[cherry-pick 54036] add host memory stats (#54037)

* add host memory stats

* add ut

paddle/fluid/memory/allocation/buddy_allocator.cc

@@ -59,6 +59,8 @@ BuddyAllocator::BuddyAllocator(
 #endif
   }
 #endif
+  VLOG(1) << "min_chunk_size_: " << min_chunk_size_
+          << ", max_chunk_size_:" << max_chunk_size_;
 }
 
 BuddyAllocator::~BuddyAllocator() {
@@ -228,7 +230,7 @@ void* BuddyAllocator::SystemAlloc(size_t size) {
   size_t index = 0;
   void* p = system_allocator_->Alloc(&index, size);
 
-  VLOG(10) << "Allocated " << p << " from system allocator.";
+  VLOG(8) << "Allocated " << p << " size " << size << " from system allocator.";
 
   if (p == nullptr) return nullptr;
 
@@ -258,8 +260,8 @@ BuddyAllocator::PoolSet::iterator BuddyAllocator::RefillPool(
 
   if (p == nullptr) return pool_.end();
 
-  VLOG(10) << "Creating and inserting new block " << p
-           << " from system allocator";
+  VLOG(8) << "Creating and inserting new block " << p << " size "
+          << allocate_bytes << " from system allocator";
 
   static_cast<MemoryBlock*>(p)->Init(&cache_,
                                      MemoryBlock::FREE_CHUNK,

paddle/fluid/memory/allocation/cpu_allocator.cc

@@ -16,6 +16,7 @@
 
 #include <stdlib.h>
 
+#include "paddle/fluid/memory/stats.h"
 #include "paddle/fluid/platform/enforce.h"
 
 namespace paddle {
@@ -25,12 +26,14 @@ namespace allocation {
 bool CPUAllocator::IsAllocThreadSafe() const { return true; }
 
 void CPUAllocator::FreeImpl(phi::Allocation *allocation) {
+  auto size = allocation->size();
   void *p = allocation->ptr();
 #ifdef _WIN32
   _aligned_free(p);
 #else
   free(p);
 #endif
+  HOST_MEMORY_STAT_UPDATE(Reserved, 0, -size);
   delete allocation;
 }
 
@@ -46,6 +49,7 @@ phi::Allocation *CPUAllocator::AllocateImpl(size_t size) {
       platform::errors::ResourceExhausted(
           "Fail to alloc memory of %ld size, error code is %d.", size, error));
 #endif
+  HOST_MEMORY_STAT_UPDATE(Reserved, 0, size);
   return new Allocation(p, size, platform::CPUPlace());
 }
 }  // namespace allocation

paddle/fluid/memory/allocation/naive_best_fit_allocator.cc

@@ -380,6 +380,7 @@ template <>
 void *Alloc<platform::CUDAPinnedPlace>(const platform::CUDAPinnedPlace &place,
                                        size_t size) {
 #if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
+  VLOG(10) << "Allocate " << size << " bytes on " << platform::Place(place);
   auto *buddy_allocator = GetCUDAPinnedBuddyAllocator();
   void *ptr = buddy_allocator->Alloc(size);
 
@@ -402,6 +403,7 @@ void Free<platform::CUDAPinnedPlace>(const platform::CUDAPinnedPlace &place,
                                      void *p,
                                      size_t size) {
 #if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
+  VLOG(10) << "Free " << size << " bytes on " << platform::Place(place);
   GetCUDAPinnedBuddyAllocator()->Free(p);
 #else
   PADDLE_THROW(platform::errors::PermissionDenied(
@@ -413,6 +415,7 @@ template <>
 uint64_t Release<platform::CUDAPinnedPlace>(
     const platform::CUDAPinnedPlace &place) {
 #if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
+  VLOG(10) << "Release on " << platform::Place(place);
   return GetCUDAPinnedBuddyAllocator()->Release();
 #else
   PADDLE_THROW(platform::errors::PermissionDenied(

paddle/fluid/pybind/pybind.cc

@@ -1935,6 +1935,8 @@ All parameter, weight, gradient are variables in Paddle.
   m.def("device_memory_stat_current_value",
         memory::DeviceMemoryStatCurrentValue);
   m.def("device_memory_stat_peak_value", memory::DeviceMemoryStatPeakValue);
+  m.def("host_memory_stat_current_value", memory::HostMemoryStatCurrentValue);
+  m.def("host_memory_stat_peak_value", memory::HostMemoryStatPeakValue);
   m.def(
       "run_cmd",
       [](const std::string &cmd,

python/paddle/fluid/tests/unittests/test_host_memory_stats.py

+# Copyright (c) 2021 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.
+
+import unittest
+
+import paddle
+from paddle.fluid import core
+
+paddle.set_device('cpu')
+
+
+class TestHostMemoryStats(unittest.TestCase):
+    def test_memory_allocated_with_pinned(self, device=None):
+        if core.is_compiled_with_cuda():
+            tensor = paddle.zeros(shape=[256])
+            tensor_pinned = tensor.pin_memory()
+            alloc_size = 4 * 256  # 256 float32 data, with 4 bytes for each one
+            memory_allocated_size = core.host_memory_stat_current_value(
+                "Allocated", 0
+            )
+            self.assertEqual(memory_allocated_size, alloc_size * 2)
+
+            def foo():
+                tensor = paddle.zeros(shape=[256])
+                tensor_pinned = tensor.pin_memory()
+                memory_allocated_size = core.host_memory_stat_current_value(
+                    "Allocated", 0
+                )
+                self.assertEqual(memory_allocated_size, alloc_size * 4)
+                max_allocated_size = core.host_memory_stat_peak_value(
+                    "Allocated", 0
+                )
+                self.assertEqual(memory_allocated_size, alloc_size * 4)
+
+            foo()
+
+            memory_allocated_size = core.host_memory_stat_current_value(
+                "Allocated", 0
+            )
+            self.assertEqual(memory_allocated_size, alloc_size * 2)
+
+            max_allocated_size = core.host_memory_stat_peak_value(
+                "Allocated", 0
+            )
+            self.assertEqual(max_allocated_size, alloc_size * 4)
+
+
+if __name__ == "__main__":
+    unittest.main()