Add Wave ops in warp reduction.

[chengjunlu]

Use the GenISA WaveAll and WaveCluster in warp horizontal reduce.

[etiotto]

I think we should have an operation in the TritonGEN dialect for WaveAll and WaveCluster. I discussed the design of that operation yesterday with @whitneywhtsang. I'm hoping she can create that operation and then we generate it in Triton.

[FMarno]

I think we should have an operation in the TritonGEN dialect for WaveAll and WaveCluster. I discussed the design of that operation yesterday with @whitneywhtsang. I'm hoping she can create that operation and then we generate it in Triton.

It will likely also be helpful to add the WavePrefix intrinsic for scans (going by the small bit of documentation).

[chengjunlu]

I think we should have an operation in the TritonGEN dialect for WaveAll and WaveCluster. I discussed the design of that operation yesterday with @whitneywhtsang. I'm hoping she can create that operation and then we generate it in Triton.

I will create a new PR based on the TritonGen operation.

[chengjunlu]

I think we should have an operation in the TritonGEN dialect for WaveAll and WaveCluster. I discussed the design of that operation yesterday with @whitneywhtsang. I'm hoping she can create that operation and then we generate it in Triton.

It will likely also be helpful to add the WavePrefix intrinsic for scans (going by the small bit of documentation).

I created an issue to track this new requirements.
#1216

[Dewei-Wang-sh]

I think we should have an operation in the TritonGEN dialect for WaveAll and WaveCluster. I discussed the design of that operation yesterday with @whitneywhtsang. I'm hoping she can create that operation and then we generate it in Triton.

@whitneywhtsang @etiotto add an op in tritongen for attention, FYI.
Dewei-Wang-sh@8670e5c#diff-22fc078e0ae2b9f9c2ee859630bfbd1092802ce694665ffa78a27d0a6464fa01

[whitneywhtsang]

I think we should have an operation in the TritonGEN dialect for WaveAll and WaveCluster. I discussed the design of that operation yesterday with @whitneywhtsang. I'm hoping she can create that operation and then we generate it in Triton.

@whitneywhtsang @etiotto add an op in tritongen for attention, FYI. Dewei-Wang-sh@8670e5c#diff-22fc078e0ae2b9f9c2ee859630bfbd1092802ce694665ffa78a27d0a6464fa01

I am working on #1214, will put you as reviewer once it is ready, so you can check if it misses any features you need for your case.

[Dewei-Wang-sh]

I think we should have an operation in the TritonGEN dialect for WaveAll and WaveCluster. I discussed the design of that operation yesterday with @whitneywhtsang. I'm hoping she can create that operation and then we generate it in Triton.

@whitneywhtsang @etiotto add an op in tritongen for attention, FYI. Dewei-Wang-sh@8670e5c#diff-22fc078e0ae2b9f9c2ee859630bfbd1092802ce694665ffa78a27d0a6464fa01

I am working on #1214, will put you as reviewer once it is ready, so you can check if it misses any features you need for your case.

great, this is the design align to our current.

[whitneywhtsang]

The operator is added in TritonGEN dialect, please update the PR to use it.