[Graphbolt] Add default prefetcher (#6541)

python/dgl/graphbolt/dataloader.py

@@ -4,6 +4,7 @@ import torch.utils.data
 import torchdata.dataloader2.graph as dp_utils
 import torchdata.datapipes as dp
 
+from .base import CopyTo
 from .feature_fetcher import FeatureFetcher
 from .item_sampler import ItemSampler
 
@@ -16,6 +17,25 @@ __all__ = [
 ]
 
 
+def _find_and_wrap_parent(
+    datapipe_graph, datapipe_adjlist, target_datapipe, wrapper, **kwargs
+):
+    """Find parent of target_datapipe and wrap it with ."""
+    datapipes = dp_utils.find_dps(
+        datapipe_graph,
+        target_datapipe,
+    )
+    for datapipe in datapipes:
+        datapipe_id = id(datapipe)
+        for parent_datapipe_id in datapipe_adjlist[datapipe_id][1]:
+            parent_datapipe, _ = datapipe_adjlist[parent_datapipe_id]
+            datapipe_graph = dp_utils.replace_dp(
+                datapipe_graph,
+                parent_datapipe,
+                wrapper(parent_datapipe, **kwargs),
+            )
+
+
 class SingleProcessDataLoader(torch.utils.data.DataLoader):
     """Single process DataLoader.
 
@@ -34,6 +54,17 @@ class SingleProcessDataLoader(torch.utils.data.DataLoader):
     # The exception is that batch_size should be None, since we already
     # have minibatch sampling and collating in ItemSampler.
     def __init__(self, datapipe):
+        datapipe_graph = dp_utils.traverse_dps(datapipe)
+        datapipe_adjlist = datapipe_graph_to_adjlist(datapipe_graph)
+        # Cut datapipe at CopyTo and wrap with prefetcher. This enables the
+        # data pipeline up to the CopyTo operation to run in a separate thread.
+        _find_and_wrap_parent(
+            datapipe_graph,
+            datapipe_adjlist,
+            CopyTo,
+            dp.iter.Prefetcher,
+            buffer_size=2,
+        )
         super().__init__(datapipe, batch_size=None, num_workers=0)
 
 
@@ -107,18 +138,22 @@ class MultiProcessDataLoader(torch.utils.data.DataLoader):
             )
 
         # (2) Cut datapipe at FeatureFetcher and wrap.
-        feature_fetchers = dp_utils.find_dps(
+        _find_and_wrap_parent(
             datapipe_graph,
+            datapipe_adjlist,
             FeatureFetcher,
+            MultiprocessingWrapper,
+            num_workers=num_workers,
         )
-        for feature_fetcher in feature_fetchers:
-            feature_fetcher_id = id(feature_fetcher)
-            for parent_datapipe_id in datapipe_adjlist[feature_fetcher_id][1]:
-                parent_datapipe, _ = datapipe_adjlist[parent_datapipe_id]
-                datapipe_graph = dp_utils.replace_dp(
+
+        # (3) Cut datapipe at CopyTo and wrap with prefetcher. This enables the
+        # data pipeline up to the CopyTo operation to run in a separate thread.
+        _find_and_wrap_parent(
             datapipe_graph,
-                    parent_datapipe,
-                    MultiprocessingWrapper(parent_datapipe, num_workers),
+            datapipe_adjlist,
+            CopyTo,
+            dp.iter.Prefetcher,
+            buffer_size=2,
         )
 
         # The stages after feature fetching is still done in the main process.