[eager] use CPUAllocator directly (#47125)

* [eager] use CPUAllocator directly

* modify pstring sizeof 48 default

* rm CPU test for NaiveBestFitAllocator

* fix Mac ci compile errors

* use UNUSED to state unused_obj

* mv UNUSED statement to allocator_facade.cc

* fix roi_align

* fix yolov3 test case

* recover original code

* recover original code

* fix trt roi_align test
Co-authored-by: Njerrywgz <jerrywgz@126.com>

paddle/fluid/memory/allocation/allocator_facade.cc

@@ -25,6 +25,7 @@
 #include "paddle/fluid/memory/allocation/stat_allocator.h"
 #include "paddle/fluid/platform/enforce.h"
 #include "paddle/fluid/platform/place.h"
+#include "paddle/phi/core/macros.h"
 
 #if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
 #include <shared_mutex>
@@ -493,8 +494,8 @@ class AllocatorFacadePrivate {
   const AllocatorMap& GetAllocatorMap() { return allocators_; }
 
   void InitNaiveBestFitCPUAllocator() {
-    allocators_[platform::CPUPlace()] =
-        std::make_shared<NaiveBestFitAllocator>(platform::CPUPlace());
+    // It is more efficient to use CPUAllocator directly.
+    allocators_[platform::CPUPlace()] = std::make_shared<CPUAllocator>();
   }
 
 #if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
@@ -1197,6 +1198,10 @@ void AllocatorFacade::RemoveMemoryPoolOfCUDAGraph(int64_t id) {
 }
 #endif
 #endif
+
+UNUSED static std::shared_ptr<NaiveBestFitAllocator> unused_obj =
+    std::make_shared<NaiveBestFitAllocator>(platform::CPUPlace());
+
 }  // namespace allocation
 }  // namespace memory
 }  // namespace paddle

paddle/phi/common/data_type.h

@@ -98,7 +98,7 @@ inline size_t SizeOf(DataType data_type) {
     case DataType::COMPLEX128:
       return 16;
     case DataType::PSTRING:
-      return 24;
+      return 48;
     case DataType::UNDEFINED:
       return 0;
     case DataType::NUM_DATA_TYPES:

python/paddle/fluid/tests/unittests/dygraph_to_static/test_yolov3.py

@@ -56,8 +56,8 @@ class FakeDataReader:
                 img = np.random.normal(
                     0.485, 0.229, [3, cfg.input_size, cfg.input_size]
                 )
-                point1 = cfg.input_size / 4
-                point2 = cfg.input_size / 2
+                point1 = 1 / 4
+                point2 = 1 / 2
                 gt_boxes = np.array([[point1, point1, point2, point2]])
                 gt_labels = np.random.randint(
                     low=0, high=cfg.class_num, size=[1]