Optimize list_to_packed to avoid for loop (#1737)

Summary:
For larger N and Mi value (e.g. N=154, Mi=238) I notice list_to_packed() has become a bottleneck for my application. By removing the for loop and running on GPU, i see a 10-20 x speedup.

Pull Request resolved: https://github.com/facebookresearch/pytorch3d/pull/1737

Reviewed By: MichaelRamamonjisoa

Differential Revision: D54187993

Pulled By: bottler

fbshipit-source-id: 16399a24cb63b48c30460c7d960abef603b115d0

pytorch3d/structures/utils.py

@@ -135,22 +135,21 @@ def list_to_packed(x: List[torch.Tensor]):
         - **item_packed_to_list_idx**: tensor of shape sum(Mi) containing the
           index of the element in the list the item belongs to.
     """
-    N = len(x)
+    if not x:
-    num_items = torch.zeros(N, dtype=torch.int64, device=x[0].device)
+        raise ValueError("Input list is empty")
-    item_packed_first_idx = torch.zeros(N, dtype=torch.int64, device=x[0].device)
+    device = x[0].device
-    item_packed_to_list_idx = []
+    sizes = [xi.shape[0] for xi in x]
-    cur = 0
+    sizes_total = sum(sizes)
-    for i, y in enumerate(x):
+    num_items = torch.tensor(sizes, dtype=torch.int64, device=device)
-        num = len(y)
+    item_packed_first_idx = torch.zeros_like(num_items)
-        num_items[i] = num
+    item_packed_first_idx[1:] = torch.cumsum(num_items[:-1], dim=0)
-        item_packed_first_idx[i] = cur
+    item_packed_to_list_idx = torch.arange(
-        item_packed_to_list_idx.append(
+        sizes_total, dtype=torch.int64, device=device
-            torch.full((num,), i, dtype=torch.int64, device=y.device)
+    )
-        )
+    item_packed_to_list_idx = (
-        cur += num
+        torch.bucketize(item_packed_to_list_idx, item_packed_first_idx, right=True) - 1
-
+    )
     x_packed = torch.cat(x, dim=0)
-    item_packed_to_list_idx = torch.cat(item_packed_to_list_idx, dim=0)
 
     return x_packed, num_items, item_packed_first_idx, item_packed_to_list_idx