[GraphBolt][CUDA] Inplace pin memory for Graph and TorchFeatureStore (#…

…6962)

python/dgl/graphbolt/impl/fused_csc_sampling_graph.py

@@ -34,6 +34,17 @@ def __init__(
     ):
         super().__init__()
         self._c_csc_graph = c_csc_graph
+        self._is_inplace_pinned = set()
+
+    def __del__(self):
+        # torch.Tensor.pin_memory() is not an inplace operation. To make it
+        # truly in-place, we need to use cudaHostRegister. Then, we need to use
+        # cudaHostUnregister to unpin the tensor in the destructor.
+        # https://github.com/pytorch/pytorch/issues/32167#issuecomment-753551842
+        for tensor in self._is_inplace_pinned:
+            assert (
+                torch.cuda.cudart().cudaHostUnregister(tensor.data_ptr()) == 0
+            )
 
     @property
     def total_num_nodes(self) -> int:
@@ -974,9 +985,33 @@ def _pin(x):
 
     def pin_memory_(self):
         """Copy `FusedCSCSamplingGraph` to the pinned memory in-place."""
+        # torch.Tensor.pin_memory() is not an inplace operation. To make it
+        # truly in-place, we need to use cudaHostRegister. Then, we need to use
+        # cudaHostUnregister to unpin the tensor in the destructor.
+        # https://github.com/pytorch/pytorch/issues/32167#issuecomment-753551842
+        cudart = torch.cuda.cudart()
 
         def _pin(x):
-            return x.pin_memory() if hasattr(x, "pin_memory") else x
+            if hasattr(x, "pin_memory_"):
+                x.pin_memory_()
+            elif (
+                isinstance(x, torch.Tensor)
+                and not x.is_pinned()
+                and x.device.type == "cpu"
+            ):
+                assert (
+                    x.is_contiguous()
+                ), "Tensor pinning is only supported for contiguous tensors."
+                assert (
+                    cudart.cudaHostRegister(
+                        x.data_ptr(), x.numel() * x.element_size(), 0
+                    )
+                    == 0
+                )
+
+                self._is_inplace_pinned.add(x)
+
+            return x
 
         self._apply_to_members(_pin)
 

python/dgl/graphbolt/impl/torch_based_feature_store.py

@@ -83,6 +83,17 @@ def __init__(self, torch_feature: torch.Tensor, metadata: Dict = None):
         # Make sure the tensor is contiguous.
         self._tensor = torch_feature.contiguous()
         self._metadata = metadata
+        self._is_inplace_pinned = set()
+
+    def __del__(self):
+        # torch.Tensor.pin_memory() is not an inplace operation. To make it
+        # truly in-place, we need to use cudaHostRegister. Then, we need to use
+        # cudaHostUnregister to unpin the tensor in the destructor.
+        # https://github.com/pytorch/pytorch/issues/32167#issuecomment-753551842
+        for tensor in self._is_inplace_pinned:
+            assert (
+                torch.cuda.cudart().cudaHostUnregister(tensor.data_ptr()) == 0
+            )
 
     def read(self, ids: torch.Tensor = None):
         """Read the feature by index.
@@ -169,14 +180,30 @@ def metadata(self):
 
     def pin_memory_(self):
         """In-place operation to copy the feature to pinned memory."""
-        self._tensor = self._tensor.pin_memory()
+        # torch.Tensor.pin_memory() is not an inplace operation. To make it
+        # truly in-place, we need to use cudaHostRegister. Then, we need to use
+        # cudaHostUnregister to unpin the tensor in the destructor.
+        # https://github.com/pytorch/pytorch/issues/32167#issuecomment-753551842
+        x = self._tensor
+        if not x.is_pinned() and x.device.type == "cpu":
+            assert (
+                x.is_contiguous()
+            ), "Tensor pinning is only supported for contiguous tensors."
+            assert (
+                torch.cuda.cudart().cudaHostRegister(
+                    x.data_ptr(), x.numel() * x.element_size(), 0
+                )
+                == 0
+            )
+
+            self._is_inplace_pinned.add(x)
 
     def to(self, device):  # pylint: disable=invalid-name
         """Copy `TorchBasedFeature` to the specified device."""
         # copy.copy is a shallow copy so it does not copy tensor memory.
         self2 = copy.copy(self)
         if device == "pinned":
-            self2.pin_memory_()
+            self2._tensor = self2._tensor.pin_memory()
         else:
             self2._tensor = self2._tensor.to(device)
         return self2

tests/python/pytorch/graphbolt/impl/test_fused_csc_sampling_graph.py

@@ -1601,10 +1601,14 @@ def test_csc_sampling_graph_to_device(device):
 def test_csc_sampling_graph_to_pinned_memory():
     # Construct FusedCSCSamplingGraph.
     graph = create_fused_csc_sampling_graph()
+    ptr = graph.csc_indptr.data_ptr()
 
     # Copy to pinned_memory in-place.
     graph.pin_memory_()
 
+    # Check if pinning is truly in-place.
+    assert graph.csc_indptr.data_ptr() == ptr
+
     is_graph_on_device_type(graph, "cpu")
     is_graph_pinned(graph)
 

tests/python/pytorch/graphbolt/impl/test_torch_based_feature_store.py

@@ -221,6 +221,9 @@ def test_torch_based_pinned_feature(dtype, idtype, shape):
     feature = gb.TorchBasedFeature(tensor)
     feature.pin_memory_()
 
+    # Check if pinning is truly in-place.
+    assert feature._tensor.data_ptr() == tensor.data_ptr()
+
     # Test read entire pinned feature, the result should be on cuda.
     assert torch.equal(feature.read(), test_tensor_cuda)
     assert feature.read().is_cuda