Merge pull request #2596 from wangkuiyi/memory_cpu_allocator

Memory CPU allocator

Author: wangkuiyi

CMakeLists.txt

@@ -27,6 +27,7 @@ if(NOT CMAKE_CROSSCOMPILING)
 endif(NOT CMAKE_CROSSCOMPILING)
 find_package(Git REQUIRED)
 find_package(Threads REQUIRED)
+find_package(Boost QUIET)
 
 include(simd)
 
@@ -110,6 +111,7 @@ include_directories("${PROJ_ROOT}")
 include_directories("${PROJ_ROOT}/paddle/cuda/include")
 include_directories("${CMAKE_CURRENT_BINARY_DIR}/proto")
 include_directories("${CMAKE_CURRENT_BINARY_DIR}/go/pserver/cclient")
+include_directories(${Boost_INCLUDE_DIRS})
 
 set(EXTERNAL_LIBS
     ${GFLAGS_LIBRARIES}

paddle/CMakeLists.txt

@@ -11,15 +11,8 @@ add_subdirectory(scripts)
 add_subdirectory(optimizer)
 add_subdirectory(strings)
 
-# Do not build go directory until go cmake is working smoothly.
-# if(CMAKE_Go_COMPILER)
-#   add_subdirectory(go)
-# endif()
-
-find_package(Boost QUIET)
-
 if(Boost_FOUND)
-  include_directories(${Boost_INCLUDE_DIRS})
+  add_subdirectory(memory)
   add_subdirectory(platform)
   add_subdirectory(framework)
 endif()

paddle/memory/.clang-format

@@ -0,0 +1,5 @@
+---
+Language:        Cpp
+BasedOnStyle:  Google
+Standard:  Cpp11 
+...

paddle/memory/CMakeLists.txt

@@ -0,0 +1 @@
+add_subdirectory(detail)

paddle/memory/README.md

@@ -97,6 +97,7 @@ class BuddyAllocator {
   struct Block {
     size_t size;
     Block* left, right;
+    size_t index; // allocator id
   };
   ...
 };

paddle/memory/detail/CMakeLists.txt

@@ -0,0 +1,7 @@
+if(${WITH_GPU})
+  nv_library(system_allocator SRCS system_allocator.cc DEPS gflags)
+  nv_test(system_allocator_test SRCS system_allocator_test.cc DEPS system_allocator gflags)
+else(${WITH_GPU})
+  cc_library(system_allocator SRCS system_allocator.cc DEPS gflags)
+  cc_test(system_allocator_test SRCS system_allocator_test.cc DEPS system_allocator gflags)
+endif(${WITH_GPU})

paddle/memory/detail/buddy_allocator.cc

@@ -0,0 +1,35 @@
+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+   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/memory/detail/buddy_allocator.h"
+
+namespace paddle {
+namespace memory {
+namespace detail {
+
+BuddyAllocator::BuddyAllocator(size_t pool_size, size_t max_pools,
+                               SystemAllocator* system_allocator)
+    : pool_size_(pool_size),
+      max_pools_(max_pools),
+      system_allocator_(system_allocator) {
+  PADDLE_ASSERT(pool_size > 0);
+  PADDLE_ASSERT(max_pools > 0);
+  PADDLE_ASSERT(system_allocator != nullptr);
+}
+
+}  // namespace detail
+}  // namespace memory
+}  // namespace paddle

paddle/memory/detail/buddy_allocator.h

@@ -0,0 +1,86 @@
+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+   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/memory/detail/system_allocator.h"
+
+#include <mutex>
+#include <vector>
+
+namespace paddle {
+namespace memory {
+namespace detail {
+
+class BuddyAllocator {
+ public:
+  BuddyAllocator(size_t pool_size, size_t max_pools,
+                 SystemAllocator* system_allocator);
+  ~BuddyAllocator();
+
+  void* Alloc(size_t size);
+  void Free(void*);
+  size_t Used();
+
+ private:
+  struct Block {
+    size_t size_;
+    Block* left_;   // left buddy
+    Block* right_;  // right buddy
+  };
+
+  // Initially, there is only one pool.  If a Alloc founds not enough
+  // memory from that pool, and there has not been max_num_pools_,
+  // create a new pool by calling system_allocator_.Alloc(pool_size_).
+  std::vector<void*> pools_;
+
+  size_t pool_size_;      // the size of each pool;
+  size_t max_num_pools_;  // the size of all pools;
+
+  SystemAllocator* system_allocator_;
+
+  std::mutex mutex_;
+
+  // Disable copy and assignment.
+  BuddyAllocator(const BuddyAllocator&) = delete;
+  BuddyAllocator& operator=(const BuddyAllocator&) = delete;
+};
+
+BuddyAllocator<CPUAllocator>* GetCPUBuddyAllocator() {
+  static BuddyAllocator<CPUAllocator>* a = nullptr;
+  if (a == nullptr) {
+    a = new BuddyAllocator<CPUAllocator>();
+  }
+  return a;
+}
+
+#ifndef PADDLE_ONLY_CPU  // The following code are for CUDA.
+
+BuddyAllocator<GPUAllocator>* GetGPUBuddyAllocator(int gpu_id) {
+  static BuddyAllocator<GPUAllocator>** as = NULL;
+  if (as == NULL) {
+    int gpu_num = platform::GetDeviceCount();
+    as = new BuddyAllocator<GPUAllocator>*[gpu_num];
+    for (int gpu = 0; gpu < gpu_num; gpu++) {
+      as[gpu] = new BuddyAllocator<GPUAllocator>();
+    }
+  }
+  return as[gpu_id];
+}
+
+#endif  // PADDLE_ONLY_CPU
+
+}  // namespace detail
+}  // namespace memory
+}  // namespace paddle

paddle/memory/detail/system_allocator.cc

@@ -0,0 +1,90 @@
+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+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/memory/detail/system_allocator.h"
+
+#include <stdlib.h>    // for malloc and free
+#include <sys/mman.h>  // for mlock and munlock
+
+#include "gflags/gflags.h"
+#include "paddle/platform/assert.h"
+#include "paddle/platform/cuda.h"
+
+// If use_pinned_memory is true, CPUAllocator calls mlock, which
+// returns pinned and locked memory as staging areas for data exchange
+// between host and device.  Allocates too much would reduce the amount
+// of memory available to the system for paging.  So, by default, we
+// should set false to use_pinned_memory.
+DEFINE_bool(use_pinned_memory, false,
+            "If set, allocate cpu/gpu pinned memory.");
+
+namespace paddle {
+namespace memory {
+namespace detail {
+
+void* CPUAllocator::Alloc(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;
+
+  void* p = malloc(size);
+  if (p != nullptr && FLAGS_use_pinned_memory) {
+    mlock(p, size);
+  }
+  return p;
+}
+
+void CPUAllocator::Free(void* p, size_t size) {
+  if (p != nullptr && FLAGS_use_pinned_memory) {
+    munlock(p, size);
+  }
+  free(p);
+}
+
+#ifndef PADDLE_ONLY_CPU
+
+void* GPUAllocator::Alloc(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;
+  }
+
+  void* p = 0;
+  cudaError_t result =
+      FLAGS_use_pinned_memory ? cudaMallocHost(&p, size) : cudaMalloc(&p, size);
+  if (result != cudaSuccess) {
+    cudaGetLastError();  // clear error if there is any.
+  }
+  return result == cudaSuccess ? p : nullptr;
+}
+
+void GPUAllocator::Free(void* p, size_t size) {
+  // Purposefully allow cudaErrorCudartUnloading, because
+  // that is returned if you ever call cudaFree after the
+  // driver has already shutdown. This happens only if the
+  // process is terminating, in which case we don't care if
+  // cudaFree succeeds.
+  cudaError_t err = FLAGS_use_pinned_memory ? cudaFreeHost(p) : cudaFree(p);
+  if (err != cudaErrorCudartUnloading) {
+    platform::throw_on_error(err, "cudaFree{Host} failed");
+  }
+}
+
+#endif  // PADDLE_ONLY_CPU
+
+}  // namespace detail
+}  // namespace memory
+}  // namespace paddle

paddle/memory/detail/system_allocator.h

@@ -0,0 +1,53 @@
+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+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 <stddef.h>  // for size_t
+
+namespace paddle {
+namespace memory {
+namespace detail {
+
+// SystemAllocator is the parent class of CPUAllocator and
+// GPUAllocator.  A BuddyAllocator object uses a SystemAllocator*
+// pointing to the underlying system allocator.  An alternative to
+// this class hierarchy is to pass a system allocator class to
+// BuddyAllocator as a template parameter.  This approach makes
+// BuddyAllocator a class template, and it's very complicated
+// algorithm would make the buddy_allocator.h messy.
+class SystemAllocator {
+ public:
+  virtual ~SystemAllocator() {}
+  virtual void* Alloc(size_t size) = 0;
+  virtual void Free(void* p, size_t size) = 0;
+};
+
+class CPUAllocator : public SystemAllocator {
+ public:
+  virtual void* Alloc(size_t size);
+  virtual void Free(void* p, size_t size);
+};
+
+#ifndef PADDLE_ONLY_CPU
+class GPUAllocator : public SystemAllocator {
+ public:
+  virtual void* Alloc(size_t size);
+  virtual void Free(void* p, size_t size);
+};
+#endif  // PADDLE_ONLY_CPU
+
+}  // namespace detail
+}  // namespace memory
+}  // namespace paddle