Speed issues
de-gozaru opened this issue · 1 comments
de-gozaru commented
Hi @rusty1s ,
I was comparing the speed of torch_cluster.knn_graph
on the GPU with the function bellow:
def knn(x, k):
inner = -2 * torch.matmul(x.transpose(2, 1), x)
xx = torch.sum(x**2, dim=1, keepdim=True)
pairwise_distance = -xx - inner - xx.transpose(2, 1)
idx = pairwise_distance.topk(k=k, dim=-1)[1] # (batch_size, num_points, k)
return idx
and found that torch_cluster.knn_graph
is way slower, around x8 slower, is this something expected?
Thank you in advance!
rusty1s commented