fix recycle dep (#474)

dipu/torch_dipu/csrc_dipu/runtime/core/allocator/DIPUCachingAllocator.h

@@ -8,8 +8,10 @@
 #include <c10/core/Allocator.h>
 #include <c10/core/Device.h>
 
-#include "../DIPUEvent.h"
+#include "csrc_dipu/runtime/core/DIPUEvent.h"
+
 #include "DIPUAsyncResourcePool.h"
+#include "DIPUCachingAllocatorUtils.h"
 #include "DIPURawAllocator.h"
 
 namespace dipu {
@@ -137,24 +139,6 @@ void setAllocator(const std::string name, c10::DeviceType device_type,
 
 c10::Allocator* getAllocator(c10::DeviceType device_type);
 
-size_t memoryReserved(const c10::Device& device);
-
-size_t memoryAllocated(const c10::Device& device);
-
-size_t maxMemoryReserved(const c10::Device& device);
-
-size_t maxMemoryAllocated(const c10::Device& device);
-
-void emptyCachedMem();
-
-void initCachedAllocator();
-
-void releaseAllDeviceMem();
-
-void recordStream(const c10::DataPtr& ptr, DIPUStream stream);
-
-void recordStream(const at::Tensor& tensor, DIPUStream stream);
-
 namespace {  // For internal implementation only
 
 struct AllocatorRegisterer {

dipu/torch_dipu/csrc_dipu/runtime/core/allocator/DIPUCachingAllocatorUtils.h

@@ -0,0 +1,28 @@
+
+#pragma once
+
+#include <c10/core/Allocator.h>
+#include <c10/core/Device.h>
+
+#include "csrc_dipu/runtime/core/DIPUStream.h"
+
+namespace dipu {
+size_t memoryReserved(const c10::Device& device);
+
+size_t memoryAllocated(const c10::Device& device);
+
+size_t maxMemoryReserved(const c10::Device& device);
+
+size_t maxMemoryAllocated(const c10::Device& device);
+
+void emptyCachedMem();
+
+void initCachedAllocator();
+
+void releaseAllDeviceMem();
+
+void recordStream(const c10::DataPtr& ptr, DIPUStream stream);
+
+void recordStream(const at::Tensor& tensor, DIPUStream stream);
+
+}  // namespace dipu

dipu/torch_dipu/csrc_dipu/runtime/core/guardimpl/DIPUGuardImpl.h

@@ -7,13 +7,9 @@
 
 #include <csrc_dipu/base/basedef.h>
 #include <csrc_dipu/runtime/core/DIPUStream.h>
+#include <csrc_dipu/runtime/core/allocator/DIPUCachingAllocatorUtils.h>
 #include <csrc_dipu/runtime/devproxy/deviceproxy.h>
-
 namespace dipu {
-// seems DIPUCachingAllocator.h and this class has Cycle reference? need
-// refactor?
-void recordStream(const c10::DataPtr& ptr, DIPUStream stream);
-
 struct DIPUGuardImpl : public c10::impl::DeviceGuardImplInterface {
   static constexpr at::DeviceType static_type = dipu::DIPU_DEVICE_TYPE;
   DIPUGuardImpl() {}