[hotfix] Address the performance issue of topk. (#2628)

* upd

* fix

* upd

* upd

* udp

* upd

* upd

* upd

* upd

* upd

* upd
Co-authored-by: Jinjing Zhou <VoVAllen@users.noreply.github.com>

python/dgl/backend/pytorch/tensor.py

@@ -227,32 +227,40 @@ def randint(shape, dtype, ctx, low, high):
 
 def pad_packed_tensor(input, lengths, value, l_min=None):
     old_shape = input.shape
-    if isinstance(lengths, th.Tensor):
-        max_len = as_scalar(lengths.max())
+    device = input.device
+    if not is_tensor(lengths):
+        lengths = th.tensor(lengths, dtype=th.int64, device=device)
     else:
-        max_len = builtins.max(lengths)
+        lengths = lengths.to(device)
+    max_len = as_scalar(lengths.max())
 
     if l_min is not None:
         max_len = builtins.max(max_len, l_min)
 
     batch_size = len(lengths)
-    device = input.device
     x = input.new(batch_size * max_len, *old_shape[1:])
     x.fill_(value)
-    index = []
-    for i, l in enumerate(lengths):
-        index.extend(range(i * max_len, i * max_len + l))
-    index = th.tensor(index).to(device)
-    return scatter_row(x, index, input).view(batch_size, max_len, *old_shape[1:])
+    index = th.ones(len(input), dtype=th.int64, device=device)
+    cum_lengths = th.cumsum(lengths, 0)
+    index[cum_lengths[:-1]] += (max_len - lengths[:-1])
+    index = th.cumsum(index, 0) - 1
+    x[index] = input
+    return x.view(batch_size, max_len, *old_shape[1:])
 
 def pack_padded_tensor(input, lengths):
-    batch_size, max_len = input.shape[:2]
+    max_len = input.shape[1]
     device = input.device
-    index = []
-    for i, l in enumerate(lengths):
-        index.extend(range(i * max_len, i * max_len + l))
-    index = th.tensor(index).to(device)
-    return gather_row(input.view(batch_size * max_len, -1), index)
+    if not is_tensor(lengths):
+        lengths = th.tensor(lengths, dtype=th.int64, device=device)
+    else:
+        lengths = lengths.to(device)
+    input = input.view(-1, *input.shape[2:])
+    out_len = lengths.sum().item()
+    index = th.ones(out_len, dtype=th.int64, device=device)
+    cum_lengths = th.cumsum(lengths, 0)
+    index[cum_lengths[:-1]] += (max_len - lengths[:-1])
+    index = th.cumsum(index, 0) - 1
+    return input[index]
 
 def boolean_mask(input, mask):
     if 'bool' not in str(mask.dtype):

python/dgl/readout.py

@@ -485,6 +485,38 @@ READOUT_ON_ATTRS = {
     'edges': ('edata', 'batch_num_edges', 'number_of_edges'),
 }
 
+def _topk_torch(keys, k, descending, x):
+    """Internal function to take graph-wise top-k node/edge features according to
+    the rank given by keys, this function is PyTorch only.
+
+    Parameters
+    ----------
+    keys : Tensor
+        The key for ranking.
+    k : int
+        The :math:`k` in "top-:math:`k`".
+    descending : bool
+        Indicates whether to return the feature corresponding to largest or
+        smallest elements.
+    x : Tensor
+        The padded feature with shape (batch, max_len, *)
+
+    Returns
+    -------
+    sorted_feat : Tensor
+        A tensor with shape :math:`(batch, k, *)`.
+    sorted_idx : Tensor
+        A tensor with shape :math:`(batch, k)`.
+    """
+    import torch as th
+    batch_size, max_len = x.shape[0], x.shape[1]
+    topk_indices = keys.topk(k, -1, largest=descending)[1]  # (batch_size, k)
+    x = x.view((batch_size * max_len), -1)
+    shift = th.arange(0, batch_size, device=x.device).view(batch_size, 1) * max_len
+    topk_indices_ = topk_indices + shift
+    x = x[topk_indices_].view(batch_size, k, -1)
+    return th.masked_fill(x, th.isinf(x), 0), topk_indices
+
 def _topk_on(graph, typestr, feat, k, descending, sortby, ntype_or_etype):
     """Internal function to take graph-wise top-k node/edge features of
     field :attr:`feat` in :attr:`graph` ranked by keys at given
@@ -534,27 +566,27 @@ def _topk_on(graph, typestr, feat, k, descending, sortby, ntype_or_etype):
     if F.ndim(data[feat]) > 2:
         raise DGLError('Only support {} feature `{}` with dimension less than or'
                        ' equal to 2'.format(typestr, feat))
-
     feat = data[feat]
     hidden_size = F.shape(feat)[-1]
     batch_num_objs = getattr(graph, batch_num_objs_attr)(ntype_or_etype)
     batch_size = len(batch_num_objs)
-
     length = max(max(F.asnumpy(batch_num_objs)), k)
     fill_val = -float('inf') if descending else float('inf')
-    feat_ = F.pad_packed_tensor(feat, batch_num_objs, fill_val, l_min=k)
+    feat_ = F.pad_packed_tensor(feat, batch_num_objs, fill_val, l_min=k)  # (batch_size, l, d)
 
+    if F.backend_name == 'pytorch' and sortby is not None:
+        # PyTorch's implementation of top-K
+        keys = feat_[..., sortby]  # (batch_size, l)
+        return _topk_torch(keys, k, descending, feat_)
+    else:
+        # Fallback to framework-agnostic implementation of top-K
         if sortby is not None:
             keys = F.squeeze(F.slice_axis(feat_, -1, sortby, sortby+1), -1)
             order = F.argsort(keys, -1, descending=descending)
         else:
             order = F.argsort(feat_, 1, descending=descending)
-
         topk_indices = F.slice_axis(order, 1, 0, k)
 
-    # zero padding
-    feat_ = F.pad_packed_tensor(feat, batch_num_objs, 0, l_min=k)
-
         if sortby is not None:
             feat_ = F.reshape(feat_, (batch_size * length, -1))
             shift = F.repeat(F.arange(0, batch_size) * length, k, -1)
@@ -566,9 +598,9 @@ def _topk_on(graph, typestr, feat, k, descending, sortby, ntype_or_etype):
                     F.cat([F.arange(0, hidden_size)] * batch_size * k, -1)
             shift = F.copy_to(shift, F.context(feat))
             topk_indices_ = F.reshape(topk_indices, (-1,)) * hidden_size + shift
-
-    return F.reshape(F.gather_row(feat_, topk_indices_), (batch_size, k, -1)),\
-           topk_indices
+        out = F.reshape(F.gather_row(feat_, topk_indices_), (batch_size, k, -1))
+        out = F.replace_inf_with_zero(out)
+        return out, topk_indices
 
 def topk_nodes(graph, feat, k, *, descending=True, sortby=None, ntype=None):
     """Return a graph-level representation by a graph-wise top-k on