Okwan Kwon

Results 38 comments of Okwan Kwon

@MaheshRavishankar do you know which benchmarks have the pattern? Do you happen to have an example input?

Passing it to Mahesh for future coordination with a low priority.

> (where possible the best solution here is still to fuse this into producers - when we are doing that we should only need to pad/swizzle in very rare cases)...

Yes, I carefully chose "investigate a fusion opportunity" instead of "do pattern matching" according to our design principles. ;) Thanks Ben!

@MaheshRavishankar is working on https://github.com/google/iree/pull/8970. But, we need a vertical fusion of unrelated operations (some calls it horizontal fusion) to fuse reductions.

![Image](https://user-images.githubusercontent.com/4314836/177612823-dbbc58d5-24aa-4d35-a625-0a9224358188.png) The work is being done with two parts, which is a little bit different from the approach above. 1. fuse reduction + elementwise: There are two groups of fusions...

The first part is being done at https://github.com/iree-org/iree/issues/9523.

Unlike a layer normalization, Softmax does not have a dependency to the original input. Softmax ![Image](https://user-images.githubusercontent.com/4314836/177617411-9a461bad-8111-430a-ae35-e85d5e971364.png) LayerNom ![Image](https://user-images.githubusercontent.com/4314836/177617512-73b408b5-63ed-46d3-8f5b-d3468ad8bacf.png)

A PR for the first part is under review. https://github.com/iree-org/iree/pull/9693