2 files changed, 7 insertions, 1 deletions
diff --git a/tensorflow/stream_executor/host/host_gpu_executor.cc b/tensorflow/stream_executor/host/host_gpu_executor.cc
index 3cd97b3cf1..8adf739b17 100644
--- a/tensorflow/stream_executor/host/host_gpu_executor.cc
+++ b/tensorflow/stream_executor/host/host_gpu_executor.cc
@@ -93,7 +93,7 @@ bool HostExecutor::MemcpyDeviceToDevice(Stream *stream,
   // the nature of the HostExecutor) memcpy  on the stream (HostStream)
   // associated with the HostExecutor.
   AsHostStream(stream)->EnqueueTask(
-      [src_mem, dst_mem, size]() { memcpy(src_mem, dst_mem, size); });
+      [src_mem, dst_mem, size]() { memcpy(dst_mem, src_mem, size); });
   return true;
 }
 
diff --git a/tensorflow/stream_executor/stream.cc b/tensorflow/stream_executor/stream.cc
index ca1b8e28e6..2c495c99e1 100644
--- a/tensorflow/stream_executor/stream.cc
+++ b/tensorflow/stream_executor/stream.cc
@@ -268,6 +268,12 @@ Stream::~Stream() {
   VLOG_CALL();
 
   temporary_memory_manager_.ForceDeallocateAll();
+  // Ensure the stream is completed.
+  auto status = BlockHostUntilDone();
+  if (!status.ok()) {
+    LOG(WARNING) << "Error blocking host until done in stream destructor: "
+                 << status;
+  }
 
   if (allocated_) {
     parent_->DeallocateStream(this);