMakeLists.txt            |   6 +-
 paddle/fluid/memory/detail/CMakeLists.txt     |  10 +-
 paddle/fluid/memory/detail/buddy_allocator.h  |  14 +-
 paddle/fluid/memory/detail/memory_block.cc    |  95 +++++++------
 paddle/fluid/memory/detail/memory_block.h     | 125 ++++++++++++------
 paddle/fluid/memory/detail/meta_cache.cc      |   7 +-
 paddle/fluid/memory/detail/meta_cache.h       |  64 ---------
 paddle/fluid/memory/detail/meta_data.cc       |  26 ++--
 paddle/fluid/memory/detail/meta_data.h        |  54 --------
 .../fluid/memory/detail/system_allocator.cc   |  22 +--
 paddle/fluid/memory/detail/system_allocator.h |   8 +-
 .../memory/detail/system_allocator_test.cc    |  18 +--
 12 files changed, 180 insertions(+), 269 deletions(-)
 delete mode 100644 paddle/fluid/memory/detail/meta_cache.h
 delete mode 100644 paddle/fluid/memory/detail/meta_data.h

diff --git a/paddle/fluid/memory/CMakeLists.txt b/paddle/fluid/memory/CMakeLists.txt
index 8b3043af7a..3b73e79eb3 100644
--- a/paddle/fluid/memory/CMakeLists.txt
+++ b/paddle/fluid/memory/CMakeLists.txt
@@ -7,11 +7,7 @@ cc_library(paddle_memory
         DEPS
         memory
         memcpy
-        meta_data
-        meta_cache
-        memory_block
-        buddy_allocator
-        system_allocator)
+        buddy_allocator)
 
 cc_test(memory_test SRCS memory_test.cc DEPS place paddle_memory)
 
diff --git a/paddle/fluid/memory/detail/CMakeLists.txt b/paddle/fluid/memory/detail/CMakeLists.txt
index b9c3fc31c1..1636b120ae 100644
--- a/paddle/fluid/memory/detail/CMakeLists.txt
+++ b/paddle/fluid/memory/detail/CMakeLists.txt
@@ -1,3 +1,5 @@
+cc_library(memory_block SRCS memory_block.cc meta_data.cc meta_cache.cc)
+
 if(${WITH_GPU})
   nv_library(system_allocator SRCS system_allocator.cc DEPS gflags cpu_info gpu_info)
 else(${WITH_GPU})
@@ -6,10 +8,4 @@ endif(${WITH_GPU})
 
 cc_test(system_allocator_test SRCS system_allocator_test.cc DEPS system_allocator)
 
-cc_library(meta_data SRCS meta_data.cc)
-
-cc_library(meta_cache SRCS meta_cache.cc)
-
-cc_library(memory_block SRCS memory_block.cc)
-
-cc_library(buddy_allocator SRCS buddy_allocator.cc DEPS glog)
+cc_library(buddy_allocator SRCS buddy_allocator.cc DEPS memory_block system_allocator glog)
diff --git a/paddle/fluid/memory/detail/buddy_allocator.h b/paddle/fluid/memory/detail/buddy_allocator.h
index a4ee70c258..2f39d774d6 100644
--- a/paddle/fluid/memory/detail/buddy_allocator.h
+++ b/paddle/fluid/memory/detail/buddy_allocator.h
@@ -14,18 +14,18 @@ limitations under the License. */
 
 #pragma once
 
-#include "paddle/fluid/memory/detail/meta_cache.h"
-#include "paddle/fluid/memory/detail/meta_data.h"
+#include <mutex>  // NOLINT
+#include <set>
+#include <tuple>
+#include <unordered_map>
+#include <vector>
+
+#include "paddle/fluid/memory/detail/memory_block.h"
 #include "paddle/fluid/memory/detail/system_allocator.h"
 #include "paddle/fluid/platform/assert.h"
 #include "paddle/fluid/platform/cpu_info.h"
 #include "paddle/fluid/platform/gpu_info.h"
 
-#include <mutex>
-#include <set>
-#include <unordered_map>
-#include <vector>
-
 namespace paddle {
 namespace memory {
 namespace detail {
diff --git a/paddle/fluid/memory/detail/memory_block.cc b/paddle/fluid/memory/detail/memory_block.cc
index 07123f2669..f744450f48 100644
--- a/paddle/fluid/memory/detail/memory_block.cc
+++ b/paddle/fluid/memory/detail/memory_block.cc
@@ -13,17 +13,15 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "paddle/fluid/memory/detail/memory_block.h"
-#include "paddle/fluid/memory/detail/meta_cache.h"
-#include "paddle/fluid/memory/detail/meta_data.h"
 #include "paddle/fluid/platform/assert.h"
 
 namespace paddle {
 namespace memory {
 namespace detail {
 
-void MemoryBlock::init(MetadataCache& cache, Type t, size_t index, size_t size,
+void MemoryBlock::init(MetadataCache* cache, Type t, size_t index, size_t size,
                        void* left_buddy, void* right_buddy) {
-  cache.store(this, Metadata(t, index, size - sizeof(Metadata), size,
+  cache->save(this, Metadata(t, index, size - sizeof(Metadata), size,
                              static_cast<MemoryBlock*>(left_buddy),
                              static_cast<MemoryBlock*>(right_buddy)));
 }
@@ -32,115 +30,112 @@ MemoryBlock::Type MemoryBlock::type(MetadataCache& cache) const {
   return cache.load(this).type;
 }
 
-size_t MemoryBlock::size(MetadataCache& cache) const {
+size_t MemoryBlock::size(const MetadataCache& cache) const {
   return cache.load(this).size;
 }
 
-size_t MemoryBlock::total_size(MetadataCache& cache) const {
+size_t MemoryBlock::index(const MetadataCache& cache) const {
+  return cache.load(this).index;
+}
+
+size_t MemoryBlock::total_size(const MetadataCache& cache) const {
   return cache.load(this).total_size;
 }
 
-MemoryBlock* MemoryBlock::left_buddy(MetadataCache& cache) const {
+bool MemoryBlock::has_left_buddy(const MetadataCache& cache) const {
+  return left_buddy(cache) != nullptr;
+}
+
+bool MemoryBlock::has_right_buddy(const MetadataCache& cache) const {
+  return right_buddy(cache) != nullptr;
+}
+
+MemoryBlock* MemoryBlock::left_buddy(const MetadataCache& cache) const {
   return cache.load(this).left_buddy;
 }
 
-MemoryBlock* MemoryBlock::right_buddy(MetadataCache& cache) const {
+MemoryBlock* MemoryBlock::right_buddy(const MetadataCache& cache) const {
   return cache.load(this).right_buddy;
 }
 
-void MemoryBlock::split(MetadataCache& cache, size_t size) {
+void MemoryBlock::split(MetadataCache* cache, size_t size) {
   // make sure the split fits
-  PADDLE_ASSERT(total_size(cache) >= size);
+  PADDLE_ASSERT(total_size(*cache) >= size);
 
   // bail out if there is no room for another partition
-  if (total_size(cache) - size <= sizeof(Metadata)) {
+  if (total_size(*cache) - size <= sizeof(Metadata)) {
     return;
   }
 
   // find the position of the split
   void* right_partition = reinterpret_cast<uint8_t*>(this) + size;
 
-  size_t remaining_size = total_size(cache) - size;
+  size_t remaining_size = total_size(*cache) - size;
 
   // Add the new block as a buddy
-  auto metadata = cache.load(this);
+  auto metadata = cache->load(this);
 
   // Write the metadata for the new block
   auto new_block_right_buddy = metadata.right_buddy;
 
-  cache.store(
+  cache->save(
       static_cast<MemoryBlock*>(right_partition),
-      Metadata(FREE_CHUNK, index(cache), remaining_size - sizeof(Metadata),
+      Metadata(FREE_CHUNK, index(*cache), remaining_size - sizeof(Metadata),
                remaining_size, this, new_block_right_buddy));
 
   metadata.right_buddy = static_cast<MemoryBlock*>(right_partition);
   metadata.size = size - sizeof(Metadata);
   metadata.total_size = size;
 
-  cache.store(this, metadata);
+  cache->save(this, metadata);
 
   // Write metadata for the new block's right buddy
   if (new_block_right_buddy != nullptr) {
-    auto buddy_metadata = cache.load(new_block_right_buddy);
+    auto buddy_metadata = cache->load(new_block_right_buddy);
 
     buddy_metadata.left_buddy = static_cast<MemoryBlock*>(right_partition);
 
-    cache.store(new_block_right_buddy, buddy_metadata);
+    cache->save(new_block_right_buddy, buddy_metadata);
   }
 }
 
-void MemoryBlock::merge(MetadataCache& cache, MemoryBlock* right_buddy) {
+void MemoryBlock::merge(MetadataCache* cache, MemoryBlock* right_buddy) {
   // only free blocks can be merged
-  PADDLE_ASSERT(type(cache) == FREE_CHUNK);
-  PADDLE_ASSERT(right_buddy->type(cache) == FREE_CHUNK);
+  PADDLE_ASSERT(type(*cache) == FREE_CHUNK);
+  PADDLE_ASSERT(right_buddy->type(*cache) == FREE_CHUNK);
 
-  auto metadata = cache.load(this);
+  auto metadata = cache->load(this);
 
   // link this->buddy's buddy
-  metadata.right_buddy = right_buddy->right_buddy(cache);
+  metadata.right_buddy = right_buddy->right_buddy(*cache);
 
   // link buddy's buddy -> this
   if (metadata.right_buddy != nullptr) {
-    auto buddy_metadata = cache.load(metadata.right_buddy);
+    auto buddy_metadata = cache->load(metadata.right_buddy);
 
     buddy_metadata.left_buddy = this;
 
-    cache.store(metadata.right_buddy, buddy_metadata);
+    cache->save(metadata.right_buddy, buddy_metadata);
   }
 
-  metadata.size += right_buddy->total_size(cache);
-  metadata.total_size += right_buddy->total_size(cache);
+  metadata.size += right_buddy->total_size(*cache);
+  metadata.total_size += right_buddy->total_size(*cache);
 
-  cache.store(this, metadata);
-  cache.store(right_buddy, Metadata(INVALID_CHUNK, 0, 0, 0, nullptr, nullptr));
+  cache->save(this, metadata);
+  cache->save(right_buddy, Metadata(INVALID_CHUNK, 0, 0, 0, nullptr, nullptr));
 }
 
-void MemoryBlock::mark_as_free(MetadataCache& cache) {
+void MemoryBlock::mark_as_free(MetadataCache* cache) {
   // check for double free or corruption
-  PADDLE_ASSERT(type(cache) != FREE_CHUNK);
-  PADDLE_ASSERT(type(cache) != INVALID_CHUNK);
-
+  PADDLE_ASSERT(type(*cache) != FREE_CHUNK);
+  PADDLE_ASSERT(type(*cache) != INVALID_CHUNK);
   set_type(cache, FREE_CHUNK);
 }
 
-void MemoryBlock::set_type(MetadataCache& cache, Type t) {
-  auto metadata = cache.load(this);
-
+void MemoryBlock::set_type(MetadataCache* cache, Type t) {
+  auto metadata = cache->load(this);
   metadata.type = t;
-
-  cache.store(this, metadata);
-}
-
-bool MemoryBlock::has_left_buddy(MetadataCache& cache) const {
-  return left_buddy(cache) != nullptr;
-}
-
-bool MemoryBlock::has_right_buddy(MetadataCache& cache) const {
-  return right_buddy(cache) != nullptr;
-}
-
-size_t MemoryBlock::index(MetadataCache& cache) const {
-  return cache.load(this).index;
+  cache->save(this, metadata);
 }
 
 void* MemoryBlock::data() const {
diff --git a/paddle/fluid/memory/detail/memory_block.h b/paddle/fluid/memory/detail/memory_block.h
index 72b40b7317..5e83a2f899 100644
--- a/paddle/fluid/memory/detail/memory_block.h
+++ b/paddle/fluid/memory/detail/memory_block.h
@@ -11,7 +11,6 @@ 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. */
-
 #pragma once
 
 #include <cstddef>
@@ -20,10 +19,11 @@ namespace paddle {
 namespace memory {
 namespace detail {
 
-// Forward Declarations
+// Forward declaration.
 class MetadataCache;
 
-/*! \brief A class used to interpret the contents of a memory block */
+// MemoryBlock represents Each allocated memory block, which contains
+// Metadata and the payload.
 class MemoryBlock {
  public:
   enum Type {
@@ -33,57 +33,96 @@ class MemoryBlock {
     INVALID_CHUNK  // memory is invalid
   };
 
- public:
-  void init(MetadataCache& cache, Type t, size_t index, size_t size,
+  // init saves the Metadata of the memory block in a MetadataCache.
+  // If it is a CPU memory block, the MetadataCache writes the
+  // Metadata to the beginning of the block; or, if it is a GPU memory
+  // block, the MetadataCache writes the Meatadata to a std::map in
+  // the CPU.
+  void init(MetadataCache* cache, Type t, size_t index, size_t size,
             void* left_buddy, void* right_buddy);
 
- public:
-  /*! \brief The type of the allocation */
-  Type type(MetadataCache& cache) const;
-
-  /*! \brief The size of the data region */
-  size_t size(MetadataCache& cache) const;
-
-  /*! \brief An index to track the allocator */
-  size_t index(MetadataCache& cache) const;
-
-  /*! \brief The total size of the block */
-  size_t total_size(MetadataCache& cache) const;
-
-  /*! \brief Check the left buddy of the block */
-  bool has_left_buddy(MetadataCache& cache) const;
-
-  /*! \brief Check the right buddy of the block */
-  bool has_right_buddy(MetadataCache& cache) const;
+  // All these accessors returns fields in the Metadata of the memory
+  // block.  They all need a MetadataCache instance as their first
+  // parameter because they read the Metadata from the cache.
+  Type type(const MetadataCache& cache) const;
+  size_t size(const MetadataCache& cache) const;
+  size_t index(const MetadataCache& cache) const;
+  size_t total_size(const MetadataCache& cache) const;
+  bool has_left_buddy(const MetadataCache& cache) const;
+  bool has_right_buddy(const MetadataCache& cache) const;
+  MemoryBlock* left_buddy(const MetadataCache& cache) const;
+  MemoryBlock* right_buddy(const MetadataCache& cache) const;
 
-  /*! \brief Get the left buddy */
-  MemoryBlock* left_buddy(MetadataCache& cache) const;
+  // Split the allocation into left/right blocks.
+  void split(MetadataCache* cache, size_t size);
 
-  /*! \brief Get the right buddy */
-  MemoryBlock* right_buddy(MetadataCache& cache) const;
+  // Merge left and right blocks together.
+  void merge(MetadataCache* cache, MemoryBlock* right_buddy);
 
- public:
-  /*! \brief Split the allocation into left/right blocks */
-  void split(MetadataCache& cache, size_t size);
-
-  /*! \brief Merge left and right blocks together */
-  void merge(MetadataCache& cache, MemoryBlock* right_buddy);
+  // Mark the allocation as free.
+  void mark_as_free(MetadataCache* cache);
 
-  /*! \brief Mark the allocation as free */
-  void mark_as_free(MetadataCache& cache);
+  // Change the type of the allocation.
+  void set_type(MetadataCache* cache, Type t);
 
-  /*! \brief Change the type of the allocation */
-  void set_type(MetadataCache& cache, Type t);
-
- public:
-  /*! \brief Get a pointer to the memory block's data */
   void* data() const;
-
-  /*! \brief Get a pointer to the memory block's metadata */
   MemoryBlock* metadata() const;
 
+ private:
+  // Metadata describes a MemoryBlock.
+  struct Metadata {
+    Metadata(MemoryBlock::Type t, size_t i, size_t s, size_t ts, MemoryBlock* l,
+             MemoryBlock* r);
+    Metadata();
+
+    // Updates guard_begin and guard_end by hashes of the Metadata object.
+    void update_guards();
+
+    // Checks that guard_begin and guard_end are hashes of the Metadata object.
+    bool check_guards() const;
+
+    // TODO(gangliao): compress this
+    size_t guard_begin = 0;
+    MemoryBlock::Type type = MemoryBlock::INVALID_CHUNK;
+    size_t index = 0;
+    size_t size = 0;
+    size_t total_size = 0;
+    MemoryBlock* left_buddy = nullptr;
+    MemoryBlock* right_buddy = nullptr;
+    size_t guard_end = 0;
+  };
+};
+
+// A cache for accessing memory block meta-data that may be expensive
+// to access directly.  This class exists to unify the metadata format
+// between GPU and CPU allocations. It should be removed when the CPU
+// can access all GPU allocations directly via UVM.
+class MetadataCache {
  public:
-  static size_t overhead();
+  explicit MetadataCache(bool uses_gpu);
+
+  // Disable copying and assignment.
+  MetadataCache(const MetadataCache&) = delete;
+  MetadataCache& operator=(const MetadataCache&) = delete;
+
+  // Returns the Metadata for a memory block.  When MetadataCache is
+  // used to manage CPU memory, the Metadata resides at the beginning
+  // of the memory block; when used to manage GPU memory, the
+  // Meatadata resides in CPU memory indexed by cache_.
+  Metadata load(const MemoryBlock* memory_block) const;
+
+  // Saves the Metadata of a memory block into the cache.  For CPU
+  // memory block, writes the Metadata to the beginning of the memory
+  // block; whereas for GPU memory, writes it to cache_.
+  void save(MemoryBlock* memory_block, const Metadata& meta_data);
+
+  // For GPU memory block, erases its Metadata from cache_.
+  void invalidate(MemoryBlock* memory_block);
+
+ private:
+  typedef std::unordered_map<const MemoryBlock*, Metadata> MetadataMap;
+  MetadataMap cache_;
+  bool uses_gpu_;
 };
 
 }  // namespace detail
diff --git a/paddle/fluid/memory/detail/meta_cache.cc b/paddle/fluid/memory/detail/meta_cache.cc
index 43249e842a..6bcb3a659e 100644
--- a/paddle/fluid/memory/detail/meta_cache.cc
+++ b/paddle/fluid/memory/detail/meta_cache.cc
@@ -12,7 +12,6 @@ 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/fluid/memory/detail/meta_cache.h"
 #include "glog/logging.h"
 #include "paddle/fluid/memory/detail/memory_block.h"
 #include "paddle/fluid/platform/assert.h"
@@ -23,7 +22,7 @@ namespace detail {
 
 MetadataCache::MetadataCache(bool uses_gpu) : uses_gpu_(uses_gpu) {}
 
-Metadata MetadataCache::load(const MemoryBlock* block) {
+Metadata MetadataCache::load(const MemoryBlock* block) const {
   if (uses_gpu_) {
     auto existing_metadata = cache_.find(block);
     PADDLE_ASSERT(existing_metadata->second.check_guards());
@@ -36,8 +35,8 @@ Metadata MetadataCache::load(const MemoryBlock* block) {
   }
 }
 
-void MetadataCache::store(MemoryBlock* block,
-                          const Metadata& original_metadata) {
+void MetadataCache::save(MemoryBlock* block,
+                         const Metadata& original_metadata) {
   auto metadata = original_metadata;
 
   metadata.update_guards();
diff --git a/paddle/fluid/memory/detail/meta_cache.h b/paddle/fluid/memory/detail/meta_cache.h
deleted file mode 100644
index 3283d756a6..0000000000
--- a/paddle/fluid/memory/detail/meta_cache.h
+++ /dev/null
@@ -1,64 +0,0 @@
-/* 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. */
-
-#pragma once
-
-#include "paddle/fluid/memory/detail/memory_block.h"
-#include "paddle/fluid/memory/detail/meta_data.h"
-
-#include <unordered_map>
-
-namespace paddle {
-namespace memory {
-namespace detail {
-
-/**
- *  \brief A cache for accessing memory block meta-data that may be expensive
- *         to access directly.
- *
- *  \note  This class exists to unify the metadata format between GPU and CPU
- *         allocations. It should be removed when the CPU can access all GPU
- *         allocations directly via UVM.
- */
-class MetadataCache {
- public:
-  explicit MetadataCache(bool uses_gpu);
-
- public:
-  /*! \brief Load the associated metadata for the specified memory block. */
-  Metadata load(const MemoryBlock* memory_block);
-
-  /*! \brief Store the associated metadata for the specified memory block. */
-  void store(MemoryBlock* memory_block, const Metadata& meta_data);
-
-  /*! \brief Indicate that the specified metadata will no longer be used. */
-  void invalidate(MemoryBlock* memory_block);
-
- public:
-  MetadataCache(const MetadataCache&) = delete;
-  MetadataCache& operator=(const MetadataCache&) = delete;
-
- private:
-  bool uses_gpu_;
-
- private:
-  typedef std::unordered_map<const MemoryBlock*, Metadata> MetadataMap;
-
- private:
-  MetadataMap cache_;
-};
-
-}  // namespace detail
-}  // namespace memory
-}  // namespace paddle
diff --git a/paddle/fluid/memory/detail/meta_data.cc b/paddle/fluid/memory/detail/meta_data.cc
index ad862af170..02f5235d14 100644
--- a/paddle/fluid/memory/detail/meta_data.cc
+++ b/paddle/fluid/memory/detail/meta_data.cc
@@ -12,10 +12,10 @@ 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/fluid/memory/detail/meta_data.h"
-
 #include <functional>
 
+#include "paddle/fluid/memory/detail/memory_block.h"
+
 namespace paddle {
 namespace memory {
 namespace detail {
@@ -37,25 +37,29 @@ Metadata::Metadata()
       left_buddy(nullptr),
       right_buddy(nullptr) {}
 
+namespace {
+
 template <class T>
-inline void hash_combine(std::size_t& seed, const T& v) {
+inline void hash_combine(std::size_t* seed, const T& v) {
   std::hash<T> hasher;
-  seed ^= hasher(v) + 0x9e3779b9 + (seed << 6) + (seed >> 2);
+  (*seed) ^= hasher(v) + 0x9e3779b9 + ((*seed) << 6) + ((*seed) >> 2);
 }
 
-inline size_t hash(const Metadata* metadata, size_t initial_seed) {
+inline size_t hash(const Metadata& metadata, size_t initial_seed) {
   size_t seed = initial_seed;
 
-  hash_combine(seed, (size_t)metadata->type);
-  hash_combine(seed, metadata->index);
-  hash_combine(seed, metadata->size);
-  hash_combine(seed, metadata->total_size);
-  hash_combine(seed, metadata->left_buddy);
-  hash_combine(seed, metadata->right_buddy);
+  hash_combine(&seed, static_cast<size_t>(metadata.type));
+  hash_combine(&seed, metadata.index);
+  hash_combine(&seed, metadata.size);
+  hash_combine(&seed, metadata.total_size);
+  hash_combine(&seed, metadata.left_buddy);
+  hash_combine(&seed, metadata.right_buddy);
 
   return seed;
 }
 
+}  // namespace
+
 void Metadata::update_guards() {
   guard_begin = hash(this, 1);
   guard_end = hash(this, 2);
diff --git a/paddle/fluid/memory/detail/meta_data.h b/paddle/fluid/memory/detail/meta_data.h
deleted file mode 100644
index 14895ee872..0000000000
--- a/paddle/fluid/memory/detail/meta_data.h
+++ /dev/null
@@ -1,54 +0,0 @@
-/* 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. */
-
-#pragma once
-
-#include "paddle/fluid/memory/detail/memory_block.h"
-
-#include <stddef.h>
-
-namespace paddle {
-namespace memory {
-namespace detail {
-
-class Metadata {
- public:
-  Metadata(MemoryBlock::Type t, size_t i, size_t s, size_t ts, MemoryBlock* l,
-           MemoryBlock* r);
-  Metadata();
-
- public:
-  /*! \brief Update the guards when metadata is changed */
-  void update_guards();
-
-  /*! \brief Check consistency to previous modification */
-  bool check_guards() const;
-
- public:
-  // TODO(gangliao): compress this
-  // clang-format off
-  size_t            guard_begin = 0;
-  MemoryBlock::Type type        = MemoryBlock::INVALID_CHUNK;
-  size_t            index       = 0;
-  size_t            size        = 0;
-  size_t            total_size  = 0;
-  MemoryBlock*      left_buddy  = nullptr;
-  MemoryBlock*      right_buddy = nullptr;
-  size_t            guard_end   = 0;
-  // clang-format on
-};
-
-}  // namespace detail
-}  // namespace memory
-}  // namespace paddle
diff --git a/paddle/fluid/memory/detail/system_allocator.cc b/paddle/fluid/memory/detail/system_allocator.cc
index a45f8c33ee..d539052916 100644
--- a/paddle/fluid/memory/detail/system_allocator.cc
+++ b/paddle/fluid/memory/detail/system_allocator.cc
@@ -13,16 +13,16 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "paddle/fluid/memory/detail/system_allocator.h"
-#include "paddle/fluid/platform/assert.h"
-#include "paddle/fluid/platform/cpu_info.h"
-#include "paddle/fluid/platform/enforce.h"
-#include "paddle/fluid/platform/gpu_info.h"
 
 #include <stdlib.h>    // for malloc and free
 #include <sys/mman.h>  // for mlock and munlock
 #include <algorithm>   // for std::max
 
 #include "gflags/gflags.h"
+#include "paddle/fluid/platform/assert.h"
+#include "paddle/fluid/platform/cpu_info.h"
+#include "paddle/fluid/platform/enforce.h"
+#include "paddle/fluid/platform/gpu_info.h"
 
 // If use_pinned_memory is true, CPUAllocator calls mlock, which
 // returns pinned and locked memory as staging areas for data exchange
@@ -35,13 +35,13 @@ namespace paddle {
 namespace memory {
 namespace detail {
 
-void* CPUAllocator::Alloc(size_t& index, size_t size) {
+void* CPUAllocator::Alloc(size_t* index, size_t size) {
   // According to http://www.cplusplus.com/reference/cstdlib/malloc/,
   // malloc might not return nullptr if size is zero, but the returned
   // pointer shall not be dereferenced -- so we make it nullptr.
   if (size <= 0) return nullptr;
 
-  index = 0;  // unlock memory
+  *index = 0;  // unlock memory
 
   void* p;
 
@@ -56,7 +56,7 @@ void* CPUAllocator::Alloc(size_t& index, size_t size) {
 
   if (p != nullptr) {
     if (FLAGS_use_pinned_memory) {
-      index = 1;
+      *index = 1;
       mlock(p, size);  // lock memory
     }
   }
@@ -75,7 +75,7 @@ bool CPUAllocator::UseGpu() const { return false; }
 
 #ifdef PADDLE_WITH_CUDA
 
-void* GPUAllocator::Alloc(size_t& index, size_t size) {
+void* GPUAllocator::Alloc(size_t* index, size_t size) {
   // CUDA documentation doesn't explain if cudaMalloc returns nullptr
   // if size is 0.  We just make sure it does.
   if (size <= 0) return nullptr;
@@ -93,7 +93,7 @@ void* GPUAllocator::Alloc(size_t& index, size_t size) {
   }
 
   if (result == cudaSuccess) {
-    index = 0;
+    *index = 0;
     gpu_alloc_size_ += size;
     return p;
   } else {
@@ -133,7 +133,7 @@ bool GPUAllocator::UseGpu() const { return true; }
 
 // PINNED memory allows direct DMA transfers by the GPU to and from system
 // memory. It’s locked to a physical address.
-void* CUDAPinnedAllocator::Alloc(size_t& index, size_t size) {
+void* CUDAPinnedAllocator::Alloc(size_t* index, size_t size) {
   if (size <= 0) return nullptr;
 
   // NOTE: here, we use CUDAPinnedMaxAllocSize as the maximum memory size
@@ -154,7 +154,7 @@ void* CUDAPinnedAllocator::Alloc(size_t& index, size_t size) {
   cudaError_t result = cudaMallocHost(&p, size);
 
   if (result == cudaSuccess) {
-    index = 1;  // PINNED memory
+    *index = 1;  // PINNED memory
     cuda_pinnd_alloc_size_ += size;
     return p;
   } else {
diff --git a/paddle/fluid/memory/detail/system_allocator.h b/paddle/fluid/memory/detail/system_allocator.h
index e3c50ef648..a0386a2dad 100644
--- a/paddle/fluid/memory/detail/system_allocator.h
+++ b/paddle/fluid/memory/detail/system_allocator.h
@@ -29,14 +29,14 @@ namespace detail {
 class SystemAllocator {
  public:
   virtual ~SystemAllocator() {}
-  virtual void* Alloc(size_t& index, size_t size) = 0;
+  virtual void* Alloc(size_t* index, size_t size) = 0;
   virtual void Free(void* p, size_t size, size_t index) = 0;
   virtual bool UseGpu() const = 0;
 };
 
 class CPUAllocator : public SystemAllocator {
  public:
-  virtual void* Alloc(size_t& index, size_t size);
+  virtual void* Alloc(size_t* index, size_t size);
   virtual void Free(void* p, size_t size, size_t index);
   virtual bool UseGpu() const;
 };
@@ -46,7 +46,7 @@ class GPUAllocator : public SystemAllocator {
  public:
   explicit GPUAllocator(int gpu_id) : gpu_id_(gpu_id) {}
 
-  virtual void* Alloc(size_t& index, size_t size);
+  virtual void* Alloc(size_t* index, size_t size);
   virtual void Free(void* p, size_t size, size_t index);
   virtual bool UseGpu() const;
 
@@ -58,7 +58,7 @@ class GPUAllocator : public SystemAllocator {
 
 class CUDAPinnedAllocator : public SystemAllocator {
  public:
-  virtual void* Alloc(size_t& index, size_t size);
+  virtual void* Alloc(size_t* index, size_t size);
   virtual void Free(void* p, size_t size, size_t index);
   virtual bool UseGpu() const;
 
diff --git a/paddle/fluid/memory/detail/system_allocator_test.cc b/paddle/fluid/memory/detail/system_allocator_test.cc
index 3e1926f632..95467a6f60 100644
--- a/paddle/fluid/memory/detail/system_allocator_test.cc
+++ b/paddle/fluid/memory/detail/system_allocator_test.cc
@@ -22,11 +22,11 @@ limitations under the License. */
 
 DECLARE_bool(use_pinned_memory);
 
-void TestAllocator(paddle::memory::detail::SystemAllocator& a, size_t size) {
+void TestAllocator(paddle::memory::detail::SystemAllocator* a, size_t size) {
   bool freed = false;
   {
     size_t index;
-    void* p = a.Alloc(index, size);
+    void* p = a->Alloc(index, size);
     if (size > 0) {
       EXPECT_NE(p, nullptr);
     } else {
@@ -36,7 +36,7 @@ void TestAllocator(paddle::memory::detail::SystemAllocator& a, size_t size) {
     int* i = static_cast<int*>(p);
     std::shared_ptr<int> ptr(i, [&](void* p) {
       freed = true;
-      a.Free(p, size, index);
+      a->Free(p, size, index);
     });
   }
   EXPECT_TRUE(freed);
@@ -45,21 +45,21 @@ void TestAllocator(paddle::memory::detail::SystemAllocator& a, size_t size) {
 TEST(CPUAllocator, NoLockMem) {
   FLAGS_use_pinned_memory = false;
   paddle::memory::detail::CPUAllocator a;
-  TestAllocator(a, 2048);
-  TestAllocator(a, 0);
+  TestAllocator(&a, 2048);
+  TestAllocator(&a, 0);
 }
 
 TEST(CPUAllocator, LockMem) {
   FLAGS_use_pinned_memory = true;
   paddle::memory::detail::CPUAllocator a;
-  TestAllocator(a, 2048);
-  TestAllocator(a, 0);
+  TestAllocator(&a, 2048);
+  TestAllocator(&a, 0);
 }
 
 #ifdef PADDLE_WITH_CUDA
 TEST(GPUAllocator, Alloc) {
   paddle::memory::detail::GPUAllocator a(0);
-  TestAllocator(a, 2048);
-  TestAllocator(a, 0);
+  TestAllocator(&a, 2048);
+  TestAllocator(&a, 0);
 }
 #endif

From 478055bd9f24ace509a092df7b197bd5d557006d Mon Sep 17 00:00:00 2001
From: Yi Wang <yi.wang.2005@gmail.com>
Date: Sat, 7 Apr 2018 17:41:10 -0700
Subject: [PATCH 02/16] Update CMakeLists.txt

---
 paddle/fluid/platform/CMakeLists.txt | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/paddle/fluid/platform/CMakeLists.txt b/paddle/fluid/platform/CMakeLists.txt
index 6780b8cc6d..49ff589b77 100644
--- a/paddle/fluid/platform/CMakeLists.txt
+++ b/paddle/fluid/platform/CMakeLists.txt
@@ -43,7 +43,7 @@ ENDIF()
 # memcpy depends on device_context, here add deps individually for
 # avoiding cycle dependencies
 cc_library(device_context SRCS device_context.cc DEPS memory buddy_allocator
-    system_allocator memory_block meta_data meta_cache place eigen3 ${GPU_CTX_DEPS} ${MKLDNN_CTX_DEPS})
+    place eigen3 ${GPU_CTX_DEPS} ${MKLDNN_CTX_DEPS})
 nv_test(device_context_test SRCS device_context_test.cu DEPS device_context gpu_info)
 
 nv_test(cudnn_helper_test SRCS cudnn_helper_test.cc DEPS dynload_cuda)

From 67ba884d2a9f564ac9cb154dbfa5234e244069d5 Mon Sep 17 00:00:00 2001
From: Yi Wang <yi.wang.2005@gmail.com>
Date: Sat, 7 Apr 2018 17:48:07 -0700
Subject: [PATCH 03/16] Update CMakeLists

---
 paddle/fluid/memory/CMakeLists.txt                    | 11 +++++------
 paddle/fluid/memory/{memory.cc => malloc.cc}          |  0
 paddle/fluid/memory/{memory.h => malloc.h}            |  0
 .../fluid/memory/{memory_test.cc => malloc_test.cc}   |  1 -
 paddle/fluid/memory/pinned_memory_test.cu             |  1 -
 paddle/fluid/platform/CMakeLists.txt                  |  4 ++--
 6 files changed, 7 insertions(+), 10 deletions(-)
 rename paddle/fluid/memory/{memory.cc => malloc.cc} (100%)
 rename paddle/fluid/memory/{memory.h => malloc.h} (100%)
 rename paddle/fluid/memory/{memory_test.cc => malloc_test.cc} (99%)

diff --git a/paddle/fluid/memory/CMakeLists.txt b/paddle/fluid/memory/CMakeLists.txt
index 3b73e79eb3..c928a8e493 100644
--- a/paddle/fluid/memory/CMakeLists.txt
+++ b/paddle/fluid/memory/CMakeLists.txt
@@ -1,15 +1,14 @@
 add_subdirectory(detail)
 
-cc_library(memory SRCS memory.cc DEPS place enforce)
+cc_library(malloc SRCS malloc.cc DEPS buddy_allocator place enforce)
 cc_library(memcpy SRCS memcpy.cc DEPS place)
 
-cc_library(paddle_memory
+cc_library(memory
         DEPS
-        memory
-        memcpy
-        buddy_allocator)
+        malloc
+        memcpy)
 
-cc_test(memory_test SRCS memory_test.cc DEPS place paddle_memory)
+cc_test(malloc_test SRCS malloc_test.cc DEPS malloc)
 
 #if (WITH_GPU)
 #   nv_test(pinned_memory_test SRCS pinned_memory_test.cu  DEPS place paddle_memory)
diff --git a/paddle/fluid/memory/memory.cc b/paddle/fluid/memory/malloc.cc
similarity index 100%
rename from paddle/fluid/memory/memory.cc
rename to paddle/fluid/memory/malloc.cc
diff --git a/paddle/fluid/memory/memory.h b/paddle/fluid/memory/malloc.h
similarity index 100%
rename from paddle/fluid/memory/memory.h
rename to paddle/fluid/memory/malloc.h
diff --git a/paddle/fluid/memory/memory_test.cc b/paddle/fluid/memory/malloc_test.cc
similarity index 99%
rename from paddle/fluid/memory/memory_test.cc
rename to paddle/fluid/memory/malloc_test.cc
index 9fbbe62559..6d68eed1a8 100644
--- a/paddle/fluid/memory/memory_test.cc
+++ b/paddle/fluid/memory/malloc_test.cc
@@ -18,7 +18,6 @@ limitations under the License. */
 
 #include "gtest/gtest.h"
 #include "paddle/fluid/memory/detail/memory_block.h"
-#include "paddle/fluid/memory/detail/meta_data.h"
 #include "paddle/fluid/platform/cpu_info.h"
 #include "paddle/fluid/platform/gpu_info.h"
 #include "paddle/fluid/platform/place.h"
diff --git a/paddle/fluid/memory/pinned_memory_test.cu b/paddle/fluid/memory/pinned_memory_test.cu
index a000001f41..0d898f59ee 100644
--- a/paddle/fluid/memory/pinned_memory_test.cu
+++ b/paddle/fluid/memory/pinned_memory_test.cu
@@ -15,7 +15,6 @@ limitations under the License. */
 #include <unordered_map>
 
 #include "paddle/fluid/memory/detail/memory_block.h"
-#include "paddle/fluid/memory/detail/meta_data.h"
 #include "paddle/fluid/memory/memcpy.h"
 #include "paddle/fluid/memory/memory.h"
 
diff --git a/paddle/fluid/platform/CMakeLists.txt b/paddle/fluid/platform/CMakeLists.txt
index 49ff589b77..917bdc64ab 100644
--- a/paddle/fluid/platform/CMakeLists.txt
+++ b/paddle/fluid/platform/CMakeLists.txt
@@ -42,12 +42,12 @@ ENDIF()
 
 # memcpy depends on device_context, here add deps individually for
 # avoiding cycle dependencies
-cc_library(device_context SRCS device_context.cc DEPS memory buddy_allocator
+cc_library(device_context SRCS device_context.cc DEPS malloc
     place eigen3 ${GPU_CTX_DEPS} ${MKLDNN_CTX_DEPS})
 nv_test(device_context_test SRCS device_context_test.cu DEPS device_context gpu_info)
 
 nv_test(cudnn_helper_test SRCS cudnn_helper_test.cc DEPS dynload_cuda)
-nv_test(transform_test SRCS transform_test.cu DEPS paddle_memory place device_context)
+nv_test(transform_test SRCS transform_test.cu DEPS memory place device_context)
 
 cc_library(device_tracer SRCS device_tracer.cc DEPS boost profiler_proto ${GPU_CTX_DEPS})
 cc_library(profiler SRCS profiler.cc DEPS device_context device_tracer)

From eebb20532478b1d5f983943c7d08fa13e48d25d7 Mon Sep 17 00:00:00 2001
From: Yi Wang <yi.wang.2005@gmail.com>
Date: Sat, 7 Apr 2018 17:50:10 -0700
Subject: [PATCH 04/16] Update CMakeLists

---
 paddle/fluid/framework/CMakeLists.txt | 8 ++++----
 paddle/fluid/inference/CMakeLists.txt | 2 +-
 paddle/fluid/operators/CMakeLists.txt | 2 +-
 paddle/fluid/pybind/CMakeLists.txt    | 4 ++--
 4 files changed, 8 insertions(+), 8 deletions(-)

diff --git a/paddle/fluid/framework/CMakeLists.txt b/paddle/fluid/framework/CMakeLists.txt
index a473ed7400..3840bbe83b 100644
--- a/paddle/fluid/framework/CMakeLists.txt
+++ b/paddle/fluid/framework/CMakeLists.txt
@@ -7,9 +7,9 @@ cc_test(ddim_test SRCS ddim_test.cc DEPS ddim)
 nv_test(dim_test SRCS dim_test.cu DEPS ddim)
 
 if(WITH_GPU)
-  nv_library(tensor SRCS tensor.cc tensor_util.cu DEPS ddim place paddle_memory device_context framework_proto)
+  nv_library(tensor SRCS tensor.cc tensor_util.cu DEPS ddim place memory device_context framework_proto)
 else()
-  cc_library(tensor SRCS tensor.cc tensor_util.cc DEPS ddim place paddle_memory device_context framework_proto)
+  cc_library(tensor SRCS tensor.cc tensor_util.cc DEPS ddim place memory device_context framework_proto)
 endif()
 
 cc_test(tensor_test SRCS tensor_test.cc DEPS tensor)
@@ -21,9 +21,9 @@ endif()
 
 cc_test(eigen_test SRCS eigen_test.cc DEPS tensor)
 
-nv_test(mixed_vector_test SRCS mixed_vector_test.cu DEPS place paddle_memory device_context init)
+nv_test(mixed_vector_test SRCS mixed_vector_test.cu DEPS place memory device_context init)
 cc_library(lod_tensor SRCS lod_tensor.cc DEPS ddim place tensor framework_proto recordio)
-cc_test(lod_tensor_test SRCS lod_tensor_test.cc DEPS lod_tensor paddle_memory)
+cc_test(lod_tensor_test SRCS lod_tensor_test.cc DEPS lod_tensor memory)
 nv_test(lod_tensor_gpu_test SRCS lod_tensor_test.cu DEPS lod_tensor init)
 
 cc_library(reader SRCS reader.cc DEPS lod_tensor ddim)
diff --git a/paddle/fluid/inference/CMakeLists.txt b/paddle/fluid/inference/CMakeLists.txt
index aff427310f..f417f62f3f 100644
--- a/paddle/fluid/inference/CMakeLists.txt
+++ b/paddle/fluid/inference/CMakeLists.txt
@@ -1,4 +1,4 @@
-set(FLUID_CORE_MODULES proto_desc paddle_memory lod_tensor executor prune init)
+set(FLUID_CORE_MODULES proto_desc memory lod_tensor executor prune init)
 
 cc_library(paddle_fluid_api
     SRCS io.cc
diff --git a/paddle/fluid/operators/CMakeLists.txt b/paddle/fluid/operators/CMakeLists.txt
index 84eabab563..5ff987ad8b 100644
--- a/paddle/fluid/operators/CMakeLists.txt
+++ b/paddle/fluid/operators/CMakeLists.txt
@@ -263,7 +263,7 @@ 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(beam_search_decode_op_test SRCS beam_search_decode_op_test.cc DEPS lod_tensor)
 cc_test(beam_search_op_test SRCS beam_search_op_test.cc DEPS lod_tensor beam_search_op)
-cc_test(strided_memcpy_test SRCS strided_memcpy_test.cc DEPS tensor paddle_memory)
+cc_test(strided_memcpy_test SRCS strided_memcpy_test.cc DEPS tensor memory)
 cc_test(save_load_op_test SRCS save_load_op_test.cc DEPS save_op load_op)
 cc_test(save_load_combine_op_test SRCS save_load_combine_op_test.cc DEPS save_combine_op load_combine_op)
 nv_test(nccl_op_test SRCS nccl_op_test.cu.cc DEPS nccl_op gpu_info device_context)
diff --git a/paddle/fluid/pybind/CMakeLists.txt b/paddle/fluid/pybind/CMakeLists.txt
index 787925d9f8..884289a7fd 100644
--- a/paddle/fluid/pybind/CMakeLists.txt
+++ b/paddle/fluid/pybind/CMakeLists.txt
@@ -2,13 +2,13 @@ if(WITH_PYTHON)
   if(WITH_AMD_GPU)
     hip_library(paddle_pybind SHARED
       SRCS pybind.cc exception.cc protobuf.cc const_value.cc recordio.cc
-      DEPS pybind python backward proto_desc paddle_memory executor prune init profiler feed_fetch_method
+      DEPS pybind python backward proto_desc memory executor prune init profiler feed_fetch_method
            parallel_executor
       ${GLOB_OP_LIB})
   else()
     cc_library(paddle_pybind SHARED
       SRCS pybind.cc exception.cc protobuf.cc const_value.cc recordio.cc
-      DEPS pybind python backward proto_desc paddle_memory executor prune init profiler feed_fetch_method
+      DEPS pybind python backward proto_desc memory executor prune init profiler feed_fetch_method
            parallel_executor
       ${GLOB_OP_LIB})
     if(NOT APPLE AND NOT ANDROID)

From a2730d1e0b4b3177bca69c89a6e6d7e367bfb4f5 Mon Sep 17 00:00:00 2001
From: Yi Wang <yi.wang.2005@gmail.com>
Date: Sat, 7 Apr 2018 18:02:51 -0700
Subject: [PATCH 05/16] Rename Metadata into MemoryBlock::Desc

---
 paddle/fluid/memory/detail/CMakeLists.txt     |  2 +-
 paddle/fluid/memory/detail/memory_block.cc    | 28 ++++++-----
 paddle/fluid/memory/detail/memory_block.h     | 49 +++++++++----------
 .../{meta_data.cc => memory_block_desc.cc}    | 12 ++---
 4 files changed, 47 insertions(+), 44 deletions(-)
 rename paddle/fluid/memory/detail/{meta_data.cc => memory_block_desc.cc} (84%)

diff --git a/paddle/fluid/memory/detail/CMakeLists.txt b/paddle/fluid/memory/detail/CMakeLists.txt
index 1636b120ae..c725dba5e9 100644
--- a/paddle/fluid/memory/detail/CMakeLists.txt
+++ b/paddle/fluid/memory/detail/CMakeLists.txt
@@ -1,4 +1,4 @@
-cc_library(memory_block SRCS memory_block.cc meta_data.cc meta_cache.cc)
+cc_library(memory_block SRCS memory_block.cc memory_block_desc.cc meta_cache.cc)
 
 if(${WITH_GPU})
   nv_library(system_allocator SRCS system_allocator.cc DEPS gflags cpu_info gpu_info)
diff --git a/paddle/fluid/memory/detail/memory_block.cc b/paddle/fluid/memory/detail/memory_block.cc
index f744450f48..79f65f1eb8 100644
--- a/paddle/fluid/memory/detail/memory_block.cc
+++ b/paddle/fluid/memory/detail/memory_block.cc
@@ -21,9 +21,10 @@ namespace detail {
 
 void MemoryBlock::init(MetadataCache* cache, Type t, size_t index, size_t size,
                        void* left_buddy, void* right_buddy) {
-  cache->save(this, Metadata(t, index, size - sizeof(Metadata), size,
-                             static_cast<MemoryBlock*>(left_buddy),
-                             static_cast<MemoryBlock*>(right_buddy)));
+  cache->save(
+      this, MemoryBlock::Desc(t, index, size - sizeof(MemoryBlock::Desc), size,
+                              static_cast<MemoryBlock*>(left_buddy),
+                              static_cast<MemoryBlock*>(right_buddy)));
 }
 
 MemoryBlock::Type MemoryBlock::type(MetadataCache& cache) const {
@@ -63,7 +64,7 @@ void MemoryBlock::split(MetadataCache* cache, size_t size) {
   PADDLE_ASSERT(total_size(*cache) >= size);
 
   // bail out if there is no room for another partition
-  if (total_size(*cache) - size <= sizeof(Metadata)) {
+  if (total_size(*cache) - size <= sizeof(MemoryBlock::Desc)) {
     return;
   }
 
@@ -78,13 +79,13 @@ void MemoryBlock::split(MetadataCache* cache, size_t size) {
   // Write the metadata for the new block
   auto new_block_right_buddy = metadata.right_buddy;
 
-  cache->save(
-      static_cast<MemoryBlock*>(right_partition),
-      Metadata(FREE_CHUNK, index(*cache), remaining_size - sizeof(Metadata),
-               remaining_size, this, new_block_right_buddy));
+  cache->save(static_cast<MemoryBlock*>(right_partition),
+              MemoryBlock::Desc(FREE_CHUNK, index(*cache),
+                                remaining_size - sizeof(MemoryBlock::Desc),
+                                remaining_size, this, new_block_right_buddy));
 
   metadata.right_buddy = static_cast<MemoryBlock*>(right_partition);
-  metadata.size = size - sizeof(Metadata);
+  metadata.size = size - sizeof(MemoryBlock::Desc);
   metadata.total_size = size;
 
   cache->save(this, metadata);
@@ -122,7 +123,8 @@ void MemoryBlock::merge(MetadataCache* cache, MemoryBlock* right_buddy) {
   metadata.total_size += right_buddy->total_size(*cache);
 
   cache->save(this, metadata);
-  cache->save(right_buddy, Metadata(INVALID_CHUNK, 0, 0, 0, nullptr, nullptr));
+  cache->save(right_buddy,
+              MemoryBlock::Desc(INVALID_CHUNK, 0, 0, 0, nullptr, nullptr));
 }
 
 void MemoryBlock::mark_as_free(MetadataCache* cache) {
@@ -139,12 +141,14 @@ void MemoryBlock::set_type(MetadataCache* cache, Type t) {
 }
 
 void* MemoryBlock::data() const {
-  return const_cast<Metadata*>(reinterpret_cast<const Metadata*>(this)) + 1;
+  return const_cast<MemoryBlock::Desc*>(
+             reinterpret_cast<const MemoryBlock::Desc*>(this)) +
+         1;
 }
 
 MemoryBlock* MemoryBlock::metadata() const {
   return const_cast<MemoryBlock*>(reinterpret_cast<const MemoryBlock*>(
-      reinterpret_cast<const Metadata*>(this) - 1));
+      reinterpret_cast<const MemoryBlock::Desc*>(this) - 1));
 }
 
 }  // namespace detail
diff --git a/paddle/fluid/memory/detail/memory_block.h b/paddle/fluid/memory/detail/memory_block.h
index 5e83a2f899..f0a06577a5 100644
--- a/paddle/fluid/memory/detail/memory_block.h
+++ b/paddle/fluid/memory/detail/memory_block.h
@@ -23,9 +23,8 @@ namespace detail {
 class MetadataCache;
 
 // MemoryBlock represents Each allocated memory block, which contains
-// Metadata and the payload.
-class MemoryBlock {
- public:
+// MemoryBlock::Desc and the payload.
+struct MemoryBlock {
   enum Type {
     FREE_CHUNK,    // memory is free and idle
     ARENA_CHUNK,   // memory is being occupied
@@ -33,17 +32,17 @@ class MemoryBlock {
     INVALID_CHUNK  // memory is invalid
   };
 
-  // init saves the Metadata of the memory block in a MetadataCache.
+  // init saves the MemoryBlock::Desc of the memory block in a MetadataCache.
   // If it is a CPU memory block, the MetadataCache writes the
-  // Metadata to the beginning of the block; or, if it is a GPU memory
+  // MemoryBlock::Desc to the beginning of the block; or, if it is a GPU memory
   // block, the MetadataCache writes the Meatadata to a std::map in
   // the CPU.
-  void init(MetadataCache* cache, Type t, size_t index, size_t size,
+  void init(MemoryBlock::DescCache* cache, Type t, size_t index, size_t size,
             void* left_buddy, void* right_buddy);
 
-  // All these accessors returns fields in the Metadata of the memory
+  // All these accessors returns fields in the MemoryBlock::Desc of the memory
   // block.  They all need a MetadataCache instance as their first
-  // parameter because they read the Metadata from the cache.
+  // parameter because they read the MemoryBlock::Desc from the cache.
   Type type(const MetadataCache& cache) const;
   size_t size(const MetadataCache& cache) const;
   size_t index(const MetadataCache& cache) const;
@@ -68,12 +67,11 @@ class MemoryBlock {
   void* data() const;
   MemoryBlock* metadata() const;
 
- private:
-  // Metadata describes a MemoryBlock.
-  struct Metadata {
-    Metadata(MemoryBlock::Type t, size_t i, size_t s, size_t ts, MemoryBlock* l,
-             MemoryBlock* r);
-    Metadata();
+  // MemoryBlock::Desc describes a MemoryBlock.
+  struct Desc {
+    Desc(MemoryBlock::Type t, size_t i, size_t s, size_t ts, MemoryBlock* l,
+         MemoryBlock* r);
+    Desc();
 
     // Updates guard_begin and guard_end by hashes of the Metadata object.
     void update_guards();
@@ -94,9 +92,10 @@ class MemoryBlock {
 };
 
 // A cache for accessing memory block meta-data that may be expensive
-// to access directly.  This class exists to unify the metadata format
-// between GPU and CPU allocations. It should be removed when the CPU
-// can access all GPU allocations directly via UVM.
+// to access directly.  This class exists to unify the
+// MemoryBlock::Desc format between GPU and CPU allocations. It should
+// be removed when the CPU can access all GPU allocations directly via
+// UVM.
 class MetadataCache {
  public:
   explicit MetadataCache(bool uses_gpu);
@@ -105,22 +104,22 @@ class MetadataCache {
   MetadataCache(const MetadataCache&) = delete;
   MetadataCache& operator=(const MetadataCache&) = delete;
 
-  // Returns the Metadata for a memory block.  When MetadataCache is
-  // used to manage CPU memory, the Metadata resides at the beginning
+  // Returns the MemoryBlock::Desc for a memory block.  When MetadataCache is
+  // used to manage CPU memory, the MemoryBlock::Desc resides at the beginning
   // of the memory block; when used to manage GPU memory, the
   // Meatadata resides in CPU memory indexed by cache_.
-  Metadata load(const MemoryBlock* memory_block) const;
+  MemoryBlock::Desc load(const MemoryBlock* memory_block) const;
 
-  // Saves the Metadata of a memory block into the cache.  For CPU
-  // memory block, writes the Metadata to the beginning of the memory
+  // Saves the MemoryBlock::Desc of a memory block into the cache.  For CPU
+  // memory block, writes the MemoryBlock::Desc to the beginning of the memory
   // block; whereas for GPU memory, writes it to cache_.
-  void save(MemoryBlock* memory_block, const Metadata& meta_data);
+  void save(MemoryBlock* memory_block, const MemoryBlock::Desc& meta_data);
 
-  // For GPU memory block, erases its Metadata from cache_.
+  // For GPU memory block, erases its MemoryBlock::Desc from cache_.
   void invalidate(MemoryBlock* memory_block);
 
  private:
-  typedef std::unordered_map<const MemoryBlock*, Metadata> MetadataMap;
+  typedef std::unordered_map<const MemoryBlock*, MemoryBlock::Desc> MetadataMap;
   MetadataMap cache_;
   bool uses_gpu_;
 };
diff --git a/paddle/fluid/memory/detail/meta_data.cc b/paddle/fluid/memory/detail/memory_block_desc.cc
similarity index 84%
rename from paddle/fluid/memory/detail/meta_data.cc
rename to paddle/fluid/memory/detail/memory_block_desc.cc
index 02f5235d14..0ad2ef4d7c 100644
--- a/paddle/fluid/memory/detail/meta_data.cc
+++ b/paddle/fluid/memory/detail/memory_block_desc.cc
@@ -20,8 +20,8 @@ namespace paddle {
 namespace memory {
 namespace detail {
 
-Metadata::Metadata(MemoryBlock::Type t, size_t i, size_t s, size_t ts,
-                   MemoryBlock* l, MemoryBlock* r)
+MemoryBlock::Desc::Desc(MemoryBlock::Type t, size_t i, size_t s, size_t ts,
+                        MemoryBlock* l, MemoryBlock* r)
     : type(t),
       index(i),
       size(s),
@@ -29,7 +29,7 @@ Metadata::Metadata(MemoryBlock::Type t, size_t i, size_t s, size_t ts,
       left_buddy(l),
       right_buddy(r) {}
 
-Metadata::Metadata()
+MemoryBlock::Desc::Desc()
     : type(MemoryBlock::INVALID_CHUNK),
       index(0),
       size(0),
@@ -45,7 +45,7 @@ inline void hash_combine(std::size_t* seed, const T& v) {
   (*seed) ^= hasher(v) + 0x9e3779b9 + ((*seed) << 6) + ((*seed) >> 2);
 }
 
-inline size_t hash(const Metadata& metadata, size_t initial_seed) {
+inline size_t hash(const MemoryBlock::Desc& metadata, size_t initial_seed) {
   size_t seed = initial_seed;
 
   hash_combine(&seed, static_cast<size_t>(metadata.type));
@@ -60,12 +60,12 @@ inline size_t hash(const Metadata& metadata, size_t initial_seed) {
 
 }  // namespace
 
-void Metadata::update_guards() {
+void MemoryBlock::Desc::update_guards() {
   guard_begin = hash(this, 1);
   guard_end = hash(this, 2);
 }
 
-bool Metadata::check_guards() const {
+bool MemoryBlock::Desc::check_guards() const {
   return guard_begin == hash(this, 1) && guard_end == hash(this, 2);
 }
 

From 903403b2e1cb4c0259d1324712d228485da3c7d8 Mon Sep 17 00:00:00 2001
From: Yi Wang <yi.wang.2005@gmail.com>
Date: Sat, 7 Apr 2018 18:25:46 -0700
Subject: [PATCH 06/16] Fix errors

---
 paddle/fluid/memory/detail/memory_block.cc | 2 +-
 paddle/fluid/memory/detail/memory_block.h  | 5 +++--
 2 files changed, 4 insertions(+), 3 deletions(-)

diff --git a/paddle/fluid/memory/detail/memory_block.cc b/paddle/fluid/memory/detail/memory_block.cc
index 79f65f1eb8..f34b922b25 100644
--- a/paddle/fluid/memory/detail/memory_block.cc
+++ b/paddle/fluid/memory/detail/memory_block.cc
@@ -27,7 +27,7 @@ void MemoryBlock::init(MetadataCache* cache, Type t, size_t index, size_t size,
                               static_cast<MemoryBlock*>(right_buddy)));
 }
 
-MemoryBlock::Type MemoryBlock::type(MetadataCache& cache) const {
+MemoryBlock::Type MemoryBlock::type(const MetadataCache& cache) const {
   return cache.load(this).type;
 }
 
diff --git a/paddle/fluid/memory/detail/memory_block.h b/paddle/fluid/memory/detail/memory_block.h
index f0a06577a5..5cceba659b 100644
--- a/paddle/fluid/memory/detail/memory_block.h
+++ b/paddle/fluid/memory/detail/memory_block.h
@@ -13,7 +13,8 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 #pragma once
 
-#include <cstddef>
+#include <cstdint>
+#include <unordered_map>
 
 namespace paddle {
 namespace memory {
@@ -37,7 +38,7 @@ struct MemoryBlock {
   // MemoryBlock::Desc to the beginning of the block; or, if it is a GPU memory
   // block, the MetadataCache writes the Meatadata to a std::map in
   // the CPU.
-  void init(MemoryBlock::DescCache* cache, Type t, size_t index, size_t size,
+  void init(MetadataCache* cache, Type t, size_t index, size_t size,
             void* left_buddy, void* right_buddy);
 
   // All these accessors returns fields in the MemoryBlock::Desc of the memory

From ccd83ef6b9347c72d26c818477785dab067dc370 Mon Sep 17 00:00:00 2001
From: Yi Wang <yi.wang.2005@gmail.com>
Date: Sat, 7 Apr 2018 19:26:30 -0700
Subject: [PATCH 07/16] Update

---
 .../fluid/memory/detail/memory_block_desc.cc  |  6 ++---
 paddle/fluid/memory/detail/meta_cache.cc      | 27 +++++++++----------
 2 files changed, 16 insertions(+), 17 deletions(-)

diff --git a/paddle/fluid/memory/detail/memory_block_desc.cc b/paddle/fluid/memory/detail/memory_block_desc.cc
index 0ad2ef4d7c..393dd9209c 100644
--- a/paddle/fluid/memory/detail/memory_block_desc.cc
+++ b/paddle/fluid/memory/detail/memory_block_desc.cc
@@ -61,12 +61,12 @@ inline size_t hash(const MemoryBlock::Desc& metadata, size_t initial_seed) {
 }  // namespace
 
 void MemoryBlock::Desc::update_guards() {
-  guard_begin = hash(this, 1);
-  guard_end = hash(this, 2);
+  guard_begin = hash(*this, 1);
+  guard_end = hash(*this, 2);
 }
 
 bool MemoryBlock::Desc::check_guards() const {
-  return guard_begin == hash(this, 1) && guard_end == hash(this, 2);
+  return guard_begin == hash(*this, 1) && guard_end == hash(*this, 2);
 }
 
 }  // namespace detail
diff --git a/paddle/fluid/memory/detail/meta_cache.cc b/paddle/fluid/memory/detail/meta_cache.cc
index 6bcb3a659e..b86e4f38c4 100644
--- a/paddle/fluid/memory/detail/meta_cache.cc
+++ b/paddle/fluid/memory/detail/meta_cache.cc
@@ -22,29 +22,28 @@ namespace detail {
 
 MetadataCache::MetadataCache(bool uses_gpu) : uses_gpu_(uses_gpu) {}
 
-Metadata MetadataCache::load(const MemoryBlock* block) const {
+MemoryBlock::Desc MetadataCache::load(const MemoryBlock* block) const {
   if (uses_gpu_) {
-    auto existing_metadata = cache_.find(block);
-    PADDLE_ASSERT(existing_metadata->second.check_guards());
-    return existing_metadata->second;
+    auto existing_desc = cache_.find(block);
+    PADDLE_ASSERT(existing_desc->second.check_guards());
+    return existing_desc->second;
   } else {
-    auto* meta = reinterpret_cast<const Metadata*>(block);
-    VLOG(10) << "Load MetaData type=" << meta->type;
-    PADDLE_ASSERT(meta->check_guards());
-    return *reinterpret_cast<const Metadata*>(block);
+    auto* desc = reinterpret_cast<const MemoryBlock::Desc*>(block);
+    VLOG(10) << "Load MemoryBlock::Desc type=" << desc->type;
+    PADDLE_ASSERT(desc->check_guards());
+    return *reinterpret_cast<const MemoryBlock::Desc*>(block);
   }
 }
 
 void MetadataCache::save(MemoryBlock* block,
-                         const Metadata& original_metadata) {
-  auto metadata = original_metadata;
-
-  metadata.update_guards();
+                         const MemoryBlock::Desc& original_desc) {
+  auto desc = original_desc;
+  desc.update_guards();
 
   if (uses_gpu_) {
-    cache_[block] = metadata;
+    cache_[block] = desc;
   } else {
-    *reinterpret_cast<Metadata*>(block) = metadata;
+    *reinterpret_cast<MemoryBlock::Desc*>(block) = desc;
   }
 }
 

From e98e2d36bd230c4e4954d955aaf8520d38a359b2 Mon Sep 17 00:00:00 2001
From: Yi Wang <yi.wang.2005@gmail.com>
Date: Sat, 7 Apr 2018 21:52:11 -0700
Subject: [PATCH 08/16] Update

---
 paddle/fluid/memory/detail/buddy_allocator.cc | 19 ++++++++++---------
 1 file changed, 10 insertions(+), 9 deletions(-)

diff --git a/paddle/fluid/memory/detail/buddy_allocator.cc b/paddle/fluid/memory/detail/buddy_allocator.cc
index 8768378386..d1f0d1755d 100644
--- a/paddle/fluid/memory/detail/buddy_allocator.cc
+++ b/paddle/fluid/memory/detail/buddy_allocator.cc
@@ -46,7 +46,8 @@ inline size_t align(size_t size, size_t alignment) {
 
 void* BuddyAllocator::Alloc(size_t unaligned_size) {
   // adjust allocation alignment
-  size_t size = align(unaligned_size + sizeof(Metadata), min_chunk_size_);
+  size_t size =
+      align(unaligned_size + sizeof(MemoryBlock::Desc), min_chunk_size_);
 
   // acquire the allocator lock
   std::lock_guard<std::mutex> lock(mutex_);
@@ -103,7 +104,7 @@ void BuddyAllocator::Free(void* p) {
     return;
   }
 
-  block->mark_as_free(cache_);
+  block->mark_as_free(&cache_);
 
   total_used_ -= block->total_size(cache_);
   total_free_ += block->total_size(cache_);
@@ -122,7 +123,7 @@ void BuddyAllocator::Free(void* p) {
                                    right_buddy));
 
       // merge its right buddy to the block
-      block->merge(cache_, right_buddy);
+      block->merge(&cache_, right_buddy);
     }
   }
 
@@ -139,7 +140,7 @@ void BuddyAllocator::Free(void* p) {
                                    left_buddy->total_size(cache_), left_buddy));
 
       // merge the block to its left buddy
-      left_buddy->merge(cache_, block);
+      left_buddy->merge(&cache_, block);
       block = left_buddy;
     }
   }
@@ -163,13 +164,13 @@ size_t BuddyAllocator::Used() { return total_used_; }
 
 void* BuddyAllocator::SystemAlloc(size_t size) {
   size_t index = 0;
-  void* p = system_allocator_->Alloc(index, size);
+  void* p = system_allocator_->Alloc(&index, size);
 
   VLOG(10) << "Allocated " << p << " from system allocator.";
 
   if (p == nullptr) return nullptr;
 
-  static_cast<MemoryBlock*>(p)->init(cache_, MemoryBlock::HUGE_CHUNK, index,
+  static_cast<MemoryBlock*>(p)->init(&cache_, MemoryBlock::HUGE_CHUNK, index,
                                      size, nullptr, nullptr);
 
   return static_cast<MemoryBlock*>(p)->data();
@@ -187,7 +188,7 @@ BuddyAllocator::PoolSet::iterator BuddyAllocator::RefillPool() {
 
   // Allocate a new maximum sized block
   size_t index = 0;
-  void* p = system_allocator_->Alloc(index, max_chunk_size_);
+  void* p = system_allocator_->Alloc(&index, max_chunk_size_);
 
   if (p == nullptr) return pool_.end();
 
@@ -238,11 +239,11 @@ void* BuddyAllocator::SplitToAlloc(BuddyAllocator::PoolSet::iterator it,
 
   VLOG(10) << "Split block (" << block << ", " << block->total_size(cache_)
            << ") into";
-  block->split(cache_, size);
+  block->split(&cache_, size);
 
   VLOG(10) << "Left block (" << block << ", " << block->total_size(cache_)
            << ")";
-  block->set_type(cache_, MemoryBlock::ARENA_CHUNK);
+  block->set_type(&cache_, MemoryBlock::ARENA_CHUNK);
 
   // the rest of memory if exist
   if (block->has_right_buddy(cache_)) {

From 45bc4538c4250d67d0f5368b15e2fd8f751bddc9 Mon Sep 17 00:00:00 2001
From: Yi Wang <yi.wang.2005@gmail.com>
Date: Sat, 7 Apr 2018 22:27:24 -0700
Subject: [PATCH 09/16] Update paddle_memory in CMakeLists.txt files

---
 cmake/generic.cmake                | 12 ++++++------
 paddle/fluid/memory/CMakeLists.txt |  2 +-
 paddle/testing/CMakeLists.txt      |  2 +-
 3 files changed, 8 insertions(+), 8 deletions(-)

diff --git a/cmake/generic.cmake b/cmake/generic.cmake
index e8bc285bdc..c4c9f77df8 100644
--- a/cmake/generic.cmake
+++ b/cmake/generic.cmake
@@ -244,11 +244,11 @@ function(cc_test TARGET_NAME)
     cmake_parse_arguments(cc_test "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
     add_executable(${TARGET_NAME} ${cc_test_SRCS})
     # Support linking flags: --whole-archive (Linux) / -force_load (MacOS)
-    target_circle_link_libraries(${TARGET_NAME} ${cc_test_DEPS} paddle_gtest_main paddle_memory gtest gflags glog)
+    target_circle_link_libraries(${TARGET_NAME} ${cc_test_DEPS} paddle_gtest_main memory gtest gflags glog)
     if("${cc_test_DEPS}" MATCHES "ARCHIVE_START")
       list(REMOVE_ITEM cc_test_DEPS ARCHIVE_START ARCHIVE_END)
     endif()
-    add_dependencies(${TARGET_NAME} ${cc_test_DEPS} paddle_gtest_main paddle_memory gtest gflags glog)
+    add_dependencies(${TARGET_NAME} ${cc_test_DEPS} paddle_gtest_main memory gtest gflags glog)
     add_test(NAME ${TARGET_NAME}
              COMMAND ${TARGET_NAME} ${cc_test_ARGS}
              WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR})
@@ -311,8 +311,8 @@ function(nv_test TARGET_NAME)
     set(multiValueArgs SRCS DEPS)
     cmake_parse_arguments(nv_test "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
     cuda_add_executable(${TARGET_NAME} ${nv_test_SRCS})
-    target_link_libraries(${TARGET_NAME} ${nv_test_DEPS} paddle_gtest_main paddle_memory gtest gflags glog)
-    add_dependencies(${TARGET_NAME} ${nv_test_DEPS} paddle_gtest_main paddle_memory gtest gflags glog)
+    target_link_libraries(${TARGET_NAME} ${nv_test_DEPS} paddle_gtest_main memory gtest gflags glog)
+    add_dependencies(${TARGET_NAME} ${nv_test_DEPS} paddle_gtest_main memory gtest gflags glog)
     add_test(${TARGET_NAME} ${TARGET_NAME})
   endif()
 endfunction(nv_test)
@@ -387,8 +387,8 @@ function(hip_test TARGET_NAME)
     endif()
     add_executable(${TARGET_NAME} ${_cmake_options} ${_generated_files} ${_sources})
     set_target_properties(${TARGET_NAME} PROPERTIES LINKER_LANGUAGE HIP)
-    target_link_libraries(${TARGET_NAME} ${hip_test_DEPS} paddle_gtest_main paddle_memory gtest gflags)
-    add_dependencies(${TARGET_NAME} ${hip_test_DEPS} paddle_gtest_main paddle_memory gtest gflags)
+    target_link_libraries(${TARGET_NAME} ${hip_test_DEPS} paddle_gtest_main memory gtest gflags)
+    add_dependencies(${TARGET_NAME} ${hip_test_DEPS} paddle_gtest_main memory gtest gflags)
     add_test(${TARGET_NAME} ${TARGET_NAME})
   endif()
 endfunction(hip_test)
diff --git a/paddle/fluid/memory/CMakeLists.txt b/paddle/fluid/memory/CMakeLists.txt
index c928a8e493..709fc7e12e 100644
--- a/paddle/fluid/memory/CMakeLists.txt
+++ b/paddle/fluid/memory/CMakeLists.txt
@@ -11,5 +11,5 @@ cc_library(memory
 cc_test(malloc_test SRCS malloc_test.cc DEPS malloc)
 
 #if (WITH_GPU)
-#   nv_test(pinned_memory_test SRCS pinned_memory_test.cu  DEPS place paddle_memory)
+#   nv_test(pinned_memory_test SRCS pinned_memory_test.cu  DEPS place memory)
 #endif()
diff --git a/paddle/testing/CMakeLists.txt b/paddle/testing/CMakeLists.txt
index 77f84cd43b..a1f446817e 100644
--- a/paddle/testing/CMakeLists.txt
+++ b/paddle/testing/CMakeLists.txt
@@ -6,6 +6,6 @@ if(WITH_TESTING)
   add_library(paddle_test_util STATIC TestUtil.cpp)
   add_dependencies(paddle_test_util paddle_proto ${external_project_dependencies})
   if(NOT MOBILE_INFERENCE)
-    cc_library(paddle_gtest_main SRCS paddle_gtest_main.cc DEPS init paddle_memory gtest gflags)
+    cc_library(paddle_gtest_main SRCS paddle_gtest_main.cc DEPS init memory gtest gflags)
   endif()
 endif()

From 770cec4e316b0a0b35c85fe263048fe5d1ffcbea Mon Sep 17 00:00:00 2001
From: Yi Wang <yi.wang.2005@gmail.com>
Date: Sat, 7 Apr 2018 22:33:41 -0700
Subject: [PATCH 10/16] Fix compilation errors

---
 paddle/fluid/memory/detail/buddy_allocator.cc |  2 +-
 paddle/fluid/memory/malloc.cc                 |  2 +-
 paddle/fluid/memory/memory.h                  | 18 ++++++++++++++++++
 3 files changed, 20 insertions(+), 2 deletions(-)
 create mode 100644 paddle/fluid/memory/memory.h

diff --git a/paddle/fluid/memory/detail/buddy_allocator.cc b/paddle/fluid/memory/detail/buddy_allocator.cc
index d1f0d1755d..4194ba1979 100644
--- a/paddle/fluid/memory/detail/buddy_allocator.cc
+++ b/paddle/fluid/memory/detail/buddy_allocator.cc
@@ -195,7 +195,7 @@ BuddyAllocator::PoolSet::iterator BuddyAllocator::RefillPool() {
   VLOG(10) << "Creating and inserting new block " << p
            << " from system allocator";
 
-  static_cast<MemoryBlock*>(p)->init(cache_, MemoryBlock::FREE_CHUNK, index,
+  static_cast<MemoryBlock*>(p)->init(&cache_, MemoryBlock::FREE_CHUNK, index,
                                      max_chunk_size_, nullptr, nullptr);
 
   // gpu fallback allocation
diff --git a/paddle/fluid/memory/malloc.cc b/paddle/fluid/memory/malloc.cc
index 2c13dbc6d5..0c74f62de5 100644
--- a/paddle/fluid/memory/malloc.cc
+++ b/paddle/fluid/memory/malloc.cc
@@ -12,7 +12,7 @@ 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/fluid/memory/memory.h"
+#include "paddle/fluid/memory/malloc.h"
 
 #include "glog/logging.h"
 
diff --git a/paddle/fluid/memory/memory.h b/paddle/fluid/memory/memory.h
new file mode 100644
index 0000000000..8d904e3be5
--- /dev/null
+++ b/paddle/fluid/memory/memory.h
@@ -0,0 +1,18 @@
+/* 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. */
+
+#pragma once
+
+#include "paddle/fluid/memory/malloc.h"
+#include "paddle/fluid/memory/memcpy.h"

From e309bcd43a38b8f2201a665362351f9349395cf4 Mon Sep 17 00:00:00 2001
From: Yi Wang <yi.wang.2005@gmail.com>
Date: Sat, 7 Apr 2018 22:51:53 -0700
Subject: [PATCH 11/16] Update

---
 paddle/fluid/memory/detail/system_allocator_test.cc | 2 +-
 paddle/fluid/memory/malloc_test.cc                  | 8 ++++----
 2 files changed, 5 insertions(+), 5 deletions(-)

diff --git a/paddle/fluid/memory/detail/system_allocator_test.cc b/paddle/fluid/memory/detail/system_allocator_test.cc
index 95467a6f60..268260142c 100644
--- a/paddle/fluid/memory/detail/system_allocator_test.cc
+++ b/paddle/fluid/memory/detail/system_allocator_test.cc
@@ -26,7 +26,7 @@ void TestAllocator(paddle::memory::detail::SystemAllocator* a, size_t size) {
   bool freed = false;
   {
     size_t index;
-    void* p = a->Alloc(index, size);
+    void* p = a->Alloc(&index, size);
     if (size > 0) {
       EXPECT_NE(p, nullptr);
     } else {
diff --git a/paddle/fluid/memory/malloc_test.cc b/paddle/fluid/memory/malloc_test.cc
index 6d68eed1a8..d39466ef60 100644
--- a/paddle/fluid/memory/malloc_test.cc
+++ b/paddle/fluid/memory/malloc_test.cc
@@ -12,7 +12,7 @@ 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/fluid/memory/memory.h"
+#include "paddle/fluid/memory/malloc.h"
 
 #include <unordered_map>
 
@@ -27,7 +27,7 @@ inline bool is_aligned(void const *p) {
 }
 
 size_t align(size_t size, paddle::platform::CPUPlace place) {
-  size += sizeof(paddle::memory::detail::Metadata);
+  size += sizeof(paddle::memory::detail::MemoryBlock::Desc);
   size_t alignment = paddle::platform::CpuMinChunkSize();
   size_t remaining = size % alignment;
   return remaining == 0 ? size : size + (alignment - remaining);
@@ -85,7 +85,7 @@ TEST(BuddyAllocator, CPUMultAlloc) {
 #ifdef PADDLE_WITH_CUDA
 
 size_t align(size_t size, paddle::platform::CUDAPlace place) {
-  size += sizeof(paddle::memory::detail::Metadata);
+  size += sizeof(paddle::memory::detail::MemoryBlock::Desc);
   size_t alignment = paddle::platform::GpuMinChunkSize();
   size_t remaining = size % alignment;
   return remaining == 0 ? size : size + (alignment - remaining);
@@ -141,7 +141,7 @@ TEST(BuddyAllocator, GPUMultAlloc) {
 }
 
 size_t align(size_t size, paddle::platform::CUDAPinnedPlace place) {
-  size += sizeof(paddle::memory::detail::Metadata);
+  size += sizeof(paddle::memory::detail::MemoryBlock::Desc);
   size_t alignment = paddle::platform::CUDAPinnedMinChunkSize();
   size_t remaining = size % alignment;
   return remaining == 0 ? size : size + (alignment - remaining);

From e9ba79c880acf53a46b8f88bccf2a2a90272d349 Mon Sep 17 00:00:00 2001
From: Yi Wang <yi.wang.2005@gmail.com>
Date: Sat, 7 Apr 2018 23:38:18 -0700
Subject: [PATCH 12/16] Update

---
 paddle/fluid/inference/CMakeLists.txt            | 12 ++++++------
 paddle/fluid/inference/tests/book/CMakeLists.txt |  2 +-
 2 files changed, 7 insertions(+), 7 deletions(-)

diff --git a/paddle/fluid/inference/CMakeLists.txt b/paddle/fluid/inference/CMakeLists.txt
index f417f62f3f..8171e7c116 100644
--- a/paddle/fluid/inference/CMakeLists.txt
+++ b/paddle/fluid/inference/CMakeLists.txt
@@ -11,13 +11,13 @@ cc_library(paddle_fluid DEPS ${fluid_modules})
 # Create shared library
 cc_library(paddle_fluid_shared SHARED
     SRCS io.cc
-    DEPS ARCHIVE_START ${GLOB_OP_LIB} ${FLUID_CORE_MODULES} ARCHIVE_END)
+    DEPS ${GLOB_OP_LIB} ${FLUID_CORE_MODULES})
 set_target_properties(paddle_fluid_shared PROPERTIES OUTPUT_NAME paddle_fluid)
-if(NOT APPLE)
-  # TODO(liuyiqun): Temporarily disable the link flag because it is not support on Mac.
-  set(LINK_FLAGS "-Wl,--version-script ${CMAKE_CURRENT_SOURCE_DIR}/paddle_fluid.map")
-  set_target_properties(paddle_fluid_shared PROPERTIES LINK_FLAGS "${LINK_FLAGS}")
-endif()
+# if(NOT APPLE)
+#   # TODO(liuyiqun): Temporarily disable the link flag because it is not support on Mac.
+#   set(LINK_FLAGS "-Wl,--version-script ${CMAKE_CURRENT_SOURCE_DIR}/paddle_fluid.map")
+#   set_target_properties(paddle_fluid_shared PROPERTIES LINK_FLAGS "${LINK_FLAGS}")
+# endif()
 
 if(WITH_TESTING)
   add_subdirectory(tests/book)
diff --git a/paddle/fluid/inference/tests/book/CMakeLists.txt b/paddle/fluid/inference/tests/book/CMakeLists.txt
index 6ed77adb9d..04522c0e12 100644
--- a/paddle/fluid/inference/tests/book/CMakeLists.txt
+++ b/paddle/fluid/inference/tests/book/CMakeLists.txt
@@ -17,7 +17,7 @@ function(inference_test TARGET_NAME)
     string(REGEX REPLACE "^_$" "" arg "${arg}")
     cc_test(test_inference_${TARGET_NAME}${arg}
         SRCS test_inference_${TARGET_NAME}.cc
-        DEPS ARCHIVE_START paddle_fluid ARCHIVE_END
+        DEPS paddle_fluid
         ARGS --dirname=${PYTHON_TESTS_DIR}/book/${TARGET_NAME}${arg}.inference.model)
     set_tests_properties(test_inference_${TARGET_NAME}${arg}
         PROPERTIES DEPENDS test_${TARGET_NAME})

From 0564e74fe5d58b4a4a01d89a3401e68693bf425d Mon Sep 17 00:00:00 2001
From: Yi Wang <yi.wang.2005@gmail.com>
Date: Sat, 7 Apr 2018 23:38:35 -0700
Subject: [PATCH 13/16] Update

---
 paddle/fluid/inference/CMakeLists.txt | 10 +++++-----
 1 file changed, 5 insertions(+), 5 deletions(-)

diff --git a/paddle/fluid/inference/CMakeLists.txt b/paddle/fluid/inference/CMakeLists.txt
index 8171e7c116..4b19839631 100644
--- a/paddle/fluid/inference/CMakeLists.txt
+++ b/paddle/fluid/inference/CMakeLists.txt
@@ -13,11 +13,11 @@ cc_library(paddle_fluid_shared SHARED
     SRCS io.cc
     DEPS ${GLOB_OP_LIB} ${FLUID_CORE_MODULES})
 set_target_properties(paddle_fluid_shared PROPERTIES OUTPUT_NAME paddle_fluid)
-# if(NOT APPLE)
-#   # TODO(liuyiqun): Temporarily disable the link flag because it is not support on Mac.
-#   set(LINK_FLAGS "-Wl,--version-script ${CMAKE_CURRENT_SOURCE_DIR}/paddle_fluid.map")
-#   set_target_properties(paddle_fluid_shared PROPERTIES LINK_FLAGS "${LINK_FLAGS}")
-# endif()
+if(NOT APPLE)
+  # TODO(liuyiqun): Temporarily disable the link flag because it is not support on Mac.
+  set(LINK_FLAGS "-Wl,--version-script ${CMAKE_CURRENT_SOURCE_DIR}/paddle_fluid.map")
+  set_target_properties(paddle_fluid_shared PROPERTIES LINK_FLAGS "${LINK_FLAGS}")
+endif()
 
 if(WITH_TESTING)
   add_subdirectory(tests/book)

From 080e442671a97c16d468650cea62cc4b9510621b Mon Sep 17 00:00:00 2001
From: Yi Wang <yi.wang.2005@gmail.com>
Date: Sun, 8 Apr 2018 00:18:30 -0700
Subject: [PATCH 14/16] Update

---
 paddle/fluid/inference/tests/book/CMakeLists.txt | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/paddle/fluid/inference/tests/book/CMakeLists.txt b/paddle/fluid/inference/tests/book/CMakeLists.txt
index 04522c0e12..5c06081b61 100644
--- a/paddle/fluid/inference/tests/book/CMakeLists.txt
+++ b/paddle/fluid/inference/tests/book/CMakeLists.txt
@@ -17,7 +17,7 @@ function(inference_test TARGET_NAME)
     string(REGEX REPLACE "^_$" "" arg "${arg}")
     cc_test(test_inference_${TARGET_NAME}${arg}
         SRCS test_inference_${TARGET_NAME}.cc
-        DEPS paddle_fluid
+        DEPS ${ARCHIVE_BEGIN} paddle_fluid ${ARCHIVE_END}
         ARGS --dirname=${PYTHON_TESTS_DIR}/book/${TARGET_NAME}${arg}.inference.model)
     set_tests_properties(test_inference_${TARGET_NAME}${arg}
         PROPERTIES DEPENDS test_${TARGET_NAME})

From e831bd43b069b972585d2cf0dfe58ade639efb38 Mon Sep 17 00:00:00 2001
From: Yi Wang <yi.wang.2005@gmail.com>
Date: Sun, 8 Apr 2018 09:39:54 -0700
Subject: [PATCH 15/16] Add ARCHIVE_START/END back

---
 paddle/fluid/inference/CMakeLists.txt | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/paddle/fluid/inference/CMakeLists.txt b/paddle/fluid/inference/CMakeLists.txt
index 4b19839631..81d0567f05 100644
--- a/paddle/fluid/inference/CMakeLists.txt
+++ b/paddle/fluid/inference/CMakeLists.txt
@@ -11,7 +11,7 @@ cc_library(paddle_fluid DEPS ${fluid_modules})
 # Create shared library
 cc_library(paddle_fluid_shared SHARED
     SRCS io.cc
-    DEPS ${GLOB_OP_LIB} ${FLUID_CORE_MODULES})
+    DEPS ${ARCHIVE_START} ${GLOB_OP_LIB} ${FLUID_CORE_MODULES} ${ARCHIVE_START})
 set_target_properties(paddle_fluid_shared PROPERTIES OUTPUT_NAME paddle_fluid)
 if(NOT APPLE)
   # TODO(liuyiqun): Temporarily disable the link flag because it is not support on Mac.

From f31a0da3632bea8f6647559595b745875441bdf6 Mon Sep 17 00:00:00 2001
From: Yi Wang <yi.wang.2005@gmail.com>
Date: Sun, 8 Apr 2018 10:49:14 -0700
Subject: [PATCH 16/16] Restore inference CMakeLists.txt

---
 paddle/fluid/inference/CMakeLists.txt            | 2 +-
 paddle/fluid/inference/tests/book/CMakeLists.txt | 2 +-
 2 files changed, 2 insertions(+), 2 deletions(-)

diff --git a/paddle/fluid/inference/CMakeLists.txt b/paddle/fluid/inference/CMakeLists.txt
index 81d0567f05..f417f62f3f 100644
--- a/paddle/fluid/inference/CMakeLists.txt
+++ b/paddle/fluid/inference/CMakeLists.txt
@@ -11,7 +11,7 @@ cc_library(paddle_fluid DEPS ${fluid_modules})
 # Create shared library
 cc_library(paddle_fluid_shared SHARED
     SRCS io.cc
-    DEPS ${ARCHIVE_START} ${GLOB_OP_LIB} ${FLUID_CORE_MODULES} ${ARCHIVE_START})
+    DEPS ARCHIVE_START ${GLOB_OP_LIB} ${FLUID_CORE_MODULES} ARCHIVE_END)
 set_target_properties(paddle_fluid_shared PROPERTIES OUTPUT_NAME paddle_fluid)
 if(NOT APPLE)
   # TODO(liuyiqun): Temporarily disable the link flag because it is not support on Mac.
diff --git a/paddle/fluid/inference/tests/book/CMakeLists.txt b/paddle/fluid/inference/tests/book/CMakeLists.txt
index 5c06081b61..6ed77adb9d 100644
--- a/paddle/fluid/inference/tests/book/CMakeLists.txt
+++ b/paddle/fluid/inference/tests/book/CMakeLists.txt
@@ -17,7 +17,7 @@ function(inference_test TARGET_NAME)
     string(REGEX REPLACE "^_$" "" arg "${arg}")
     cc_test(test_inference_${TARGET_NAME}${arg}
         SRCS test_inference_${TARGET_NAME}.cc
-        DEPS ${ARCHIVE_BEGIN} paddle_fluid ${ARCHIVE_END}
+        DEPS ARCHIVE_START paddle_fluid ARCHIVE_END
         ARGS --dirname=${PYTHON_TESTS_DIR}/book/${TARGET_NAME}${arg}.inference.model)
     set_tests_properties(test_inference_${TARGET_NAME}${arg}
         PROPERTIES DEPENDS test_${TARGET_NAME})