Index add new

[JackCaoG]

This is to fix the compilation error introduced by pytorch/pytorch#65993. Will support alpha properly in a separate pr.

[bdhirsh]

not supporting the alpha param for now sounds good to me.

I think you'll actually need to change the index_add_ lowering to index_add though, now that it's a structured kernel in core. That's probably better anyway, since we prefer pytorch/xla's lowerings to be functional, right? I think the current state is just a byproduct of the fact that the old index_add kernel in core was composite. I see some test failures, but my guess of what's happening now is:

• you call aten::index_add
• XLA doesn't have a lowering: we hit a code-generated kernel that redispatches to aten::index_add.out
• XLA also doesn't have a lowering for index_add.out so things break

[JackCaoG]

not supporting the alpha param for now sounds good to me.

I think you'll actually need to change the index_add_ lowering to index_add though, now that it's a structured kernel in core. That's probably better anyway, since we prefer pytorch/xla's lowerings to be functional, right? I think the current state is just a byproduct of the fact that the old index_add kernel in core was composite. I see some test failures, but my guess of what's happening now is:

• you call aten::index_add
• XLA doesn't have a lowering: we hit a code-generated kernel that redispatches to aten::index_add.out
• XLA also doesn't have a lowering for index_add.out so things break

Good point, I will do that instead

[bdhirsh]

LGTM. Didn't realize we already have an (unused?) index_add implementation lying around 😛

[bdhirsh]

merging now that pytorch/pytorch#65993 has gone through