Add pinned memory to host memory stats (#43096)

* Add pinned memory to HostMemoryStats

* Add macro for WrapStatAllocator

* Fix CI errors

paddle/fluid/memory/allocation/allocator_facade.cc

@@ -931,9 +931,15 @@ class AllocatorFacadePrivate {
 
   void WrapStatAllocator() {
     for (auto& pair : allocators_) {
+      // Now memory stats is only supported for CPU and GPU
+      const platform::Place& place = pair.first;
+      if (platform::is_cpu_place(place) ||
+          platform::is_cuda_pinned_place(place) ||
+          platform::is_gpu_place(place)) {
         pair.second = std::make_shared<StatAllocator>(pair.second);
       }
     }
+  }
 
 #if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
   // a standalone CUDA allocator to support multi-stream GC in new executor

paddle/fluid/memory/allocation/pinned_allocator.cc

@@ -13,7 +13,7 @@
 // limitations under the License.
 
 #include "paddle/fluid/memory/allocation/pinned_allocator.h"
-
+#include "paddle/fluid/memory/stats.h"
 namespace paddle {
 namespace memory {
 namespace allocation {
@@ -24,6 +24,7 @@ void CPUPinnedAllocator::FreeImpl(phi::Allocation *allocation) {
 #else
   PADDLE_ENFORCE_GPU_SUCCESS(cudaFreeHost(allocation->ptr()));
 #endif
+  HOST_MEMORY_STAT_UPDATE(Reserved, 0, -allocation->size());
   delete allocation;
 }
 phi::Allocation *CPUPinnedAllocator::AllocateImpl(size_t size) {
@@ -33,6 +34,7 @@ phi::Allocation *CPUPinnedAllocator::AllocateImpl(size_t size) {
 #else
   PADDLE_ENFORCE_GPU_SUCCESS(cudaHostAlloc(&ptr, size, cudaHostAllocPortable));
 #endif
+  HOST_MEMORY_STAT_UPDATE(Reserved, 0, size);
   return new Allocation(ptr, size, platform::CUDAPinnedPlace());
 }
 }  // namespace allocation

paddle/fluid/memory/allocation/stat_allocator.h

@@ -45,11 +45,13 @@ class StatAllocator : public Allocator {
     phi::Allocator::AllocationPtr allocation =
         underlying_allocator_->Allocate(size);
 
-    if (platform::is_cpu_place(allocation->place())) {
-      HOST_MEMORY_STAT_UPDATE(Allocated, allocation->place().GetDeviceId(),
+    const platform::Place& place = allocation->place();
+    if (platform::is_cpu_place(place) ||
+        platform::is_cuda_pinned_place(place)) {
+      HOST_MEMORY_STAT_UPDATE(Allocated, place.GetDeviceId(),
                               allocation->size());
     } else {
-      DEVICE_MEMORY_STAT_UPDATE(Allocated, allocation->place().GetDeviceId(),
+      DEVICE_MEMORY_STAT_UPDATE(Allocated, place.GetDeviceId(),
                                 allocation->size());
     }
     return allocation.release();

paddle/fluid/memory/detail/system_allocator.cc

@@ -211,6 +211,7 @@ void* CUDAPinnedAllocator::Alloc(size_t* index, size_t size) {
   if (result == gpuSuccess) {
     *index = 1;  // PINNED memory
     cuda_pinnd_alloc_size_ += size;
+    HOST_MEMORY_STAT_UPDATE(Reserved, 0, size);
     return p;
   } else {
     LOG(WARNING) << "cudaHostAlloc failed.";
@@ -255,6 +256,7 @@ void CUDAPinnedAllocator::Free(void* p, size_t size, size_t index) {
             err));
   }
 #endif
+  HOST_MEMORY_STAT_UPDATE(Reserved, 0, -size);
 }
 
 bool CUDAPinnedAllocator::UseGpu() const { return false; }