SetDeviceId in StreamSafeCUDAAllocation (#49080)

paddle/fluid/memory/allocation/stream_safe_cuda_allocator.cc

@@ -16,6 +16,7 @@
 #include <thread>
 
 #include "paddle/fluid/platform/profiler/event_tracing.h"
+#include "paddle/phi/backends/gpu/gpu_info.h"
 
 #ifdef PADDLE_WITH_CUDA
 #include "paddle/fluid/platform/device/gpu/cuda/cuda_graph.h"
@@ -43,6 +44,9 @@ void StreamSafeCUDAAllocation::RecordStream(gpuStream_t stream) {
     return;
   }
 
+  std::call_once(once_flag_,
+                 [this] { phi::backends::gpu::SetDeviceId(place_.device); });
+
   std::lock_guard<SpinLock> lock_guard(outstanding_event_map_lock_);
 #ifdef PADDLE_WITH_CUDA
   if (UNLIKELY(platform::CUDAGraph::IsThisThreadCapturing())) {
@@ -63,6 +67,9 @@ bool StreamSafeCUDAAllocation::CanBeFreed() {
   }
 #endif
 
+  std::call_once(once_flag_,
+                 [this] { phi::backends::gpu::SetDeviceId(place_.device); });
+
   RecordGraphCapturingStreams();
 
   for (auto it = outstanding_event_map_.begin();
@@ -259,6 +266,8 @@ uint64_t StreamSafeCUDAAllocator::ProcessUnfreedAllocationsAndRelease() {
   return underlying_allocator_->Release(place_);
 }
 
+std::once_flag StreamSafeCUDAAllocation::once_flag_;
+
 std::map<platform::Place, std::vector<StreamSafeCUDAAllocator*>>
     StreamSafeCUDAAllocator::allocator_map_;
 SpinLock StreamSafeCUDAAllocator::allocator_map_lock_;

paddle/fluid/memory/allocation/stream_safe_cuda_allocator.h

@@ -45,6 +45,7 @@ class StreamSafeCUDAAllocation : public Allocation {
   gpuStream_t GetOwningStream() const;
 
  private:
+  static std::once_flag once_flag_;
   void RecordGraphCapturingStreams();
   void RecordStreamWithNoGraphCapturing(gpuStream_t stream);
   DecoratedAllocationPtr underlying_allocation_;