ENH: Add buddy allocator Free

paddle/memory/detail/CMakeLists.txt

 if(${WITH_GPU})
-  nv_library(system_allocator SRCS system_allocator.cc DEPS gflags)
-  nv_test(system_allocator_test
-    SRCS system_allocator_test.cc
-    DEPS system_allocator gpu_info gflags)
+  nv_library(system_allocator SRCS system_allocator.cc DEPS gflags gpu_info)
 else(${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})
+cc_test(system_allocator_test SRCS system_allocator_test.cc DEPS system_allocator)
 
-cc_library(metadata SRCS metadata.cc)
+cc_library(meta_data SRCS meta_data.cc)
 cc_library(buddy_allocator SRCS buddy_allocator.cc)

paddle/memory/detail/buddy_allocator.cc

@@ -75,10 +75,49 @@ void* BuddyAllocator::Alloc(size_t unaligned_size) {
 }
 
 void BuddyAllocator::Free(void* p) {
+  // Point back to metadata
   auto block = static_cast<MemoryBlock*>(p)->metadata();
 
-  // acquire the allocator lock
+  // Acquire the allocator lock
   std::lock_guard<std::mutex> lock(mutex_);
+
+  DLOG(INFO) << "Free from address " << block;
+
+  if (block->type(cache_) == MemoryBlock::HUGE_CHUNK) {
+    DLOG(INFO) << "Free directly from system allocator";
+    system_allocator_->Free(block, block->total_size(cache_),
+                            block->index(cache_));
+
+    // Invalidate GPU allocation from cache
+    if (system_allocator_->UseGpu()) {
+      cache_.erase(block);
+    }
+    return;
+  }
+
+  block->mark_as_free(cache_);
+
+  total_used_ -= block->total_size(cache_);
+  total_free_ += block->total_size(cache_);
+
+  // Trying to merge the right buddy
+  if (block->has_right_buddy(cache_)) {
+    DLOG(INFO) << "Merging this block " << block << " with its right buddy "
+               << block->right_buddy(cache_);
+  }
+
+  // Trying to merge the left buddy
+  if (block->has_left_buddy(cache_)) {
+    DLOG(INFO) << "Merging this block " << block << " with its left buddy "
+               << block->left_buddy(cache_);
+  }
+
+  // Dumping this block into pool
+  DLOG(INFO) << "Inserting free block (" << block << ", "
+             << block->total_size(cache_) << ")";
+  pool_.insert({block->index(cache_), block->total_size(cache_), block});
+
+  // TODO(gangliao): Clean up if existing too much free memory
 }
 
 void* BuddyAllocator::SystemAlloc(size_t size) {

paddle/memory/detail/buddy_allocator.h

@@ -14,7 +14,7 @@
 
 #pragma once
 
-#include "paddle/memory/detail/metadata.h"
+#include "paddle/memory/detail/meta_data.h"
 #include "paddle/memory/detail/system_allocator.h"
 #include "paddle/platform/assert.h"
 #include "paddle/platform/cpu_info.h"

paddle/memory/detail/memory_block.h

@@ -14,7 +14,7 @@
 
 #pragma once
 
-#include "paddle/memory/detail/metadata.h"
+#include "paddle/memory/detail/meta_data.h"
 
 #include <cstddef>
 #include <unordered_map>

paddle/memory/detail/metadata.cc → paddle/memory/detail/meta_data.cc

@@ -12,7 +12,7 @@
    See the License for the specific language governing permissions and
    limitations under the License. */
 
-#include "paddle/memory/detail/metadata.h"
+#include "paddle/memory/detail/meta_data.h"
 
 #include <functional>
 

paddle/platform/cpu_info.h

@@ -28,15 +28,5 @@ size_t CpuMinChunkSize();
 //! Get the maximum chunk size for buddy allocator.
 size_t CpuMaxChunkSize();
 
-int GetCurrentDeviceId(void) {
-  int device_id;
-  throw_on_error(cudaGetDevice(&device_id), "cudaGetDevice failed");
-  return device_id;
-}
-
-void SetDeviceId(int device_id) {
-  throw_on_error(cudaSetDevice(device_id), "cudaSetDevice failed");
-}
-
 }  // namespace platform
 }  // namespace paddle

paddle/platform/gpu_info.cc

@@ -31,6 +31,19 @@ int GpuDeviceCount() {
   return count;
 }
 
+int GetCurrentDeviceId() {
+  int device_id;
+  throw_on_error(
+      cudaGetDevice(&device_id),
+      "cudaGetDevice failed in paddle::platform::GetCurrentDeviceId");
+  return device_id;
+}
+
+void SetDeviceId(int id) {
+  throw_on_error(cudaSetDevice(id),
+                 "cudaSetDevice failed in paddle::platform::SetDeviceId");
+}
+
 void GpuMemoryUsage(size_t& available, size_t& total) {
   throw_on_error(cudaMemGetInfo(&available, &total),
                  "cudaMemGetInfo failed in paddle::platform::GetMemoryUsage");

paddle/platform/gpu_info.h

@@ -24,6 +24,12 @@ namespace platform {
 //! Get the total number of GPU devices in system.
 int GpuDeviceCount();
 
+//! Get the current GPU device id in system.
+int GetCurrentDeviceId();
+
+//! Set the GPU device id for next execution.
+void SetDeviceId(int device_id);
+
 //！Get the memory usage of current GPU device.
 void GpuMemoryUsage(size_t& available, size_t& total);