Why torch.sparse.cuda() is much faster than torch.cuda()?

1

The opt torch.sparse.cuda() is much faster than torch.cuda(), why?

2

Hi,

I don’t think torch.cuda() is a function? Could you be clearer on what you compare and how please?

3

Thanks for your replying.

a=torch.FloatTensor()

b=torch.sparse.FloatTensor()

b.cuda() is much faster than a.cuda() even though a.size() is roughly equal to b.size().

发自我的小米手机

在 Alban D via PyTorch Forums noreply@discuss.pytorch.org，2019年10月28日 22:08写道：

4

Hi,

How many non-zero elements are in the sparse Tensor? The whole point of the sparse tensor is to only save the non-zero values, so there is potentially much less things to transfer to the gpu.

5

Hi, i have another question. When i use sparse tensor cuda() in dfferent models, the same tensor takes different time (one is based on nn.Module, another is an C++ cuda extension i wrote). What can affect the time ?

6

I have known the reason. The cuda() is asynchronous.

7

I have a kernel that needs one array of floats (for input) and one array of ints (for labels)? How to do that?

8

Hi,

I’m not sure what the problem is. Just pass these as arguments? There is no restriction about types.
Do you have a code sample that shows what you’re trying to do?