Non-blocking gather_csr_cuda()

[rBenke]

Is it possible to make gather_csr_cuda() without cpu-gpu sync?

I can only guess that the problem is in line 248 in csrc/cuda/segment_csr_cuda.cu:

sizes[dim] = indptr.flatten()[-1].cpu().data_ptr<int64_t>()[0];

[rusty1s]

We query the last element of indptr to determine the output size of the out tensor. I don't think there is any real workaround besides passing in an out tensor as part of the input arguments (already supported). Would that work for you?

[rBenke]

I'd have to look more closely at PyG, but in general, what I'm trying to achieve is a fully non-blocking GAT forward pass (with csr data layout). I'd appreciate any hint if you know how to do this, or I'll take some time to figure it out next week.

[github-actions]

This issue had no activity for 6 months. It will be closed in 2 weeks unless there is some new activity. Is this issue already resolved?