[GraphBolt][CUDA] Add CooperativeConv and minor fixes. (#7797)

python/dgl/graphbolt/impl/__init__.py

@@ -15,3 +15,4 @@
 from .gpu_graph_cache import *
 from .cpu_feature_cache import *
 from .cpu_cached_feature import *
+from .cooperative_conv import *

python/dgl/graphbolt/impl/cooperative_conv.py

@@ -0,0 +1,109 @@
+"""Graphbolt cooperative convolution."""
+from typing import Dict, Union
+
+import torch
+
+from ..sampled_subgraph import SampledSubgraph
+from ..subgraph_sampler import all_to_all, convert_to_hetero, revert_to_homo
+
+__all__ = ["CooperativeConvFunction", "CooperativeConv"]
+
+
+class CooperativeConvFunction(torch.autograd.Function):
+    """Cooperative convolution operation from Cooperative Minibatching.
+
+    Implements the `all-to-all` message passing algorithm
+    in Cooperative Minibatching, which was initially proposed in
+    `Deep Graph Library PR#4337<https://github.com/dmlc/dgl/pull/4337>`__ and
+    was later first fully described in
+    `Cooperative Minibatching in Graph Neural Networks
+    <https://arxiv.org/abs/2310.12403>`__.
+    Cooperation between the GPUs eliminates duplicate work performed across the
+    GPUs due to the overlapping sampled k-hop neighborhoods of seed nodes when
+    performing GNN minibatching. This reduces the redundant computations across
+    GPUs at the expense of communication.
+    """
+
+    @staticmethod
+    def forward(
+        ctx,
+        subgraph: SampledSubgraph,
+        tensor: Union[torch.Tensor, Dict[str, torch.Tensor]],
+    ):
+        """Implements the forward pass."""
+        counts_sent = convert_to_hetero(subgraph._counts_sent)
+        counts_received = convert_to_hetero(subgraph._counts_received)
+        seed_inverse_ids = convert_to_hetero(subgraph._seed_inverse_ids)
+        seed_sizes = convert_to_hetero(subgraph._seed_sizes)
+        ctx.save_for_backward(
+            counts_sent, counts_received, seed_inverse_ids, seed_sizes
+        )
+        outs = {}
+        for ntype, typed_tensor in convert_to_hetero(tensor).items():
+            out = typed_tensor.new_empty(
+                (sum(counts_sent[ntype]),) + typed_tensor.shape[1:]
+            )
+            all_to_all(
+                torch.split(out, counts_sent[ntype]),
+                torch.split(
+                    typed_tensor[seed_inverse_ids[ntype]],
+                    counts_received[ntype],
+                ),
+            )
+            outs[ntype] = out
+        return revert_to_homo(out)
+
+    @staticmethod
+    def backward(
+        ctx, grad_output: Union[torch.Tensor, Dict[str, torch.Tensor]]
+    ):
+        """Implements the forward pass."""
+        (
+            counts_sent,
+            counts_received,
+            seed_inverse_ids,
+            seed_sizes,
+        ) = ctx.saved_tensors
+        outs = {}
+        for ntype, typed_grad_output in convert_to_hetero(grad_output).items():
+            out = typed_grad_output.new_empty(
+                (sum(counts_received[ntype]),) + typed_grad_output.shape[1:]
+            )
+            all_to_all(
+                torch.split(out, counts_received[ntype]),
+                torch.split(typed_grad_output, counts_sent[ntype]),
+            )
+            i = out.new_empty(2, out.shape[0], dtype=torch.int64)
+            i[0] = torch.arange(
+                out.shape[0], device=typed_grad_output.device
+            )  # src
+            i[1] = seed_inverse_ids[ntype]  # dst
+            coo = torch.sparse_coo_tensor(
+                i, 1, size=(seed_sizes[ntype], i.shape[1])
+            )
+            outs[ntype] = torch.sparse.mm(coo, out)
+        return None, revert_to_homo(outs)
+
+
+class CooperativeConv(torch.nn.Module):
+    """Cooperative convolution operation from Cooperative Minibatching.
+
+    Implements the `all-to-all` message passing algorithm
+    in Cooperative Minibatching, which was initially proposed in
+    `Deep Graph Library PR#4337<https://github.com/dmlc/dgl/pull/4337>`__ and
+    was later first fully described in
+    `Cooperative Minibatching in Graph Neural Networks
+    <https://arxiv.org/abs/2310.12403>`__.
+    Cooperation between the GPUs eliminates duplicate work performed across the
+    GPUs due to the overlapping sampled k-hop neighborhoods of seed nodes when
+    performing GNN minibatching. This reduces the redundant computations across
+    GPUs at the expense of communication.
+    """
+
+    def forward(
+        self,
+        subgraph: SampledSubgraph,
+        x: Union[torch.Tensor, Dict[str, torch.Tensor]],
+    ):
+        """Implements the forward pass."""
+        return CooperativeConvFunction.apply(subgraph, x)

python/dgl/graphbolt/impl/neighbor_sampler.py

@@ -601,17 +601,18 @@ def _seeds_cooperative_exchange_2(minibatch):
                 typed_seeds.split(typed_counts_sent),
             )
             seeds_received[ntype] = typed_seeds_received
-        subgraph._seeds_received = seeds_received
+            counts_sent[ntype] = typed_counts_sent
+            counts_received[ntype] = typed_counts_received
+        minibatch._seed_nodes = seeds_received
         subgraph._counts_sent = revert_to_homo(counts_sent)
         subgraph._counts_received = revert_to_homo(counts_received)
         return minibatch
 
     @staticmethod
     def _seeds_cooperative_exchange_3(minibatch):
-        subgraph = minibatch.sampled_subgraphs[0]
         nodes = {
             ntype: [typed_seeds]
-            for ntype, typed_seeds in subgraph._seeds_received.items()
+            for ntype, typed_seeds in minibatch._seed_nodes.items()
         }
         minibatch._unique_future = unique_and_compact(
             nodes, 0, 1, async_op=True
@@ -627,6 +628,11 @@ def _seeds_cooperative_exchange_4(minibatch):
         }
         minibatch._seed_nodes = revert_to_homo(unique_seeds)
         subgraph = minibatch.sampled_subgraphs[0]
+        sizes = {
+            ntype: typed_seeds.size(0)
+            for ntype, typed_seeds in unique_seeds.items()
+        }
+        subgraph._seed_sizes = revert_to_homo(sizes)
         subgraph._seed_inverse_ids = revert_to_homo(inverse_seeds)
         return minibatch
 
@@ -831,6 +837,16 @@ class NeighborSampler(NeighborSamplerImpl):
     gpu_cache_threshold : int, optional
         Determines how many times a vertex needs to be accessed before its
         neighborhood ends up being cached on the GPU.
+    cooperative: bool, optional
+        Boolean indicating whether Cooperative Minibatching, which was initially
+        proposed in
+        `Deep Graph Library PR#4337<https://github.com/dmlc/dgl/pull/4337>`__
+        and was later first fully described in
+        `Cooperative Minibatching in Graph Neural Networks
+        <https://arxiv.org/abs/2310.12403>`__. Cooperation between the GPUs
+        eliminates duplicate work performed across the GPUs due to the
+        overlapping sampled k-hop neighborhoods of seed nodes when performing
+        GNN minibatching.
     asynchronous: bool
         Boolean indicating whether sampling and compaction stages should run
         in background threads to hide the latency of CPU GPU synchronization.
@@ -986,6 +1002,16 @@ class LayerNeighborSampler(NeighborSamplerImpl):
     gpu_cache_threshold : int, optional
         Determines how many times a vertex needs to be accessed before its
         neighborhood ends up being cached on the GPU.
+    cooperative: bool, optional
+        Boolean indicating whether Cooperative Minibatching, which was initially
+        proposed in
+        `Deep Graph Library PR#4337<https://github.com/dmlc/dgl/pull/4337>`__
+        and was later first fully described in
+        `Cooperative Minibatching in Graph Neural Networks
+        <https://arxiv.org/abs/2310.12403>`__. Cooperation between the GPUs
+        eliminates duplicate work performed across the GPUs due to the
+        overlapping sampled k-hop neighborhoods of seed nodes when performing
+        GNN minibatching.
     asynchronous: bool
         Boolean indicating whether sampling and compaction stages should run
         in background threads to hide the latency of CPU GPU synchronization.

python/dgl/graphbolt/subgraph_sampler.py

@@ -16,6 +16,7 @@
 __all__ = [
     "SubgraphSampler",
     "all_to_all",
+    "convert_to_hetero",
     "revert_to_homo",
 ]
 
@@ -89,6 +90,13 @@ def revert_to_homo(d: dict):
     return list(d.values())[0] if is_homogenous else d
 
 
+def convert_to_hetero(item):
+    """Utility function to convert homogenous data to heterogenous with a single
+    node type."""
+    is_heterogenous = isinstance(item, dict)
+    return item if is_heterogenous else {"_N": item}
+
+
 @functional_datapipe("sample_subgraph")
 class SubgraphSampler(MiniBatchTransformer):
     """A subgraph sampler used to sample a subgraph from a given set of nodes
@@ -251,6 +259,8 @@ def _seeds_cooperative_exchange_2(minibatch, group=None):
                 group,
             )
             seeds_received[ntype] = typed_seeds_received
+            counts_sent[ntype] = typed_counts_sent
+            counts_received[ntype] = typed_counts_received
         minibatch._seed_nodes = seeds_received
         minibatch._counts_sent = revert_to_homo(counts_sent)
         minibatch._counts_received = revert_to_homo(counts_received)
@@ -275,6 +285,11 @@ def _seeds_cooperative_exchange_4(minibatch):
             ntype: typed_inv[0] for ntype, typed_inv in inverse_seeds.items()
         }
         minibatch._seed_nodes = revert_to_homo(unique_seeds)
+        sizes = {
+            ntype: typed_seeds.size(0)
+            for ntype, typed_seeds in unique_seeds.items()
+        }
+        minibatch._seed_sizes = revert_to_homo(sizes)
         minibatch._seed_inverse_ids = revert_to_homo(inverse_seeds)
         return minibatch
 

tests/python/pytorch/graphbolt/test_dataloader.py

@@ -6,6 +6,7 @@
 
 import dgl
 import dgl.graphbolt
+import dgl.graphbolt as gb
 import pytest
 import torch
 import torch.distributed as thd
@@ -194,5 +195,23 @@ def test_gpu_sampling_DataLoader(
                 if sampler_name == "LayerNeighborSampler":
                     assert torch.equal(edge_feature, edge_feature_ref)
     assert len(list(dataloader)) == N // B
+
+    if asynchronous and cooperative:
+        for minibatch in minibatches:
+            x = torch.ones((minibatch.node_ids().size(0), 1), device=F.ctx())
+            for subgraph in minibatch.sampled_subgraphs:
+                x = gb.CooperativeConvFunction.apply(subgraph, x)
+                x, edge_index, size = subgraph.to_pyg(x)
+                x = x[0]
+                one = torch.ones(
+                    edge_index.shape[1], dtype=x.dtype, device=x.device
+                )
+                coo = torch.sparse_coo_tensor(
+                    edge_index.flipud(), one, size=(size[1], size[0])
+                )
+                x = torch.sparse.mm(coo, x)
+            assert x.shape[0] == minibatch.seeds.shape[0]
+            assert x.shape[1] == 1
+
     if thd.is_initialized():
         thd.destroy_process_group()