[GPU] reduction numerical correctness issue

[IanWood1]

What happened?

Some changes during dispatch creation (encountered here #18341) causes f16 SDXL to produce incorrect results under aggressive fusion. This only occurs when the linalg.generic ops are at a higher dimensionality.

Steps to reproduce your issue

1. Download a repro of the correct dispatch as well as the problematic dispatch
2. Compile/Run both dispatches to generate output. I used this to generate input and this to compile/run each dispatch (the flags are the same as mi300 sdxl CI flags).
3. Compare the outputs

Version information

I was using b197555, but current main should also work

cc @bangtianliu You said you might want to take a look + I think the work you're doing might solve this issue

[IanWood1]

Also, here are the logs with mlir-print-ir-after-all:

• https://gist.github.com/IanWood1/759c49ecaab9d1c5974edf35dfd70ebb
• https://gist.github.com/IanWood1/ab7b4f419ec4022f24e89726cd28b1f6

[bangtianliu]

Sure, will take a look after I am done with my current task.

[IanWood1]

Okay, great. BTW the "good" version has terrible performance on the sdxl benchmarks. I think it goes down the pipeline you fixed a bit ago for #17948, causing the final generic op to be done sequentially. Hopefully with the new pipeline that won't be an issue.