Crashes on inputs with numel == 0

Hi, thanks so much for providing this package–it’s very useful. I am running torch 1.0.1.post2 on Ubuntu 16.04 with NVIDIA Quadro P4000 and with torch_scatter installed recently via pip. Please see my minimum working example below which demonstrates what I feel is a bug. Notice that num_vals == 0.

import torch
from torch_scatter import scatter_mean, scatter_max

num_vals = 0
m, n, k = 4, 5, 6

src = torch.zeros((m, n, num_vals)).cuda()
index = torch.zeros((m, n, num_vals)).long().cuda()
kwargs = dict(src=src, index=index, dim=2, dim_size=k)

for _ in range(2):
    out = scatter_mean(**kwargs)
    
for _ in range(2):
    out, argmax = scatter_max(**kwargs)

• scatter_mean, which I understand is mostly just a wrapper for torch.scatter_add_, handles the tensor.numel() == 0 case gracefully.

• But scatter_max fails with RuntimeError: cuda runtime error (9) : invalid configuration argument at /.../aten/src/THC/generic/THCTensorMath.cu:14.

I guess the culprit is #define BLOCKS(N) (N + THREADS - 1) / THREADS in scatter_kernel.cu which doesn’t work so well for N == 0.

[rusty1s]

Fixed in master 😃

[rusty1s]

Great! Fixed in master 😃